## Background
Published on YouTube by Bg2 Pod on June 11, 2026, this episode features Brad Gerstner sitting down with Gavin Baker, Andrew Fox, and Clark Tang. The show notes describe a deep-dive discussion on the upcoming SpaceX IPO, Starship reusability, Starlink, xAI, orbital data centers, and the broader AI CapEx boom.
## TL;DR
The panel explores the multifaceted investment thesis for SpaceX, examining its core rocket business alongside its rapid emergence as a major AI hyperscaler ("Elon Web Services") and the economics of orbital compute. They analyze recent frontier model releases like Fable 5 and Mythos, as well as the strategic value of xAI's Cursor acquisition in driving coding intelligence. Finally, the hosts address Wall Street's skepticism around AI CapEx sustainability, arguing that soaring monetization per gigawatt and agentic compute demand fully justify the multi-trillion-dollar infrastructure buildout.
## Atmosphere & style
The discussion carries an energetic, high-conviction, and insider-analyst vibe, blending sharp venture-capital intellectual rigor with fast-paced Silicon Valley camaraderie. The speakers move fluidly between complex technical metrics and bold macro forecasts, exuding excitement about frontier AI acceleration while maintaining a grounded, self-aware approach to market risks and post-IPO drawdowns. Listening feels like sitting in on an elite, behind-the-scenes investment committee meeting where deep technical domain expertise meets ambitious market strategy.
## Key takeaways
- [00:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=0s) Set at $135 per share, SpaceX's IPO values the company at $1.77 trillion, with market expectations projecting forward revenue to reach $160 billion by 2028.
- [03:05](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=185s) xAI's cloud compute deals achieve industry-leading operating profit per gigawatt, driven by Elon Musk's ability to deploy data center clusters in just 122 days.
- [04:05](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=245s) xAI's acquisition of Cursor provided vast reserves of proprietary coding tokens, enabling the rapid development of Pareto-dominant coding models like Composer 2.5.
- [07:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=420s) Starship's rapid two-stage reusability is designed to reduce launch costs below $250/kg, enabling high-frequency launches and making orbital compute clusters economically viable.
- [10:38](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=638s) Starlink broadband and Direct-to-Cell services are projected to grow revenue from $10 billion to $50 billion by 2028 while representing less than 1% of total global market penetration.
- [13:28](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=808s) SpaceX has rapidly emerged as a major AI hyperscaler ("Elon Web Services") by leasing raw processing capacity to external tech giants like Google and Anthropic.
- [21:05](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1265s) Launching compute into orbit eliminates terrestrial land and power constraints, reducing non-silicon infrastructure CapEx from $25B per gigawatt on the ground down to $5B in space.
- [31:58](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1918s) xAI has secured up to 20% of early Nvidia Vera Rubin capacity and is integrating Cursor data directly into the pre-training pipeline for Grok 4.3.
- [37:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2220s) SpaceX's ~50% founder ownership by Elon Musk, 365-day lock-up, and decade-long history of semi-annual private tender offers cushion post-IPO selling pressure and stock volatility.
- [43:56](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2636s) Static snapshot benchmarks fail to evaluate frontier models properly, as long-running inference compute allows continuous reasoning on complex, multi-month problems.
- [47:37](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2857s) While open-source models capture high total token volumes through routing, over 90% of economic value and high-value workloads accrue exclusively to frontier models.
- [56:53](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3413s) Nvidia retains hardware dominance because its higher token-per-watt throughput delivers superior revenue per gigawatt compared to hyperscaler custom ASICs.
- [01:03:03](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3783s) Global AI CapEx is projected to reach $1.5 trillion by 2027, driven by an infrastructure race and high-margin monetization scaling toward $30B–$40B+ per gigawatt.
- [01:07:09](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4029s) Global agentic AI adoption sits below 0.2%, but power-law compute usage by active agentic workflows is driving persistent structural shortages of processing power.
- [01:11:52](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4312s) While software and internet sectors have lagged, public equity markets remain supported by AI megacaps despite rising inflation and macroeconomic headwinds.
- [01:18:13](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4693s) The next $1 trillion in tech revenue is expected to be generated in just 4 to 5 years by three private powerhouses: SpaceX, Anthropic, and OpenAI.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=0s) Introduction: SpaceX IPO, Mythos, and the AI Race
The hosts kick off the episode in Silicon Valley, diving into the most pressuring developments across technology, markets, and artificial intelligence. Brad Gerstner welcomes Gavin Baker and Andrew Fox from Atreides Management, alongside Altimeter partner Clark Tang, to break down the upcoming SpaceX IPO, recent AI model releases, and key supply chain takeaways from Taiwan.
The conversation focuses heavily on SpaceX’s valuation dynamics, market expectations, and the broader implications of superintelligence. Institutional consensus increasingly views SpaceX not merely as a space exploration company, but as an essential compound bet on both space technology and the burgeoning AI future.
* **IPO Valuation & Share Price:** Set at $135 per share, valuing SpaceX at $1.77 trillion.
* **Forward Revenue Targets:** Street expectations—including reports from Goldman Sachs and *The Wall Street Journal*—project $160 billion in revenue by 2028.
* **Core Investment Thesis:** "If you're AI pilled, that means we got to build a lot more compute than the world thinks... I don't know another entrepreneur or another business that's a better bet on the future than SpaceX."
---
[00:03](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3s) Terrestrial AI Compute & Operating Profit per Gigawatt
Gavin Baker outlines the first critical value lever driving SpaceX’s valuation: its aggressive expansion into terrestrial AI data centers. Clark Tang’s analysis reveals that xAI’s cloud computing deal with Google achieves a higher operating profit per gigawatt than deals structured by Anthropic, Meta, Google, or OpenAI. Similarly, xAI's arrangement involving Anthropic yields profit margins second only to Anthropic's own internal metrics.
```
Operating Profit per Gigawatt (Relative Efficiency)
xAI / Google Deal ████████████████████████████████ 100%
Anthropic (Internal)█████████████████████████ 78%
Meta ████████████████ 50%
OpenAI ██████████████ 44%
```
A key driver of this financial efficiency is Elon Musk’s speed of execution. As Nvidia’s Jensen Huang has noted, Musk brings data center clusters online in just 122 days—dramatically undercutting standard industry timelines. Because rapid construction minimizes daily labor overheads for electricians and plumbers, speed directly translates into structural cost reduction.
* **Projected Returns:** Altimeter’s internal calculations show an estimated 55% Internal Rate of Return (IRR) on xAI's compute infrastructure investments.
* **Capital Efficiency:** Borrowing capital at 6–8% to invest in assets yielding a 55% IRR creates an exceptionally high-margin math equation for scaling compute.
* **The Cost of Speed:** "Speed is literally cost because every day you're paying electricians and plumbers... they're now monetizing them at arguably the highest rate."
---
[00:04](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4s) Coding Models, Pareto Efficiency, and Cursor Data
The second major variable is the rapid progression of AI intelligence per dollar, illustrated by Pareto efficiency curves for software development models. Within just a ten-day span, the frontier model landscape shifted dramatically with updates to Anthropic's Opus 4.7/4.8, OpenAI's Codex, and the introduction of Fable and Mythos. Baker stresses that in frontier AI, nearly all revenue accrues to whichever model sits on the absolute Pareto frontier for a given workload.
```
Pareto Efficiency in AI Coding Models
High ▲ ● Opus 4.8 / Mythos
│ ●
│ ● Composer 2.5
│ ● Composer 2
│ ● Opus 4.7
Intelligence │ ● Codex
│
Low └───────────────────────────────────────────────────► High
Compute / Cost
```
xAI’s acquisition of Cursor has proven to be a strategic masterstroke in training proprietary coding models. Cursor and Anthropic possess vast reserves of proprietary coding tokens—far exceeding what is publicly available on the internet. By leveraging these proprietary datasets alongside ChemK 0.25, supervised fine-tuning, and reinforcement learning (RL) across three weeks in the Colossus 2 cluster, xAI produced Composer 2.5, achieving Pareto-dominant coding performance.
* **Data Dominance:** Cursor’s dataset contains more proprietary coding tokens than exist across the entire public web.
* **Rapid Training Cycles:** Composer 2.5 achieved Pareto dominance after just 3 weeks of training on the Colossus 2 cluster.
* **Market Monetization:** All high-value market revenue tends to flow exclusively to models occupying the top edge of the Pareto curve.
---
[00:07](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=7s) The Core Rocket Business: Rapid Reusability & Economics
Andrew Fox breaks down the foundational pillar of SpaceX: its launch business and the push toward true rapid reusability. Traditional aerospace economics functioned like "boarding a plane, flying to California, getting off, and having the plane explode." SpaceX’s core breakthrough—and the key enabler for orbital compute—is turning rockets into reusable assets operating at airline-like cadence.
To realize the full economic promise of Starship, SpaceX aims to fly both stages (booster and ship) 30 to 50 times before requiring major retrofits. Amortizing the capital cost of the vehicle across dozens of flights reduces launch costs by orders of magnitude, making massive low-Earth orbit (LEO) constellations and orbital compute clusters economically viable.
* **Launch Cadence Expansion:** Moving from ~160–165 launches per year up into the high hundreds, with projections targeting thousands of annual launches within three to five years.
* **Second-Stage Recovery Target:** SpaceX plans to attempt full recovery of Starship’s second stage later this year, with full re-flight attempts planned for next year.
* **Operational Goal:** Achieving a launch cadence of 2 to 3 Starship launches per day.
Starlink’s broadband business remains at a relatively early phase of its global market penetration. Despite its rapid growth, Starlink currently serves less than 1% of total global households across residential, maritime, aviation, and vehicular markets. However, with the deployment of Starship V3 and rapid reusability, Starlink is poised to scale into hundreds of millions of active terminals worldwide.
The expansion is further bolstered by Starlink Direct-to-Cell technology, designed to eliminate mobile dead zones globally. The hosts joke about losing cellular coverage right on Sand Hill Road in Silicon Valley—the epicenter of tech capital—highlighting the massive consumer demand for reliable satellite-to-phone connectivity.
* **Revenue Trajectory:** Projections show Starlink/connectivity revenues scaling from $10 billion to $50 billion by 2028.
* **Market Share:** A $50 billion revenue run-rate represents roughly 0.3% global telecom market penetration.
* **Value Proposition:** "Better, faster, cheaper has been a winning formula... wherever I am in the world, Starlink is the best connection."
---
[13:28](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=808s) "Elon Web Services" (EWS) & The Colossus Advantage
Clark Tang highlights the unexpected emergence of "Elon Web Services" (EWS). Until recently, SpaceX model projections only accounted for Starlink communications and xAI's proprietary model development. Few analysts anticipated SpaceX becoming a major third-party AI compute hyperscaler, leasing raw processing capacity to external tech giants like Anthropic and Google.
Musk’s ability to build massive infrastructure at unprecedented speed was demonstrated by the Colossus cluster. While traditional enterprise supercomputers take three years to plan and a full year to bring online after delivery, Musk's team brought a 100,000 GPU cluster online in just 19 days. This speed, paired with top-tier power procurement and highly favorable financing rates, creates a self-reinforcing advantage over traditional cloud hyperscalers.
* **Infrastructure Speed:** Built and deployed a functional 100,000-GPU supercomputer cluster in 19 days vs. an industry standard 3–4 year lifecycle.
* **Execution Delta:** Jensen Huang famously categorized Musk's execution capabilities as an "N of 1."
* **Competitive Moat:** Superior access to power, rapid site construction, and low cost of capital allow EWS to capture high operating margins while undercutting traditional compute providers.
[16:16](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=976s) Monetization Premiums and "Elon Web Services"
The panel opens by exploring why hyperscalers like Google are paying top dollar for SpaceX’s infrastructure. Gavin Baker and Andrew Fox highlight that Google is paying a steep premium for terrestrial compute, acting as an upfront "call option" to lock in first-in-line access to future space-based compute. SpaceX can stand up massive, coherent, single-location compute faster than almost anyone else, making them an irresistible partner. This shift marks a dramatic transformation: in a mere 30 days, SpaceX has transformed from a non-entity in cloud hosting into the number-four AI hyperscaler globally, eclipsing legacy players like Oracle and standing shoulder-to-shoulder with specialized neo-clouds like CoreWeave, Nebius, and Iren.
```
AI Hyperscaler Scale (Compute Capacity)
[ Top 3 Hyperscalers ] --> [ #4 SpaceX / "Elon Web Services" ] --> [ Oracle, Neo-Clouds ]
```
The speakers argue that data center design is far from a commoditized business. Just as Elon Musk engineered electric vehicles and reusable rockets from first principles, he approached data center design with a fresh, unconventional lens. By rejecting standard industry assumptions, SpaceX constructed operational facilities in an astonishing 122 days. As power equipment suppliers like GE Vernova face constrained inventory for gas combustion engines, they increasingly prioritize proven, fast-moving builders like xAI over unproven neo-cloud startups—because in this market, speed to energize directly translates to cash flow.
- **Fast Build Times:** SpaceX erected functional data center infrastructure in 122 days by applying first-principles engineering.
- **Hyperscaler Status:** Post-Google deal, SpaceX effectively jumped into the top four AI compute providers worldwide.
- **Power Bottlenecks:** Turbines, switchgears, and transformers are scarce; suppliers favor high-speed operators who monetize GPUs faster.
---
[21:05](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1265s) Ground vs. Space Compute Economics
Moving beyond terrestrial builds, the conversation turns to the economics of orbital data centers and whether space-based compute is essential for SpaceX's near-term IPO valuation. Brad Gerstner notes that current global capacity sits around 20 to 30 gigawatts. While terrestrial expansion can carry xAI and SpaceX far, orbital compute offers a massive structural cost advantage. Clark Tang breaks down the math: on Earth, building out a gigawatt of compute capacity costs roughly $60 billion overall—$35 billion for GPUs/silicon and $25 billion for land, shell, power, and cooling infrastructure. In contrast, by taking land and terrestrial power constraints out of the equation, launching compute into orbit drops the non-silicon infrastructure CapEx down to approximately $5 billion per gigawatt.
```
Total Cost per Gigawatt (Terrestrial vs. Orbital)
Terrestrial: [ Silicon: $35B ] + [ Land/Power/Cooling: $25B ] = $60B
Orbital: [ Silicon: $35B ] + [ Launch/Satellite: $5B ] = $40B (5x savings on facility CapEx)
```
This drastic cost reduction hinges entirely on Starship V3's rapid two-stage reusability. Payload launch costs are projected to drop from $1,500/kg on Falcon 9 down to $250/kg or lower on Starship, eventually asymptoting toward the raw cost of fuel. Each Starship launch carrying 100 metric tons can place about 5 megawatts of compute capacity into orbit. While hardware degradation—such as melting GPUs or failing optical lasers—remains a operational risk, orbital compute achieves a 5x CapEx reduction on half the bill-of-materials if satellite failure rates stay reasonably low.
- **Terrestrial Cost Breakdown:** $20B–$25B per gigawatt for shell, switchgears, transformers, and cooling; $35B for GPUs.
- **Orbital Cost Breakdown:** ~$5B per gigawatt launch/satellite CapEx using Starship.
- **Starship Payload Metrics:** ~5 MW of compute capacity delivered per 100-metric-ton Starship launch.
- **Launch Cost Trajectory:** Decreasing from $1,500/kg on Falcon 9 to <$250/kg on Starship.
---
[29:25](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1765s) The Cursor Acquisition and xAI's Frontier Model Leap
The discussion pivots to xAI's software strategy and its game-changing acquisition of Cursor. Just months prior, xAI was seen as a solid contender, but acquiring Cursor—a 700-to-800-person team tracking toward an estimated $10 billion annual revenue run rate—radically upgraded xAI's execution capabilities. Gavin Baker explains that while Cursor's Composer 2.5 was Pareto-dominant using Moonshot's Kimi K2.5 as a base model, xAI is now training Grok 4.3 (a 1.5-trillion parameter model) and injecting Cursor's proprietary coding data directly into the pre-training pipeline rather than relying solely on post-training reinforcement learning.
> "Coding may be the fastest path to AGI and ASI because if a model's good at coding, you can write code to do anything."
Referencing Replit founder Amjad Masad's insights, the panel notes that automated coding capability represents a "bitter lesson-adjacent" shortcut toward Artificial Superintelligence (ASI). By pairing Cursor's elite engineering team with SpaceX's virtually unlimited compute supply, xAI has secured a powerful flywheel: monetizing raw compute to external customers while simultaneously training frontier-class models internally.
- **Cursor Scale:** 700–800 employees with projections exiting the year at up to $10 billion in revenue.
- **Grok 4.3 Specs:** 1.5-trillion parameter frontier model integrating Cursor data into pre-training.
- **Base Model Progression:** Moving from Kimi K2.5 base integrations to native Grok 4.3 infrastructure.
[31:58](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1918s) The Underrated Upside of xAI and "Elon Web Services"
Clark Tang addresses how xAI has managed to overcome early compute bottlenecks through sheer speed and aggressive infrastructure scaling. A year and a half ago, xAI lagged behind competitors in graphics processing units (GPUs), lacking substantial Nvidia H100 capacity. However, they rapidly built the original Colossus cluster and followed it up with Colossus 2 at a scale far exceeding anyone else in the industry. Looking ahead to Nvidia's next-generation Vera Rubin architecture, internal conversations suggest xAI has already secured up to 20% of early capacity during a period of extreme supply scarcity.
```
xAI Infrastructure Acceleration Timeline
---------------------------------------------------------------------
1.5 Years Ago: Lagging behind in H100 capacity
Colossus 1: Rapidly deployed to catch up on frontier training
Colossus 2: Scaled beyond industry peers
Vera Rubin: Secured ~20% of initial next-gen capacity
```
Brad Gerstner draws a direct historical parallel between Elon Musk’s strategy and Jeff Bezos building Amazon Web Services (AWS). Bezos initially built massive computing infrastructure to handle peak holiday traffic on Black Friday, leaving Amazon with excess capacity during the rest of the year. Investors in 2009–2010 hated AWS because it consumed vast free cash flow, yet Bezos ended up digging one of the greatest corporate "gold mines" in history. Similarly, if xAI over-procurs compute for its own frontier model training, that scarce capacity can be monetized to third parties at best-in-class operating margins and rapid payback periods.
- **Vera Rubin Allocation:** xAI has secured up to **20%** of early Vera Rubin capacity.
