A16z's Baker Says AI Supply Crunch Will Persist Through 2028
a16z
Gavin Baker argues that physical bottlenecks and political pushback will constrain compute supply through 2028, driving up token prices and giving Nvidia an unassailable financial flywheel.
## Background
Published by a16z on YouTube on August 31, 2026, this episode features host David George in conversation with Gavin Baker. According to the page description, the discussion explores why AI demand is dramatically outrunning compute supply, the economics of infrastructure buildouts, and how value will be captured across labs, applications, clouds, and NVIDIA.
## TL;DR
Gavin Baker and David George argue that the current AI buildout is an 'And' world rather than a zero-sum game, with massive value creation occurring simultaneously across frontier labs, open-source models, hyperscalers, and hardware providers. Despite public equity volatility, fundamental compute demand and token usage are accelerating dramatically, driven by early-stage market diffusion and rapid shifts toward autonomous execution. Physical supply constraints, capital payback structures, and regulatory hurdles mean the primary risk through 2028 is a severe global compute shortage rather than overbuilding. The conversation highlights how vertical integration, multi-model routing, orbital data centers, and innovative financing models solidify NVIDIA's position at the center of the AI ecosystem.
## Atmosphere & style
The discussion is fast-paced, highly analytical, and intellectually energetic, blending institutional finance precision with optimistic tech-forward ambition. David George acts as an engaged anchor, prompting deep economic inquiries while Gavin Baker delivers sharp, confident, and anecdote-rich arguments. The narrative style feels like an insider masterclass—grounded in real-world firm metrics and supply-chain realities, yet comfortably expanding into forward-looking visions of orbital compute and robotic automation.
## Key takeaways
- [01:06](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=66s) There is a striking paradox in the AI landscape: while public equities face drawdowns, underlying compute demand and AI adoption across labs, open source, and enterprise tools are accelerating with no negative data points.
- [02:41](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=161s) AI value capture is an "And" world rather than a zero-sum winner-take-all market, allowing frontier labs, hyperscalers, Neoclouds, open-source models, and application layers to thrive simultaneously.
- [05:50](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=350s) Frontier AI lab economics differ from traditional software due to the direct coupling of revenue and compute, leading to voluntary revenue volatility when labs pivot gigawatts from inference to next-gen training.
- [11:44](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=704s) Capital payback periods for AI compute infrastructure are exceptionally fast, with Neoclouds achieving full hardware payback in 9 to 10 months backed by low-cost institutional debt financing.
- [14:34](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=874s) The demand side of AI remains in its earliest diffusion stages, with current industry revenues driven by under 30 million heavy users out of 1.5 billion global knowledge workers.
- [29:30](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=1770s) Data center opposition and physical bottlenecks run the risk of creating global compute shortages rather than oversupply, which could invert token pricing and cause severe compute inequality.
- [34:21](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2061s) Orbital data centers in sun-synchronous orbits represent a viable and economically attractive swing capacity for global compute, avoiding terrestrial land, power, and labor inflation.
- [47:38](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2858s) The enterprise software landscape is shifting toward multi-model routing and open-source base models fine-tuned on private data, rather than total reliance on a single closed frontier lab.
- [59:50](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3590s) To capture the software abstraction layer long-term, software vendors must remain vertically integrated down to the physical compute fleet to avoid fatal margin compression.
- [01:00:26](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3626s) NVIDIA maintains systemic dominance by offering a comprehensive 9-chip full-stack ecosystem, locking up 70%–80% of fab capacity, and providing residual value guarantees that make its clusters highly bankable.
- [01:09:17](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=4157s) Non-GPU custom ASICs face severe real-world hardware execution risks and must compete against integrated multi-chip system architectures rather than single compute dies.
- [01:12:10](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=4330s) The underlying deal structures of hardware sales—ranging from direct equity investments to performance-indexed warrants—reveal whether customers adopt chips out of genuine organic preference or financial engineering.
## Detailed walkthrough
### [00:00](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=0s) Intro and the AI Acceleration Paradox
Gavin Baker opens the discussion by asserting that when the history of the 21st century is written, it will likely be remembered as the "Age of Elon and Jensen"—referring to Elon Musk and Jensen Huang—because of how fundamentally they are altering the fabric of human society and civilization.
David George introduces the central puzzle: despite macro stock market jitters and public tech equities dropping, the fundamental AI landscape is accelerating rapidly. David asks Gavin whether, during his time spending the summer on the West Coast, he has found anyone who can present a credible, quantitative bear case for AI. Gavin confirms he asks every founder and executive a standard question: *"Can you tell me one quantitative data point in your business that's getting worse? Just one."*
Through July and August 2026, he hasn't found a single person with a negative data point. While Anthropic might be in a temporary quiet period, the rest of the ecosystem—including OpenAI, open-source models, and xAI's Grok (especially after Grokbot)—is experiencing a dramatic acceleration. This creates a striking paradox: public stocks are experiencing significant drawdowns, but underlying AI adoption and compute demand are ramping up faster than ever.
---
[02:41](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=161s) The "And" World: Why Everyone Can Win
David and Gavin address a common misconception among institutional LPs and public investors: the assumption that AI is a zero-sum market where one frontier lab or architecture must destroy the others. Echoing a perspective shared by Eric Vishria on Patrick O'Shaughnessy's podcast, Gavin argues that we are in an "And" world, not an "Or" world.
In this scenario, almost every major player can capture massive value simultaneously:
- **Frontier Labs:** OpenAI and Anthropic continue to scale revenue rapidly.
- **Hyperscalers & Hardware Providers:** SpaceX/xAI and Meta expand infrastructure; Google wins by selling vast quantities of TPUs; NVIDIA remains at the absolute center of the universe.
- **Clouds & Infra:** NeoClouds and inference clouds on top of them thrive on token volume.
- **Open Source & Applications:** Open-source N-1 models thrive alongside application companies that execute effectively.
Discussing Anthropic's trajectory, Gavin and David note that the company likely rebased and cleaned up its accounting definitions to align revenue reporting with OpenAI. The frontier model competition has evolved into strategic gamesmanship around model releases. Frontier labs hold back advanced checkpoints, waiting for competitors to drop new models (e.g., waiting for OpenAI to release Astra) before releasing their counter-punch (e.g., "Fable 5.1") just hours later.
While Anthropic's culture explicitly screens for mission-aligned employees rather than mercenaries—even asking candidates how they would feel if equity went to zero—Gavin points out that "you can't afford the compute you want for your mission if the equity goes to zero." Anthropic has effectively become an "accidental enterprise company," driving massive enterprise adoption as a byproduct of its core research mission.
---
[05:50](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=350s) The Economics of Compute: Payback Periods and Capital Allocation
David and Gavin analyze the unique capital allocation model of frontier AI labs compared to traditional cloud and internet companies. In traditional software, infrastructure costs and revenue generation were decoupled, allowing companies like Google or Meta to generate massive free cash flow smoothly.
In AI frontier labs, revenue and compute are directly coupled. Gavin illustrates this dynamic with a concrete scenario:
* **Hypothetical Lab Allocation:** Suppose a lab controls 10 gigawatts of power.
* **Inference Allocation:** They allocate 8 GW to inference and 2 GW to training/research.
* **Monetization Rate:** Labs currently monetize inference at roughly $100 million per gigawatt (or ~$60 billion per GW annually in extreme utilization scenarios). At 8 GW monetized, that yields $480 billion in annualized revenue.
* **The Pivot Risk:** If the lab hits a major research breakthrough, they might abruptly shift from 8 GW of inference down to 2 GW, reallocating 8 GW back to training next-generation models. Annualized revenue instantly drops from $480 billion to $120 billion.
Gavin emphasizes that public markets will have to adjust to this volatility, as frontier labs will prioritize long-term model supremacy over smooth quarterly free cash flow. Frontier labs are generating huge *operating* cash flow, but they instantly re-invest it into GPUs, XPUs, and heavy first-party product token subsidies.
---
[11:44](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=704s) Capital Payback Timelines and Wall Street Financing
Despite the massive capital expenditures (CapEx), the unit economics of AI compute infrastructure remain extraordinarily lucrative:
- **Nebius & CoreWeave Paybacks:** Based on public disclosures, Neoclouds like Nebius achieve full hardware payback in just **9 to 10 months**. When bringing a 1-gigawatt site online (costing ~$50 billion), upfront customer commitments cover 50–60% ($25–$30 billion) immediately. Monetizing remaining capacity on the spot market at $2 to $8 per GPU hour accelerates paybacks even further.
- **SpaceX / xAI Cluster Speed:** SpaceX and xAI build massive clusters significantly faster than legacy providers, compressing their payback timelines well inside the sub-one-year mark.
- **Low Cost of Capital:** Financial giants like Blackstone, KKR, and Apollo are aggressively financing GPU and data center buildouts at low costs of capital. They recognize that the useful operating lives of these chips are continually extended as model capabilities advance and token monetization per gigawatt increases.
---
[14:34](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=874s) Demand-Side Diffusion: The 30 Million User Baseline
Addressing the bear argument that supply overbuild will destroy compute economics, David and Gavin turn to the demand side. Current AI revenues across the industry (totaling tens of billions of dollars) are being generated by an extremely narrow slice of the global workforce.
- **Current Heavy User Base:** Gavin estimates there are fewer than **30 million heavy, paying AI users** globally—and notes he would "take the under" on that figure, suggesting the true number of power users may be under 10 million.
- **Enterprise Power Law:** Within enterprises, token consumption follows an extreme power law. While traditional banks spend ~1% of software budgets on tokens and tech-forward firms spend high single digits, the top 1% of engineers consume 10x to 100x more tokens than the median engineer.
- **The Diffusion Horizon:** With over **1.5 billion global knowledge workers**, AI adoption is still in its earliest stages of market diffusion.
- **Portfolio Token Spend:** Leading startups in the Andreessen Horowitz (a16z) portfolio now spend high single digits to 10% of their total human compensation budgets strictly on LLM token consumption.
[016:04] Internal Token Consumption and the "Grokbot" Shift
David George and Gavin Baker look at how quickly AI usage is ramping up inside actual firms. Gavin highlights his own firm's internal metrics: their token consumption exploded 100x between March and August. After getting access to Grokbot Enterprise, just two heavy users threatened to drive token spend another 10x to 20x higher in a single month—not through wasteful usage, but through immensely productive, high-value workflows.
The two reflect on their own habits and how younger generations naturally adapt to AI. Gavin, who is 50, jokes about coaxing his parents onto iPhones and iPads, noting that 23-year-olds today possess an effortless fluency with AI tools. While Gavin spent hours writing scripts with Claude Code to build personal utilities—like family calendar apps—he found that Grokbot could generate far superior versions in a mere 7 to 12 seconds.
- **Token Explosion**: Gavin's internal firm spend grew 100x from March to August, with projections of another 10x–20x increase following Enterprise bot deployment.
- **Micro-apps Built in Seconds**: Gavin recounts sharing on Patrick O'Shaughnessy's podcast how he built custom summarizers for podcasts, Substack, and X, as well as X sentiment trackers. Tools like Grokbot executed these in 7–12 seconds, rendering manual Claude Code workflows obsolete.
[019:16] Shift to Autonomous Action and Historical Technology Bubbles
David notes that early AI tools (like Claude Code or Codex) were largely reactive—functioning as reading aids, summarizers, and knowledge enhancers. The true paradigm shift occurs when bots move from passive synthesis to proactive recommendation and autonomous execution. Gavin shares that he is currently running head-to-head "horse races" between Grokbot, Codex, and portfolio company Town to analyze his workflows and recommend automated actions.
The conversation turns to the history of financial bubbles. Gavin observes that every transformational infrastructure shift—from canals and railroads to automobiles, radio, television, steel mills, and the internet—triggers market over-excitement, overvaluation, and eventual overbuilding. However, he notes a critical distinction today: much of the current AI buildout is funded via corporate operating cash flow rather than short-term debt, which lowers the risk of immediate debt-driven liquidations.
- **Reactive vs. Autonomous AI**: Moving beyond 20%–90% AI-assisted coding toward agents that actively recommend and execute operational tasks.
- **Historical Analogy Corrective**: Gavin admits he previously claimed on a podcast that the South Sea Bubble was tied to calculating longitude at sea, before realizing it was more akin to a speculative tulip mania.
- **Debt vs. Cash Flow**: Debt-funded infrastructure buildouts demand immediate 3-year ROI, whereas cash-flow-funded investments allow companies to tolerate longer payoff horizons.
[022:38] Physical Constraints, Regulation, and the Public Narrative
Gavin emphasizes that physical limitations—such as copper availability and wafer production (mentioning conversations with Jazz at Watson Wafers)—act as natural speed bumps for the AI buildout. Rather than viewing these delays as purely negative, he argues that constraints, rising real interest rates, and regulatory friction serve as healthy societal guardrails to prevent runaway overbuilding.
The discussion grows heated over the public perception and political pushback surrounding data centers. Gavin criticizes doom-mongering narratives like those of Eliezer Yudkowsky ("if we build it, everyone will die") and references a recent public X exchange with Anthropic’s Dario Amodei and Sholto Douglas. Gavin argues that the AI industry needs to communicate its tangible benefits—such as curing cancer and lifting up working-class Americans—rather than letting fear dominate the discourse.
- **Macro Constraints**: Raw material shortages (copper mines, silicon wafers) and higher real interest rates slow the buildout naturally.
- **Working-Class Economic Revival**: Data center construction creates lucrative non-college career paths for electricians, plumbers, and HVAC technicians, while boosting local town tax bases by 10x rather than merely doubling them.
- **Debunked Myths**: Gavin and David agree that fears over data center water consumption have been completely debunked, noting that natural gas provides clean, behind-the-meter power.
[025:38] Data Center Economic Impacts and Re-Industrializing America
Addressing anti-data-center critics, Gavin brings up Loudoun County, Virginia—the highest per-capita income county in the United States, which also happens to house the world's highest density of data centers. When opponents sarcastically suggest putting data centers in rich neighborhoods, the reality is that the wealthiest county already embraced them to immense economic benefit.