- **Model Benchmark:** Grok 4.3 sits on the Pareto frontier as the most intelligent **500B parameter model** in the world.
- **Frontier AI Players:** Currently dominated by four entities: xAI (SpaceX AI), Google (Gemini 3.1 Pro), Anthropic, and OpenAI.
---
[35:05](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2105s) Sizing Up the IPO Bull & Bear Case
Brad Gerstner highlights the stark contrast between Wall Street's skepticism and the long-term fundamental bull case surrounding the SpaceX initial public offering (IPO). Bears look at last year's $18 billion in revenue and scoff at investment bank forecasts projecting $160 billion in three to four years, noting that almost no company in history has achieved an 8x revenue expansion at that scale. However, when evaluated from first principles, Starlink’s expansion, terrestrial AI compute building, and the model integration following the Cursor acquisition make that growth trajectory entirely plausible.
The panel emphasizes that institutional investors view SpaceX not as a short-term trade, but as a "must-buy, set-it-and-forget-it" core asset. For investors who believe the world will require vastly more compute and that AI models will yield unprecedented economic value, SpaceX represents the premier vehicle for capturing both space infrastructure and the AI future under one roof.
- **Revenue Trajectory:** Bears question the jump from **$18 billion** to a projected **$160 billion** over 3–4 years (an **8x growth**).
- **Implied Valuation:** The discussion frames the IPO valuation around **$1.77 trillion**.
- **Consensus Perspective:** SpaceX is regarded as an indispensable institutional ballast combining space dominance with frontier compute.
---
[37:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2220s) Post-IPO Volatility, Lock-ups, and Share Supply
Addressing historical market patterns, Brad notes a widely circulated chart showing that past tech giants—including Facebook, Twitter, Alibaba, and Shopify—suffered average maximum drawdowns exceeding 50% following their IPOs. Gavin Baker agrees that while post-IPO markets are inherently bouncy, SpaceX represents an unprecedented situation due to its sheer scale, rapid inclusion into major market indices, and unique share structure.
> "He and Gwynne and the team are great at business... I don't know that any of the Mag 7 could have moved that quickly to adjust the business that they did. It's exceptionally entrepreneurial at scale." — *Brad Gerstner*
Gavin points out that Elon Musk owns approximately 50% of SpaceX and is subject to a 365-day lock-up, eliminating any concern of founder liquidation. Furthermore, because SpaceX has offered semi-annual liquidity tender offers for its employees and early investors over the last decade, much of the historic selling pressure has already been absorbed organically over nearly 20 liquidity cycles.
| Metric / Dimension | SpaceX IPO Structure & Context |
|:--- |:--- |
| **Historic Tech Drawdown Avg.** | >50% max post-IPO drawdown across major tech listings |
| **Elon Musk Ownership & Lock-up** | ~50% ownership, locked up for 365/366 days |
| **Revenue Multiple Shift** | Multiple compressed from 100x trailing revenue down to ~39x after adding $29B in contracts in one month |
| **Private Liquidity Frequency** | Tender offers provided to employees/investors every 6 months for ~10 years (~20 opportunities) |
| **Comparative Capital Raised** | $250B combined private AI capital raise equates to roughly **1% of the Magnificent 7** total market cap |
---
[43:56](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2636s) Fable 5, Mythos, and the Breakdown of Snapshot Benchmarks
The discussion shifts to Anthropic's release of Fable 5—a production variant of their flagship Mythos model outfitted with safety classifiers, cybersecurity limits, and biological/chemical safeguards that fail back to Opus 4.8 when triggered. Reflecting on a tweet by Noam Brown, Gavin Baker argues that traditional, static "snapshot benchmarks" have become obsolete for evaluating modern frontier models.
Gavin details the profound implication of long-running inference compute: humanity does not actually know how intelligent these models are because no one has allowed a model like Mythos to run continuously on a single problem for an entire year. Just as a human genius like Albert Einstein was limited by physical fatigue, sleep, and distraction, an unconstrained frontier model spending weeks or months of continuous compute on a single domain could solve previously intractable scientific problems.
> "Imagine if Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn't have to eat, he doesn't have to sleep, he doesn't have to relax... and he thought for 1 year. We might already have solved a lot of these intractable problems." — *Gavin Baker*
- **Anthropic Model Stack:** Fable 5 utilizes safety/cyber classifiers and falls back to **Opus 4.8** when safety guardrails are triggered.
- **Inference Shift:** Intelligence is increasingly a function of time, tokens, and inference compute, making traditional snapshot benchmarks uninformative.
- **Market Dynamics:** Contrary to expectations that open-source models would commoditize AI with cheap tokens, frontier tokens are capturing the vast majority of industry revenue six months into the year.
[47:37](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2857s) Frontier Models vs. Open Source Economics
The panel explores whether open-source models are closing the performance gap with frontier systems or if the leading models are continually extending their leads. Clark Tang points out that while the debate over being "three to six months behind" has persisted since the inception of LLM training, empirical evidence shows that almost all revenue continues to accrue at the frontier. Every time a new frontier model lands, it unlocks entirely new categories of use cases—such as long-running, multi-agent orchestration—that were impossible with previous-generation models.
Clark shares a personal example from his team's workflow after testing Claude 3.5 Sonnet (referred to during the session alongside Claude 3.5 Haiku and "Fable 5"). He loaded seven complex research models into the system, instructing it to establish a unified view based on intersecting assumptions like TSMC supply chain capacity. The model analyzed the contradictions across these distinct inputs seamlessly. Furthermore, after dumping three years of his personal notes into the system, the model cross-referenced his past ideas to identify which sources proved to have the highest historical signal.
```
Frontier Lead ──> Unlocks Complex Workflows (Coding, Multi-Agent) ──> Captures ~90%+ Revenue
Open Source ──> 6-Month Technical Lag ──> Drives High Token Volume
```
- **Stripe Codebase Case Study:** Anthropic demonstrated refactoring a 50-million-line Ruby codebase at Stripe in a single day, a task that previously required weeks of engineering team effort.
- **Token Consumption Disconnect:** Gavin Baker notes that two things can be true at once: the vast majority of total tokens consumed globally may be open source, yet 90% or more of actual economic value and dollar revenue stays with frontier model developers.
- **Enterprise Survey Data:** A survey of 300 enterprises showed that companies are aggressively adopting model routers to direct routine queries to open-source models, while keeping high-value, critical tasks (like code generation) strictly on frontier models.
---
[51:21](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3081s) Model Routing, Enterprise Usage, and Regional Dynamics
Gavin highlights a case study by legal tech company Harvey, which used proprietary legal datasets to conduct reinforcement learning and supervised fine-tuning via Fireworks on an open-source model. By deploying an intelligent router to select when to hit their fine-tuned model versus a frontier API, Harvey achieved superior performance over Claude 3 Opus at a significantly lower cost.
The hosts draw a sharp distinction between enterprise usage patterns in Western tech hubs versus Asia. In Silicon Valley, the prevailing ethos leans toward closed-source cloud APIs where all traffic flows through a single centralized intelligence provider. In contrast, Asian markets operate under a strong belief in cost optimization—matching specific workloads to the precise model required so as not to overspend on compute.
> "You don't need Albert Einstein to book you a trip. You don't need Albert Einstein to do KYC."
- **US Enterprise Preferences:** US financial institutions like JPMorgan remain reluctant to use Chinese open-source models, choosing instead to wait for competitive US open-source models for back-office automation.
- **Jensen's Routing Thesis:** Nvidia CEO Jensen Huang argues that model routing will naturally expand open-source adoption. Once open-source models gain stable long-running agentic capabilities, they will claim a larger share of enterprise spend.
- **Hardware Bullishness:** Gavin emphasizes that strong open-source model performance is inherently bullish for compute and hardware providers; when software margins shrink, a higher percentage of the total budget flows directly into raw infrastructure.
---
[56:53](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3413s) Nvidia's Strategic Playbook and the ASIC Threat
The discussion shifts to how Nvidia might respond if open-source models edge closer to frontier capabilities. Gavin posits that Nvidia could leverage its substantial resources to become a primary provider of frontier-level open-source AI. By releasing powerful models like Nemotron, Nvidia can suppress the margins of alternative custom ASICs (Application-Specific Integrated Circuits), making it economically difficult for competitors to justify the high R&D cost of bespoke chips.
Gavin highlights that if hyperscalers continue attempting to bypass Nvidia with custom chips, Jensen Huang could choose to compete directly with his own customers by scaling up Nvidia's cloud infrastructure and foundation model offerings much faster than the market expects.
- **Nvidia vs. Custom ASICs:** Rather than losing market share as many predicted, Nvidia has maintained or even expanded its share of total compute deployed over recent years.
- **Tokens Per Watt Economics:** In a power-constrained market, the core metric is revenue generated per gigawatt. Nvidia's hardware often delivers higher token throughput per watt than custom ASICs, offsetting the higher initial chip cost with superior operating revenue.
- **Emerging ASIC Competitors:** Clark notes new entrants in the ASIC space discussed during his trip to Taiwan, including MediaTek's custom chips and Broadcom's TPU design work (such as the V8I series).
---
[01:00:48](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3648s) Power Allocation and Hyperscaler Custom Silicon
Clark and Gavin review a breakdown of power allocations across major chip platforms for upcoming large-scale cluster deployments. Out of a theoretical 27-gigawatt pipeline being discussed across industry reports for major labs like OpenAI:
Gavin expresses skepticism that Broadcom or alternative ASIC providers will actually execute on that full allocation once physical deployments begin. When datacenters face hard power caps, choosing a chip with lower token output per watt directly reduces top-line revenue.
The panel reflects on custom silicon efforts among frontier labs. Gavin remarks that while Microsoft and Meta's custom ASIC programs have experienced setbacks, OpenAI's internal custom chip effort (internally code-named "Jalapeno") produced impressive technical performance. However, because it runs at lower operational temperatures than Nvidia GPUs, it requires significantly more cooling infrastructure, consuming additional power.
- **Focus vs. Integration:** Brad Gerstner questions whether designing custom hardware is the highest and best use of time for frontier AI labs.
- **The Intelligence Race Window:** Brad argues that if the timeline to achieving superintelligence is compressed into the next two to three years via recursive self-improvement loops, frontier labs must maintain hyper-focus on model development rather than spreading resources into hardware manufacturing.
[01:03:03](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3783s) The AI CapEx Explosion: $1.5 Trillion CapEx vs. $300 Billion Revenue
Brad Gerstner shifts the discussion back to the staggering macro numbers driving the AI boom, asking whether the massive capital expenditure (CapEx) expansion actually "maths" out against current revenue returns. Google recently raised its capital spending to $80 billion, and overall free cash flow across the Hyperscalers/Mag 7 is down sharply—roughly 80% from just a few years ago—as money gets poured directly into data centers and silicon. Looking at a Morgan Stanley chart, Gerstner notes their 2027 CapEx forecast was raised from $950 billion to $1.1 trillion. When accounting for private players and neoclouds like SpaceX and CoreWeave, Gerstner estimates the actual 2027 CapEx number will sit closer to **$1.5 trillion**.
Against that $1.5 trillion in annual CapEx, combined 2027 revenue across all major AI labs is projected at around $300 billion. Gerstner poses the core concern haunting Wall Street: can the market sustain $1.5 trillion in annual CapEx for only $300 billion in inference revenue, or will a sudden loss of confidence trigger a massive pullback across the semiconductor complex? Gavin Baker and Andrew Fox immediately push back on the bear thesis, emphasizing that $300 billion in revenue is likely a significant undercount. Baker points out that at 60% to 70% gross margins, that revenue profile already yields substantial cash generation, adding that he expects the industry to end this year "well over $200 billion in inference revenue." They note that NVIDIA CEO Jensen Huang was deemed "outlandish" two years ago for predicting a $1 trillion AI buildout, yet his estimates turned out to be far too conservative.
The group outlines how tech leaders like Elon Musk, Sundar Pichai, Sam Altman, and Anthropic's Dario Amodei are all continually taking "the over" on AI revenue trajectory. Amodei previously noted that frontier labs could generate low hundreds of billions by 2028 and hit trillions in revenue before 2030. Furthermore, Baker clarifies that a large chunk of CapEx—roughly 35% to 50% depending on the lab—is dedicated to training next-generation frontier models rather than generating immediate inference revenue today. Combined with the "prisoner's dilemma" where opting out of the CapEx race is an existential threat to tech giants, the panel agrees the underlying economics remain extremely sound.
- **2027 Global AI CapEx:** Morgan Stanley estimates $1.1T (Mag 7); Gerstner's total estimate including SpaceX and CoreWeave is **$1.5 trillion**.
- **Mag 7 Free Cash Flow:** Down ~80% from peak levels due to massive infrastructure spending.
- **Projected Inference Revenue:** Industry estimates ~$300 billion by 2027; Gavin Baker expects >$200 billion by the end of this year alone.
- **CapEx Split:** ~35% of CapEx is earmarked for pure non-revenue frontier training, with the remainder going toward active inference capacity.
[01:07:09](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4029s) Token Economics, Gigawatt Monetization, and Rational Economic Actors
Clark Tang highlights a complete reversal in narrative regarding AI pricing dynamics over the last year. Going into the year, consensus held that token pricing and compute costs would experience a smooth, deflationary downward path. Instead, supply shortages and booming demand for test-time compute and agentic workflows have pushed monetization rates per watt sharply upward. Tang references a metric from Alex at Well Rock: less than **0.2% of the global population** currently uses AI in an agentic way. Yet, even a single active power-user can consume "500 CPU cores in a VM instance" alongside five GPUs running 24/7. As agent adoption expands to a wider percentage of humanity, structural compute shortages are poised to persist for years.
```
+-----------------------------------------------------------------------+
| AI MONETIZATION & INFRASTRUCTURE SHIFT |
+-----------------------------------------------------------------------+
| Metric / Trend | Early Year Baseline | Current Level |
+--------------------------+-----------------------+--------------------+
| Monetization / Gigawatt | ~$20 Billion | $30B – $40B+ |
| Agentic AI Adoption | < 0.2% Global Pop. | Scaling rapidly |
| Infra Cost Dynamics | Expected Deflation | Rising Value/Watt |
+--------------------------+-----------------------+--------------------+
```
Tang and Fox explain how this bottleneck fundamentally transforms lab unit economics. Monetization per gigawatt has surged from roughly $20 billion at the start of the year to between **$30 billion and $40+ billion per gigawatt** today. Because data centers represent a heavy fixed-cost base, every dollar of incremental yield flows straight to pure operating margin. Far from suffering low-margin compression, labs like Anthropic have experienced "accidental profitability" simply because they couldn't spend money on additional compute fast enough.
Addressing critics like Chamath Palihapitiya, who argue that massive AI CapEx lacks proven return on investment (ROI) and represents a "token maxi" bubble, Gerstner counters with a fundamental market reality: millions of independent businesses and everyday consumers are voluntarily opening their wallets. Rational economic actors wouldn't continuously increase their spend on API tokens and AI subscriptions unless it was actively creating measurable value and productivity gains in their operations.
> "The first argument against the frontier labs was they'll never generate revenue... Then it was like, even if they generate revenue, it'll be really shitty gross margins... To me, the best evidence as to why I think this revenue can continue is why are millions of independent businesses and consumers all choosing to do the same thing? They're not dumb." — **Brad Gerstner**
- **Monetization Surge:** Revenue per gigawatt of compute capacity increased from ~$20B to **$30B–$40B+**.
- **Agent Resource Demands:** A single agentic power-user can draw on 500 CPU cores and 5 GPUs continuously.
- **Global Penetration:** Less than **0.2%** of the world's population currently uses agentic AI.
Transitioning to portfolio management and broader equity markets, Gerstner and Baker conduct a transparent "market check" on current valuations and risk exposure. The market has displayed wild internal dispersion: while semiconductor stocks have "ripped" with widespread doubles and triples, the software sector is down ~8% on the year and broader internet stocks are down ~16%. Key indices like the S&P 500 and Nasdaq remain elevated almost entirely on the back of AI- and compute-linked megacaps. Given the recent run-up in asset prices and emerging macroeconomic risks, Altimeter has prudently dialed back its risk exposure from "large" to "medium-small."
```
+-----------------------------------------------------------------------+
| YEAR-TO-DATE MARKET DISPERSION |
+-----------------------------------------------------------------------+
| Asset Class / Sector | YTD Performance Trend |
+-----------------------------------+-----------------------------------+
| Semiconductors & AI Compute | Massive Rally (2x - 3x gains) |
| Software Index | Down ~8% |
| Internet Sector | Down ~16% |
| Broad Market (S&P 500 / Nasdaq) | Up (Driven by AI Megacaps) |
+-----------------------------------+-----------------------------------+
```
Gavin Baker uses a vivid running metaphor to describe the stock market's current trajectory. Coming out of 2022, equities were like a runner with downhill momentum and pent-up energy. Over the past few months, however, stocks have sprinted up an extremely steep hill, with certain AI bottlenecks going "straight up a cliff." Baker warns that these stocks are now exhausted and need to rest—either by hanging out in their harness at the top of the cliff or retracing downward. Interestingly, major market pillars like NVIDIA and Broadcom have actually been relative laggards during the most recent leg of the rally.
The discussion turns to macro headwinds that would typically crush equity markets: inflation re-accelerating, geopolitical conflict in the Middle East, high energy prices, and elevated interest rates. CPI prints recently ticked back above **4.2%**, with core inflation holding firm. Under normal conditions, investors would want nothing to do with such a tape. However, underlying AI revenue growth and compute demand have completely outweighed macro drag.
Baker notes a potential short-term seasonal lull, mentioning that AI usage historically slows during summer months as college students go off-campus—though Gerstner jokes that his own kid is currently "burning compute 24/7" building multi-agent swarms and a SpaceX DCF model ahead of attending the SpaceX IPO on Friday. Looking ahead, Baker stresses remaining hyper-vigilant: "I always assume a bullet is coming for me... It's the bullet you don't see that gets you."
- **Sector Performance:** Software down **8%**, Internet down **15-16%**, while AI semiconductor plays surged 2x to 3x.
- **Inflation Metrics:** Headline CPI printed at **4.2%**, signaling ongoing pressure on interest rates and core PCE.
- **Altimeter Exposure:** Reduced from "large" to "medium-small" to harvest gains and manage macroeconomic risks.