Gavin further argues that opposition campaigns against US data centers are actively hindering a historic re-industrialization of America. Thanks to cheap domestic natural gas ($2–$3 compared to $20–$25 in Europe and Asia), the US holds a massive energy cost advantage for manufacturing and compute infrastructure.
- **Loudoun County Model**: Demonstrates that high data center concentration directly correlates with massive municipal tax revenue and regional wealth.
- **Energy Cost Disparity**: US natural gas costs $2–$3 per unit versus $20–$25 in Europe and Asia, giving America a distinct structural advantage in power-intensive industries.
- **Corporate Storytelling**: Gavin commends Meta for its long-standing practice (started by Sheryl Sandberg on earnings calls) of spotlighting specific small businesses—like a two-woman cake bakery in Des Moines expanding to 15 locations—to ground tech adoption in human stories. He urges OpenAI, Anthropic, SpaceX, Nvidia, AMD, and Broadcom to adopt the same strategy.
[029:30] The Risk of Underbuilding and "Compute Inequality"
David and Gavin conclude that the primary macro risk through 2028 is not an oversupply of compute, but a severe global shortage. Political opposition, permitting delays, and supply chain bottlenecks mean that scheduled capacity builds will likely lag behind surging demand.
Citing commentary from Dwarkesh Patel, Gavin discusses the counterintuitive scenario where token prices could actually skyrocket 10x due to supply constraints. If anti-data-center "degrowthers" block capacity expansions, the result won't be a halt to AI, but severe "compute inequality"—where only well-capitalized megacorporations and the wealthy can afford access to frontier intelligence, locking out ordinary consumers.
> "The consequence of these data center degrowthers may be real compute inequality where big companies and wealthy people can afford compute... and that happened because you wouldn't let us build data centers." — **Gavin Baker**
- **Supply Deficit**: Available compute capacity is projected to remain severely constrained through at least 2028.
- **Price Inversion**: Instead of intelligence becoming a dirt-cheap commodity, supply bottlenecks could force token prices upward.
- **Consumer Surplus vs. Ad Models**: Mass-market consumer access requires cheap compute or advertising-supported models; a structural supply shortage threatens to delay affordable AI access for the broader public.
[31:51](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=1911s) Open-Source Tokens and the Economics of Compute
Gavin Baker addresses common misconceptions surrounding open-source AI, highlighting that "open source" tokens are not fundamentally free to produce. On a physical level, generating a token from an open-source or open-weights model requires the exact same amount of compute as a proprietary frontier model of comparable size—the only difference lies in the margin or markup charged on top of the base infrastructure. David George and Gavin point to examples like Kimi, whose open-weights license stipulates a 30% revenue share, demonstrating that open weights do not imply a lack of monetization. Furthermore, on a per-task basis, reasoning models or less optimized architectures can be far more token-hungry, making them expensive to run at scale despite their open nature.
Gavin expresses deep admiration for tech leaders like Jensen Huang and Elon Musk, framing them as generational figures who are fundamentally altering human civilization. Drawing an analogy to the Victorian Era, Gavin predicts that history will view this period as the "Age of Elon and Jensen." He emphasizes the massive consumer surplus provided by initiatives like SpaceX's Starlink, which brings low-cost internet access to the poorest and most remote communities on Earth—a feat that traditional telecom infrastructure could never justify economically due to deployment costs and low willingness to pay. David agrees, emphasizing that future incremental internet capacity will be built in space rather than on Earth.
[34:21](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2061s) Demystifying Orbital Data Centers and SpaceX's Engineering Reality
David and Gavin address the growing discussion surrounding orbital compute, dispelling wild science-fiction caricatures of floating space fortresses. David clarifies that orbital data centers are not "big buildings in space" or "the Death Star," but rather compact payloads roughly the size of a small airplane or a cluster of a few people standing together, anchored by massive solar wings. By deploying compute in a sun-synchronous orbit, the hardware's radiators remain permanently shadowed by the rack, enabling passive cooling into deep space. Gavin notes how skepticism from physics PhDs on social media usually evaporates once they consult with SpaceX engineers. While critics spend hours pondering the problem, SpaceX has thousands of world-class engineers spending hundreds of thousands of collective hours proving that orbital compute is a solved engineering challenge—and actually simpler than a Starlink satellite, which requires complex moving phased-array antennas.
```
+-----------------------------------------------------------------------------------+
| ORBITAL DATA CENTER MODEL |
| |
| [ Solar Wings ] <--- (Continuous Sunlight for Power) |
| | |
| +-----------+ |
| | GPU Rack | <--- Payload (Size of small airplane / NVIDIA NVLink cluster) |
| +-----------+ |
| | |
| [ Radiator Panel ] <--- (Kept in permanent shadow via Sun-Synchronous Orbit) |
+-----------------------------------------------------------------------------------+
```
While initial cost curves seem daunting, Gavin points out that every major Musk venture—from Falcon 9 reusability to the Tesla Model 3—followed a dramatic cost reduction trajectory. Coupled with severe self-inflicted power and land supply crunches for terrestrial data centers, orbital compute serves as a critical "swing capacity" for global AI demand.
- **Orbital Thermal Management:** Sun-synchronous orbits keep the cooling radiator continuously shielded from the sun by the server rack, radiating heat directly into cold space.
- **Starlink Complexity vs. Orbital Compute:** Orbital compute nodes do not require dynamic phased-array tracking antennas, making their mechanical design simpler than standard Starlink broadband satellites.
[37:06](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2226s) The Economics of Orbital Compute vs. Terrestrial Expansion
Gavin breaks down the unit economics comparing terrestrial and orbital data center builds. On Earth, building a gigawatt-scale data center costs roughly $50 billion, with approximately $35 billion allocated to the IT/GPU equipment and $15 billion spent on power, cooling, land, and labor. While the $35 billion IT cost remains largely constant whether in space or on land, the $15 billion terrestrial infrastructure cost is heavily inflationary due to rising labor rates for skilled trades (e.g., electricians) and escalating raw material prices like copper.
In contrast, space eliminates the need for ongoing terrestrial power and cooling infrastructure by utilizing solar panels and deep-space radiation. The primary trade-off is the cost of orbital launch. With full Starship reusability, launch costs for a gigawatt payload are projected to drop under $1 billion, instantly flipping the economic equation in favor of space. Gavin notes that while latency and speed-of-light constraints mean frontier model training will still occur on Earth, an increasing fraction of world compute will migrate to orbit. Elon Musk and Jensen Huang have co-designed a space-optimized NVIDIA Rubin rack targeted for launch in late 2027 or 2028.
- **Terrestrial Cost Breakdown:** ~$35B for GPUs/IT hardware; ~$15B for power, cooling, real estate, and local labor.
- **Orbital Cost Breakdown:** ~$35B for GPUs/IT hardware; <$1B for Starship launch capacity once full reusability is scaled.
- **Launch Schedule:** Co-designed NVIDIA Rubin orbital racks are scheduled for Starship deployment starting around late 2027 to 2028.
[39:46](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2386s) Starlink's Expansion, Grok, and SpaceX's Strategic Moat
Looking beyond compute, Gavin outlines Starlink's expanding total addressable market (TAM). Combining global mobile coverage with fixed satellite broadband unlocks an 800 to 900 billion dollar wireless market, pushing Starlink's total potential market opportunity close to $2 trillion. Furthermore, xAI's rapid execution with Grok and Grokbot demonstrates a "heads-you-win, tails-you-win" dynamic. If xAI's first-party applications succeed, their heavy compute investments are immediately monetized; if they overbuild, high compute scarcity allows them to achieve sub-six-month payback periods by renting capacity or bundling services. Gavin envisions potential product bundles combining Starlink satellite connectivity, Grok AI access, and X advertising subscriptions, mirroring how enterprise clouds were historically scaled.
David and Gavin express excitement over SpaceX's hardware manufacturing and launch footprint, citing the development of Starbase Louisiana designed to support thousands of launches per year. Gavin predicts the emergence of orbital starbases across multiple global coasts and international regions—such as the Middle East, Europe, or East Asia—enabling a future with over 5,000 launches annually.
- **Total Addressable Market:** Starlink mobile and broadband satellite communications address a near-$2 trillion global market.
- **Compute Payback:** High market demand for AI compute yields sub-six-month payback periods on excess GPU infrastructure.
- **Launch Frequency:** Starship launch pads are being engineered for rapid turnaround, aiming for more than two launches per pad per day.
[44:04](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2644s) Asteroid Mining, Planetary Zonation, and the Robotic Frontier
When asked about the most futuristic 10-to-20-year vision for SpaceX, Gavin highlights orbital asteroid mining as an inevitable economic reality rather than science fiction. Pointing to metallic asteroids like 16 Psyche—which holds more gold, silver, platinum, and precious metals than the entire accessible Earth's crust—Gavin outlines a workflow where Starship and a permanent lunar base are used to capture asteroids. These space rocks can be moved into a stable geosynchronous orbit above an American-owned Pacific atoll. Earth-side delivery of refined materials could be handled autonomously via Tesla Optimus humanoid robots with zero risk to human life, effectively providing unlimited raw materials.
> "I think in the future Earth is going to be zoned residential... all heavy industry will take place in outer space." — Jeff Bezos (cited by Gavin Baker)
Gavin underscores Jeff Bezos's vision that Earth will eventually be zoned "residential," shifting all pollutive heavy industrial processes, mining, and power-intensive computing into space. In the near term (within the next 6 to 8 years), Gavin expects fleets of Starships to land on Mars. Modified "Mars Colonial Transporters" will deploy Optimus robots to lay out solar arrays, set up battery banks, assemble compute racks, and establish high-bandwidth Starlink communications. Streaming 4K video feeds of autonomous robots carrying out construction on the Martian surface will mark an unprecedented triumph for technology and human exploration.
- **Asteroid 16 Psyche:** Contains reserves of platinum-group metals and gold exceeding the total mined quantities in Earth's crust.
- **Mars Timeline:** Uncrewed Starship cargo missions and robotic infrastructure deployment targeted within an 8-year window.
Here is a detailed walkthrough of this portion of the conversation between David George and Gavin Baker:
[47:38](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2858s) Open Source, NVIDIA's Meta-Strategy, and Microsoft's Survival Bet
The conversation opens with Gavin Baker and David George reflecting on Microsoft's strategic position in the AI landscape. While companies like Apple made what Gavin views as extreme bets against the immediate future, Microsoft has taken a more adaptive approach. Although CEO Satya Nadella pledged 18 months ago that Microsoft would build its own top-tier competitive frontier models—a push that ultimately failed—the overall market environment has unexpectedly shifted to favor Microsoft's strategy.
Gavin argues that the world is moving away from a single winner-take-all model and toward an ensemble of models along a Pareto frontier. In this emerging paradigm, no single model is the best at everything. For the top 10,000 global enterprises, the future likely involves deploying high-quality open-source base models customized with proprietary data. Crucially, Gavin notes that chip companies like NVIDIA are uniquely positioned to finance massive open-source training runs—potentially dropping $50 billion to $100 billion without breaking a sweat—to ensure open-source models stay on the heels of closed frontier labs. NVIDIA's acquisition of Poolside was a deliberate move to bring top-tier American open-source AI talent under its roof.
- **Microsoft's Pivot:** Failed to build an in-house frontier model to rival OpenAI, but benefits hugely if open-source and multi-model routing become the enterprise standard.
- **NVIDIA's Training Bankroll:** Because custom ASIC labs create weird incentives across the supply chain, NVIDIA can trivially fund $50B–$100B open-source training runs to drive compute demand.
- **Data Protection & Fine-Tuning:** Enterprise customers fear handing over their core intellectual property (IP) and context to frontier labs—especially given shifting Zero Data Retention (ZDR) policies—making local RL (reinforcement learning) and SFT (supervised fine-tuning) on open-source bases far more attractive.
> "If intelligence is like a super important input into your business, you want to own and control your intelligence, its capabilities, its cost."
---
[51:07](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3067s) Multi-Model Routing, "Grokbot," and the Corporate Chessboard
Gavin highlights how leading-edge implementations already rely on intelligent routing across diverse models behind a single interface. He points out that system setups like xAI's "Grokbot" orchestrate a blend of models behind the scenes—combining fast execution engines like Gemini 3.7 Flash and Grok 4.6 with heavy reasoning models like Claude Opus. In these architectures, the most capable frontier model acts as the strategic planner, while cheaper, specialized models handle lower-cost execution.
This multi-model reality creates a friendlier environment for Microsoft and application providers compared to a duopoly dominated by just two closed frontier labs. Gavin gives immense credit to Meta, noting how remarkable it is that Meta re-entered the top tier after being written off, while Google's Gemini temporarily lost its dominant luster in capability debates. The speakers frame this dynamic as the highest-stakes game of corporate chess ever played.
- **Model Ensembles in Production:** Systems dynamically route tasks: top-tier frontier models handle high-level reasoning and planning, while smaller models execute sub-tasks cost-effectively.
- **Notable Implementations:** Product architectures like Grokbot, Cursor, Harvey, and Fireworks Nexus are proving the commercial viability of multi-model abstraction layers.
- **Fireworks Nexus Model:** Allows enterprises (e.g., Goldman Sachs, Morgan Stanley, JPMorgan, Fidelity) to select open-source models, perform RL on their private data, and serve them seamlessly behind a single router alongside frontier options.
> "This is kind of like the highest stakes game of corporate chess ever played... You've seen some people come out of the game, others come back in."
---
[54:01](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3241s) Who Wins the "Abstraction Layer of Intelligence"?