<FollowUp label="Want to dive into the final section on Fable 5, Mythos, and the open-source debate?" query="Walk me through the next part of the transcript covering Fable 5, Mythos, and open-source models."/>
[01:17:22](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4642s) Fable Capabilities and the Accelerating Pace of Creative Destruction
Reflecting on Noam Brown's insights and observing the sheer capabilities demonstrated by Fable, Brad Gerstner notes that it remains nearly impossible to take a long-term bearish stance on the technology. Turning to the younger investors in the room, Brad offers a perspective on how unprecedented the current tech landscape actually is. Newer entrants to the industry might assume this hyper-accelerated environment is business as usual, but the steepening curve of creative destruction and the scaling advantages currently at play are beyond anything historical models predicted.
Brad frames this shift by looking at the staggering numbers over the past two decades. While it took the Magnificent 7 over 20 years to generate their first trillion dollars in revenue, they added another $1 trillion of revenue in just the last seven years alone—a feat that created roughly $17 trillion in market value.
* **Revenue Acceleration:** The Magnificent 7 added $1T in revenue over the last 7 years, generating an estimated $17T in market capitalization.
* **First Trillion Milestone:** It took over two decades for the same group of companies to reach their first $1T in total revenue.
---
[01:18:13](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4693s) The Next $1 Trillion and the $10 Trillion Prize
Looking ahead, Brad emphasizes that the rate of compounding is doubling down again. Forecasts suggest that the next $1 trillion in tech revenue won't require seven massive conglomerates or two decades; instead, it is projected to come from just three companies—SpaceX, Anthropic, and OpenAI—over the next four to five years. While market volatility and temporary bumps in the road are inevitable, the trajectory points toward historic highs because the ultimate prize is so massive: AI and frontier technologies are poised to transform 5% to 15% of global GDP, which at a 10% baseline represents a staggering $10 trillion market opportunity.
This technological expansion comes with a societal mandate. Brad stresses the need to ensure America maintains its lead while actively evolving the social contract so that economic progress "lifts the floor" and brings everyone along for the ride. Gavin Baker voices strong appreciation for Brad's advocacy surrounding "Trump Accounts" (investing accounts for children), highlighting how giving young people an equity stake to compound over their lifetimes is critical for broader economic alignment.
* **Next $1T Projection:** $1T in new revenue is expected to be generated in just 4 to 5 years by only three private powerhouses: SpaceX, Anthropic, and OpenAI.
* **Global GDP Impact:** Transforming 10% of global GDP represents an addressable market value of roughly $10T.
* **Collaborative Culture:** Gavin and Brad wrap up the episode by highlighting the deep collegiality, shared open-source research philosophy, and mutual respect between Atreides Management and Altimeter Capital, with Gavin noting that team members like Clark Tang and Andrew Fox ("Foxy") spend extensive time collaborating behind the scenes.
## Notable quotes
> [00:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=0s) If you're AI pilled, that means we got to build a lot more compute than the world thinks... I don't know another entrepreneur or another business that's a better bet on the future than SpaceX.
> [03:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=180s) Speed is literally cost because every day you're paying electricians and plumbers... they're now monetizing them at arguably the highest rate.
> [10:38](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=638s) Better, faster, cheaper has been a winning formula... wherever I am in the world, Starlink is the best connection.
> [29:25](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=1765s) Coding may be the fastest path to AGI and ASI because if a model's good at coding, you can write code to do anything. — Gavin Baker
> [37:00](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2220s) He and Gwynne and the team are great at business... I don't know that any of the Mag 7 could have moved that quickly to adjust the business that they did. It's exceptionally entrepreneurial at scale. — Brad Gerstner
> [43:56](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=2636s) Imagine if Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn't have to eat, he doesn't have to sleep, he doesn't have to relax... and he thought for 1 year. We might already have solved a lot of these intractable problems. — Gavin Baker
> [51:21](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=3081s) You don't need Albert Einstein to book you a trip. You don't need Albert Einstein to do KYC.
> [01:07:09](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4029s) The first argument against the frontier labs was they'll never generate revenue... Then it was like, even if they generate revenue, it'll be really shitty gross margins... To me, the best evidence as to why I think this revenue can continue is why are millions of independent businesses and consumers all choosing to do the same thing? They're not dumb. — Brad Gerstner
> [01:11:52](https://www.youtube.com/watch?v=Tx9jT2c6e3U&t=4312s) I always assume a bullet is coming for me... It's the bullet you don't see that gets you. — Gavin Baker
## Mentioned
SpaceX, xAI, Google, Anthropic, OpenAI, Meta, Nvidia, Cursor, Starlink, Elon Musk, Brad Gerstner, Gavin Baker, Andrew Fox, Clark Tang, Jensen Huang, Gwynne Shotwell, Noam Brown, Amjad Masad, Sundar Pichai, Sam Altman, Dario Amodei, Chamath Palihapitiya, Atreides Management, Altimeter Capital, Starship, Falcon 9, Colossus, Colossus 2, Fable 5, Mythos, ChatGPT 5.5, Grok 4.3, Opus 4.7, Opus 4.8, Codex, Composer 2, Composer 2.5, Kimi K2.5, Claude 3.5 Sonnet, Claude 3.5 Haiku, Nemotron, ChemK 0.25, Vera Rubin, H100, Goldman Sachs, The Wall Street Journal, Morgan Stanley, GE Vernova, Oracle, CoreWeave, Nebius, Iren, Replit, Moonshot, Stripe, Harvey, Fireworks, JPMorgan, Broadcom, AMD, Cerebras, MediaTek, TSMC
Altimeter and Atreides investors argue that surging compute monetization and unparalleled build speeds justify the massive AI capital expenditure cycle, positioning SpaceX as a premier dual bet on space infrastructure and frontier compute scaling.
SpaceX IPO Anchored by Launch and Starlink Scaling (0:00)
SpaceX's planned initial public offering targets a $135 share price, implying a $1.77 trillion valuation, alongside Wall Street projections of $160 billion in revenue by 2028, up from $18 billion. Andrew Fox and Brad Gerstner attribute this target to the fundamental transformation of launch economics through Starship, where rapid reusability across 30 to 50 flights per vehicle is projected to compress launch costs from $1,500 per kilogram on Falcon 9 to below $250 per kilogram, eventually approaching the cost of fuel.
Simultaneously, Starlink broadband and Direct-to-Cell services serve as the near-term cash engine. While currently serving under 1% of global households, Starlink connectivity revenues are forecasted to grow from $10 billion to $50 billion by 2028, capturing roughly 0.3% of the global telecommunications market with superior latency and unit costs.
Execution Speed Drives Terrestrial and Orbital Compute Arbitrage (21:05)
Clark Tang highlights that SpaceX and xAI's rapid infrastructure deployment yields estimated 55% internal rates of return on compute assets. Elon Musk's operational pace—standing up the 100,000-GPU Colossus cluster in 19 days and standard facilities in 122 days—sharply reduces labor overhead and equipment financing drag, enabling xAI's hosting deals to generate higher operating profit per gigawatt than arrangements managed by Meta, Google, or Anthropic.
Looking beyond terrestrial limits, Tang compares the physical economics of ground versus space compute. On Earth, a gigawatt of compute capacity costs roughly $60 billion—$35 billion for silicon and $25 billion for land, power, and cooling. Starship's capacity to lift 100 metric tons (roughly 5 megawatts of compute) per flight could reduce non-silicon infrastructure CapEx to approximately $5 billion per gigawatt, yielding a 5x facility CapEx reduction, provided hardware degradation from radiation and thermal constraints remains manageable.
Frontier Model Hegemony and Proprietary Coding Data (29:25)
Gavin Baker emphasizes that frontier software value adheres strictly to Pareto efficiency, where the top-performing model captures almost all high-value revenue. While open-source models dominate aggregate token volume, the panel notes that over 90% of industry dollar revenue concentrates among leading frontier labs like Anthropic, OpenAI, Google, and xAI. Enterprises increasingly implement model routers to offload simple queries to cheap open-source systems while routing high-stakes tasks to frontier models.
xAI's acquisition of Cursor reinforces this dynamic by integrating massive proprietary coding datasets directly into pre-training for models like the 1.5-trillion parameter Grok 4.3, rather than relying solely on post-training adjustments. Coding capabilities are viewed by the panel as a critical shortcut toward artificial superintelligence, creating a self-reinforcing flywheel between compute infrastructure and software monetization.
CapEx Sustainability, Silicon Efficiency, and Macro Risks (63:03)
Addressing Wall Street skepticism over a projected $1.5 trillion in global AI capital expenditures by 2027 against an estimated $300 billion in direct inference revenue, Baker and Fox argue revenue is substantially underestimated. They note that roughly 35% to 50% of current CapEx is dedicated to non-revenue research and pre-training. Furthermore, Clark Tang observes that monetization per gigawatt has doubled from roughly $20 billion to between $30 billion and $40 billion or more, as compute-intensive agentic workflows begin to scale.
Power bottlenecks also reinforce Nvidia's competitive moat against custom ASICs, as superior tokens-per-watt efficiency yields higher top-line revenue per capped data center megawatt. Despite structural optimism, Gerstner notes sharp market dispersion—semiconductors surging while software declined 8%—and persistent 4.2% CPI inflation, prompting Altimeter to trim equity exposure from large to medium-small.
While the panel projects that the next $1 trillion in tech revenue could be generated in four to five years by SpaceX, OpenAI, and Anthropic alone, near-term equity markets face significant macroeconomic drag from inflation and severe valuation dispersion that warrant disciplined risk management.
[0:00] If you're AI pilled, that means we got to build a lot more compute than the world thinks... I don't know another entrepreneur or another business that's a better bet on the future than SpaceX.
[3:00] Speed is literally cost because every day you're paying electricians and plumbers... they're now monetizing them at arguably the highest rate.
[10:38] Better, faster, cheaper has been a winning formula... wherever I am in the world, Starlink is the best connection.
[29:25] Gavin Baker: Coding may be the fastest path to AGI and ASI because if a model's good at coding, you can write code to do anything.
[37:00] Brad Gerstner: He and Gwynne and the team are great at business... I don't know that any of the Mag 7 could have moved that quickly to adjust the business that they did. It's exceptionally entrepreneurial at scale.
[43:56] Gavin Baker: Imagine if Albert Einstein had just thought about fundamental physics 24 hours a day. He doesn't have to eat, he doesn't have to sleep, he doesn't have to relax... and he thought for 1 year. We might already have solved a lot of these intractable problems.
[51:21] You don't need Albert Einstein to book you a trip. You don't need Albert Einstein to do KYC.
[67:09] Brad Gerstner: The first argument against the frontier labs was they'll never generate revenue... Then it was like, even if they generate revenue, it'll be really shitty gross margins... To me, the best evidence as to why I think this revenue can continue is why are millions of independent businesses and consumers all choosing to do the same thing? They're not dumb.
[71:52] Gavin Baker: I always assume a bullet is coming for me... It's the bullet you don't see that gets you.
Original transcript
[0:00] And I think we're all pretty AI pilled.
[0:02] And if you're AI pilled, that means we
[0:03] got to build a lot more compute than the
[0:05] world thinks. And that these models are
[0:07] going to be a lot more valuable than
[0:08] people think. You combine that with
[0:10] their core business, I don't know
[0:12] another entrepreneur or another business
[0:14] that's a better bet on the future,
[0:17] right, than SpaceX. And so I think for
[0:19] most institutional investors, it's a
[0:21] must buy, a must own, a set it and
[0:24] forget it, right, in order to have a a
[0:27] real bet on both the space and the AI
[0:29] future.
[0:40] All right, here we go.
[0:42] Early morning Silicon Valley, BG2 is
[0:45] back. We're chopping it up on all things
[0:47] tech and markets. To do that, I have
[0:49] none other than GB in the house, Gavin
[0:51] Baker from a Treaties. He's brought his
[0:53] main guy, Andrew Fox. And of course, I
[0:55] had to draft Clark Tang into the mix, my
[0:58] partner,
[0:59] um to talk to to talk about some of the
[1:00] big questions of the day. You know, how
[1:02] should we be thinking about the SpaceX
[1:04] IPO? You know, what are the big levers?
[1:06] There are big numbers out there for
[1:07] what's going to happen over the course
[1:09] of the next few years. So, let's break
[1:10] that down a bit, help simplify it for
[1:12] folks.
[1:13] I Mythos launched yesterday. I want to
[1:16] talk a little bit about like who's up,
[1:18] who's down in the race for
[1:19] superintelligence. Where are we? What
[1:21] did we learn with the Mythos launch? And
[1:23] Clark was in uh in in in Taiwan last
[1:25] week um with Jensen at Computex and GTC.
[1:28] So, what was our takeaway there? What's
[1:30] going on with GPUs, memory? Where are
[1:32] the bottlenecks? And where do we go from
[1:34] here? To start everything off, um you
[1:37] know, maybe just kick it over to you,
[1:38] Gavin, talking about the SpaceX IPO. The
[1:41] IPO is in 2 days. Uh you're a big
[1:43] shareholder. Congratulations. We're also
[1:45] a shareholder. Uh you know, we also we
[1:47] expect to be buying in the IPO. The Wall
[1:49] Street Journal's reporting
[1:51] um you know, the Goldman Sachs are both
[1:54] saying 160 billion in revenue in 2028.
[1:57] Um we know that the IPO is $135 a share,
[2:01] 1.77 trillion. Um so
[2:05] when we think about kind of what the big
[2:08] levers are, there's so many moving parts
[2:11] in this IPO. Um nobody's better than you
[2:13] at just breaking it down, simplifying
[2:15] it. What are the key levers that we
[2:16] ought to be thinking about that you're
[2:18] thinking about over the course of the
[2:19] next few years?
[2:20] >> Sure. Um so great to be here. Thank you
[2:22] for having me. I thought we're going to
[2:23] call it B BGG B,
[2:26] but [laughter] we can stick with BG2.
[2:27] I've been your house.
[2:28] >> hey.
[2:28] All subject to revision.
[2:29] >> That's okay. That's okay. So I think
[2:31] there's two
[2:33] big levers or variables that I think
[2:36] people should focus on. And you know,
[2:37] I'm not going to comment on where I
[2:39] think um those variables go.
[2:41] But one is um you guys have this chart.
[2:44] Um did you did you post this on X?
[2:46] >> I did I did before and then we also
[2:48] included a new addition with uh XAI's
[2:52] new deals as well.
[2:53] >> Yeah.
[2:53] >> Yeah.
[2:54] >> So Clark, who I've known for many years,
[2:57] um
[2:57] uh made a did a great analysis here.
[3:00] And he shows that XAI's deal um with
[3:04] Google for cloud computing
[3:07] uh generates more operating profit per
[3:08] gigawatt um than Anthropic, than Meta,
[3:11] than Google, than OpenAI. Uh their deal
[3:14] um
[3:15] >> [clears throat]
[3:15] >> actually with uh Anthropic
[3:18] also generates probably more operating
[3:21] profit than anyone but um Anthropic.
[3:25] And so you know, um the your your
[3:27] colleague at Altimeter, Freida, also she
[3:29] calculated a 55% IRR
[3:31] >> Mhm.
[3:31] >> on Claude's one.
[3:32] >> Mhm.
[3:33] >> You know, if you can borrow money at 6,
[3:34] 7, 8% and invest in something with a 55%
[3:38] ARR,
[3:39] I'm not the most sophisticated thinker,
[3:40] but that math maths.
[3:41] >> Right.
[3:42] >> And so I think the most important
[3:44] variable, one of the two most important,
[3:46] is how quickly they bring on terrestrial
[3:48] data centers.
[3:49] >> Mhm.
[3:50] >> We do know from Jensen that uh Elon
[3:52] brings data centers up faster than
[3:54] anyone 122 days. Speed is literally cost
[3:57] because every day you're paying
[3:58] electricians and plumbers. That's cost.
[4:01] And they're now monetizing them at
[4:04] arguably the highest rate. And so I
[4:06] think, you know, everybody should run
[4:09] their own math on that, but that is a
[4:10] massive variable.
[4:12] >> Yes.
[4:12] >> Truly massive variable. The second thing
[4:15] is, you know, we have a chart and it's
[4:17] wildly out of date now. It's kind of
[4:19] freaking amazing. This chart is I think
[4:21] is this chart from 10 days ago?
[4:23] But it like the 10 or 12 days since this
[4:26] chart since we made this chart
[4:28] which shows the Pareto curves for Opus
[4:30] 4.7 for coding
[4:32] for Codex from OpenAI. And now we've had
[4:36] Opus 4.8. It was already out of date.
[4:38] And now we have Fable
[4:40] >> Totally.
[4:40] >> and Mythos, which is freaking wild. In
[4:42] 10 days
[4:43] >> Yes.
[4:43] >> like we would have had to update the
[4:44] chart twice.
[4:45] >> Right.
[4:46] >> But what the Pareto curve sure shows is
[4:48] how much intelligence you can get for a
[4:50] given amount of cost. And I do think
[4:52] being all revenue will accrue to the
[4:54] Pareto curve. All at least kind of
[4:56] frontier model revenue will accrue to
[4:57] the Pareto curve. And this is Pareto
[4:59] curve for code coding. And what I think
[5:01] is so impressive
[5:04] is that
[5:05] you can see in the chart that Composer 2
[5:07] was Pareto dominant
[5:09] or, you know, at the lowest level of
[5:11] intelligence with very little training.
[5:14] This just reflects, and I know you know
[5:15] Cursor well. I think you know Cursor
[5:16] [ __ ] a lot better than I do.
[5:18] A vast amount better than I do.
[5:19] [laughter]
[5:21] But my understanding is that Cursor and
[5:24] Anthropic have more tokens of
[5:26] proprietary coding data than anyone
[5:28] else. And they have more tokens of
[5:30] proprietary coding data than exist on
[5:32] the public internet.
[5:33] And so they fed Cursor fed
[5:36] um used ChemK 0.25, used their own
[5:39] private data, did some RL, some
[5:41] supervised fine-tuning
[5:43] and they got a really good model. And
[5:44] then they spent 3 weeks in the Colossus
[5:46] 2 cluster and they got a model that
[5:50] 12 days ago was Pareto dominant with
[5:52] Composer 2.5. That's on their own
[5:54] benchmark.
[5:55] Um Cursor bench, so maybe take it with a
[5:58] grain of salt. But I think this just
[6:00] suggests that the Cursor data is very
[6:03] valuable for coding and when it is
[6:07] trained, you know, to Chinchilla optimal
[6:09] or beyond Chinchilla optimal with
[6:10] reinforcement learning,
[6:12] you know, I think it suggests that XAI
[6:14] and SpaceX AI has a shot of being a real
[6:16] player in coding.
[6:17] >> I mean, I think one of the interesting
[6:18] things is, you know, we
[6:21] you know, the way he answered the
[6:22] question, right? We didn't talk about
[6:23] launch.
[6:24] Right? We didn't talk about Starlink or
[6:26] communications. Those up until really 6
[6:29] months ago were the business.
[6:31] >> Yeah.
[6:31] >> Right? And, you know, and then we merged
[6:34] in x.ai and we merged in Cursor and then
[6:37] we announced these deals where it was
[6:39] very clear he was kind of building AWS
[6:41] right under our nose, you know, in in in
[6:43] in terms of this. But what I want to do
[6:45] is go to go to Fox. Give us the
[6:47] breakdown. Three big lines of business,
[6:49] right? We've got the the the
[6:51] communication Starlink launch business,
[6:53] we've got the, you know, AI compute
[6:55] business and then I want to come back to
[6:56] x.ai that you were just clicking on. But
[6:58] if we just go to the core business, what
[7:01] do we have to assume goes right in the
[7:03] core business both with launch and with
[7:05] Starlink in order to achieve the numbers
[7:07] that are out there?
[7:08] >> Yeah, sure. So, look, I think
[7:10] the thing that's foundational to
[7:11] everything is the launch business.
[7:13] >> Right.
[7:13] >> Right? This is the kind of
[7:15] crown jewel of SpaceX. Um it's something
[7:18] that no one else really has, notably
[7:20] reusability.
[7:21] >> Right.
[7:21] >> And soon rapid reusability. Right? This
[7:24] is, I think, what you need to believe in
[7:27] to get to the economics in AI that make
[7:31] orbital compute something that's very
[7:33] economically attractive.
[7:34] >> Right.
[7:35] >> Outside of the idea that we are in
[7:36] shortage for power, shortage for chips.
[7:40] Right? Um so, I think rapid reusability
[7:42] is the main thing that we're watching
[7:44] for and I think most people should watch
[7:45] for. Um,
[7:47] you know, Elon talks about it a lot, but
[7:49] getting these rockets to fly at a
[7:50] cadence that's comparable to an airline,
[7:52] right? And and Gavin has used this
[7:53] analogy before, but
[7:55] um, the old rocket industry was kind of
[7:57] like, imagine boarding a plane, flying
[8:00] to California, getting off the plane,
[8:01] the plane explodes
[8:03] after. Um, so I think what SpaceX are
[8:06] ultimately trying to achieve is have a
[8:08] Starship fly both stages, not just the
[8:11] booster. Um,
[8:13] uh,
[8:14] 30, 40, 50 times
[8:17] before you have to retrofit that ship.
[8:20] Um,
[8:20] and when you do that, you're amortizing
[8:23] the cost of the vehicle over many
[8:25] flights, right? And that's what brings
[8:27] the cost down significantly.
[8:29] Um,
[8:29] >> But that's a really hard problem to
[8:31] solve.
[8:32] >> Extremely difficult and look, I think
[8:35] the company, you know, have
[8:37] been loud and clear, they're going to
[8:38] attempt to bring back the second stage
[8:40] >> Right.
[8:41] >> of Starship later this year.
[8:42] >> Right.
[8:43] >> Um, and then make it reusable, you know,
[8:46] re-fly the second stage next year. Um,
[8:49] and from there, ramp up the cadence. But
[8:52] at the end of the day, driving down the
[8:54] cost of launch
[8:56] is what enables all of these other
[8:58] businesses and is what makes them so
[9:00] attractive relative to incumbents.
[9:03] >> So how many how many times are we
[9:04] Starship just launched Starship 3, you
[9:06] know, just launched. How many launches
[9:09] are you know, do you think kind of the
[9:10] consensus out there is assuming, you
[9:13] know, two or three years from now? Like
[9:14] what is the launch cadence? Are we
[9:16] launching one of these every day or we
[9:18] launching one of these every week or
[9:19] every month? Like where are we in terms
[9:21] of expectations?
[9:22] >> Yeah, so look, I think expectations for
[9:24] now, you know, we're going from, you
[9:27] know, call it 160 165 launches last year
[9:30] up into the high hundreds of launches in
[9:32] several years and, you know, getting
[9:34] into the thousands of launches probably
[9:36] in the next 3 years thereafter.
[9:38] >> Okay.
[9:38] >> Um I think the company have aspirations.
[9:41] >> Thousands of launches, you're launching,
[9:43] you're doing two or three launches a
[9:44] day.
[9:45] >> Right.
[9:45] >> Right. And then talk to us a little bit,
[9:48] what is this enable? Obviously, you
[9:50] know, I'm here in Silicon Valley. I
[9:51] can't even I can't even keep a call on
[9:54] Sand Hill Road, two decades into the
[9:57] mobile revolution. I mean, it's the
[9:58] craziest thing. It's like a third world.
[10:01] >> It's It's a major business problem when
[10:03] you're freaking out here. [laughter]
[10:04] >> It's crazy. It's crazy. Right by the
[10:06] Starwood dead zone. And I'm like, how
[10:08] can this possibly be? So almost like
[10:10] it's a joke. It's the epicenter of
[10:11] technology in America and you can't
[10:13] maintain a call. Okay, so we're all
[10:16] going to switch to Starlink mobile when
[10:17] it comes along because I don't want to
[10:18] lose that call on Sand Hill Road. So
[10:21] walk me through a little bit just again
[10:23] high-level. Um it's a big portion of the
[10:26] revenue growth expected in the business
[10:28] over the course of the next two to three
[10:30] years. My hunch is a lot of this is
[10:32] driven uh by uh by direct to cell
[10:35] connectivity. Walk me through a little
[10:37] bit those economics.
[10:38] >> Yeah, so look, it's actually
[10:39] interesting. Um
[10:41] the broadband business is still very
[10:43] early stage when you think about um the
[10:45] percent of households that have actually
[10:47] been penetrated to date. You look at the
[10:49] percent of global households with
[10:51] Starlink, it's less than 1%. Uh and
[10:53] that's the broadband, you know, you kind
[10:55] of have a base terminal at your house,
[10:57] on your car, on your boat, um and then
[11:00] airlines now as well. Um
[11:02] so I actually think broadband can scale
[11:04] to hundreds of millions of terminals,
[11:06] hundreds of millions of users. And today
[11:09] the subscriber base
[11:09] >> Hundreds of millions if if they get
[11:13] rapid reusability of Starship, which is
[11:16] really hard.
[11:17] Um
[11:18] you know, if if there's not competition.
[11:21] Um hundreds of millions, um it's
[11:22] possible.
[11:24] But
[11:25] maybe
[11:25] >> I I always say around here, it's funny.
[11:27] I love seeing PM and kind of analysts in
[11:29] in this situation. It's exactly what I
[11:31] do with Clark. Clark will say something,
[11:32] I'll say the future is a distribution of
[11:34] unknown probabilities. It's either more
[11:36] likely or less likely, so give me the
[11:37] distribution. Are we talking 20% 30%?
[11:40] It's hilarious. It's the same
[11:41] >> Well, no, 100% same thing. And like I've
[11:43] watched Elon do many hard things and
[11:44] this is a really hard thing. So, I think
[11:46] it's reasonable to think that they're
[11:48] going to succeed with rapid reusability,
[11:50] but just I just think it's important to
[11:52] acknowledge that like
[11:54] orbital compute
[11:56] you know, Starlinks, you know, Starlink
[11:58] V3, Starlink direct to cell, we need we
[12:02] need first reusability for Starship V3
[12:04] and then rapid reusability unlocks a lot
[12:06] of this.
[12:07] >> Right. When I see when I see the models
[12:09] that the banks are putting out there,
[12:11] right? And Wall Street Journal,
[12:12] everybody's reported on these. These
[12:13] same things have been widely leaked.
[12:16] They they they largely have the revenue
[12:18] on connectivity, so let's call it
[12:20] Starlink direct to cell, etc. Going
[12:22] from, you know, let's call it 10 billion
[12:24] to 50 billion uh by 2028. And so, I'm
[12:28] not asking you guys to react to, you
[12:30] know, to tell me your specific numbers,
[12:32] but when I'm talking to Clark all I'm
[12:34] trying to size up is order of magnitude.
[12:37] Do we think we can 5x the business over
[12:40] the course of the next 3 years? Is there
[12:41] enough TAM both in terms of broadband
[12:44] and direct to consumer? And I think the
[12:45] answer to that is yes.
[12:46] >> Yeah, here's what I just say very simply
[12:48] is I have um
[12:50] I I travel with Starlink. Um
[12:53] I'm I'm a big video gamer and very
[12:55] consistently wherever I am in the world,
[12:57] Starlink is the best connection.
[12:58] >> Yes.
[12:59] >> It's the fastest, it's lowest latency
[13:01] and I do think once they get to rapid
[13:04] reusability, it's also going to be
[13:05] they're going to have the cheapest cost
[13:08] per gigabyte or megabyte delivered um
[13:11] and better faster, cheaper has been a
[13:14] winning formula.
[13:15] And so, 50 billion, that's, you know,
[13:18] 0.3%
[13:19] penetration of the global telecom
[13:21] market. Now, maybe there's some
[13:22] deflation with Starlink pricing.
[13:24] Um but that's the way I'd frame it up.
[13:26] >> Yeah. I like betting on better, faster,
[13:28] cheaper.
[13:28] >> Um, Clark, I would say probably the
[13:30] biggest surprise of the last six six
[13:32] weeks
[13:34] is that Elon, you know, we talked about
[13:36] it on all-in podcast, we called it EWS,
[13:38] Elon web services, right? That that he
[13:41] struck these huge deals with Anthropic
[13:44] and Google.
[13:45] I don't even think people were thinking
[13:47] about SpaceX in the AI compute game,
[13:50] right? We If you looked at the models as
[13:53] of a few months ago, it was
[13:55] connectivity, so Starlink, and then it
[13:57] was x.ai, the model.
[13:59] But this whole category of taking all of
[14:02] this compute, which he's uniquely good
[14:04] at standing up, right? And then
[14:06] reselling it in a way that's highly
[14:08] profitable was not in a lot of people's
[14:10] forecast. Now it's a major component of
[14:13] the forecast. You know, you and I did
[14:15] this podcast with Jensen, where Jensen
[14:18] said Elon is an N of 1.
[14:20] >> What they achieved is is singular. Never
[14:22] been done before. Just to put in
[14:24] perspective, 100,000 GPUs, that's
[14:27] you know, easily the fastest
[14:29] supercomputer on the planet as one
[14:30] cluster.
[14:31] Um, a supercomputer uh,
[14:34] that you would build would take normally
[14:37] 3 years to plan.
[14:39] >> Right.
[14:39] >> And then they deliver the equipment, and
[14:42] it takes 1 year
[14:44] to get it all working.
[14:46] Yes. We're talking about 19 days.
[14:49] >> Wow.
[14:49] >> N of 1 is right. Elon is an N of 1.
[14:52] >> And his ability to secure supply, stand
[14:55] up the supply, you know, deploy it in a
[14:58] way that's uh, you know, coherent and
[15:00] effective for both himself and I guess
[15:02] now for others. So, walk us through kind
[15:05] of that. It looks to me again like this
[15:08] is a major component of the revenue
[15:09] story.
[15:10] >> Totally. I mean, so we we were all at
[15:13] the macro hard data center, and it was
[15:16] just very evident the amount of
[15:18] engineering that was that had gone into
[15:20] building these sites.
[15:22] Um,
[15:23] you people always talk about Google and
[15:24] their ability to to build a TPU and sell
[15:27] the TPU to Anthropic to generate
[15:29] revenues for AI.
[15:31] I think it's a pretty similar dynamic
[15:33] here with Elon able to secure power, uh
[15:37] build these sites faster than anyone
[15:39] else and also be able now to monetize it
[15:42] to um to the this massive AI market
[15:45] that's ahead of us.
[15:47] Um
[15:48] if you look at the the relationships
[15:51] that he's forged with a lot of his
[15:53] suppliers, you know, be it Jensen, be
[15:55] it, you know, all of these different um
[15:58] different sites that actually want xAI
[16:00] as a tenant. Um his his ability to
[16:03] finance these deals at at very
[16:05] attractive uh financing rates relative
[16:07] to a lot of the other players in in the
[16:10] space, you know, these are advantages
[16:12] that compound over time. And when you've
[16:14] built the credibility to stand up these
[16:16] sites and monetize at these levels, um
[16:19] you know, it's very it's actually a very
[16:20] attractive uh
[16:22] very attractive proposition for for a
[16:24] lot of folks involved. Um and actually,
[16:27] you know,
[16:28] if you look at the these deals in
[16:29] particular,
[16:31] Gavin, you you pointed out, but, you
[16:32] know, they're they're actually
[16:33] monetizing, you know, perhaps better
[16:36] than
[16:37] other players in the space by selling
[16:38] this infrastruc-
[16:39] >> a lot higher.
[16:40] >> Um Google is obviously paying SpaceX a
[16:42] huge premium for this compute. Fox, you
[16:45] said something that I thought was really
[16:47] important, which is, you know, it may
[16:49] very well be that in order to get, you
[16:52] know, first in line on space compute,
[16:54] which Google certainly wants to do, that
[16:57] they're willing to pay a premium for
[16:58] their terrestrial compute. And so, to
[17:01] me, that's how you kind of square the
[17:02] circle as to why the premium. Any
[17:04] thoughts?
[17:05] >> Yeah, look, I think there's some of that
[17:06] embedded there, but um look, at the end
[17:08] of the day, SpaceX can stand up compute
[17:11] quickly.
[17:12] They can stand it up coherently. And
[17:14] they can stand up a lot of it in one
[17:15] place and have it readily available. So,
[17:17] look, I think that's most of the
[17:18] premium, but outside of that certainly
[17:21] people are going to space over time.
[17:22] >> I have to pay a little call option to
[17:24] get first in line for space.
[17:25] >> There you go. Good one.
[17:25] >> We've all been investing in the neo
[17:27] cloud space. So, like there's a
[17:28] fundamental belief around this table I
[17:30] that that we lack the compute needed to
[17:34] continue to push the frontier on
[17:36] intelligence. So, we have to build a lot
[17:38] of compute, okay? Now there's a there's
[17:40] competition going on. On one end you
[17:42] have the hyper scalers who are building
[17:44] out that capability. Then we have AI
[17:46] dedicated clouds that are building out
[17:48] that capability. And now literally in a
[17:50] matter of weeks, right, we have a a you
[17:53] know a giant that's emerged in this
[17:55] category which is SpaceX.
[17:58] The question to you Gavin is can they
[18:00] consolidate this market, right? Because
[18:03] if I think about a marketplace, Elon has
[18:05] a unique ability to get the supply. He
[18:07] has a unique ability to cut deals on the
[18:09] other side and nobody can stand it up
[18:12] like he can stand it up. So, I think
[18:14] there might be a real consolidation in
[18:16] the AI compute market where you have the
[18:18] hyper scalers on the one hand and on the
[18:20] other hand, you know, he may emerge as
[18:23] the largest, strongest player in the AI
[18:25] compute market.
[18:27] >> Yeah, so I think they're are they the
[18:29] number four number five hyper scaler
[18:30] today after the Google deal?
[18:32] >> Um it will be number four.
[18:35] >> Kind of wild.
[18:36] >> Yeah.
[18:36] >> In 30 days
[18:38] we went from not being an AI hyper
[18:40] scaler to being number four. And we
[18:43] passed a lot of companies including
[18:45] Oracle.
[18:45] >> Coreweave is a huge business, right?
[18:47] That we're we're investors in, you know,
[18:49] and have been investors in, right? But
[18:51] there are a lot of other players, the
[18:52] Nebius's of the world, the Iron's of the
[18:54] world. And I would say that they're
[18:55] probably 50 neo labs being funded in
[18:57] Silicon Valley right now as we speak
[18:59] because of the shortage in compute.
[19:01] >> Absolutely.
[19:03] So, that's
[19:04] kind of crazy in 30 days. That's just
[19:06] extraordinary.
[19:08] What I would say is there I think there
[19:09] is a belief that these data centers are
[19:11] commodities.
[19:12] >> Mhm.
[19:13] >> And I do not share that belief.
[19:15] Um I don't think anybody around this
[19:17] table shares that belief.
[19:19] And in the same way that Elon was able
[19:21] to re-engineer a rocket from first
[19:23] principles and make it reusable,
[19:25] he engineered an electric car from first
[19:27] principles. You know, everyone else was
[19:28] trying to, you know, make an electric
[19:30] car like an internal combustion engine
[19:32] car and he thought about it differently.
[19:34] And um
[19:36] I think he looked at data center design
[19:39] from first principles and he designed
[19:41] something fundamentally different. And I
[19:43] did actually ask the team. I said, "Hey
[19:45] guys, maybe I'd be a little less public
[19:49] about things that are very obvious to
[19:51] you [laughter]
[19:52] >> Right.
[19:52] >> about how to design a data center, but
[19:54] are revelations to other people
[19:57] because I think what you're doing is
[19:59] maybe um
[20:00] more differentiated than you perhaps
[20:02] realize cuz what you're doing is so
[20:03] logical to you,
[20:05] but maybe lot not logical to everyone
[20:07] else. And that's how he was able to do
[20:09] it in 122 days.
[20:10] >> Yeah, to I mean to that point, Brad,
[20:12] yesterday we were meeting one of our
[20:13] portfolio companies and we were talking
[20:15] about behind the meter and we're, you
[20:16] know, really thinking about it. There's
[20:18] only maybe two or three two or three
[20:21] players now that can actually reliably
[20:23] engineer behind the meter data center.
[20:26] And you know, there's real engineering
[20:27] work that goes into all of this. So, if
[20:29] you think about this,
[20:30] if you're a gas combustion if you're
[20:31] Vernova and you say we only have a
[20:33] certain number of gas combustion
[20:35] engines. Now, we can sell them to x.ai
[20:38] or we can sell them to one of these
[20:40] startup neo clouds. Who are you going to
[20:42] sell them to?
[20:43] >> Well, and there's another dynamic.
[20:44] Everyone starts making more money when
[20:46] the GPUs get energized and sold faster.
[20:49] So, literally speed is money for all of
[20:51] the suppliers. Power, land, turbines.
[20:55] So,
[20:56] I think it's we'll we'll see.
[20:58] >> Right.
[20:58] >> Hey Brad, man.