David and Gavin debate who will capture the ultimate prize in enterprise software: becoming the trusted "abstraction layer of intelligence" for organizations and consumers. Gavin compares the difficulty of building this software layer to running a massive American retail empire. On paper, starting a retail chain sounds simple—just open 1,000 clean, well-lit stores across 50 states, stock regional preferences, and manage employees with 100% annual turnover to unlock a $50 billion business. Yet, in reality, only a handful of companies in American history have ever pulled it off because execution across messy real-world constraints is brutally difficult.
Similarly, making model orchestration, continuous RL, zero-data-leakage security, and automated model swapping seamless for an enterprise sounds easy in pitch decks, but is extraordinarily complex to deliver in practice. Coding tools like Cursor succeeded because they stayed obsessively product-focused rather than chasing grand visions of creating a "digital deity" or ASI (Artificial Superintelligence). They treated AI as a practical engineering problem, meeting developers where they were.
- **The Difficulty of Knowledge Work:** Unlike coding—which is uniquely verifiable and perfectly documented—general enterprise knowledge work is unstructured and messy, making widespread automation far harder to pull off smoothly.
- **The $500M DIY Fallacy:** Kirkland & Ellis committing $500 million to build its own internal AI system validates the massive market size, but Gavin cautions that building a custom system isn't a one-time capex spend—models require continuous, expensive upgrades and maintenance.
- **The Vertical Battleground:** Multi-billion-dollar incumbent platforms (Microsoft, Databricks, Snowflake, Salesforce, Workday, Palantir) and hyper-focused vertical agents (Harvey in legal, Cognition/Cursor in code) are on a collision course to own this orchestration layer.
> "Everybody else in the lab space had this like 'we're creating a digital deity, AGI and ASI'... and the Cursor guys were just like, 'we want to make great product.'"
---
[59:50](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3590s) The Scale Advantage: Vertical Integration & Enterprise Value
Addressing the economics of software and cloud providers, Gavin emphasizes that long-term margins and market dominance will belong to those who are vertically integrated. If a vendor does not own its underlying compute fleet over the multi-decade horizon, it cannot remain the low-cost provider and will struggle to capture the abstraction layer against vertically integrated giants.
Because compute infrastructure represents the fundamental engine of value creation in the AI era, Gavin notes that he increasingly evaluates hyperscale technology companies through a specific valuation metric: Enterprise Value to Net Property, Plant, and Equipment (EV / Net PP&E). In this framework, Net PP&E acts as a direct proxy for an enterprise's physical fleet of compute, serving as an AI-era equivalent to traditional "Price-to-Book" ratios.
- **Cost Floor:** Non-integrated software layers riding on third-party cloud infrastructure will suffer margin compression against full-stack competitors who own their data centers and silicon.
- **EV / Net PP&E Metric:** Financial valuation of hyperscalers is shifting toward measuring how efficiently a company monetizes its balance-sheet investment in physical compute capacity.
---
[01:00:26](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3626s) Jensen Huang, NVIDIA's Supremacy, and the "Michael Jordan" Rule
The conversation pivots to NVIDIA and its CEO, Jensen Huang. Gavin expresses deep admiration for Jensen’s strategy of keeping NVIDIA vertically integrated as a full-stack system provider (offering nine distinct chip families including CPUs, GPUs, switch fabrics, and specialized networking) while remaining horizontally open to ecosystem integration.
Gavin offers a stern warning to semiconductor startups and rival accelerator companies: never attack NVIDIA or Jensen head-on. With every 1% market share in the AI chip space representing roughly $100 billion in enterprise value, niche player startups should focus on securing their tiny slice of the giant pie and finding ways to plug directly into NVIDIA's ecosystem. Gavin draws a vivid analogy to playing against Michael Jordan during his prime with the Chicago Bulls—when a young, overconfident opponent makes the fatal mistake of talking trash during a routine regular-season game, only to unleash Jordan's ruthless, game-destroying competitive drive.
- **NVIDIA's Multi-Tiered Product Suite:** NVIDIA has expanded far beyond generic GPUs into a 9-chip portfolio spanning CPUs, multiple GPU architectures, Ethernet switches, and multi-dimensional interconnect topologies (scale-up, scale-out, scale-across, and scale-in).
- **The $100B Rule of Thumb:** In today's compute market, capturing just a single percentage point of AI accelerator market share translates to roughly $100 billion in value, making head-to-head competition with NVIDIA unnecessary and reckless.
- **Financability of Compute:** NVIDIA data center clusters have achieved a unique status in global finance—they are fully bankable, highly liquid, and standard collateral for massive debt financing.
> "My number one thing is if you're a semiconductor CEO, the only thing you should ever say is: 'Thank you, Jensen. Thank you for creating this opportunity... How can we work with you?'"
> "Have you ever seen game tape of the Chicago Bulls when Jordan... is a little bored, and then somebody who's kind of young decides, 'I'm going to talk to him because we're beating him'? And then he just looks... Don't do that. Just be like, 'Hey Michael, man, I'm so happy to be on the court with you.' That's the move."
[01:03:12](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3792s) The Financial Flywheel of NVIDIA Data Centers
The discussion begins with a breakdown of how massive NVIDIA data center builds are financed. David and Gavin note that for a $50 billion project, an entity might put up a $15 billion equity check while financing the remaining $35 billion through major private equity firms like Blackstone, KKR, and Apollo. Crucially, this is not circular financing; these firms rigorously underwrite each deal. To sweeten the arrangement, NVIDIA offers a residual value guarantee (RVG). As long as this RVG amount remains lower than the gross profit dollars NVIDIA earns from selling chips into that data center, it becomes a net-present-value (NPV) positive, low-risk setup for Jensen Huang. NVIDIA then caps it off with a revenue share, making its data centers the most easily financable in the entire industry. By contrast, building out custom infrastructure like Google's TPUs requires at least double the equity check and comes with significantly higher interest rates.
This cost-of-capital advantage makes plugging into the NVIDIA ecosystem irresistible. Gavin highlights that Jensen is taking this strategy a step further by securing land, power, and shell companies, and then matching them directly with offtake agreements. Without these strategic RVGs and financing setups, the compute landscape would be almost entirely monopolized by Anthropic and OpenAI, who possess the deepest pockets. By providing these structured guarantees, Jensen effectively democratizes compute, enabling a broader array of players to compete. Gavin points out that while Jensen is a ruthless competitor, his business incentives—maintaining a fragmented model landscape where no single AI lab dominates—happen to align perfectly with broader American interests.
Furthermore, Gavin addresses the misconception that open-source AI represents a threat to NVIDIA's core business. In reality, Jensen benefits immensely from open-source models. When open-source models lower application-level gross margins—say from 90% down to 40% per token—the cost per token drops, driving exponential growth in overall token consumption. In a compute-constrained environment, higher token volume directly translates into even greater demand for NVIDIA hardware.
- A standard $50 billion data center build typical of NVIDIA's ecosystem requires roughly $15 billion in equity and $35 billion in debt.
- Major institutional underwriters involved in financing these builds include Blackstone, KKR, Apollo, Goldman Sachs, Morgan Stanley, and JP Morgan.
- Alternative compute projects (such as Google TPU deployments) generally require at least double the equity input and carry higher borrowing rates.
- Lower token margins (e.g., dropping from 90% to 40%) boost overall token consumption, accelerating hardware demand.
[01:06:00](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3960s) Supply Chain Lock-in and the Hard Realities of Hardware
NVIDIA's dominance is reinforced by Jensen Huang's foresight in locking up critical supply chain capacity long before competitors realized its necessity. Gavin estimates that NVIDIA controls 70% to 80% of total fab capacity alongside sweeping allocations across DRAM, NAND, optical lasers, capacitors, and rack assembly components. Fifteen years ago, Jensen was making $2 billion to $3 billion bets every two years; today, he makes multi-hundred-billion-dollar bets that drag the entire global supply chain and financial system along with him.
David shares reflections from his own firm's venture investments in semiconductor startups, echoing Elon Musk's warning that "a lot of people are going to learn a hard lesson in hardware." Designing a chip on paper or running flawless simulations is faster than ever, but real-world execution remains brutally unforgiving. Founders often experience the exhilaration of taping out a chip and receiving the physical silicon back from the lab, only to face the crushing reality of plugging it in and discovering it does not work at all. Fixing a fundamental hardware flaw requires hundreds of millions or billions of dollars in fresh capital and pushes timelines back by years—assuming the company can even survive to raise another round.
The conversation touches on Cerebras as a case study in hardware perseverance. While the company famously weathered major challenges across its early chip iterations, David clarifies that Cerebras's initial struggle was less about dead silicon and more about finding product-market fit for its ultra-large wafer-scale designs, which they eventually achieved. Nevertheless, the sheer difficulty of hardware manufacturing underscores why Elon Musk and X.AI chose to plug heavily into NVIDIA's supply chain rather than attempting to build everything from scratch—a decision Gavin calls a "very high Elo move."
- NVIDIA holds an estimated 70%–80% lock on critical fab capacity and key system components (DRAM, NAND, lasers, capacitors, and rack housing).
- Hardware startups face high binary execution risks where a single failed tape-out can require two years and hundreds of millions to a billion dollars to rectify.
[01:09:17](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=4157s) Custom ASICs, "Jalapeño", and the Federal Reserve of AI
David and Gavin assess the landscape of custom ASICs developed by internal cloud and lab teams attempting to rival GPUs. Gavin pays tribute to a newly surfaced internal ASIC codenamed "Jalapeño", calling it the first truly impressive non-TPU/Trainium ASIC he has seen developed in a short timeframe. While frontier labs designing custom chips possess the distinct advantage of knowing their own model architectures and research roadmaps, NVIDIA maintains a structural advantage by working with every major player across the industry.
Because open-source models—such as China's DeepSeek, Kimi, and Qwen—are evolving in radically different architectural directions, general-purpose GPUs remain essential for training and running varied workloads. A custom ASIC might achieve performance parity with one specific workload or single chip type, but an enterprise needs an integrated system of eight or nine specialized chips and interconnects to match NVIDIA at the system level. Drawing on insights from industry analyst Dylan Patel (SemiAnalysis), Gavin notes that Jensen acts as the "Central Bank of AI" or the "Federal Reserve of AI," controlling the liquidity of compute.
Using a sports analogy, David and Gavin joke that chip designers shouldn't "talk trash about Michael Jordan" or "tug on Superman's cape" by declaring victory over NVIDIA after launching a single successful ASIC. The TPU team at Google and the creators of Jalapeño have made impressive strides, but taking on NVIDIA means competing against an integrated system ecosystem backed by unmatched financial and supply chain momentum.
- "Jalapeño" is cited as a notable custom ASIC achievement outside of Google's TPU and AWS's Trainium efforts.
- Frontier AI models from China (DeepSeek, Kimi, Qwen) highlight how diverse architectural paths necessitate flexible, general-purpose GPUs over narrow ASICs.
- Competing with NVIDIA at the system level requires an entire suite of complementary chips (often 8 to 9 distinct components per rack architecture), not just a single compute die.
[01:12:10](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=4330s) Decoding Customer Preference Through Chip Deal Structures
In a supply-constrained environment where old chips like the NVIDIA H100 retain high pricing and any TSMC allocation sells out instantly, standard sales volumes do not reflect true organic customer preference. To determine what customers actually prefer, Gavin breaks down the hierarchy of deal structures that chip suppliers use to incentivize adoption:
1. **Vendor Investment Deals:** The chip manufacturer directly invests equity into the customer. Amazon and Google utilized this model by investing in Anthropic to drive adoption of Trainium and TPUs. This setup overcomes the cold-start problem of new hardware, and as long as the cash invested is less than the gross profit earned on the hardware sold, the supplier stays in the black.
2. **Residual Value Guarantees (RVGs):** The chip manufacturer pairs third-party debt financing (e.g., Blackstone or Apollo) with residual guarantees. Provided the guarantee liability remains below gross profits, the chip company captures upside via revenue-sharing without taking balance-sheet risk.
3. **Performance-Indexed Warrants:** Hardware suppliers grant warrants to customers tied to a guaranteed fixed price per million tokens. If chip performance outpaces stock growth, the deal is value-accretive; however, unhedged warrant giveaways can easily turn into net-present-value negative disasters if equity dilution outstrips hardware margins.
By analyzing where companies land on this spectrum of deal mechanics, observers can infer which hardware solutions customers choose out of genuine performance demand versus which ones require heavy financial engineering to clear the market. The episode concludes with David thanking Gavin for an insightful discussion on the macroeconomic and structural realities driving the AI revolution.
- High demand keeps legacy GPU hardware (such as the NVIDIA H100) priced high, making raw sales figures a noisy metric for actual customer preference.
- Key deal structures used to overcome hardware adoption cold-start problems include direct customer equity investments, third-party debt paired with RVGs, and token-priced warrant structures.
## Notable quotes
> [01:06](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=66s) Can you tell me one quantitative data point in your business that's getting worse? Just one. — Gavin Baker
> [02:41](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=161s) you can't afford the compute you want for your mission if the equity goes to zero. — Gavin Baker
> [29:30](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=1770s) The consequence of these data center degrowthers may be real compute inequality where big companies and wealthy people can afford compute... and that happened because you wouldn't let us build data centers. — Gavin Baker
> [44:04](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2644s) I think in the future Earth is going to be zoned residential... all heavy industry will take place in outer space. — Gavin Baker
> [47:07](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=2827s) If intelligence is like a super important input into your business, you want to own and control your intelligence, its capabilities, its cost. — Gavin Baker
> [51:07](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3067s) This is kind of like the highest stakes game of corporate chess ever played... You've seen some people come out of the game, others come back in. — Gavin Baker
> [54:01](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3241s) Everybody else in the lab space had this like 'we're creating a digital deity, AGI and ASI'... and the Cursor guys were just like, 'we want to make great product.' — Gavin Baker
> [01:00:26](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3626s) My number one thing is if you're a semiconductor CEO, the only thing you should ever say is: 'Thank you, Jensen. Thank you for creating this opportunity... How can we work with you?' — Gavin Baker
> [01:02:06](https://www.youtube.com/watch?v=FGC4ofTcg2k&t=3726s) Have you ever seen game tape of the Chicago Bulls when Jordan... is a little bored, and then somebody who's kind of young decides, 'I'm going to talk to him because we're beating him'? And then he just looks... Don't do that. Just be like, 'Hey Michael, man, I'm so happy to be on the court with you.' That's the move. — Gavin Baker
[1:06] Gavin Baker: Can you tell me one quantitative data point in your business that's getting worse? Just one.