[20:58] >> And but but this is just we're just
[21:00] talking terrestrial. I do I do want to
[21:02] hit on and then you can flip it back on
[21:04] me. Talk to me, okay, So, let's let's
[21:07] assume, right, that they continue to
[21:10] build out the terrestrial landscape.
[21:13] They continue to find buyers for that.
[21:15] Um, walk us through, you know, what this
[21:18] unlocks, you know, and how this is
[21:19] related to space data centers because I
[21:21] think, you know, once you start talking
[21:23] terrafab capacity and beyond. So, we're
[21:25] talking 1,000 gigs, right? And this year
[21:27] what what we're doing, 25 or 30 gigs
[21:29] just to put it all in perspective.
[21:31] >> 20, yeah.
[21:31] >> Right?
[21:32] >> 20, 25 gigs.
[21:33] >> Okay, so so once we start scaling up,
[21:36] walk us through,
[21:38] do we have to have space data centers in
[21:41] order to get excited about buying the
[21:43] IPO, right? And then there's obviously
[21:46] this this debate in the world. I I heard
[21:48] Jeff Bezos say, you know, I think it's
[21:49] more like 6 years, but Elon's going to
[21:51] say three because if if he says six,
[21:53] then it will take even longer. So, say
[21:55] three and we may get it in four or five.
[21:57] But are space data centers integral and
[22:00] essential to, you know, the IPO? And
[22:02] what do you think the timeline is, Erin?
[22:04] There are you guys.
[22:05] >> So, I don't think I think if you think
[22:07] about those variables around what
[22:10] Crusher could mean for XAI.
[22:13] And we do have an existence proof that
[22:15] once you really get on that Pareto
[22:17] frontier,
[22:18] revenue can scale rapidly and it's
[22:20] called Entropic. And there does seem to
[22:22] be an exhaust There seems to be a lot of
[22:24] demand for coding. And I do think I'm
[22:26] John Massad posted something very
[22:28] interesting.
[22:29] >> The The founder of Replit.
[22:30] >> The founder of Replit. It he called it
[22:32] bitter lesson adjacent that coding may
[22:35] be the fastest path to AGI and ASI
[22:38] because if you really go to coding, you
[22:40] can write code if a model's good at
[22:42] coding to do anything. So, I think
[22:43] that's a profound point and I think
[22:45] coding is going to continue to be very
[22:46] important. So, I think if you think
[22:48] about that variable,
[22:49] if you think about Starlink direct to
[22:51] cell enabled by Starlink V3, and you
[22:54] think about how quickly they can or
[22:57] cannot bring on terrestrial compute, I I
[23:00] think orbital compute is is
[23:02] is necessary for the IPO valuation,
[23:06] but it's certainly important and it's
[23:08] >> Well, maybe another way to say it is you
[23:11] think you
[23:12] you may think we're going to get to ASI
[23:15] faster than we're going to get to
[23:17] orbital compute. That may take us from
[23:19] 300 IQ to 400 IQ, 500 IQ, and beyond.
[23:24] Um and the ability to scale it up to
[23:25] consume 10% of, you know, global GDP,
[23:28] but maybe maybe that's where we should
[23:30] move next.
[23:31] >> No, no, I think on the orbital compute,
[23:32] I think Foxy would be great our Clark to
[23:34] lay out the math from first principles
[23:37] on, you know, Clark has this great chart
[23:39] on, you know, the gigawatts it costs,
[23:41] you know, the dollars per gigawatt.
[23:42] >> Right.
[23:42] >> Walk us through the economic case.
[23:45] >> Yeah. Yeah, so I mean, on this point of
[23:48] is orbital key to investing here? I
[23:51] don't think it is and I'll first point
[23:53] I'll make is what are the implied
[23:55] monetization rates
[23:57] based on expectations today for the AI
[23:59] business? And you know, I think you
[24:00] threw out the $160 billion number that's
[24:02] been leaked out there that people are
[24:03] talking about.
[24:04] The implied monetization rate on that
[24:06] number is something like $14 billion per
[24:08] gigawatt per year for the AI business.
[24:11] They just signed Anthropic at 22 to 23.
[24:14] They just signed Google at 50.
[24:15] >> Right.
[24:16] >> Right. So, I I think you can invest
[24:18] behind the AI business terrestrially and
[24:20] still be excited about it. But with
[24:22] orbital
[24:23] >> an important point. Excited about it if
[24:25] they can get the land and the power.
[24:26] >> Right. But but but I mean I I think for
[24:28] most investors, right? They get They
[24:30] have an easier time getting their head
[24:32] around how SpaceX wins terrestrially.
[24:35] Like can they go get land, power, and
[24:37] chips? The answer to that is high
[24:39] probability yes, okay? And what we're
[24:41] saying is at the rate they're monetizing
[24:43] that, that gets you to the numbers that
[24:45] are being leaked out there before you
[24:47] even have to take the leap of faith that
[24:49] they're going to extend the lead with
[24:50] orbital data centers. But take us there
[24:52] on that, too.
[24:52] >> Sure. Yeah, so so look, with orbital, I
[24:55] think the key thing is um two-stage
[24:58] reusability.
[24:59] >> Yeah.
[24:59] >> And beyond that, rapid two-stage
[25:02] reusability.
[25:02] >> Yeah.
[25:03] >> So, today with Starship, they've shown
[25:05] that they can successfully reland the
[25:07] booster.
[25:09] The second stage, we'll see what happens
[25:11] later this year. I think they're
[25:12] attempting to bring that back and then
[25:14] make it reusable by next year.
[25:16] Um
[25:17] but the thing that's important about
[25:19] two-stage reusability when it comes to
[25:20] the economics for orbital compute,
[25:23] right, is the cost per kg comes down
[25:26] significantly. You know, we're talking
[25:27] about going from $1,500 per kg on
[25:30] Falcon, somewhere in that range, to 250
[25:33] per kg, something lower. Um and the more
[25:36] that you can reuse the rocket,
[25:38] the more that price comes down.
[25:39] >> Right.
[25:40] >> Right, cuz you're just depreciating the
[25:41] cost of the launch.
[25:43] And eventually, you asymptote to the
[25:44] cost of the fuel.
[25:46] >> Right.
[25:46] >> Right.
[25:47] Assuming you can use a rocket for
[25:49] forever.
[25:49] >> Yes.
[25:49] >> Right, which will take a very long time
[25:51] for us to to really achieve that. But,
[25:53] um and at that point, we're talking
[25:54] about something well south of 250 per
[25:56] kg.
[25:58] So,
[25:59] then you look at the specs of these AI
[26:01] satellites. You know, Elon did a great
[26:03] >> Yeah, that pod that that pod was
[26:04] incredible that he laid out the other
[26:06] day, the specs on the satellites.
[26:07] >> It was really great because I think they
[26:10] are finally showing people, here's how
[26:13] you could viably design one of these
[26:16] satellites.
[26:17] And how heavy is the satellite? How many
[26:19] could you fit into a Starship launch?
[26:21] And when you back into the numbers, you
[26:23] get to something like 5 MW of capacity
[26:27] per Starship launch.
[26:28] >> Right.
[26:29] >> There's 100 metric tons in one of those
[26:31] Starships. So, you can back into the
[26:33] math of how much will it cost per
[26:36] gigawatt
[26:37] to launch these satellites into space.
[26:40] >> Right.
[26:40] >> Launch this compute into space. Um
[26:43] and the math that you get to before you
[26:45] account for things like
[26:47] bad GPUs, bad satellites, right, these
[26:50] will all be things that happen.
[26:52] But the math you get to is it's about $5
[26:54] billion per gigawatt of CapEx to put
[26:58] these in space.
[26:59] >> Right.
[27:00] >> For comparison, terrestrially,
[27:03] talk about the switch gears, the
[27:05] generators, the transformers, the shell,
[27:08] getting the power, that today is about
[27:10] 25 20 to 25 billion per gigawatt.
[27:13] So we're talking about a 5x reduction
[27:16] in cost
[27:18] on half of your bill of materials
[27:20] >> Right.
[27:20] >> for the data center.
[27:21] >> Right.
[27:21] >> Which is a huge number.
[27:22] >> Yeah, just just very simply, I mean just
[27:23] to say that put it cost $60 to put a
[27:26] gigawatt
[27:28] on the ground today. And we'll call it
[27:30] 35 of that is are the GPUs and the
[27:33] silicon that's doing the training and
[27:34] the inference.
[27:35] And 25 billion is the land, the shell,
[27:38] the power, and the cooling.
[27:40] I would hypothesize that those elements
[27:42] are probably going to be inflationary,
[27:44] so that 25 billion may not go down.
[27:47] And because space, power, cooling are
[27:52] effectively free in space, and when I
[27:54] say space, I mean land. You know,
[27:57] there's no land in space, but there is
[27:58] space.
[28:00] Um
[28:02] you're you're talking about putting a
[28:04] gigawatt into space for 30 billion and
[28:07] having lower operating costs. Now the
[28:09] dynamic versus 60 billion that's
[28:11] inflationary, and that third and that 30
[28:13] billion, that five may be deflationary
[28:16] over time.
[28:17] But what we need to consider
[28:19] is you know, the reliability and the
[28:21] maintenance. And so as long as you know,
[28:25] everybody can do the math,
[28:26] but as long as these satellites in space
[28:30] aren't failing at an at an astronomical
[28:33] rate, the math maths. As you can see,
[28:36] and by the way, we know GPUs melt and
[28:38] lasers fail. We know this happens in
[28:41] data centers,
[28:42] particularly during big training runs.
[28:44] And yeah, I mean GPUs melt.
[28:47] Um so as long as the reliability and
[28:50] maintenance is not dramatically lower,
[28:52] the math is there once we have
[28:54] reusability and then rapid reusability
[28:56] for Starship V3.
[28:57] >> I when you
[28:58] when we look at this, okay, so we we
[29:00] went through Starlink and we said,
[29:02] "Okay, like it it just stands to reason
[29:04] we're going to have direct to cell on
[29:06] Starlink." Like the assumptions there
[29:07] are, you know, again, seem like you can
[29:10] get your head around. Then when it comes
[29:11] to building terrestrial data centers,
[29:13] again, not a hard one to think that
[29:15] based on these couple deals that Elon's
[29:17] going to build a much bigger Starlink's
[29:18] going to build or SpaceX is going to
[29:20] build a much bigger business there. And
[29:22] then you have this call option on space
[29:24] that would drop the price even further.
[29:25] The one thing we haven't talked about is
[29:27] their model, right? And I find this
[29:30] surprising, right? Six Six months ago,
[29:33] x.ai was competing, they were doing
[29:34] pretty well, but they've done something
[29:36] dramatic over the course of the past
[29:37] couple
[29:38] couple months, which is they bought
[29:40] Cursor, right? Cursor is 700 800 people
[29:44] was already doing incredibly well from a
[29:46] revenue perspective. Our own
[29:48] projections were that they could exit
[29:50] this year at up to $10 billion
[29:52] of revenue, so they were growing very
[29:54] fast
[29:56] one of the leading coding agents, but
[29:58] they also had this incredible team with
[29:59] the potential, right, to really build a
[30:01] frontier level model, but they were
[30:03] compute constrained. So all of a sudden,
[30:05] they get bought by X. X has massive
[30:07] compute that they can now train on.
[30:11] And when I think about the revenue in AI
[30:13] that like if I look at that line item in
[30:15] the models having it go from $10 billion
[30:18] to $150 billion, yes, a lot of that will
[30:21] be the core weave type business that
[30:23] they have, but the question is how much
[30:25] of that is going to be the core x.ai
[30:28] business that's really powered by the
[30:30] new team from Cursor. So any thoughts on
[30:31] that, Kevin?
[30:32] >> Right now, so Composer 2.5 was Pareto
[30:35] dominant 12 days ago. It was trained on
[30:37] the Kimmy K2.5 base model.
[30:39] >> Right.
[30:40] >> Now, what's happening is the Grok 4.3
[30:43] 1.5 trillion parameter model is
[30:45] training.
[30:46] One would hypothesize based on scaling
[30:49] laws that that will might be a better
[30:52] base model. And then the cursor data is
[30:55] being injected into the pre-training
[30:57] process, not just reinforcement
[30:58] learning.
[30:59] And we'll see, and I think that is going
[31:01] to be a very important data point when
[31:03] that comes out. And I just think
[31:06] everyone should keep in mind that once
[31:07] you are at multiple places on that
[31:09] Pareto curve, if you have compute, you
[31:12] can scale really rapidly.
[31:14] >> You know, that that to me is if I had to
[31:16] say what the one piece that's being lost
[31:19] in the story,
[31:20] right? Like it's easy for everybody to
[31:22] get excited about the deals with
[31:23] Anthropic because you can put your hands
[31:25] around that. You know how much revenue
[31:26] it is. I see debate about, you know, the
[31:28] 90-day termination and how long they
[31:30] last and what multiple do you put on
[31:32] those revenues. But I think the thing
[31:34] that's getting lost is I think they've
[31:35] dramatically advanced their capability
[31:38] when it comes to building a frontier
[31:40] model. People outside Silicon Valley may
[31:42] not know, you know, Michael and the team
[31:44] at Cursor as well. This is an
[31:45] extraordinary team that he just
[31:48] downloaded, right, into SpaceX. SpaceX
[31:51] was already building good models. And
[31:53] what they have is
[31:54] they have this way to monetize compute
[31:56] that gives you this call option that you
[31:58] can pull all that compute in-house,
[32:01] right, to train a model and then to run
[32:02] the model. I suspect if there's an
[32:04] upside surprise, if we went around the
[32:06] table, I'd say this is the place that's
[32:08] getting the least amount of attention
[32:10] and could have the biggest upside
[32:11] surprise. Any any thoughts, Clark, on
[32:14] what you think is being overlooked or
[32:17] areas that you think are misunderstood
[32:19] about the business today?
[32:20] >> I I would say I would say
[32:23] what the last few weeks have proven is
[32:25] that Elon, um,
[32:28] their team can stand up all this
[32:29] compute. Actually, if you just, you
[32:32] know, went back 1 and 1/2 years, you
[32:34] know, they were behind in the race to
[32:36] stand up compute. They were you know
[32:38] they they didn't have that many H100s.
[32:40] They brought in Colossus. Then they
[32:42] brought in Colossus 2 at a scale much
[32:44] larger than anyone else. And now you
[32:46] know as we gear for Vera Rubin
[32:49] you know
[32:50] from you know a lot of my conversations
[32:52] it looks like they've you know secured
[32:54] maybe up to 20% of Vera Rubin capacity
[32:57] especially in the early days of you know
[33:00] when you know these these chips are very
[33:02] scarce that that they're going to have a
[33:06] a
[33:07] a lead on all of this because you know
[33:09] people think that they can stand up this
[33:11] compute better. So I think they'll all
[33:13] you know what what the last few weeks
[33:14] have actually shown is that Elon you
[33:16] know Elon will
[33:18] take you know take a shot at hitting the
[33:21] frontier but if it you know if for
[33:23] whatever reason
[33:24] um they
[33:26] they have over procured some capacity
[33:28] this is a very scarce asset that they've
[33:31] shown that they can monetize at actually
[33:33] you know best in class margins and
[33:35] payback periods.
[33:36] >> The irony is like you know
[33:38] you and I've been doing this long enough
[33:40] to know I mean that's why Bezos built
[33:41] AWS. Right? He had to build capacity for
[33:45] Black Friday.
[33:46] >> Yeah.
[33:46] >> Right? But then the rest of the year he
[33:47] sat on all this capacity they had to
[33:49] build and he figured out a really
[33:50] incredible way to monetize this. And by
[33:52] the way investors at the time 2009 2010
[33:56] when he was building out the capability
[33:57] around AWS hated it.
[33:59] >> Of course.
[33:59] >> Because he was consuming all that free
[34:01] cash flow. My meanwhile he was digging
[34:03] the biggest gold mine in the history of
[34:05] the world. One of the biggest.
[34:07] >> One of the biggest.
[34:07] >> Among them among them at the time was
[34:09] probably the biggest.
[34:10] >> Yeah Google search might want to have a
[34:12] we'll have a we'll have a discussion.
[34:14] >> [clears throat]
[34:14] >> By the way I do think it is important.
[34:16] Grok 4.3 I think the cursor if they
[34:18] acquire it that may end up being very
[34:20] important. But Grok 4.3 was on the
[34:23] Pareto frontier and has of 10 or 12 days
[34:26] ago and this these things move fast. But
[34:28] most intelligent 500 billion parameter
[34:30] model in the world. And they were on the
[34:33] frontier and there are four companies on
[34:35] the frontier.
[34:36] xAI, SpaceX AI, Google one with Gemini
[34:40] 3.1 Pro, and then the rest of it was
[34:42] dominated by Anthropic and OpenAI. But
[34:43] they were on the Pareto frontier and now
[34:45] we'll see what they do with Cursor.
[34:46] >> Yeah. Um I want to come back to that in
[34:48] a second.
[34:49] >> way, man, I want to ask you some
[34:50] questions.
[34:50] >> go go go. What do you think? So you
[34:52] think the biggest source of potential
[34:53] upside is the model?
[34:55] >> Yes.
[34:55] >> What do you think?
[34:56] >> I think that's the I think that's the
[34:57] thing that's least talked about.
[34:58] >> Least talked about.
[34:59] >> Right? And so, listen.
[35:02] When I look at the bull bear case on the
[35:04] IPO, right? The bears are looking at
[35:06] last year's revenue. Say it was $18
[35:08] billion
[35:09] and they're looking at the forecast from
[35:11] the banks of $160 billion, you know, 3
[35:13] years from now and they're saying,
[35:14] "Listen, not many companies in the
[35:16] history of the world have basically 8x
[35:18] their revenue over 3 to 4 years." Right?
[35:20] So that's where, you know, I think and
[35:23] people get nervous about the valuation.
[35:25] When I look at this, again, when you
[35:27] break it down as an analyst first
[35:29] principles, part by part, which is what
[35:31] I tried to do here, right? When you look
[35:33] at Starlink, it looks totally doable.
[35:36] When I look at what they're building in
[35:37] AI compute terrestrially, looks totally
[35:40] doable over the course of next 3 years.
[35:42] When I look at the model itself after
[35:44] the acquisition of Cursor, you know,
[35:45] combining those things around the
[35:47] compute they have, that looks to me like
[35:49] it could be an upside surprise. So I
[35:51] would say that I think that uh you know,
[35:53] in the IPO, but I think when you look
[35:55] back 3 years from now, there's a decent
[35:57] chance that everybody's like, "Oh my
[35:59] god, that was super obvious." Right?
[36:02] Even though today all of these things
[36:04] have risk associated and back to where
[36:06] we started. I'm not, you know, none of
[36:07] us are here to pump the IPO at 1.77
[36:10] trillion. It's really to just break it
[36:12] down as we do inside our shop and to
[36:14] say, "What is that distribution of
[36:15] future probabilities? What's the
[36:17] probability that it's higher from here?
[36:19] What's the prob" And I think we're all
[36:21] pretty AI pilled. And if you're AI
[36:23] pilled, that means we got to build a lot
[36:24] more compute than the world thinks and
[36:26] that these models are going to be a lot
[36:28] more valuable than people think. You
[36:29] combine that with their core business. I
[36:31] don't know another entrepreneur or
[36:33] another business that's a better bet on
[36:36] the future,
[36:37] right, than SpaceX. And so I think for
[36:39] most institutional investors, it's a
[36:41] must buy, a must own, a set it and
[36:44] forget it, right, in order to have a
[36:47] real bet on both the space and the AI
[36:49] future.
[36:49] >> From your lips to God's ears.
[36:51] >> I mean, listen, I I again, I think that
[36:53] I I I think that you're going to have to
[36:54] wait, but you know, we had this chart
[36:56] last week, right, that came out.
[36:58] Everybody was sending around Twitter,
[37:00] conveniently timed, and you know, it's
[37:02] like shows the average max drawdown post
[37:05] IPO for like 20 companies from Facebook,
[37:08] Twitter, Alibaba, Shopify is, you know,
[37:11] over 50%. And so maybe that again will
[37:14] will will end this section here. You
[37:16] know, Gavin, you and I've been doing
[37:17] this a long time. We know it's going to
[37:19] be bouncy around the IPO. Um,
[37:21] you know, how do you as a manager try to
[37:25] try to manage that? Um, do you try to
[37:27] trade around the IPO? Do you set it kind
[37:30] of and forget it? I would say from an
[37:32] Altimeter perspective, what we tend to
[37:34] do is we take a base position that we
[37:36] set and forget, right? And then we may
[37:38] size up or size down depending upon how
[37:41] the market reacts in, you know, in a
[37:43] particular moment. Um, but any thoughts
[37:47] on on this chart or
[37:49] you know, how how people you guys are
[37:51] thinking about it in particular. You
[37:52] obviously own a lot going into it.
[37:54] >> First agree with absolutely everything
[37:56] you said and I actually think about it
[37:57] the same way, set it and forget it.
[37:58] You've talked about you have ballast,
[38:00] you move around and you move the ballast
[38:02] to one side of the ship when you want to
[38:03] the ship to lean into the wind to go
[38:05] faster and you move it to the other side
[38:07] when you don't want the ship to tip
[38:08] over. I think that's a great analogy.
[38:09] Think about all
[38:11] important companies in the portfolio the
[38:13] same way. So 100% agree. I mean, this
[38:16] this chart is a bummer. What I would say
[38:19] is, you know, this data on IPOs, but
[38:20] what I would just say is this is a
[38:23] really unprecedented situation.
[38:24] >> Yes. We've never had an IPO this big.
[38:28] We've never had an IPO that's going to
[38:29] go into an index this quickly.
[38:32] We
[38:34] simply do not know how much selling
[38:36] there will be from investors.
[38:39] I would hazard a guess. I mean, I'm I
[38:42] don't know.
[38:43] But Elon, I don't think he needs
[38:45] liquidity and I think he owns What does
[38:47] he own, Foxy?
[38:49] >> It's
[38:50] 50%
[38:51] >> 50% of the company.
[38:52] >> way, he's locked up for 365 days or 366
[38:55] days. So, we know he's not selling,
[38:57] right?
[38:57] >> So, I just think it's an unprecedented
[38:59] situation and the right answer
[39:01] >> Yeah.
[39:02] >> is I don't know what's going to happen
[39:04] in the short term. And the right answer
[39:06] that I would just, you know, encourage
[39:07] every investor making their own decision
[39:09] is to just think exactly [clears throat]
[39:10] the way you articulated it. We have
[39:13] these different levers. We have these
[39:14] different variables. Think about each
[39:15] one of them from first principles. Make
[39:17] your own decision.
[39:19] Do your own due diligence. Be
[39:21] thoughtful. But, there are a lot of
[39:22] variables here and that it is a little
[39:24] funny to me that uh you know, it was 100
[39:27] times trailing TTM revenue. Well, after
[39:30] the deals they signed, I think it's at
[39:31] 39 times.
[39:33] >> That can change fast.
[39:34] >> So, they added $29 billion in a month.
[39:36] >> Yes. Now, it's
[39:37] >> [laughter]
[39:37] >> By the way, have you ever seen that
[39:39] happen?
[39:39] >> Never. Never. And you know, it just goes
[39:41] to show
[39:42] first um Elon is not only a great
[39:45] engineer.
[39:46] He and Gwen and the team are great at
[39:49] business.
[39:50] >> And Brad,
[39:50] >> They they they understand what needs to
[39:52] be done to raise the capital to get to
[39:55] the next phase. They have a long-term
[39:57] mission in the business. And so, to me,
[39:59] again, what we saw in the course of the
[40:01] last few weeks with cursor, what we saw
[40:03] with these deals that they cut, I don't
[40:05] know that any of the mag seven could
[40:07] have moved that quickly to adjust the
[40:09] business that they did. It's
[40:11] exceptionally entrepreneurial at scale,
[40:14] which we very rarely see in businesses.
[40:16] Two other things I would just say
[40:18] >> you a hug, Brad?
[40:19] >> Two Two other things I I I I would just
[40:21] say. Number one is people talk a lot
[40:23] about the total amount of capital being
[40:25] raised. If you add up the capital here,
[40:27] right, for Anthropic what they may
[40:30] raise, what OpenAI may raise, what, you
[40:32] know, SpaceX may raise, let's call it
[40:34] $250 billion.
[40:36] That's 1% of the Mag 7.
[40:38] Okay, it's 1% of the Mag 7.
[40:40] >> I Yeah. And
[40:41] And we will as well. You know, like that
[40:43] to me is like a bet on the future that
[40:45] we all believe in. And so, if I said,
[40:47] "Where are we out of consensus? What is
[40:49] our variant perception?" We actually
[40:51] think it's going to be bigger, faster,
[40:52] and we've thought that for a couple
[40:53] years. Um so, first, it's only 1% of the
[40:56] Mag 7 market cap. And then you
[40:58] referenced it, the amount of selling. Um
[41:01] I've got a chart we'll post here. This
[41:03] is, you know, the the dribble share
[41:05] release for SpaceX shareholders. You
[41:08] know, so there's not a lot that can be
[41:09] released um up until after the first
[41:12] earnings. This We saw this in the
[41:14] Cerebras IPO. Um there's a version of it
[41:17] here in this IPO. And so, again, I think
[41:19] the banks have been thoughtful here,
[41:21] knowing that this is a very large IPO.
[41:23] And I'm not saying that won't trade
[41:24] down. Like there's possibility, you
[41:26] know, these things trade down. But
[41:28] again, for me, telescope out, is there
[41:31] any company better positioned as a bet
[41:32] on the future? I think what they've
[41:34] shown over the course of last 5 weeks,
[41:36] they're
[41:36] they're they're probably number one. But
[41:39] let's move on.
[41:39] >> No, no, can I just say one thing about
[41:41] the employees? I think another thing
[41:42] that's unprecedented here is the
[41:44] employees
[41:45] >> Yeah.
[41:45] >> and to a large degree the investors here
[41:48] have had liquidity every 6 months.
[41:50] >> Exactly.
[41:51] >> the last 10 years.
[41:52] >> Yes.
[41:52] >> So, if you're a SpaceX employee or
[41:54] former employee, and you wanted to sell
[41:57] you've had whatever that is,
[41:59] close to 20 chances. And it is a matter
[42:02] of historical record that large
[42:04] investors have been able to sell. So
[42:08] I would think a lot of the people
[42:10] >> they've
[42:11] >> chosen to own it. Now, there's a new
[42:13] valuation and we'll see what they do,
[42:14] but just this is utterly unprecedented
[42:17] and we'll see.
[42:17] >> Yeah, I know. It's It's It's a great
[42:18] point. We have in fact called these
[42:20] companies quasi-public. Um you and I
[42:22] both know that SpaceX and I'd put
[42:25] Anthropic in in in this category as
[42:27] well, Databricks in this category. These
[42:29] things in many ways have been more
[42:30] liquid over the course of the past 3
[42:32] years than some public biotech companies
[42:34] we know. Right? And so there's a
[42:36] continuum of liquidity here. We We treat
[42:39] it as a binary, private versus public,
[42:41] but it's really about this continuum.
[42:43] You know, let's keep going on models.
[42:45] You know, um Anthropic launched Fable 5,
[42:47] which you referenced um yesterday, which
[42:49] is basically Mythos um with some
[42:52] classifiers and safeguards um around
[42:55] cyber and biology, chemistry,
[42:57] um and distillation. When those things
[42:59] get triggered, it fails back [snorts] to
[43:01] Opus 4.8. Um you know, there was a
[43:03] Copart tweet about this yesterday. He
[43:06] said, you know, it sold on all the
[43:07] benchmarks, but what really makes it
[43:09] special is long-running tasks. Okay? You
[43:13] retweeted our good friend, you know,
[43:15] Noam Brown. Um you know, ChatGPT 5.5
[43:19] also exhibited these capabilities.
[43:21] Um you know, it it led Noam, right, to
[43:24] suggest that it's not very relevant to
[43:27] do these snapshot benchmarks anymore.
[43:29] Yeah, like the x-axis has to be time or
[43:32] tokens or compute because we can solve
[43:34] most problems now if we just let these
[43:36] frontier models for a very long uh point
[43:39] in time. So, Gavin, what is this new
[43:42] class of model, right, Fable Fable 5,
[43:46] ChatGPT 5.5? What does it mean for the
[43:49] race in superintelligence? Who's up?
[43:51] Who's down? Who's still on the frontier?
[43:54] Um give us your thoughts.
[43:56] >> I mean, it's hard to say that
[43:58] Anthropic's not up.
[43:59] >> Yeah.
[43:59] >> Like after
[44:01] the revenue numbers they've put up,
[44:03] after the Fable 5 release, and Mythos is
[44:05] evidently even better.
[44:07] But I just think that Gnome Brown post
[44:10] from yesterday, polynomial, is so
[44:13] profound.
[44:15] And just the idea that we do not know
[44:17] how smart these models are.
[44:19] And we made
[44:20] >> Say more about that. Why don't we know
[44:21] how smart they are?
[44:22] >> Because nobody has run Mythos for a year
[44:26] continuously. And we may never know how
[44:29] smart each generation of models actually
[44:31] is or was, but because we don't have
[44:34] time to appropriately evaluate their
[44:36] intelligence before the next model comes
[44:38] out. I mean, this is a profound
[44:39] statement. And just just imagine, okay?
[44:42] So, I always say like when you think
[44:44] about FSD,
[44:45] just imagine a human being who never
[44:48] gets distracted, never gets tired, never
[44:51] talks on the phone in the car, never
[44:53] drinks and drives, never yells at their
[44:55] kids, never has to go to the backseat to
[44:57] give their baby a bottle.
[44:59] And like of course you would think that
[45:01] over time that is superior to humans who
[45:03] are distracted.
[45:04] I don't know how long How long can you
[45:06] think deeply about one topic, Brad?
[45:08] >> What do you Give me an hour. Give me an
[45:09] [laughter] hour. Give me an hour.
[45:11] >> A BIT.
[45:13] THAT MAKES me feel terrible cuz I think
[45:15] I can think deeply about one topic
[45:17] continuously before having a stray
[45:18] thought enter my mind
[45:20] for like maybe 5 minutes. Then I can
[45:22] come back to that.
[45:23] Imagine if Albert Einstein
[45:26] had been able instead of, you know, and
[45:28] maybe that maybe I
[45:29] maybe he could think for 3 hours at a
[45:31] time. Clearly an exceptional intellect.
[45:34] But imagine Albert Einstein had just
[45:35] thought about fundamental physics
[45:38] 24 hours a day.
[45:40] He doesn't have to eat, he doesn't have
[45:42] to sleep, he doesn't have to relax, he
[45:43] doesn't drink,
[45:45] >> never gets old,
[45:46] >> never gets old,
[45:47] >> never has diminished intelligence,
[45:49] >> and he thought for 1 year. I mean, we
[45:51] might already, you know,
[45:53] >> have solved a lot of these intractable
[45:54] problems.
[45:55] >> So, I just think that's an extraordinary
[45:58] thought. And just my takeaway was
[46:01] however
[46:02] bullish I was on compute before then,
[46:05] I'm just a lot more bullish.
[46:06] >> Right. Right. Right. So, so, so that is
[46:09] a, you know, we saw when
[46:11] that was probably what really unlocked
[46:14] Opus 4.6. It was the first really
[46:17] long-running model that could maintain
[46:19] that context, maintain that memory, um
[46:22] solve some of these longer-running
[46:23] problems, right? For us, the signal was
[46:26] in January. We knew we felt like that
[46:29] was a big moment, but then when you
[46:30] started to see the revenue go up, we
[46:33] knew that lots of people were voting
[46:35] independently, that that was a profound
[46:37] moment that they became much, much more
[46:39] useful. So,
[46:41] but one of the things that the consensus
[46:44] going into this year, right? So, the big
[46:45] question going into this year was was
[46:48] the AI revenue going to show up? Were we
[46:51] going to get to these thresholds of
[46:52] intelligence that caused enterprises and
[46:55] consumers to use them more? And I think
[46:57] the consensus at the time, at least on
[46:59] this podcast, um the the the debate with
[47:02] with my with with with Bill was the
[47:05] open-source models, cheap tokens, were
[47:07] catching up on the frontier, that
[47:09] perhaps these models were beginning to
[47:11] asymptote, um that people wouldn't
[47:14] really pay for premium tokens,
[47:16] and it seems to me that the evidence on
[47:19] the field, 6 months into the year, is
[47:21] just the opposite, right? That frontier
[47:24] tokens are capturing the vast majority
[47:27] of all the revenues,
[47:28] and that in fact, if you believe in the
[47:31] long-running capabilities and more
[47:32] compute allows you to do that, they may
[47:34] actually be extending their lead, right?
[47:37] On some of these models that were built
[47:38] on distillation. So, I just open it up
[47:40] to anyone around the table, what are
[47:42] your thoughts on whether or not, you
[47:45] know, have we challenged this thesis
[47:48] that cheap open-source tokens are going
[47:50] to always, you know, close the gap on
[47:52] these frontier models, or are they
[47:54] extending their leads?
[47:55] >> I I think this debate, like this same
[47:58] debate has existed since the beginning
[48:01] of
[48:02] since we started training these models
[48:03] to begin with, which was hey, we're
[48:05] always kind of three, six months behind
[48:07] the frontier. But empirically, like you
[48:09] can just see all of the revenue has
[48:11] actually just accrued at the frontier.
[48:13] And that I think that's because every
[48:15] time we release the frontier,
[48:18] a whole new like slew of use cases
[48:21] >> Right.
[48:22] >> that that previously we could have never
[48:24] tackled before, like coding.
[48:26] Um but also just, you know, you know,
[48:28] we've we've just been locked at our desk
[48:30] for the last last day just, you know,
[48:33] hammering Claude because, you know, it's
[48:35] just fascinating the things that now we
[48:36] can do with fable five that we could
[48:38] just couldn't do with opus 48 just a day
[48:40] before.
[48:41] >> So what are some of those things, man?
[48:42] I'm curious.
[48:43] >> So So I think it's really really good at
[48:45] multi-agent orchestration. So they they
[48:47] Anthropic released a um a blog post
[48:50] about like different uh agent um six
[48:53] different agent like orchestration
[48:55] patterns that, you know, they've they've
[48:56] talked about. But really like once you
[48:58] start being able to manage all these
[49:01] agents, the harness and the model itself
[49:04] is being arled with one another, they're
[49:06] actually being,
[49:08] you know, fused closer and closer
[49:09] together, but the model can understand
[49:12] the, you know, the extent of your work.
[49:13] So, you know, one of the things, for
[49:16] instance, is um
[49:18] I just threw in like seven of our models
[49:21] and just said, "Okay, like I want to
[49:22] create a master view of like my beliefs
[49:26] given all of these assumptions of all
[49:28] these companies, TSMC capacity, like and
[49:31] then and then produce me a report on all
[49:33] this stuff." And you know, the the model
[49:35] is able to reason through all of our
[49:37] assumptions. Like actually, if you
[49:39] believe this
[49:40] >> Right. What are the contradictions
[49:41] exactly?
[49:42] >> Yeah, it's it was fascinating. And and
[49:44] and you know, before we'd never do that,
[49:46] but but now, you know, I think we're
[49:48] just step one into multi-agent
[49:50] orchestration. We're going to do this
[49:52] even further and that's one example.
[49:54] I've also dumped all my all my notes
[49:56] into it and it's reason across all my
[49:58] notes from the last 3 years and said,
[50:01] you know, here are some of your ideas
[50:03] that were consistent. Here are like, you
[50:05] know, the sources that were actually the
[50:07] highest signal to what actually played
[50:09] out, you know, and then it is actually
[50:11] just super fascinating what you could do
[50:13] and we've just blown through our blown
[50:16] through our limits.
[50:16] >> mean it's it's it's unlocking all this.
[50:18] I mean like they gave examples yesterday
[50:20] and the release Anthropic did, you know,
[50:21] 50 million line Ruby code base at Stripe
[50:24] that was, you know, refactored in a day
[50:26] versus many weeks with many people. You
[50:29] think about where this is impacting
[50:30] biology and life sciences just across
[50:32] the spectrum.
[50:34] Um
[50:34] and to me it really gets back to this
[50:37] fundamental point. Number one, if you
[50:38] believe this to be true about
[50:39] long-running agents, then we're going to
[50:41] produce and consume more tokens in the
[50:44] future as far as the eye can see. So the
[50:46] world this gets me back to, you know,
[50:48] terrafab and space orbital and all this
[50:51] because
[50:52] we we we may in fact unlock real
[50:55] thresholds of intelligence, but we're
[50:56] going to have to let these horses run
[50:59] for a long time in order to get there.