[2:41] Gavin Baker: you can't afford the compute you want for your mission if the equity goes to zero.
[29:30] Gavin Baker: The consequence of these data center degrowthers may be real compute inequality where big companies and wealthy people can afford compute... and that happened because you wouldn't let us build data centers.
[44:04] Gavin Baker: I think in the future Earth is going to be zoned residential... all heavy industry will take place in outer space.
[47:07] Gavin Baker: If intelligence is like a super important input into your business, you want to own and control your intelligence, its capabilities, its cost.
[51:07] Gavin Baker: This is kind of like the highest stakes game of corporate chess ever played... You've seen some people come out of the game, others come back in.
[54:01] Gavin Baker: Everybody else in the lab space had this like 'we're creating a digital deity, AGI and ASI'... and the Cursor guys were just like, 'we want to make great product.'
[60:26] Gavin Baker: My number one thing is if you're a semiconductor CEO, the only thing you should ever say is: 'Thank you, Jensen. Thank you for creating this opportunity... How can we work with you?'
[62:06] Gavin Baker: Have you ever seen game tape of the Chicago Bulls when Jordan... is a little bored, and then somebody who's kind of young decides, 'I'm going to talk to him because we're beating him'? And then he just looks... Don't do that. Just be like, 'Hey Michael, man, I'm so happy to be on the court with you.' That's the move.
Original transcript
[0:00] when the history of the 21st century is
[0:01] written, you know, there was like the
[0:02] Victorian age. I think this will be like
[0:04] the age of Elon and Jensen because they
[0:06] are fundamentally altering the fabric of
[0:09] human society and civilization.
[0:10] >> What happens if there's like a massive
[0:12] supply shortage?
[0:13] >> Every time you've had a real profound
[0:16] new technology, you get a bubble because
[0:17] the markets get really excited and they
[0:20] get ahead of themselves. Things get
[0:21] overvalued. That overvaluation leads to
[0:24] an overbuild. One of the things that I
[0:26] think has been correct but ineffective
[0:29] is this idea that we need to stay ahead
[0:30] of China.
[0:31] >> You're opposed to data centers. Well,
[0:33] you know what? It's probably the best
[0:35] thing that has ever happened to
[0:37] workingass Americans. We are
[0:38] re-industrializing America and it's
[0:40] awesome.
[0:41] >> Assume that you're right. There's not a
[0:43] physics reason why this can't work.
[0:45] >> An increasing fraction of the world's
[0:48] compute is going to be in orbit. This
[0:50] sounds crazy, but asteroid mining is
[0:52] going to be a very real thing. It has
[0:54] more gold, silver, platinum, every
[0:57] precious metal in it that exists in the
[0:58] earth's crust.
[0:59] >> Every LP conversation that we have
[1:00] starts with like, how's this all going
[1:02] to go wrong?
[1:06] >> Gavin, uh, you've been out here hanging
[1:09] out on the West Coast over the summer
[1:11] and you've been talking about the fact
[1:13] that you're like trying to find someone
[1:14] to make you to give you like a a bearish
[1:17] case, like to make your sentiment more
[1:19] negative. Um, have you found anybody?
[1:22] >> No. And I ask everyone, my standard
[1:24] question is, can you tell me one
[1:27] quantitative data point in your business
[1:30] that's getting worse? Just one. That's
[1:33] my standard question. And it's at least
[1:36] in July and August, I haven't been able
[1:38] to find a single person. Now, if we're
[1:40] being honest, you know, Anthropic is um,
[1:43] you know, in a quiet period, so maybe
[1:44] they've slowed down a little bit, but I
[1:47] do think the rest of the world
[1:50] has accelerated. You know, OpenAI is
[1:53] clearly accelerated. Open source, I
[1:54] think, has accelerated more. And then I
[1:56] do think Grock, particularly after
[1:58] Grockbot, has had a pretty experience,
[2:01] has had a pretty dramatic acceleration.
[2:04] And so AI overall, it accelerated in
[2:06] July. It accelerated in August, and it
[2:10] can't keep accelerating forever, but
[2:13] it's just kind of wild that, you know,
[2:15] public stocks have kind of fallen out of
[2:16] bed over the last, you know, two months.
[2:19] And I mean,
[2:21] you know, it's uh that, you know, they
[2:23] you can you can drown crossing a river
[2:25] that's on average 2 ft deep. And so, you
[2:27] know, there's not a lot of action at the
[2:29] index level,
[2:30] >> right?
[2:30] >> But some of these AI names are in pretty
[2:32] significant draw downs. And they b they
[2:34] bounced a little bit um in August, but
[2:37] still pretty big draw downs and things
[2:39] are broadly accelerating.
[2:41] >> Yeah.
[2:41] >> It's um you know, our friend Eric
[2:43] Fishery did a podcast with Patrick
[2:44] Oanessy and he said maybe everyone wins.
[2:47] >> Yeah. you know, Anthropic wins, OpenAI
[2:50] wins, SpaceX wins, Meta wins. Um, you
[2:54] know, Google wins by selling a lot of
[2:55] TPUs. Um, open source wins, NeoClouds
[2:59] win, inf you know, inference cloud, uh,
[3:02] the inference clouds win on top of the
[3:04] Neoclouds. Um,
[3:06] >> applications win.
[3:07] >> Yeah. Every Yeah. And that kind probably
[3:10] maybe not all applications applications
[3:12] that I think execute well and navigate
[3:14] this
[3:16] but that feels like a very possible
[3:19] scenario to me and there's so much zero
[3:23] someum thinking in the world and by the
[3:25] way on anthropic what is
[3:28] my hypothesis would be if you're
[3:31] anthropic one I think they probably tred
[3:34] up and cleaned up some accounting
[3:36] >> yes definitely
[3:36] >> you would you'd rather do Yes.
[3:39] >> So you rebased and now you're comparable
[3:41] to OpenAI.
[3:42] >> Yeah. In terms of revenue added, like in
[3:44] terms of the definition and now I think
[3:46] kind of revenue added.
[3:47] >> Exactly. So you kind of rebased and then
[3:51] they you know they did their testing the
[3:53] waters. Um
[3:56] and then you know I would hypothesize
[3:58] because they've executed well probably
[4:00] the next disclosure is a reaceleration.
[4:04] And then there's always this kind of
[4:06] funny game between the frontier model
[4:09] companies. They always have more
[4:10] advanced checkpoints. Anthropic is
[4:13] clearly waiting for OpenAI to release
[4:15] Astra.
[4:16] >> Yes.
[4:16] >> And then it's like the next
[4:17] >> the next day.
[4:18] >> Here's Fable 5.1.
[4:20] >> Yes. Exactly.
[4:21] >> Magically and just happened to be
[4:22] available several hours after Astra.
[4:25] >> Yeah.
[4:25] >> So I think they're being thoughtful um
[4:28] and you know heading heading into this
[4:29] IPO and everyone is shooting at them.
[4:32] >> Yes. everybody's shooting at them and
[4:33] they're in a quiet period so they can't
[4:35] really shoot back. Um, and so it's, you
[4:39] know, there's a lot of gamesmanship, but
[4:41] I do think
[4:42] having OpenAI anthropic be public
[4:46] companies is going to be helpful for the
[4:47] market just cuz it's,
[4:49] you know, it's such a uh powerful force
[4:52] and a lot of public investors, you know,
[4:54] you hear, oh, you know, Sarah Frier said
[4:57] this at an all hands meeting and it's on
[4:58] the cover Wall Street Journal. Okay,
[4:59] we're going to put that into our model.
[5:01] >> Yeah. And it's just a lot I think it'll
[5:03] be better for them to be public. I am a
[5:06] little um you know Anthropic is now in
[5:10] their culture interviews saying how
[5:12] would you feel if the equity went to
[5:13] zero?
[5:14] >> Yeah.
[5:14] >> Because we're looking for people who are
[5:16] mission aligned.
[5:17] >> Yeah. Mission not mercenary. Yeah.
[5:18] >> And and that's great. We we want we want
[5:21] missionaries, but we also want people to
[5:23] make money. And at the end of the day,
[5:25] you can't afford the compute you want
[5:27] for your mission if you go if the equity
[5:29] goes to zero. Like I'm no expert, but
[5:32] I'm pretty sure on that
[5:34] >> in that I do think they are
[5:36] >> they're like the accidental enterprise
[5:38] company.
[5:38] >> Oh, for sure. Oh, yeah. They're kind of
[5:40] like the accidental everything.
[5:42] >> Enterprise is just a byproduct of like
[5:44] the the the mission, the objective at
[5:46] the end. Yeah. Whereas I think open eye
[5:47] is a little more commercial and
[5:48] obviously SpaceX a little more
[5:49] commercial.
[5:50] But all of these companies like let's
[5:52] just let's let's just say um let's just
[5:54] say they have 10 gigs of power and
[5:58] they're allocating eight to inference
[6:01] and let's just say they're monetizing
[6:03] that inference at you know whatever um
[6:08] you know 60 60 billion a year. Um, so
[6:11] $480 billion a year in revenue,
[6:14] >> which is like a on a revenue payback
[6:16] basis would be like a one-year payback
[6:18] on a revenue basis, not a gross profit
[6:19] basis.
[6:20] >> Yeah. On a revenue basis. Yeah. Yeah.
[6:22] >> Um, and I've tried to use conservative
[6:23] numbers. You know, people seem to think
[6:25] and open are both monetizing at hundred
[6:27] billion dollars a gigawatt today.
[6:29] >> Yeah.
[6:31] Let's say they have a big research
[6:32] breakthrough and they decide, "Wow, it
[6:35] is to our long-term advantage
[6:38] to go from eight gigs allocated to
[6:40] inference, two gigs allocated to
[6:43] training to 8 gigs on training and then
[6:46] your revenue just went from 480 to 120."
[6:50] And I think they your annualized revenue
[6:53] and I actually think they would do that
[6:55] >> make that decision.
[6:55] >> Yeah. And this is just something that
[6:58] like public markets are going to really
[7:01] have to get used to.
[7:03] >> Yeah.
[7:04] >> It as you say, open AI may be a
[7:06] different animal. And I do think like
[7:09] the realities, you know, everybody
[7:11] everybody has these ideals about how
[7:13] they're going to manage their business,
[7:14] then they go public and the stock is
[7:16] volatile and it really impacts, you
[7:18] know, employee morale, recruiting,
[7:20] retention. So, I'd be surprised if they
[7:23] did such a dramatic cut, but a lot of
[7:26] the revenue is kind of under their
[7:28] control based on what checkpoint they
[7:30] release.
[7:31] >> Yeah.
[7:32] >> Where they price um along this, you
[7:35] know, kind of paro curve and then how
[7:37] much they allocate between training and
[7:38] inference. So, it's just
[7:42] it's going to be, you know, Meta and
[7:44] Google and these kind of internet
[7:46] companies. It was just it was pretty
[7:49] smooth fundamentally even if the stocks
[7:51] were volatile.
[7:51] >> Well, there was no like massive
[7:53] trade-off they had to make in terms of
[7:54] the cost or infrastructure to serve
[7:57] revenue side. Like they were totally
[7:59] separate
[7:59] >> 100%.
[8:00] >> Yeah. It's it's fascinating. Um so, you
[8:03] know, if you go back to Eric's point of
[8:05] like it's all going to work like I
[8:07] actually think that's a great point.
[8:08] Like I I I describe it differently. I've
[8:10] had this conversation with LPs a lot cuz
[8:11] every LP conversation that we have it's
[8:13] probably the same for you starts with
[8:14] like how's this all going to go wrong
[8:17] >> and it's like what's what's going to
[8:18] crash and I'm like this is the this is
[8:20] the and like oh are the are the large
[8:22] models screwed or the labs screwed
[8:24] because of open source and I'm like this
[8:26] is this is this is all wrong like this
[8:27] is not an or thing it's an and thing
[8:29] right like this is an and thing um
[8:32] Frontier is going to work really well
[8:33] like N minus one models are going to
[8:34] work really well open source is going to
[8:36] work really well um there's going to be
[8:38] a bunch of application companies that
[8:40] work really well. Like the clouds are
[8:42] probably going to be fine. They're
[8:43] probably going to work really well.
[8:45] >> Like the five lab companies are probably
[8:46] going to do really well.
[8:48] >> Yeah. And Nvidia is
[8:51] at the center of all of it.
[8:53] >> Yes. Yes. They're probably going to do
[8:54] pretty well.
[8:55] >> Yeah.
[8:58] The last um 26 years have taught me not
[9:00] to bet against Jensen.
[9:02] >> Yeah. He's he's he's in a pretty good
[9:03] position here. Um I want to come back to
[9:05] that. The the point that you made about
[9:07] training verse inference is an
[9:08] interesting one. It seems to me like the
[9:11] labs will decide to take all incremental
[9:15] profits and probably much more than
[9:18] their profits and invest them in
[9:20] training for a long period of time.
[9:22] Would you think that's fair? Like this
[9:24] is very different than like the clouds,
[9:26] you know, cuz like the the cloud like
[9:28] the internet companies and the clouds,
[9:29] they just end up being supply demand
[9:32] driven and they generate tons of profit
[9:34] and they can still grow a certain amount
[9:37] like but they don't have some maybe with
[9:39] the exception of Meta like some big
[9:41] long-term bet that's like a multi-year
[9:43] payoff.