[51:00] Yeah, I would just say two things I two
[51:02] things can be true.
[51:03] >> Mhm.
[51:04] >> The majority of economic value may
[51:07] continue to accrue to the frontier and
[51:09] man has it ever accrued to the frontier
[51:11] thus far and for sure the first 6 months
[51:13] of this year, but the majority of tokens
[51:16] consumed in the world may be open
[51:17] source.
[51:17] >> And they are
[51:18] >> today.
[51:18] >> Yes. I and I think that this current
[51:21] state is likely to persist. Harvey had a
[51:25] great blog post that they put out on X
[51:28] and they used and it's just amazing how
[51:31] everything gets out out of date like in
[51:33] 5 days, you know.
[51:35] But they used their own proprietary
[51:37] legal data to do reinforcement learning
[51:40] and supervised fine-tuning
[51:42] with Fireworks on an open source model
[51:44] and then And used a router and a router
[51:46] being something that picks which model
[51:47] you send which query to, and which model
[51:50] you use to check which model. And they
[51:51] got better outcomes than Opus 4 either
[51:54] 4.7 or 4.8 at a lower cost.
[51:57] >> Yes.
[51:57] >> And I think that is the future. And the
[52:00] reality is
[52:01] they were still consuming a lot of Opus,
[52:03] but a majority of the tokens they were
[52:04] processing probably were in their own
[52:07] open-source models.
[52:08] >> We heard the same thing. We did a We did
[52:10] um
[52:11] We did an enterprise survey that we'll
[52:12] post of 300 companies how which ones
[52:16] were optimizing, so these are folks who
[52:18] are kind of looking at model routing and
[52:20] saying we're going to send certain
[52:21] tokens over here, which ones are
[52:22] thinking about optimizing, which ones
[52:24] aren't optimizing yet, and then what is
[52:26] their expected use of frontier model
[52:28] tokens, right? And they're all expecting
[52:30] to consume a lot more even though
[52:32] they're already in the process of
[52:33] optimizing. Think of it in the in in the
[52:35] context of JP Morgan. If they're doing
[52:37] some back of the house stuff, right, on
[52:40] customer service or whatever, they may
[52:42] very well use an open-source model. Now,
[52:44] I think they're loath to use Chinese
[52:46] open-source models, so they're waiting
[52:47] on kind of US open-source models to, you
[52:50] know, be able to really deliver the bang
[52:52] that they need, but my hunch is for
[52:54] these enterprises, a lot of that back of
[52:56] the house stuff will get rooted there.
[52:57] That will probably be a majority of the
[52:59] tokens, but I think the really
[53:00] high-value stuff, you know, coding as an
[53:03] example, they don't want to write
[53:04] second-tier code. I think the vast
[53:06] majority of that will continue to be on
[53:07] the frontier.
[53:09] >> Um
[53:09] >> You don't need Albert Einstein to book
[53:11] you a trip. You don't need Albert
[53:13] Einstein to do KYC.
[53:15] >> But but but this is the debate we had at
[53:17] literally at this table two years ago.
[53:18] However, if you just look at the revenue
[53:21] curves, right? What bill What what folks
[53:23] concluded when they said that, they
[53:25] said, "Therefore, the frontier models
[53:28] will not accrue most of the revenue."
[53:30] And what we're seeing right now, it's
[53:31] 90% of the
[53:32] >> That has been decisively wrong. Probably
[53:34] more than 90%, and it may continue to be
[53:36] decisively wrong. Frontier might be 90%
[53:38] of the
[53:39] economic value. Open-source
[53:40] >> might be 80% of tokens.
[53:42] Something that I think is very important
[53:44] on open source
[53:45] is that you know, I think there's this
[53:48] belief that it's bearish for AI.
[53:50] It's actually it may be very bearish for
[53:52] the frontier models. There's that bear
[53:54] case you talked about. It's actually
[53:55] really bullish for compute and hardware
[53:58] because if the frontier models are
[54:00] capturing less of the margin, then
[54:02] you're going to spend more on compute.
[54:04] So, the better open source does, the
[54:06] better it is for compute providers.
[54:08] >> And I yeah, I I will say it there is a
[54:11] very
[54:12] I would say between um spending time in
[54:14] the heart of like the West, Silicon
[54:16] Valley, and also spending time in Asia,
[54:18] there is like a very
[54:20] big um
[54:21] like a deep-seated belief in one versus
[54:24] the other, which is like if you spend a
[54:25] lot of time here, it's like all closed
[54:27] source, cloud, every all traffic is
[54:30] going to go, you know, by way of this
[54:32] direction. And then you spend time in
[54:34] Asia, you know,
[54:36] the the overwhelming belief is that
[54:38] we're going to find the right model to
[54:40] the right workload, and we're not going
[54:42] to overspend.
[54:43] >> Right.
[54:43] >> And I think, you know, I would say I
[54:45] would say
[54:46] the next year is probably going to be
[54:48] the most indicative of which way this
[54:51] falls
[54:53] um because
[54:55] I think I think the reason why
[54:57] uh closed source models have captured so
[54:59] much of the value is because um the
[55:02] models actually get the intention and
[55:04] actually carry through the work. And
[55:05] this is the first year where we actually
[55:07] had agents that actually carried out
[55:10] user intention from just answering a
[55:12] chatbot request to actually producing
[55:15] useful work.
[55:16] >> Right.
[55:17] >> Um now the the the level of this
[55:19] intelligent has scaled so rapidly, and
[55:21] we continue to push against like the
[55:23] most economically valuable tasks, which
[55:24] are coding and finance and all these
[55:27] like knowledge work tasks. But like for
[55:29] the long tail of tasks, if open source
[55:32] continues to maintain a 6-month lag, we
[55:35] might actually see a lot more open
[55:37] source used for
[55:39] you know, our everyday tasks that we
[55:41] might actually
[55:42] >> basically Jensen's argument, right?
[55:43] Jensen's argument is you're going to
[55:45] have model routing
[55:46] and we're just in a moment in time where
[55:48] the frontier models gain the advantage
[55:50] can do long-running tasks that open
[55:52] source models couldn't do it very well
[55:54] and so they're accruing all of the
[55:56] value, but as soon as the open source
[55:57] models can do the long-running tasks as
[55:59] well, which is not far away that they
[56:01] too will grab a bunch a bunch of this
[56:03] revenue.
[56:04] >> Are you about to burst into reflection?
[56:06] >> I'm not.
[56:07] >> Okay. No, no, no, no are we, but I'm
[56:09] very impressed by Misha and and the team
[56:12] and what they're doing. I very much want
[56:14] a frontier open source US lab to win. We
[56:17] know that, you know, I heard you say
[56:19] recently and I believe it to be true
[56:21] Nvidia any day that they really wanted
[56:24] to, right? They already have some great
[56:25] open source models. They could
[56:27] absolutely build a frontier open source
[56:29] model whenever they chose to do it and
[56:31] so it's not a question in my mind as to
[56:33] whether or not the US is going to have a
[56:35] frontier open source model. It's just a
[56:37] question about timing and then like at
[56:39] that point in time is that you know,
[56:42] let's say let's assume they get these
[56:43] long-running capabilities.
[56:46] Have the frontier labs now achieved
[56:48] something yet again that allows them to
[56:50] keep keep the the stranglehold on the
[56:52] revenues?
[56:53] >> Yeah, and I just think it's if you're
[56:56] Wow, that's a cute ASIC you've built
[56:58] there. That is so cute. How would you
[57:01] like
[57:02] open source to join the frontier?
[57:04] >> Right.
[57:05] >> How would you like that? How do you like
[57:06] them apples? So, I mean I'm not sure
[57:08] that's the explicit calculation, but I
[57:10] do think Jensen
[57:11] >> Say more. Just double click on that for
[57:13] everybody at home.
[57:13] >> Yeah.
[57:14] >> If you were if they were to put an open
[57:15] source model out there, how does that
[57:17] impact the ASIC landscape?
[57:19] >> Well,
[57:21] you might not have the revenue
[57:23] to fund [laughter]
[57:24] to fund that the revenue of the margins
[57:26] to fund that ASIC.
[57:28] And I do think Nvidia is highly likely
[57:31] to be the world's dominant provider of
[57:33] open source AI. And I do think Jensen
[57:35] will bring open source,
[57:38] you know, right now it's whatever, 6
[57:40] months behind the frontier.
[57:41] >> Yeah.
[57:42] >> We might see it
[57:44] creep closer and closer and closer.
[57:47] And I do think Jensen has a big business
[57:49] decision. I see this, you know, chart
[57:50] here, so let's, you know, chop it up
[57:52] about Nvidia, as you say.
[57:53] But if all of his customers
[57:56] are going to compete with him,
[57:58] >> Yes.
[57:59] >> then
[58:00] why not compete with his customers? And
[58:02] we have all these neo clouds.
[58:04] >> Right.
[58:04] >> So that's a cloud computing business
[58:06] that can compete with all these cloud
[58:08] computing businesses.
[58:10] He has his own models that are really,
[58:11] really good. Nematron 3 or 3.1 was
[58:13] actually really, really cool from a
[58:15] computer efficiency perspective. And
[58:17] he's always careful to release small
[58:18] models so as to not tread on Anthropic
[58:21] and OpenAI,
[58:22] >> Right.
[58:22] >> Google's toes.
[58:23] But I do think that is a choice he is
[58:25] making.
[58:26] And just, you know, at if if the
[58:29] economics change,
[58:30] >> Right.
[58:31] >> I think Nvidia can join the frontier and
[58:33] become one of the world's largest cloud
[58:35] computing companies much faster than
[58:37] people think.
[58:37] >> Interesting. Interesting. Clark, walk us
[58:39] through this
[58:40] this chart.
[58:41] >> Yeah, so so I think one of the takeaways
[58:43] from spending time in Taiwan was there
[58:47] there is certainly a lot of excitement
[58:48] around the next wave of ASICs.
[58:52] Um, but I think I think it's like a very
[58:54] clear moment now where Nvidia
[58:57] it used to be an argument of Nvidia
[58:58] versus ASICs one or the other and, you
[59:01] know, total domination one or the other.
[59:03] Now I think it increasingly every year
[59:06] every every one assumed that Nvidia was
[59:08] going to lose share dramatically on a
[59:10] revenue scale, on a gigawatt scale, on a
[59:12] unit scale. And actually, if you
[59:14] actually look at the last few years, you
[59:16] know, they've actually maintained their
[59:19] share very, very handsomely.
[59:21] Um, actually,
[59:23] um, if you accounted for the fact that
[59:25] Anthropic
[59:26] was not really using Nvidia. They
[59:28] probably actually gain share against
[59:31] if not for in 25 26. So, I think I think
[59:35] what was very interesting though was a
[59:37] new class of
[59:39] accelerators or ASICs.
[59:41] MediaTek with their
[59:43] with their new V8T
[59:45] versus, you know, Broadcom's V8I for
[59:48] TPUs
[59:50] actually was a big topic of discussion.
[59:53] And, you know, I I think for ASICs
[59:56] the argument now is that more and more
[59:58] will look custom to the actual workload
[60:01] and that is like one vector that people
[60:03] are moving in versus Nvidia now is
[60:06] has kind of shown itself as the the
[60:09] predominant provider of compute to
[60:12] a lot of the world and for, you know,
[60:14] internal internal workloads,
[60:17] perhaps they will go more and more
[60:18] custom and more and more down the stack.
[60:21] And I I remember just, you know, 1 year
[60:23] ago when it was kind of a Broadcom or
[60:26] Nvidia battle. It seems there's a lot
[60:27] more nuance now to, you know, what type
[60:30] of accelerators will fit which workloads
[60:33] and fit which customers and fit which
[60:35] business models. Um
[60:37] and yeah,
[60:40] I thought I thought that was a
[60:42] a new topic.
[60:43] >> It's actually
[60:44] >> New realization though, I think we all
[60:45] kind of shared this view for a long
[60:47] time.
[60:48] >> Yeah, I was just shocked. I mean, I'm
[60:49] I'm out here. I did a board meeting with
[60:51] one of our companies
[60:52] and just, you know, their biggest one
[60:54] thing they emphasized is
[60:56] we thought the world would be have be
[60:59] consuming less Nvidia than it is and if
[61:01] anything, Nvidia is accelerating and
[61:03] they just continue to out execute their
[61:06] competitors. And I think a lot of people
[61:08] are indexing to this
[61:10] OpenAI gigawatt and you know, Nvidia has
[61:13] 10.
[61:14] Broadcom has 10.
[61:16] Um
[61:17] who has six?
[61:18] AMD AMD has six and they have warrants.
[61:21] And then Cerebras has
[61:24] our shared portfolio company
[61:25] has a gigawatt.
[61:27] And I just that is what's on paper.
[61:29] >> Right.
[61:30] >> What actually gets deployed, let's see.
[61:33] I will be very surprised if you know
[61:35] that 10 out of 27, what's that math?
[61:37] Let's see who's best at math. What
[61:38] percentage market share is that?
[61:40] >> 30% yeah.
[61:41] >> Yeah. I'll be very surprised if that is
[61:44] where they land. I think that is an
[61:45] extremely unlikely outcome.
[61:48] And especially as long as we're in a
[61:50] watt constrained world, if you can get
[61:53] more tokens per watt, which is literally
[61:54] revenue with Nvidia
[61:57] than a lot of alternatives just if you
[62:00] build your factory with another chip
[62:03] you may save some money, but you're
[62:04] going to have less revenue and the
[62:05] margins may be lower and that's a point
[62:07] that Jensen keeps hammering and I think
[62:09] is a really important. And by the way,
[62:11] credit where credit is due
[62:13] the most important the most one of the
[62:14] most surprising things to me in this
[62:17] ASIC landscape
[62:19] >> I'd say Meta and Microsoft have been
[62:21] probably disappointing.
[62:22] >> Yes.
[62:22] >> You know who made a good ASIC?
[62:23] >> Yes.
[62:24] >> Well, I know you know.
[62:25] >> Yes.
[62:25] >> Jalapeno
[62:26] >> Yeah, exactly.
[62:26] >> from Open AI. They made a great chip.
[62:29] >> Yes.
[62:29] >> Now, unfortunately needs to run at a
[62:31] much lower temperature than the Nvidia
[62:33] GPUs, which means you need to spend more
[62:35] money on cooling and that consumes more
[62:36] power. They made a great chip.
[62:37] >> Well, we can I mean I think the question
[62:39] there and the question for everybody is
[62:40] going to be is that the highest and best
[62:42] use of your time? Right? Like I you
[62:45] know, I tend to think that the frontier
[62:47] companies like there's this belief that
[62:49] they got to be vertical vertically
[62:50] integrated. But if you believe like I do
[62:52] that the race to super intelligence
[62:54] particularly as we get these recursive
[62:56] loops working may be over in the next
[62:58] two to three years, then I think focus
[63:00] focus focus focus. You exist to build
[63:03] the best intelligence in the world and
[63:05] to deliver the best intelligence in the
[63:07] world and you that means you have to
[63:08] have all the revenue. Because if you
[63:10] want to build out the compute that's
[63:12] going to be required to continue to push
[63:13] the frontier, you have to have the
[63:15] revenue in order to support it. So I
[63:16] think you know, subject to the focus
[63:19] question, I think they certainly did.
[63:21] This all brings me back to kind of a
[63:22] reality check, though.
[63:24] Um
[63:25] you know, we just got done talking about
[63:27] test time compute, inference time
[63:28] compute, long-running agents. This is
[63:30] really the thing that's unlocked the
[63:32] revenue this year. Um it all pushes us
[63:34] in the direction of more CapEx. Google
[63:37] just raised $80 billion,
[63:39] right? We've now taken the Mag 5 or Mag
[63:41] 7 free cash flow, you know, down
[63:44] dramatically, 80% um from just a few
[63:47] years ago. Um and Morgan Stanley, you've
[63:50] got this chart in front of you, up to
[63:51] their 2027 CapEx forecast from 950
[63:56] billion to 1.1 trillion. I mean, we were
[63:58] talking about this with Jensen. That was
[63:59] his forecast 2 years ago. You know,
[64:02] obviously, this doesn't even include
[64:03] SpaceX, CoreWeave, etc. So, I think the
[64:06] number on 2027 is likely closer to 1.5
[64:09] trillion.
[64:10] And if we compare this to the total
[64:12] incremental inference revenue, so the
[64:15] thing that the market gets worried
[64:16] about, you know, back to my Sam Altman
[64:18] podcast, you know, in October of last
[64:20] year, can we really afford to spend 1.5
[64:23] trillion of CapEx a year if we're only
[64:26] generating X amount in inference
[64:28] revenue? The thing I think that lit the
[64:29] fuse this year was Anthropic showed up
[64:32] in a major way with revenue, right? And
[64:35] so, we have, you know, the AI lab
[64:38] revenue everybody combined at around
[64:40] $300 billion next year, right? So, can't
[64:44] you know, and go roll that out to 2027
[64:48] uh or that is 2027, 300 billion. So,
[64:50] we're spending 1.5 trillion of CapEx on
[64:53] 300 billion of inference revenue. Does
[64:55] that math math for you? And what would
[64:58] cause you, you know, to to get more
[65:00] nervous again about our ability to
[65:02] continue to make these investments?
[65:04] Because the second we get nervous about
[65:05] it, the entire semi complex is going to
[65:07] come down a lot. Well, what do you think
[65:09] the gross margins are on that 300
[65:10] billion? Yeah, let's call it 50%.
[65:13] >> I I would guess they're probably a
[65:14] little bit higher than that. I might say
[65:16] 60 or 70.
[65:17] But, I mean, that math starts to math,
[65:20] and what I would just say is I think
[65:22] that 300 billion is low, man.
[65:24] >> Yeah. Yeah.
[65:24] >> I just think it's low.
[65:25] >> From your mouth to God's
[65:26] >> Yeah, yeah, exactly. I think I think we
[65:29] end this year well over 200 billion in
[65:32] inference revenue, well over.
[65:34] And so, I think the math really maths,
[65:35] and I do think we have to give uh Jensen
[65:38] >> Yeah, our friend.
[65:39] >> some credit because he said some things
[65:42] that seemed outlandish.
[65:43] >> Right.
[65:44] >> And he was conservative. He was low. He
[65:46] said a trillion 2 years ago.
[65:49] And I mean, he was really low.
[65:50] >> Right.