[9:44] >> Yeah, I think it's important to kind of
[9:45] be precise. They I for sure I don't
[9:48] think they will generate free cash flow
[9:50] anytime soon. I think they're going to
[9:52] generate a lot of operating cash flow
[9:54] and then they'll use that to buy a lot
[9:56] of, you know, GPUs,
[9:59] um, XPUs, whatever, whatever we're going
[10:01] to call them. Um, or maybe they
[10:04] subsidize heavily. Like we do know that
[10:06] that's happening at the labs.
[10:08] >> Subsidize what heavily
[10:10] >> their first party products. So token
[10:12] consumption of their first party
[10:13] products. So like they're doing all this
[10:15] research and they're spending a lot on
[10:17] data on compute
[10:18] >> and the first products that are like a
[10:20] heavy subsidy
[10:21] >> products today, right?
[10:22] >> Yeah. So it's 8 gigs of inference and
[10:24] two gigs is for internal research and
[10:26] then you know two gigs is actually
[10:29] training.
[10:29] >> Yeah. Exactly.
[10:30] >> Um and you know including probably the
[10:32] inference that goes into post- training.
[10:35] >> Yeah. I don't I I think given the belief
[10:38] systems that they all seem to have about
[10:41] scaling laws
[10:43] which continue to hold I don't think any
[10:46] of them are going to be that focused on
[10:48] generating free cash flow and you've
[10:50] seen right we saw Satcha blink.
[10:53] >> Yes.
[10:53] >> And Satcha really regrets that I think.
[10:56] >> Yeah. Yeah. um you know he kind of
[10:57] blinked I think it was last year
[11:01] you know he gave that great interview
[11:03] for Davos and they asked him about all
[11:04] the capex and he said I know I'm good
[11:06] for my 80 billion
[11:08] >> right
[11:08] >> and and I think they blinked a little
[11:11] they slowed down they regret that and
[11:14] then Daario famously he went on a
[11:16] podcast and he made and he said listen
[11:18] some people are being super
[11:19] irresponsible with their spending and
[11:21] it's a hard decision because if you
[11:24] don't spend enough you could lose a lot
[11:26] of shares But if you spend too much, you
[11:27] could go bankrupt. And like those are
[11:29] both bad things, but bankruptcy is worse
[11:31] than losing shares. So I'd rather be
[11:33] conservative. And he was conservative.
[11:35] And OpenAI was aggressive. And now
[11:37] OpenAI is back in the game.
[11:39] >> And SpaceX was aggressive.
[11:40] >> And SpaceX was aggressive.
[11:42] >> And so, you know, like there are clear
[11:44] high ROIs on those independent of supply
[11:47] demand mismatches that are happening.
[11:48] Like clearly that seems to be the right
[11:50] decision short-term and long-term.
[11:52] >> Yeah, absolutely. I mean, we we
[11:53] calculate, you know, Nebius um and
[11:56] Corweave both gave some interesting
[11:57] disclosures, but you can kind of get to
[12:00] a 9 to 10 month payback for Nebius
[12:04] because you know, okay, you bring on a
[12:06] gig, it costs 50 billion. You get you
[12:09] can get an upfront payment for 50 to 60%
[12:11] of that for customers. Yeah.
[12:13] >> So now, you know, you're talking about
[12:15] 25 or 30 billion and then you can
[12:17] monetize it if you put it into the spot
[12:20] market.
[12:20] >> The spot. Yeah. at a spot spot paybacks
[12:23] are probably much faster than nine or 10
[12:25] bucks.
[12:26] >> Yeah, you got to assume like a smoothed
[12:27] out level like two two bucks, three
[12:29] bucks even with that. It's very very Now
[12:31] you can get like five bucks or eight
[12:32] bucks and Yeah.
[12:33] >> And then SpaceX cuz they build these
[12:35] really big clusters and and I think at a
[12:37] really important point is they bring
[12:39] them on fast.
[12:40] >> Yes.
[12:41] >> They have an even faster payback and
[12:44] they can monetize at you know higher. I
[12:46] I have tried to shift um you know to
[12:49] think of pricing and you know per
[12:51] megawatt rather than per GPU because it
[12:53] seems like it's where where the world
[12:54] >> world is but like SpaceX the payback
[12:57] feels well inside of that.
[12:58] >> Yes.
[12:59] >> And I just in my career as an investor
[13:04] there haven't been that many
[13:06] opportunities where you have companies
[13:08] that could deploy tens hundreds of
[13:11] billions of dollars and get sub one-year
[13:14] paybacks.
[13:14] >> Yes. And it's kind of crazy. And then
[13:17] also like we should also talk if
[13:19] particularly if you're buying Nvidia
[13:21] GPUs to a lesser extent TPUs, you can
[13:23] finance these.
[13:24] >> Yes.
[13:25] >> And there's a very sophisticated, you
[13:27] know,
[13:27] >> Yeah. very low cost of capital to
[13:29] finance them today.
[13:30] >> Yeah. And everybody's, you know, worked
[13:31] up about, you know, circularity and it's
[13:33] like, well, I don't know. Um, I know a
[13:37] lot of smart people who work at
[13:39] Blackstone and KKR and Apollo
[13:43] and they're the ones that are financing
[13:45] >> the ones who are financing it at a
[13:46] relatively low cost
[13:47] >> at a relatively low cost. And I think
[13:50] one reason that's happening is useful
[13:52] lives just keep getting extended and has
[13:55] these models get better and better and
[13:57] better and the ROI on token spend goes
[13:59] up, you know, the monetiz monetization
[14:02] rate per gigawatt goes up. So, I mean,
[14:06] the true equity payback like might be
[14:11] way inside of a year.
[14:12] >> Yeah. Exactly. Exactly. Yeah. And look,
[14:14] there's a case you could make that the
[14:16] prices actually of all the stuff go up,
[14:19] which could make the the supply side
[14:21] economics even more compelling, right?
[14:23] Like, you know, so on the supply side,
[14:27] like that's the dynamic today. Like, it
[14:28] just is what it is. Like, there's a ton
[14:30] of data points out there that paybacks
[14:31] are within a year.
[14:32] >> Yep. Um I think it's actually
[14:34] interesting to think about the demand
[14:35] side too because the knock would be well
[14:38] in all these cycles you get some
[14:40] overbuild and then that you know
[14:42] destroys the economics of the supply
[14:43] side. The demand side today like what
[14:46] are we monet like the monetization of
[14:49] these companies which are doing call it
[14:52] 80 billion of revenue or something in
[14:53] that direction um is on the back of what
[14:57] like 30 million actual heavy paying
[15:00] users like re getting real value. I'm
[15:02] talking about like developers like
[15:04] >> I might take the under on 30 million
[15:06] >> so call it yeah actually what we see
[15:07] inside our companies is you know
[15:09] obviously there's a power law in which
[15:11] companies are spending a lot on tokens
[15:13] like old banks are probably spending 1%
[15:16] very techforward companies are spending
[15:18] high single digits but if you actually
[15:21] look at the sort of the the power law of
[15:24] what's happening of the actual engineers
[15:26] in those companies the highest spending
[15:28] engineers are spending 10 or sometimes
[15:31] is 100x more than the median engineer.
[15:33] And so, yeah, your 30 million is
[15:35] probably way overstated. It might be sub
[15:37] 10. And so, there's this question of
[15:39] like where are we at in diffusion?
[15:41] There's one and a half billion knowledge
[15:43] workers. Like, it feels like we're
[15:44] nowhere on the demand side and we're
[15:46] massively supply constrained.
[15:48] >> And what are I'm just curious across the
[15:50] A6Z portfolio if what are your best
[15:54] companies spending on tokens per month
[15:58] relative to human compensation? What
[16:01] rough range?
[16:02] >> Oh, high single digits, some at 10%,
[16:04] like some of the very AI native ones
[16:06] like 10% plus. And so, you know, and and
[16:09] then old economy companies are spending
[16:11] the ones that are probably doing a good
[16:12] job like 1%. So, it feels to me like
[16:16] when I look at the supply demand
[16:17] characteristics, it's like supply stuff
[16:21] people say, is that sustainable? Well,
[16:23] like when you pair it with the demand
[16:24] stuff, I it feels it feels specific.
[16:27] Like there could be things that
[16:28] disappoint us in terms of like diffusion
[16:30] into the real economy,
[16:32] but it feels like over a 10-year
[16:33] stretch, like we're nowhere.
[16:35] >> Yeah. Absolutely nowhere. And I just
[16:38] >> my So at a trade is our internal token
[16:41] consumption has gone up 100x from the
[16:44] month of March. March through August.
[16:46] 100x our token spend. And we just got
[16:51] access to uh Grockbot Enterprise and
[16:55] with two people using it like it looks
[16:58] like it token spend might 10 or 20x in a
[17:03] month.
[17:03] >> Yes.
[17:03] >> From August.
[17:04] >> Yes.
[17:04] >> Like like I
[17:06] >> But but it's actually extremely
[17:07] valuable. Like we have some heavy
[17:09] Grockbot users here and like it is very
[17:13] productive use. Like this is not like
[17:14] wasteful tokens, but
[17:16] >> yeah, I was and and listen like I I try
[17:19] super hard. I you know I always when I
[17:22] use AI, I just remember when my parents
[17:24] like I was trying to get them to shift
[17:26] to an iPhone and an iPad and like you
[17:29] know get them used to it and like you
[17:32] know and they did a good job. I give
[17:34] them loads of credit and but you know
[17:36] I'm 50 years old you know like how old
[17:39] are you David?
[17:40] >> 42. 42 and you see these like 23-y old
[17:42] kids and just the way they use AI,
[17:44] they're just fluent and native in it. I
[17:47] just feel like maybe in a way that no
[17:48] matter how hard I try, I will never be
[17:50] and I'm trying really hard. But, you
[17:52] know, like we got cloud code, I try I
[17:55] you know, I built some stuff, did some
[17:57] cool stuff and in like I don't know
[18:01] three minutes of type creating Grock
[18:04] bots, I had much better versions of
[18:06] everything I created, you know. You
[18:08] know, so I went on this Patrick Oannessy
[18:10] pod podcast like 5 months ago and I
[18:13] said, you know, like I love having a
[18:15] podcast summarizer. Everybody's like,
[18:16] "How'd you do it?" I was like, "Well,
[18:17] just use AI and do it."
[18:19] >> Yes. Pretty simple.
[18:20] >> It takes 10 seconds and Grockbot.
[18:23] >> Yes.
[18:23] >> It's amazing and it's so good.
[18:25] >> Yeah.
[18:25] >> And then, you know, a Substack
[18:27] summarizer, an X summarizer, um an Xs
[18:30] sentiment tracker for topics and stocks.
[18:34] >> Yeah. And like that all of those would
[18:36] have taken me I don't know hours working
[18:40] with cloud code and they each took 7 to
[18:45] 12 seconds with Grockbot and it's
[18:48] better.
[18:48] >> Yeah.
[18:48] >> So to to me Grockbot does feel like
[18:51] another um at least for me like kind of
[18:54] chat GPT moment because Claude code
[18:57] >> like I can see in the data was it was
[18:59] powerful. I did some really cool stuff
[19:01] with it that was like empowering and
[19:03] this is neat.
[19:04] >> Um
[19:06] >> you know like family calendar apps,
[19:09] things like that.
[19:09] >> Yeah.
[19:10] >> Um but this is just 10 seconds and it's
[19:14] way better than what I was able to do.
[19:16] >> Yeah. Yeah. Yeah, the cloud code thing
[19:17] like was obviously the shift in coding
[19:20] and you know our our most sophisticated
[19:24] engineers you know were doing whatever
[19:25] 20% of their code you know with with
[19:28] with AI to like you know whatever 90
[19:30] plus% and so now I think everything you
[19:33] described in what you built with cloud
[19:35] code or codec
[19:36] >> is still kind of reactive
[19:39] >> in a way right like it's it's still you
[19:40] know it's like summarizers yeah
[19:42] preparation it's all like knowledge
[19:44] enhancing which is part of your job, but
[19:46] it's not actually doing the work for
[19:48] you.
[19:48] >> Yeah. And now you have a Groc bot that
[19:51] says, "What are the recommended
[19:52] actions?" Yes. Exactly.
[19:54] >> Based on everything the other bots have
[19:56] learned today.
[19:56] >> Yeah.
[19:57] >> What recommendations do you have for me
[19:59] today? And that for sure is like
[20:01] >> and it it was so easy to build. Um I now
[20:04] have it. So I'm I'm like horse racing
[20:07] all these which is like I have uh uh
[20:09] Crockbot doing it, Codeex doing it. All
[20:12] the like action taking for just I want
[20:13] to know
[20:14] >> make me better at my job. Look at
[20:17] everything I do. Give me give me
[20:18] recommended automations you can do. I
[20:20] have Town doing it as well which is one
[20:22] of our companies very good at it.
[20:24] >> Um but and we're like kind of on the
[20:26] bleeding edge of trying to do this
[20:27] stuff.
[20:28] >> Just wait till everyone does this stuff
[20:30] >> and then and then when we actually click
[20:32] like yes go just automate this.
[20:33] >> Yeah. It feels like that's sort of
[20:35] endless token
[20:36] >> and but I do we should acknowledge like
[20:38] the the history of financial markets,
[20:42] >> you know, dating kind of back to like
[20:43] the South Sea bubble is whenever you get
[20:46] this transformational new technology.
[20:48] Um, I actually went on a podcast, I said
[20:50] I thought the South Sea bubble was
[20:52] connected to like the invention of
[20:54] longitude and the ability to sell. Turns
[20:55] out it was not. It was just it was kind
[20:58] of like a more of a tulip episode. But
[21:01] like every time you've had a real, you
[21:04] know, profound new technology, you know,
[21:05] whether it's the automobile, the TV, the
[21:07] radio, internet, the PC, um, railroads,
[21:12] >> steel mills, you get a bubble because
[21:14] the markets get really excited
[21:16] >> and they get ahead of themselves. Things
[21:19] get overvalued. That overval
[21:21] overvaluation
[21:23] leads to an overbuild. And then
[21:25] particularly if you're funding it with
[21:27] debt um and and even today a majority of
[21:30] this is still being funded out of
[21:31] operating cash flow which I think is
[21:32] really helpful. Um you know debt funded
[21:36] built buildouts they demand immediate
[21:38] ROI not an ROI in three years.
[21:40] >> Yeah. You can't be off in the time. You
[21:41] can't be off off on the time, but I'm
[21:43] just more, you know, like I um, you
[21:46] know, I talked to Jazz about how Watson
[21:48] wafers Jazz I guess and Patrick Watson
[21:50] wafers are these fundamental constraints
[21:52] and just that the buildout is so big and
[21:55] we're so early that we are
[21:58] it's like impacting the raw productive
[22:01] capacity of so many industries. you
[22:03] know, now you know, everybody in
[22:05] everybody in copper, there's like an AI
[22:07] thesis and like we're going to have to
[22:10] like think about it to like fill the,
[22:13] >> you know, if if
[22:15] >> 10% of what we just talked about comes
[22:17] true, you know, we're in this acute
[22:18] shortage with, I don't know, several
[22:20] million people are driving a crazy
[22:23] global compute shortage. What happens
[22:25] when that's 500 million? And you know,
[22:28] how many copper mines do we need to
[22:30] build to like support this? Yeah,
[22:33] >> it's kind of a wild thought. And so like
[22:35] these fundamental constraints, I think,
[22:37] are slowing us down.
[22:38] >> And I
[22:39] >> and I think that's good. I actually
[22:41] think that's good for society. And I
[22:42] would now say rates and regulation, you
[22:45] know, real rates are going up. Yes.
[22:46] >> And it just is what it is, which makes
[22:48] sense because we're like investing a
[22:49] lot. So it makes sense um that real
[22:52] rates are going up. And then regulation,
[22:54] man. It's it is like I'm kind of shocked
[22:58] at what's happening in America.
[23:00] >> We're in a really bad place.
[23:01] >> Yeah. And just, you know, I had this
[23:04] exchange with with um Schulto for from
[23:07] Anthropic and and and Daario on X last
[23:11] weekend. You know, Daario said, "Hey, I
[23:13] don't think I've been negative. You
[23:15] know, I've written I've written two
[23:16] essays. One was positive, one was
[23:18] negative." So being 50% negative and
[23:21] particularly when it's like a terrifying
[23:25] negative
[23:25] >> like an existential
[23:27] >> an existential negative everybody might
[23:28] be out of out of a job like that Eleazar
[23:31] Yukowski guy says if we build it
[23:33] everyone will die and it's like how
[23:35] about if we build it like we're going to
[23:37] cure cancer we're all going to live
[23:39] forever. I thought one of the best
[23:40] things Dario said was like what we need
[23:42] to do is stop talking about curing
[23:43] cancer and actually cure cancer.
[23:45] >> Actually cure cancer and actually make
[23:46] breakthroughs like
[23:48] >> but just somebody like that my favorite
[23:50] line in the Bible is the truth shall set
[23:52] you free.
[23:52] >> Yes. But the only person who can the
[23:56] only group that can tell the AI
[23:58] industry's truth is the AI industry.
[24:00] They need to just start telling the
[24:02] truth. Hey when we Okay, you're opposed
[24:06] to data centers. Well, you know what?
[24:07] It's probably the best thing that has
[24:09] ever happened to workingclass Americans.
[24:12] >> Yeah, exactly.
[24:13] >> You know, it's like going to college
[24:14] might be significantly NPV negative now
[24:18] because you can go learn how to be an
[24:20] electrician, a plumber, an HVAC tech,
[24:23] and make ungodly amounts of money. Yeah.
[24:26] >> So, this has been amazing for
[24:27] working-class Americans. We now have a
[24:29] lot of data that particularly with
[24:31] behind the meter power generation, when
[24:32] a data center goes in,
[24:35] it transforms a town. like tax revenue,
[24:38] it doesn't double. It like 10xes and it
[24:42] is re revitalizing all of these like
[24:45] dying small towns all over America. And
[24:48] listen, we're getting we're getting much
[24:50] better at addressing the environ
[24:52] environmental stuff. Generally, they use
[24:54] natural gas, which is a pretty clean
[24:56] fuel.
[24:57] >> The the water the water consumption
[24:58] thing is totally debunked. It's totally
[25:00] debunked. Yeah.
[25:01] >> It's nothing. It's nothing. So these are
[25:03] like really really really good and
[25:04] they're having a really positive impact
[25:06] on the world. That's without even
[25:08] considering things like curing cancer,
[25:10] but somebody needs to tell that story.
[25:12] >> It's now and and I think the problem
[25:14] with it now is like the burden of proof
[25:16] is on not curing cancer, but actually
[25:18] delivering some real tangible everyday
[25:21] American benefits beyond using chat, you
[25:24] know, or Grock to like answer your
[25:25] questions or substitute
[25:27] >> for a search engine, right? It it does
[25:29] feel like we're pretty close to that. Um
[25:33] yeah, it does. And and by the way, like
[25:36] one of the things that I think has been
[25:38] correct but ineffective is this idea
[25:41] that we need to stay ahead of China.
[25:43] >> Like it's like it is true. Like I'm very
[25:45] much like I'm a patriot. Like I believe
[25:46] that. But it's way too abstract. Yeah.
[25:48] The abstract for the average American.
[25:50] Like
[25:50] >> nobody's worried about China invading
[25:52] America.
[25:52] >> Yeah. Exactly. Like they ocean is really
[25:55] big.
[25:55] >> Yeah. like affordability and like how is
[25:57] this going to change my life for the
[25:58] better or worse, right? And so
[26:00] >> I think there's a pretty immediate
[26:01] impact you could feel like I my favorite
[26:03] is, you know, Lowden County, Virginia,
[26:05] which is like the highest uh highest per
[26:08] capita income uh zip code in the US or
[26:12] county in the US
[26:14] >> and it has the highest density of data
[26:16] centers.
[26:17] >> Yeah.
[26:17] >> And they and they make a tremendous
[26:18] amount of tax revenue from data centers.
[26:20] Like we should we should do this
[26:21] everywhere.
[26:22] >> Yeah. It was actually very funny. a
[26:23] someone very opposed to data centers
[26:25] said, "Oh, you're for data centers. I'd
[26:26] like to see them put in the highest
[26:28] income zip code and the highest, you
[26:30] know, income county." And they're like,
[26:31] "Actually, the highest income zip code
[26:34] in America and the highest income county
[26:36] has the highest per capita concentration
[26:38] of data centers." So, we've done that
[26:40] >> and it worked out really well.
[26:42] >> Yeah. But, you know, hey, don't bother
[26:43] me with the details. I'm on to my next
[26:45] talking point.
[26:45] >> That's good. That's good.
[26:46] >> And all those talking points, it's
[26:48] tragic. Like there is an organized CCP
[26:50] funded campaign. I think against data
[26:53] centers here in America like I think a
[26:55] lot of it gets laundered through Tik Tok
[26:58] and it's just tragic because the other
[27:00] thing that's happening is this is
[27:01] re-industrializing America. The
[27:03] combination of having the straight of
[27:05] foremost closed which is amazing for
[27:06] America. You know natural gas here is
[27:08] two or three bucks.
[27:09] >> It's now 25 bucks
[27:12] >> in Europe and Asia or 20 bucks or
[27:13] whatever it is. And natural gas is an
[27:16] you know important input to the cost of
[27:18] electricity which is an important input
[27:20] to almost all manufacturing processes.
[27:23] And so we have a huge cost advantage for
[27:27] that basic input now.
[27:29] >> And you have that happening and you have
[27:31] this kind of data center boom happening.
[27:33] We are re-industrializing America. And
[27:35] it's awesome. This is what everyone in
[27:37] both parties has wanted for a long time.
[27:40] >> Yeah. Exactly.
[27:41] >> Like bring industry back. small towns
[27:43] that were left behind by the steel mills
[27:45] closing. Well, data centers are bringing
[27:47] them back.
[27:47] >> Yeah. But somebody has to tell that
[27:49] truth. I mean, I try to do it on every
[27:50] podcast, but like I'm just a dude.
[27:53] >> Yeah. And like your audience is the tech
[27:55] audience that that already believes
[27:56] you're you're preaching the choir, if
[27:58] you will. Um, but yeah, the story the
[28:00] story I met Meta is probably doing the
[28:01] best job of telling that story, I would
[28:03] think.
[28:04] >> Yeah, it seems.
[28:05] >> You know, and I think one reason it's
[28:07] really wired into Meta's DNA. So, one of
[28:09] the first things they started doing as a
[28:11] public company I don't remember if it
[28:13] was on their first attorney's call, but
[28:15] Cheryl would run through Cheryl Samberg
[28:18] would run through 10 or 15 very specific
[28:21] small businesses that had started using
[28:25] Meta's advertising products and the
[28:28] impact it had on that business.
[28:29] >> Yeah. you know, this cake bakery in De
[28:33] Moines started, you know, worked with
[28:35] Meta and, you know, it was it was it was
[28:38] two women who were single mothers
[28:40] working by themselves and now they have
[28:43] 15 locations. They employ 50 people.
[28:47] >> Yeah.
[28:47] >> And this has been amazing for De Moine
[28:49] and it's been transformative for them.
[28:52] >> Yeah.
[28:52] >> And they would just run through that
[28:54] every time. And and I do think the
[28:56] entire AI industry um like I'd love to
[28:59] see, you know, everybody SpaceX,
[29:03] Enthropic, OpenAI, Google, Meta say,
[29:06] "Hey,
[29:07] >> here are real businesses and real
[29:09] Americans and like either name the
[29:11] business or get permission to if you can
[29:13] name the American or anonymize it." This
[29:15] is a really positive thing it did it it
[29:17] had on their life.
[29:18] >> Already very tangible. Yeah.
[29:19] >> Yeah. Same. Nvidia, AMD, Broadcom, all
[29:22] of them.
[29:23] >> Yeah. just run through specifics because
[29:25] the truth will set you free but only if
[29:27] you tell it.
[29:27] >> Yeah. Exactly. Exactly. Yeah. So it
[29:30] seems more likely than given that fact
[29:32] pattern if you go back to just the sort
[29:34] of macro situation that we're in that we
[29:37] we underbuild on the supply side.
[29:39] >> Oh yeah. For for like through 28. And
[29:42] and by the way like there's no capacity
[29:43] available with all the forecast builds
[29:46] that will happen through 28 which are
[29:48] probably now going to be delayed given
[29:49] the political dynamics they have. So,
[29:52] um,
[29:52] >> everybody's worried about over supply.
[29:54] I'm like more worried about
[29:55] >> massive massively under supply. Yeah,
[29:57] exactly. Which, which Okay. So, then if
[29:59] that's the scenario,
[30:01] like you could see a scenario where you
[30:02] see, you know, big price increases
[30:05] actually to access the intelligence.
[30:06] Yeah.
[30:06] >> Which is the opposite direction of where
[30:08] everybody thinks this is going to go.
[30:09] >> Yeah. Well, Dorcash had a wild point. I
[30:11] forget what it was, but he was positing
[30:14] >> um I forget the
[30:15] >> like the cost of a token could go up 10x
[30:16] or something like that. Yes. Yeah,
[30:18] >> which is crazy, but like we do live in a
[30:20] supply demand world.
[30:21] >> Like it's conceivable if the demand goes
[30:24] massively. And by the way, the whole
[30:26] premise of this that's happening so far
[30:28] is that there's a massive amount of
[30:30] consumer or user surplus being
[30:32] generated, right? So like why do people
[30:34] select the frontier tokens when they
[30:36] could use the cheaper tokens to do most
[30:38] tasks? There's many reasons why, but
[30:41] like the biggest one is because there's
[30:42] a tremendous amount of surplus even if
[30:44] you're using the frontier tokens, right?
[30:45] Absolutely. And so yeah, what happens if
[30:48] there's like a massive supply shortage?
[30:50] Well, I think that would be the, you
[30:52] know, kind of funny the consequence of
[30:55] like the these like data center
[30:58] degrowthers
[31:00] um
[31:02] may be like real compute inequality
[31:06] where big companies and wealthy people
[31:09] can afford compute and then you know two
[31:12] years from now they'll be on about that
[31:13] and it's like well that happened because
[31:15] of you. Yeah.
[31:16] >> You know that happened because you
[31:17] wouldn't let us build data centers.
[31:19] >> Yeah. And by the way, we've we've seen
[31:20] this, right? Like the path to a lowcost
[31:23] product delivered to consumers in a mass
[31:26] market is advertising. It takes a long
[31:29] time to build an advertising business.
[31:30] >> Yeah.
[31:31] >> Um as we've seen with all the, you know,
[31:33] consumer internet businesses that we've
[31:34] invested in over the years.
[31:35] >> Um and so there may be a disconnect in
[31:37] the period where you can't actually
[31:38] offer that.
[31:39] >> Yeah.
[31:39] >> And that would be a terrible outcome.
[31:41] >> That'd be a terrible outcome for the
[31:42] world. Nobody wants that. So we need to
[31:43] build a lot of data centers.
[31:44] >> Yeah. Exactly. Exactly. Yeah. like a
[31:46] compute in inequality like future that's
[31:50] that's not a good that's not a good
[31:51] future for anyone which is another
[31:52] reason open source is so important and
[31:55] just one of the things um you know I you
[31:58] know I had uh Grock make me make me like
[32:01] a meme of that like three-headed dragon
[32:04] and one of the heads is like kind of
[32:05] confused about like all of the really
[32:07] like stupid
[32:09] >> bearish AI narratives but people have
[32:11] this idea that open- source tokens are
[32:14] free they're
[32:16] And it's like it takes the exact same
[32:18] amount of compute.
[32:20] >> Yeah.