[65:51] >> And so, like, let's give the guy some
[65:52] credit and think about what he is saying
[65:54] right now.
[65:55] >> For sure, for sure. And and listen,
[65:58] I would say consistently,
[66:00] Elon's been taking the over.
[66:02] Sundar's been taking the over.
[66:05] Sam, Dario, you know, Dario did the
[66:08] podcast with Dwarkesh when he was
[66:10] talking about country geniuses in the
[66:11] data center. He said that will be here
[66:13] by 2028. He said revenues will go into
[66:16] the low hundreds of billions by 2028.
[66:19] So, let's call that, you know, 3 400
[66:21] billion of revenue by 2028. And he said
[66:23] that a while ago now, so he may even be
[66:25] revising up his number. And he said it's
[66:27] hard for me to see that there won't be
[66:29] trillions of dollars in revenue before
[66:31] 2030. And if you're on that revenue
[66:34] trajectory, if we're on a trajectory to
[66:36] 200 by the end of this year, let's call
[66:38] it 4 or 500 by next year, and a path to
[66:42] trillion plus by 2029, then the math
[66:45] maths.
[66:46] >> And we got to keep in mind that half of
[66:48] the spending is there to
[66:50] you know, for training, maybe a little
[66:52] less than half. What is it, Foxy?
[66:53] >> It's probably that depends on the lab,
[66:54] but it's I would say it's increasingly
[66:56] less than half.
[66:56] >> Yes.
[66:56] >> Okay. So, we'll call it 35% is spending
[66:59] that's not revenue generating, but it's
[67:01] going to kind of make the next model.
[67:02] So, I think the math maths.
[67:04] >> Right.
[67:04] >> And there's still this prisoner's
[67:05] dilemma where if you opted out, that may
[67:08] be an existential decision.
[67:09] >> And I think like coming into this year,
[67:11] going back to this kind of what
[67:12] narratives were violated, you know,
[67:15] I think into this year everyone expected
[67:17] token pricing, uh the price of compute,
[67:20] it's all deflationary. And it will be
[67:22] kind of a smooth line deflationary over
[67:24] time. But, I think this year what we've
[67:27] seen is the opposite. And you know, it's
[67:29] all comes back to supply-demand. The
[67:31] demand side of the equation seems to be
[67:33] far outstripping the supply. Right? And
[67:35] I think
[67:36] you look at the deals signed by SpaceX
[67:38] and others, the monetization rates per
[67:41] watt are increasing.
[67:44] Um
[67:44] and
[67:46] look, that is on a a pretty nascent
[67:48] small base of users, right? Like Alex at
[67:51] Well Rock, he has this great um
[67:54] way to frame it.
[67:55] Less than 0.2% of people on Earth are
[67:58] actually using AI in an agentic way.
[68:01] >> Right.
[68:01] >> Right? Like I'm not a technical person,
[68:04] but I'm consuming 500 CPU cores in a VM
[68:07] instance, five GPUs 24/7.
[68:10] >> Yeah.
[68:10] >> I mean, if you draw that out to any
[68:12] meaningful percentage of the population,
[68:14] I mean, we're going to be in, you know,
[68:17] this kind of shortage environment maybe
[68:18] for some time. So,
[68:20] I think that is all positive for this
[68:22] ROI question.
[68:23] >> Man, foxy, 100 to one CPU to GPU ratio.
[68:26] >> [laughter]
[68:28] >> Kind of agentic workflow.
[68:30] >> He said of course.
[68:31] >> [laughter]
[68:32] >> Five.
[68:33] >> Five, yes.
[68:33] >> Yeah.
[68:34] >> I'm being smart with my phone.
[68:35] >> Good, good, good. Excellent.
[68:37] >> I I will say also that ratio of 300 to
[68:40] 1. You know, call it 1.2, 1.5.
[68:44] Um there there is also a rate that now
[68:47] physically we can only expand
[68:50] how much we can produce and how much we
[68:53] can actually increase that spend by,
[68:55] whereas we're seeing the opposite right
[68:57] now on the on the the willingness to pay
[68:59] for these tokens. And actually like when
[69:02] the willingness to pay for these when
[69:03] the monetization per gigawatt is
[69:05] actually increasing from, you know, call
[69:07] it like 20 20 billion um in the in the
[69:11] best best of cases for at the beginning
[69:14] of the year to now like 30 to even
[69:16] pushing 40
[69:18] >> per gigawatt
[69:18] >> per gigawatt.
[69:20] Um all of that is is a very heavy fixed
[69:22] cost base, but all of that is like pure
[69:24] margin flow through now. And you're
[69:26] actually, you know, as we scale like the
[69:29] willingness to pay for for all of this
[69:32] and and now the all of this stipulated
[69:33] by like, you know, everything we're
[69:35] talking about of like how much is open
[69:37] source versus not and all of these
[69:39] different flows, but really like as
[69:40] we're climbing this curve, you know, the
[69:43] the the revenue is might actually
[69:46] outstrip our fixed cost base by by
[69:48] significant amount. And I think that's
[69:49] why all the labs are pushing, you know,
[69:52] the the the gas to the pedals because
[69:54] they all they all see like within if we
[69:57] continue this curve within like 3 years,
[69:59] you know, we're just going to be so
[70:00] short on all the computer
[70:02] >> It's a great I'm sorry, but I mean I
[70:04] like it's a great point. Like if you
[70:06] thought you were getting a when you made
[70:07] these decisions
[70:08] >> Yes.
[70:09] >> in November of 2025,
[70:12] you thought you were getting a certain
[70:13] return.
[70:13] >> Yeah.
[70:14] >> You may be getting triple that return
[70:16] today.
[70:17] >> At Tropic, no way no way did they think
[70:19] they were going to be anywhere close to
[70:20] break even.
[70:21] >> Yeah.
[70:22] >> Right? And and and and in this part of
[70:23] the curve, and the reason like I I I
[70:26] called it accidental profitability that,
[70:28] you know, people have been talking about
[70:29] that because they want to spend a lot
[70:31] more money on computer. They just had a
[70:32] hard time doing it. Now maybe with
[70:33] SpaceX, you know, they could take some
[70:35] of those dollars and and and go spend
[70:37] them other places. But that to me is,
[70:39] you know, a a fundamental change. Um the
[70:42] first argument against the frontier labs
[70:44] was they'll never generate revenue.
[70:46] Okay? And then we that got blown up.
[70:48] Then it was like even if they generate
[70:50] revenue, it'll be really shitty gross
[70:51] margins, and they'll never be able to
[70:52] get make money. And then kind of that
[70:54] that's blown up. And and you know, I
[70:56] think now, you know, people are falling
[70:58] back and they're saying, "Well, they're
[71:00] overcharging. This is token maxi." My
[71:02] good friend, you know, Chamath has said
[71:04] there's no ROI on any of this spend.
[71:05] It's all this token maxi. My best
[71:08] evidence for why we all know, of course,
[71:11] when somebody puts on this much spend
[71:12] like at Altimeter, we're not optimally
[71:14] spending every single dollar. But, the
[71:17] question is, why are millions of
[71:19] independent businesses, small, medium,
[71:21] and large, why are millions of consumers
[71:23] all choosing to do the same thing?
[71:25] They're not dumb. These are, you know,
[71:28] rational economic actors that are all
[71:30] simultaneously saying, "I want to do
[71:32] this because it makes my life better. It
[71:34] makes my business better, etc." To me,
[71:36] that is the best evidence as to why I
[71:38] think this revenue can continue.
[71:40] >> Yeah. And Clark, I think like the point
[71:42] you made is dead on cuz I mean, you want
[71:44] to own asset-heavy businesses in
[71:45] inflationary environments, and token
[71:47] pricing is going up, and supply and
[71:49] demand is tightening, so totally agree.
[71:52] >> Um you know, as we begin to
[71:54] uh find our way to the exit ramp and and
[71:57] [laughter] and wrap here, one of the
[71:58] things I you know, you and I've been
[72:00] doing this for a long time, Gavin, a
[72:02] couple decades. Um you may even sketch
[72:04] longer than me, even though I'm a little
[72:06] bit older than you. Um
[72:08] you know, we have uh I always like to do
[72:10] a market check, because I find a lot of
[72:12] time that analysts come on these things,
[72:15] and they talk their, you know, talk
[72:16] their book, and you know, there are a
[72:18] lot of people who listen to these
[72:19] things, retail investors and others.
[72:21] It's just kind of like, what do we
[72:22] really think? And so, I always
[72:23] characterize as kind of small, medium,
[72:25] and large. Like, what am I doing? Do I
[72:27] have small exposure on? Do I have medium
[72:29] exposure on? Do I have large exposure
[72:30] on? You know, and if you look at what's
[72:32] happened in the markets, semis ripped
[72:34] this year. I mean, like uh you've been
[72:36] doing this a long time. I don't I've
[72:37] never seen it before, right? I've never
[72:39] seen, you know, the doubles and the
[72:41] triples across the board like we saw.
[72:43] But, there's been huge dispersion,
[72:45] right, in the market. Internet's down
[72:47] 16%,
[72:49] uh software's down 8% on the year. You
[72:52] know, spy and and Nasdaq are up, but
[72:54] really up because of their components
[72:55] that are related to AI and compute. And
[72:58] so, the market itself has kind of
[73:00] struggled. Meanwhile, if you were in the
[73:02] stuff that we were invested in, we've
[73:04] all done pretty well. I think you know,
[73:06] I've said it a couple times. I think if
[73:07] the Anthropic revenue had not shown up
[73:09] this year, because that was the overhang
[73:11] on the market, I think the whole market
[73:13] could be down this year. Right? Um but
[73:15] that showed up. You know, we just had
[73:17] these huge months in in in April and
[73:19] May. Um for us, you know, because prices
[73:23] came up so much, because I have some
[73:25] worry about, you know, geopolitics, the
[73:27] macro backdrop with, you know, with with
[73:30] what's going on with inflation in the
[73:32] short run, and just like, you know,
[73:33] needing a little consolidation in this
[73:35] market to answer some of these
[73:37] questions, because now expectations are
[73:39] higher. You know, we dialed back from
[73:40] what I would call large for Altimeter to
[73:42] something kind of like medium small. Um
[73:45] again, it's never all or nothing for us.
[73:47] It's like, what is the risk-reward at a
[73:49] given price?
[73:51] Um and so, we think this is a, you know,
[73:53] maybe going to be a period of
[73:54] consolidation on way to much higher
[73:56] highs. Um curious just how you run the
[73:59] book, how you think about it like a
[74:00] portfolio manager.
[74:01] >> Very similarly, man. I always think
[74:03] stocks, the markets, I imagine them as
[74:05] runners. Okay?
[74:07] And like in '22,
[74:09] that runner had gone downhill. It had a
[74:11] lot of energy, man.
[74:12] Yeah, it was painful. It wasn't fun. Um
[74:15] but coming out of that, there was a lot
[74:17] of kind of pent-up upside in the market.
[74:19] And you know, the market, particularly
[74:21] last 2 months, it has run up a very
[74:23] steep hill.
[74:24] And a lot of companies, semiconductor
[74:26] companies in particular, you know,
[74:28] ironically, you know, Nvidia and
[74:30] Broadcom, they they have been laggards.
[74:32] >> Totally.
[74:33] >> And so, but a lot of these, like I do
[74:36] see a lot on X about finding the next
[74:38] bottleneck. I think that was the last
[74:40] game. That game is over.
[74:42] You've had a lot of stocks that forget
[74:44] climbing a mountain or a hill. They've
[74:47] gone straight up a cliff, okay?
[74:49] >> Yes. They're tired. They need to rest.
[74:52] And we'll see, do they just rest at the
[74:55] top of that cliff they climbed? Do they
[74:57] hang out on the in their harness for a
[74:58] while?
[75:00] We've seen some.
[75:02] Or do they need to go downhill for a
[75:03] bit? We'll see, but I'm thinking very
[75:06] similarly to you.
[75:07] But it is and I think there's, you know,
[75:09] the market is seasonal. I think there's
[75:11] real
[75:12] real concerns around inflation and
[75:14] rates.
[75:15] >> What was CPI this morning?
[75:17] >> uh 4.2. I think we added core came in at
[75:20] like 0.2 versus 0.3, so a little bit
[75:22] better.
[75:23] Um but you know, clearly we're we're
[75:25] above four again.
[75:26] And um and and there's short-term
[75:28] pressure on, you know, core PCE, etc.
[75:32] Um and we have some unknown unknowns,
[75:33] but the market, I mean, if I had told
[75:35] you the fact pattern for this year, that
[75:37] we're going to be in a war with Iran,
[75:38] that, you know, oil was going to be at
[75:40] 100 bucks, that CPI was going to be
[75:42] creeping back up, that internet was
[75:44] going to be down 15%. Software is going
[75:46] to be down 8%. You would have said, "I
[75:47] want nothing to do with that market,
[75:49] right?" And here we are. The market's
[75:52] done pretty good in the stuff that we
[75:53] traffic in because the world
[75:55] underestimated AI revenues and
[75:58] underestimated the amount of compute
[75:59] that was going to be needed.
[76:01] >> It's odd to say you know, we're heading
[76:02] into a seasonally weak period with all
[76:04] of these fears. AI has actually been
[76:06] seasonal for the last three summers.
[76:09] Token consumption is kind of plateaued,
[76:11] slowed down, and that's cuz you know,
[76:12] college kids are big AI consumers and
[76:14] they don't use as much AI, you know,
[76:16] hopefully they're all using it to learn
[76:17] and not cheat.
[76:19] But that may happen. It may not happen
[76:21] because of generative AI.
[76:23] >> is building swarms of agents, building a
[76:26] SpaceX model. He's going to the SpaceX
[76:27] IPO with me at the exchange on Friday,
[76:30] but I he had to build an AI model using
[76:32] AI agents. He had to build a model, a
[76:34] DCF before we go to the exchange. He is
[76:37] mesmerized. He is absolutely and it's
[76:40] extraordinary what he's doing.
[76:41] >> So he's one kid who's not easy to less
[76:43] computer
[76:44] >> [laughter]
[76:44] >> or something.
[76:44] >> He's burning it. He's burning it.
[76:46] >> Yeah, but you know, if token consumption
[76:48] plateaus, if open source takes some
[76:51] share, there's a Silicon data index that
[76:53] has showed, which is an index of kind of
[76:56] consumption and pricing. I think there
[76:57] may have been a little bit of a shift
[76:59] over the last 2 weeks to open source
[77:00] tokens that are cheaper. Like people
[77:02] looking at that data as bearish or not
[77:04] understanding it. But nonetheless, like
[77:07] I just think there's reasons, you know,
[77:08] to look around, be careful, be
[77:10] thoughtful. I always assume a bullet is
[77:11] coming for me. Head on [laughter] a
[77:12] swivel. It's the bullet you don't see
[77:15] that gets you. So, I'm trying to spin as
[77:16] fast as I can.
[77:18] But yeah, it's the market may need to
[77:20] take a breather. But man, when I think
[77:22] about what Noam Brown said
[77:25] and when I see the capabilities of
[77:27] Fable,
[77:29] it's just hard for me to get too
[77:31] bearish.
[77:33] >> I mean, like to me
[77:34] um and we got two, I think, of the most
[77:37] extraordinary guys of, you know, the
[77:39] next generation, you know, sitting in
[77:42] the room. We have at Altimeter, we have
[77:43] deep admiration for the work that you
[77:45] guys do. I always appreciate when you
[77:47] send me a note about the work that we do
[77:49] and we publish. Um but for the guys who
[77:52] are newer to the business, they might
[77:54] think this is the way that it kind of
[77:55] always was, right? And like this line,
[77:59] the steepening of the line of creative
[78:01] destruction, the steepening of the line
[78:04] of, you know, scale advantages. Um I
[78:07] always believed it was to it was going
[78:10] to be true. I never thought it would be
[78:11] true at this rate. I went back last
[78:14] night. In the last 7 years, we've added
[78:17] 1 trillion of revenue to the Mag 7
[78:20] in the last 7 years, okay? To get to a
[78:22] trillion, to get to the first trillion
[78:25] of, you know, took over 20 years. In the
[78:28] last 7, we had another tr- trillion and
[78:30] that added 17 trillion in market cap.
[78:33] That trillion dollars, okay?
[78:35] I The forecast now that we're going to
[78:37] add another trillion of revenue in just
[78:40] three companies SpaceX Anthropic and
[78:43] open AI over the next four to five
[78:45] years.
[78:46] Okay, like not seven companies three
[78:49] companies and in half the time right and
[78:52] so I would say that you know, we are
[78:54] going to have bumps in the road. I know
[78:56] that it's going to be like this but
[78:58] we're going to higher highs because the
[79:00] size of the prize. This is going to
[79:02] transform five ten 15% of global GDP.
[79:06] There is no doubt in my mind and 10% of
[79:09] global GDP is 10 trillion dollars. It's
[79:11] an exciting future to be a part of it's
[79:13] fun to do it with you guys. I think
[79:15] we're going to have to do our work to do
[79:16] the things to make sure America wins and
[79:18] that we evolve the social contract keep
[79:20] everybody you know lift the floor take
[79:22] everybody with us on this ride
[79:25] but it's a it's a it's a really exciting
[79:27] time to be doing what we're doing it's
[79:28] fun to be doing it with you guys.
[79:30] >> Yeah, I just want to say Brad thanks for
[79:31] having us and thank you for what you've
[79:32] done with the Trump accounts. I actually
[79:34] think it's super important for America
[79:36] for the world to give people an equity
[79:38] stake at a very young age. They they
[79:41] will see it compound over their
[79:42] lifetimes. This is a great thing you've
[79:44] done for the world. So thank you. I'd
[79:46] echo all your comments like deep
[79:48] admiration for you your team gratitude
[79:50] for the collegiality and friendship
[79:52] between our firms. I know Clark and Foxy
[79:55] they hang out like all the time.
[79:57] >> That's a people think that you know and
[79:59] there are people in our business who
[80:00] don't want to share anything. Our view
[80:02] is like we open source it
[80:04] but there are very few people who we
[80:06] actually call and ask their opinion
[80:08] because there are very few people who do
[80:09] the thousands of hours of work that we
[80:11] do
[80:12] you know that are adding to that and you
[80:14] do it and we appreciate that and you do
[80:16] as well Gavin we appreciate that. So
[80:18] with that love fest let's call it a
[80:19] wrap.
[80:20] Thanks for being here.
[80:21] >> Thank you.
[80:47] >> Mhm.