[32:20] >> All else equal to make an open source
[32:23] token as a you know Frontier token for a
[32:26] comparably sized model. Now there's a
[32:29] lot of nuances there but that's broadly
[32:31] true.
[32:32] >> It's just a question of what are the
[32:33] margins that are charged on top of that.
[32:36] And even then, the Kimmy license,
[32:39] something that I don't think a lot of
[32:40] people appreciate is the Kimmy license
[32:43] stipulates a 30% um share of any
[32:46] revenue.
[32:46] >> Yeah. Yeah. Yeah.
[32:47] >> So like Kimmy has taken a 30% cut of all
[32:50] the revenue generated on its and this is
[32:53] because it's open weights, not open
[32:54] source.
[32:54] >> Yeah. Exactly. Yeah.
[32:55] >> Yeah. But it's also extremely token
[32:57] hungry too, right? So it's more it's
[32:59] it's far even we're talking on a token
[33:01] basis, but on a task basis, it's far
[33:03] more inefficient. So it's very costly.
[33:05] >> Yeah. And I just always like Jensen,
[33:08] he's a great patriot, great American.
[33:10] Like we're so lucky to have we're lucky
[33:12] to have him and Elon like and I think
[33:14] like you know kind of when the when the
[33:16] history of the 21st centurion 21st
[33:19] century is written you know there was
[33:20] like the Victorian age I think this will
[33:22] be like the age of Elon and Jensen.
[33:24] >> Yeah. because they have they they are
[33:26] fundamentally altering kind of like the
[33:28] fabric of human society and civilization
[33:31] with AI SpaceX making humanity
[33:33] multilanetary Starlink you know bringing
[33:36] lowcost internet access to the poorest
[33:38] communities in the world which is
[33:40] amazing um which is you know something
[33:43] that people don't talk about but it's
[33:44] like an amazing you know you talked
[33:46] about consumer surplus that is an
[33:48] amazing surplus
[33:49] >> there was never there there was never
[33:51] going to be an economic case to build
[33:53] internet access in those places because
[33:55] of the cost
[33:56] >> and the willingness to pay and now you
[33:58] could
[33:59] >> without and any incremental internet
[34:02] capacity like is not going to be built
[34:04] in a traditional sense on Earth. It's
[34:06] going to come from space and so like
[34:07] that is a huge that is a huge unlock. I
[34:08] agree.
[34:09] >> It's a good thing but like we're you
[34:11] know we're like you know we should we
[34:13] should all be grateful for them because
[34:15] I do think that you know they're you the
[34:17] they're making the future as exciting
[34:19] and inspiring as possible. say we are in
[34:22] this supply crunch. Um it's so funny
[34:25] when whenever I talk about SpaceX and
[34:27] it's it's obviously near and dear to
[34:28] both our hearts. Um you know I I say
[34:31] like first of all the orbital data
[34:33] center stuff it's not like big buildings
[34:36] in space like it's helpful to actually
[34:37] think of it's like the size of an
[34:38] airplane.
[34:40] >> People are picturing like the Death Star
[34:43] like or the Pentagon floating around in
[34:45] space. That's not what it is at all.
[34:47] >> Yeah. It's a It's you know whatever the
[34:49] size of an airplane, right? Rack of 72
[34:51] whatever chips.
[34:52] >> Yeah. It's it's like five of us standing
[34:55] together is kind of roughly
[34:57] >> is like the wings
[34:59] solar wings.
[35:00] >> Yeah.
[35:00] >> And then you keep it in a suns
[35:02] synchronous orbit.
[35:03] >> So you have the radiator always in the
[35:06] shadow of the rack.
[35:08] >> That's how you cool it.
[35:09] >> And it's I like I can't it's very hard
[35:13] for me to engage. you know, there's all
[35:14] these people on X and they're like, I am
[35:17] a physics PhD and I this is impossible.
[35:22] Um, and actually there's there's there's
[35:24] a friend who's another investor who
[35:25] actually is a physics PhD who had many
[35:28] um arguments with him and he's like, I
[35:30] am a PhD and this is impossible. And
[35:32] then he goes to the SpaceX day and you
[35:35] know he talks to the SpaceX engineers.
[35:36] He's like, well, I was wrong. And so
[35:38] like if let's say you're an astrophysics
[35:42] PhD, you are brilliant. You're hanging
[35:45] 100 IQ points on me. Have you thought
[35:48] about this for an hour? Have you thought
[35:50] about it for 10 hours? Have you thought
[35:52] about for five hours? Cuz you have
[35:54] 10,000 of the world's smartest engineers
[35:56] at SpaceX who've thought about this each
[35:58] for hundreds if not thousands of hours.
[36:01] And the sum of that working with like
[36:04] very sophisticated, you know,
[36:06] engineering tools is it's a solved
[36:09] problem. And in their minds, it's
[36:10] dramatically simpler and easier.
[36:12] >> Yeah.
[36:13] >> Than a Starlink satellite cuz a Starlink
[36:15] has to have the phased arrays and move
[36:16] around.
[36:18] >> I think it's like So, okay. So, assume
[36:20] that you're right. I say it's like
[36:23] physics. There's not a physics reason
[36:26] why this can't work. Costwise, it seems
[36:30] really imposing, but kind of the history
[36:33] of the Elon companies is the cost curve
[36:36] gets dramatically better. Like when we
[36:37] first invested in SpaceX, you know,
[36:40] Starlink like was not commercially
[36:42] available and like we had all these
[36:45] questions about how the economics would
[36:47] proceed over time. The same on the
[36:49] launch side, the same with the Model 3.
[36:51] Like I I I just have to think that that
[36:53] will get solved paired with the fact
[36:55] that we're going to have massive under
[36:56] supply self-inflicted on Earth.
[36:59] >> Uh it feels clear to me at a minimum it
[37:01] will be swing capacity.
[37:03] >> Yeah.
[37:03] >> And you know in the fullness of time
[37:05] maybe it will be larger.
[37:06] >> Well no it's really simple like if we
[37:08] use 50 and it is the people the question
[37:11] people should be asking about orbital
[37:12] compute which is the one SpaceX is
[37:14] focused on is Starship reusability.
[37:17] >> Yes. Because the math is like let's
[37:19] let's just say it's 50 billion a gig and
[37:22] let's just say 35 of that is it. Yep. So
[37:24] that's the same and maybe it grows a
[37:26] little because it's it's going into
[37:27] space. The rest is power, cooling,
[37:31] labor, all sorts of things that you
[37:33] don't need in space because you have the
[37:35] so you have the solar panel and the big
[37:38] radiator. Um
[37:41] and that call that's 15 billion and
[37:44] that's probably inflationary here on
[37:46] Earth.
[37:46] >> Yeah. Because labor fundamentally feeds
[37:48] into that. We just talked about what's
[37:49] happening to, you know, electrician. Um,
[37:52] >> yeah. Comp. Yeah.
[37:53] >> Yeah. Electrician
[37:53] >> materials are all going to go.
[37:54] >> Yeah. All of it. Yeah. We're going to
[37:55] have Yeah. We're going to run out of co,
[37:57] you know, we're we're the the copper
[37:59] bulls are, you know, focused on like
[38:01] copper shortages. All of it.
[38:03] >> Yeah. So that 15 billion is
[38:04] inflationary.
[38:06] And so what you have to compare it to is
[38:08] the cost of launch. And with Starship
[38:10] reusability, that goes to under a
[38:12] billion. So the economics just instantly
[38:14] flip. Now, you're always going to train
[38:17] on Earth. There will always be
[38:19] advantages to having, you know, GPUs
[38:23] right next to each other. Like there
[38:25] are, you know, speed of light
[38:26] limitations are a real thing. Latency
[38:28] matters. So, data centers on Earth,
[38:30] they're not going anywhere. I think
[38:32] they're going to continue to be very,
[38:33] very valuable. But an increasing
[38:36] fraction of the world's compute is going
[38:39] to be in orbit. And you know, Elon said
[38:43] that he and Jensen have co-designed a
[38:44] Reuben rack
[38:46] >> and they're it's gonna launch in the
[38:48] fourth quarter of 27.
[38:49] >> Yeah.
[38:50] >> And let's just say let's just say he's
[38:52] off by two quarters.
[38:54] >> Yeah.
[38:54] >> I mean, that's that's 2028.
[38:56] >> Yeah. That's still okay. That's pretty
[38:57] soon
[38:58] >> that, you know, as Brad Gersonner says,
[38:59] like nobody's really paying attention to
[39:01] this and it's like kind of happening in
[39:03] plain sight. And it kind of to me solves
[39:06] for something, you know, mids single
[39:08] just billions today,
[39:10] >> which by the way, you know, is like
[39:12] that's just like keeping share constant.
[39:15] >> Yeah. Exactly.
[39:16] >> You know, of like what's happening with
[39:18] >> not presumably taking any share on on
[39:19] Grockbot.
[39:20] >> Yeah. Yeah. From from three billion. And
[39:22] by the way, man, I would just I'd
[39:23] probably take the over with Grockbot.
[39:25] Yeah.
[39:26] >> I bet it's like
[39:27] >> changing by the day just based on my own
[39:30] usage and the number of people who are
[39:31] hitting their usage limits. And then you
[39:33] are starting to get from you know
[39:35] Grockbot like hey we're servers are
[39:38] overloaded every once in a while and
[39:39] like they have a lot of compute. Um so
[39:42] it's just like okay you don't want to
[39:44] debate orbital data centers
[39:46] >> no problem. Well like Starlink mobile
[39:48] like they have a pretty clear credible
[39:51] plan
[39:51] >> for how that's going to work and that
[39:54] you know wireless is you know call it
[39:55] another 8 900 billion of revenue that
[39:58] they address. So your yeah your mobile
[40:00] plus your broadband whatever it's call
[40:02] it like close to two trillion of a
[40:03] market
[40:04] >> and then you have a really rapidly
[40:07] growing AI AR base.
[40:10] >> Yeah. AI AR you've got the cloud you
[40:12] know the sort of the cloud business.
[40:14] >> Yeah. Um so I don't think great you're
[40:16] an orbital computic no problem. It
[40:19] doesn't matter.
[40:20] >> Yeah. Exactly.
[40:21] >> We don't even need to. We could just
[40:22] look at things that are happening today
[40:24] with terrestrial compute, with cursor,
[40:26] with Grock, with Grockbot. By the way, I
[40:29] think X ads are, you know, we have
[40:31] telemetry.
[40:32] >> They're also growing.
[40:33] >> You know, I would expect at some point
[40:35] you'll have like a Starlink
[40:37] Grockbot
[40:39] um Xadvertising bundle. You know, kind
[40:42] of one of the ways Google built their
[40:44] cloud business as they bundled it with
[40:45] ads and like, hey, we're, you know,
[40:47] maybe you're bundling the ads with AI,
[40:49] but why not do that?
[40:51] >> Yeah. Yeah. I actually like the AI
[40:53] position that they're in because it's
[40:55] like heads you win, tails you win in the
[40:57] sense that their first party business is
[40:59] growing very fast and they they caught
[41:01] up to the frontier like very quickly.
[41:04] Yeah.
[41:04] >> Um and so they've made the very
[41:07] aggressive compute investments to enable
[41:09] that first party work.
[41:11] >> Um and that's the kind of heads you win
[41:13] and like tails you win. Say they
[41:16] overbuilt their capacity for what they
[41:18] need for inference or training. they
[41:20] have a very compelling sub six month
[41:22] payback on the compute side um you know
[41:25] with like massive scarcity supply and so
[41:27] I think that's a really good setup
[41:29] >> and there was a bare case that hey okay
[41:31] well in in a in the open AI anth
[41:35] anthropic maximalist view where they're
[41:37] the only two companies and they're
[41:38] designing their own chips then like
[41:41] where what's the room for anyone else
[41:42] well like I don't think they're going to
[41:44] have a reusable starship and multiple
[41:46] spaceports anytime soon and if the
[41:48] economics of computer such that orbital
[41:51] is where it makes sense increasingly
[41:53] going forward because Starship should be
[41:55] deflationary, you know, terrestrial
[41:57] cooling, you know, power should be
[41:59] inflationary. Well, like even in in a
[42:01] world where
[42:03] they fumble the ball with their first
[42:05] party AI applications, like they do
[42:07] still have
[42:07] >> they're a massive infrastructure
[42:08] business.
[42:09] >> Yeah. Yeah. I I'm I'm so fired up about
[42:11] the uh the Starbase Louisiana. Uh
[42:13] >> Oh, yeah.
[42:14] >> I can't wait to visit, man.
[42:16] >> So cool. Yes.
[42:17] >> Uh I was reading about it last night and
[42:19] uh yeah, it's sort of like it's now the
[42:22] they now have the infrastructure for you
[42:23] know thousands of launches a year.
[42:26] >> Yeah. And eventually I think you will
[42:28] see like these star bases in multiple
[42:31] places, multiple coasts all over the
[42:34] world.
[42:34] >> Yeah.
[42:35] >> Like you know at some point you'll
[42:36] probably see one somewhere in the Middle
[42:38] East. You'll see
[42:40] >> you know whatever European country is
[42:41] like the least bureaucratic at the time.
[42:43] You'll see one there. You know, you'll
[42:45] for I think you'll see probably one in,
[42:47] you know, whether it's Japan, South
[42:48] Korea, who knows?
[42:50] >> Yeah. Yeah. Yeah. Yeah. Yeah. It's
[42:51] pretty exciting.
[42:52] >> Yeah.
[42:52] >> Yeah. The uh the capability to do to
[42:56] call it, you know, whatever 5,000
[42:58] launches a year, like that feels very
[43:00] futuristic.
[43:01] >> Yeah. I mean, it's wild. And I do think
[43:04] a distinction that um you know, SpaceX
[43:07] really tried to kind of hammer home
[43:08] during their their IPO is there's a
[43:11] difference between reusability and and
[43:13] China. They did catch kind of a rocket
[43:15] using this um it was actually kind of
[43:17] ironic. It was this kind of juryrigged
[43:19] system of kind of wires. Yeah.
[43:21] >> That had actually been suggested on the
[43:23] SpaceX subreddit.
[43:24] >> Yes.
[43:25] >> Like seven or eight or n or no no it was
[43:28] before they landed the first Falcon. So
[43:29] it's like more than 10 years ago
[43:31] >> and like China's clearly paying close
[43:33] attention to the SpaceX subre subreddit.
[43:36] But that's very different catching that
[43:38] thing from what they're trying to do
[43:39] with Starship where you know the uh the
[43:42] booster gets caught with the things and
[43:44] then it gets moved and then the Starship
[43:46] gets caught and then it gets stacked, it
[43:48] gets fueled and just sent right back.
[43:51] Yeah. Two a day. Two a day per pad.
[43:53] >> Like those numbers add up pretty fast.
[43:55] >> And there and I do think I think they're
[43:57] engineering the pads for more than two a
[43:59] day if I
[44:00] >> Yeah. I think that's a conservative I
[44:01] think that's a conservative assumption.
[44:02] Yeah.
[44:03] >> Yeah. Um but I mean
[44:05] >> Yeah. What's the Okay, so SpaceX, like
[44:06] again, you and I have talked a ton about
[44:08] SpaceX.
[44:10] What's like the most futuristic thing
[44:11] that you think about with SpaceX? Like
[44:15] the 10-year Okay, so you and I were at
[44:17] this conference together and there was
[44:19] this whole debate about um among a small
[44:22] group of public investors of like what's
[44:23] going to be the the first 10 trillion
[44:25] company. And uh I think what you said
[44:29] was like I have no idea, but I know
[44:30] which one's going to be the first 20
[44:31] trillion dollar company. Uh, so like
[44:35] what's the most futuristic like product
[44:37] or market or technology thing about
[44:39] SpaceX that that you can think of?
[44:41] >> Look, I mean this sounds crazy, but
[44:43] asteroid mining is going to be a very
[44:45] real thing. We're going to capture, you
[44:46] know, there's asteroid psyche. It has
[44:48] more gold, silver, platinum, you know,
[44:51] every precious metal in it that exists
[44:53] in the Earth's crust.
[44:55] At some point, particularly with
[44:57] Starship, you will be, you know, and we
[45:00] may need that um lunar base to make this
[45:02] happen. You'll be able to cap capture
[45:04] these asteroids. You'll bring them into
[45:07] a stable kind of geocynchronous orbit
[45:09] over some, you know, Americanowned
[45:12] atal in the middle of the Pacific. Um,
[45:16] you know, no humans within whatever 50
[45:18] miles. you'll, you know, you can imagine
[45:21] like Optimus robots, you know, um doing
[45:24] doing the work. Yeah.
[45:25] >> Yeah. Doing the work. Um and then, you
[45:28] know, delivery to Earth is free and for
[45:30] sure some of it's going to burn up,
[45:32] >> but I think that's going to happen. And
[45:36] I always think um
[45:38] Jeff Bezos said something very
[45:40] interesting. He said, "I think in the
[45:41] future Earth is going to be zoned
[45:44] residential." And you know, somebody
[45:46] asked him, this is like 15 years ago,
[45:47] what do you mean by that? He's like all
[45:49] heavy industry will take place in outer
[45:51] space. And then this addresses the
[45:53] pollution concerns. It addresses
[45:54] everything.
[45:55] >> You know, people always get like really
[45:57] worried about, oh, you know, we still be
[45:58] able to see the stars.
[46:00] >> And it's just like I think it's hard for
[46:02] like the human mind to understand how
[46:05] big space is, how big outer space is,
[46:08] >> you know, it's
[46:09] >> we don't have to worry so much about
[46:10] emissions up there. Yeah.
[46:11] >> Yeah. Yeah. So I think that is um
[46:16] that's probably the most futuristic
[46:18] thing.
[46:18] >> But in terms of an economic application,
[46:20] but it does um
[46:24] I mean
[46:26] I I do think in the next few years
[46:29] you're going to have a fleet of
[46:30] starships land on Mars. Next few years I
[46:33] mean I don't know let's just say at the
[46:35] outside this is eight years away.
[46:37] >> Yeah. They're going to land on Mars.
[46:39] Going to have like, you know, a little
[46:41] ramp's going to come out of the PEZ
[46:43] dispenser and it's going to be a
[46:44] modified Starship, the Mars colonial
[46:46] transporter, and it's going to be wild.
[46:48] You're going to have Optimus robots
[46:50] holding American flags like walk down
[46:54] and then, you know, they're going to
[46:56] pull out a bunch of solar panels and
[46:58] batteries and racks of compute and
[47:01] they're going to set all of that up.
[47:03] they'll be dropping Starlinks,
[47:05] you know, and maybe the orbital
[47:08] mechanics don't allow this, but I think,
[47:10] you know, they'll they will figure out a
[47:11] way to have, you know, capacity. So,
[47:14] just think how crazy it is to watch like
[47:16] the views from Pathfinder,
[47:18] >> you know, or, you know, whatever these
[47:20] different, you know, Mars um
[47:22] >> rovers and stuff,
[47:22] >> rovers are and like, you know, 4K video
[47:25] through Optimus robots all over Mars and
[47:28] then after that there will be humans
[47:30] >> who can inhabit it. Yeah. Yeah. Yeah.
[47:32] That is crazy to think about.
[47:33] >> And that that's going to be an amazing
[47:35] moment for America.
[47:36] >> Yeah.
[47:36] >> Oh, I mean, think about the moon
[47:38] landing.
[47:39] >> This is a little bit bigger. Yeah.
[47:40] >> Yeah.
[47:41] >> Yeah.
[47:41] >> Um, so that seems cool. Um,
[47:45] >> that's a good one. That's That's a good
[47:46] That's a good one. Yeah. Not a lot of
[47:48] chatter about that one out there. Yeah.
[47:50] But I think it's highly likely to
[47:52] happen.
[47:52] >> Yeah. Yeah. Yeah. So, you mentioned
[47:54] Microsoft.
[47:55] >> Yeah. and the bet that they made which
[47:57] is like a little bit of you know like
[48:00] Apple's the extreme kind of bet against
[48:02] the future kind of bet they made and
[48:04] like Microsoft is kind of a gradient of
[48:06] that.
[48:06] >> Yeah.
[48:07] >> Like what's your what's your outlook for
[48:10] their decisions?
[48:12] >> Well, I do think the world has gotten a
[48:13] lot friendlier for their strategy. Um
[48:15] you know they clearly tried to make a
[48:17] frontier model. They failed.
[48:18] >> Yeah.
[48:19] >> You know Satia said we're going to have
[48:20] our own models that are very
[48:21] competitive. Like I think he said that
[48:23] 18 months ago. they don't have their own
[48:25] models that are competitive, but what
[48:28] you're seeing with um I think the future
[48:32] is an ensemble of models. You know,
[48:34] there's a paro curve. No one model is
[48:36] going to be the best at everything. And
[48:38] I think the future for certainly, you
[48:40] know, kind of the global, you know,
[48:42] 10,000 biggest companies is you're going
[48:45] to take whatever the best open source
[48:46] model is, I think probably in the in the
[48:49] very near near future that's going to be
[48:51] an NVIDIA model.
[48:52] >> Yep. The labs making AS6 create very
[48:55] interesting
[48:56] >> incentives for to get into each other's
[48:58] business
[48:58] >> incentives for Jensen and everybody's
[49:01] well oh in a world where open source
[49:02] wins who funds the training well this
[49:04] the chip companies could fund the
[49:06] training yeah
[49:06] >> it's trivial to do a 50 to$100 billion
[49:09] training run uh you know for Jensen and
[49:11] maybe soon I do wonder if this is kind
[49:14] of Google's like super long-term play
[49:17] like they they they seem to like maybe
[49:18] have opted out of the frontier race for
[49:20] now um We're going to monetize our
[49:23] compute at high rates and we're going to
[49:26] um sell TPUs externally, but that
[49:29] generates so much cash flow and open
[49:32] source is getting closer and closer and
[49:34] closer to the frontier. And it just may
[49:36] be the winner is ultimately just who has
[49:39] kind of the most cash flow to to fund
[49:40] these big training runs. But I do think
[49:43] you're going to see American open source
[49:45] led by led by Nvidia get really close to
[49:49] the frontier like they paid that
[49:51] poolside acquisition was made for a
[49:53] reason. Poolside actually had a lot of
[49:54] really good American open source talent.
[49:57] I they're they're you know they're doing
[49:58] a lot of smart things but that that is
[50:00] really good for Microsoft and at some
[50:03] level almost every application software
[50:06] company because what you can do now is
[50:08] you can take a base model and Neatron to
[50:11] date has not had a lot of post-raining.
[50:13] It's kind of been a good pre-trained
[50:15] model that you could do with what you
[50:16] want. So if you take a really good
[50:19] pre-trained base model and then instead
[50:22] of sharing your own kind of enterprise
[50:26] context that's truly your IP that's
[50:28] truly the value you know of your company
[50:31] is like you know the context embedded in
[50:33] all of your data and like sharing that
[50:35] with a frontier lab you know that may be
[50:37] hazardous for your financial health.
[50:38] Yeah, certainly with like the shift in
[50:40] the ZDR policy like Yes.
[50:43] >> Yes. And so you take a really capable
[50:46] open source model and you do a lot of RL
[50:49] and supervised fine-tuning on your own
[50:51] data. So you own it and it's your model.
[50:53] >> Yeah.
[50:54] >> And then if intelligence is like a super
[50:56] important input into your business, you
[50:59] want to own and control your
[51:02] intelligence, its capabilities, its
[51:04] cost. And then what we've seen from a
[51:07] lot of companies and you know Grockbot
[51:10] my understanding is you know I think
[51:11] it's Gemini 3.7 flash
[51:15] >> um Grock 4.6
[51:17] >> and some Opus
[51:18] >> y
[51:19] >> and what you and behind a router
[51:22] >> and you will um
[51:25] >> and I'm sure Elon is very focused on
[51:27] having it all grow as soon as possible.
[51:29] >> Yeah. Yeah. Of course. Um but I think
[51:31] what you'll see these companies do is
[51:34] they'll have their own model on their
[51:36] data and it will work with one or two
[51:38] other frontier models. Um not not you
[51:41] know necessarily but just you know
[51:43] checking each other it'll be kind of
[51:45] transparent to you the the most frontier
[51:47] for planning and then have execution run
[51:49] by everything else that's lower costed.
[51:50] Yeah, absolutely. And so I think that
[51:53] feels like a very likely future to me.
[51:57] And that's a that is a much Microsoft
[51:59] friendlier future than one in which
[52:01] there's just only two dominant frontier
[52:04] models. And it certainly looks like
[52:06] there's going to be at least three with
[52:08] Grock. I do think you got to give Meta a
[52:10] lot of credit.
[52:11] >> They've done a great job.
[52:12] >> Yeah. And I mean they were out of the
[52:13] game and they got back in the game. And
[52:15] it's just it's kind of amazing. Who
[52:16] could have imagined a year ago, you
[52:19] know, when it was like Gemini was
[52:21] ascendant exactly that this is the
[52:23] scenario
[52:23] >> Gemini wouldn't even be in the
[52:25] conversation
[52:26] >> and Muse and Meta would be significantly
[52:28] ahead of them from a capability
[52:30] perspective.
[52:31] >> Um, so it's just, you know, this is
[52:33] >> kind of like the highest stakes game of
[52:36] like corporate chess ever played.
[52:38] >> And, you know, people, you know, some
[52:39] people have made bad moves, they've made
[52:41] good moves. You seem some people come
[52:42] out of the game, others come back in.
[52:45] Um, but a future where that future where
[52:48] it's a, you know, I don't know if we're
[52:50] going to call it multimodel, a hybrid
[52:52] model, you I don't know what terminology
[52:54] the world is going to settle on, but I
[52:56] think that's the future.
[52:57] >> Yeah.
[52:58] >> And I'm actually surprised. I think the
[53:01] best broad instantiation of that today
[53:05] outside of Grockbot, outside of cursor,
[53:08] outside of you know like Harvey's done
[53:09] some cool things with that
[53:11] >> where they've done it is actually just
[53:13] the Fireworks Nexus product.
[53:14] >> Yeah.
[53:15] >> Where you can Yeah. You can
[53:18] >> choose your frontier model.
[53:20] Let us take whatever open source model
[53:22] you want, RL it for you, for your data
[53:24] for Gold Coleman Sachs, for Morgan
[53:25] Stanley, for JP Morgan, for Fidelity,
[53:27] for A16Z. You have all your own data.
[53:30] you control your intelligence and we
[53:32] make it transparent behind a router.
[53:34] >> Yeah,
[53:34] >> I think that is like a very plausible
[53:37] future and that's clearly what um Lynn
[53:42] from Fireworks, she was the first one to
[53:43] say it and then Alex Karp and Satia,
[53:46] they both kind of like
[53:47] >> Yeah, they've they've taken their own
[53:48] version of it. Yeah.
[53:49] >> Yeah. But you know, Satia's essay of
[53:51] specialized intelligence, like I think
[53:53] it's very plausible,
[53:55] >> but this stuff is really hard to do.
[53:58] Like that that sounds easy.
[54:00] >> I was it sounds easy to describe like
[54:02] the way I describe it to people is like
[54:04] who gets to be the abstraction layer to
[54:07] the organization and the users with
[54:09] intel like of of intelligence. It's like
[54:11] the most whatever vi after space or
[54:14] position that you could imagine in
[54:15] business like in the history of
[54:17] business.
[54:17] >> Yeah, for sure.
[54:18] >> Right. I think it's like the answer is
[54:19] and again.
[54:20] >> Yeah. Yes. And for sure it's Yeah. Who's
[54:22] the arbiter of intelligence for global
[54:24] enterprises and probably consumers? I
[54:27] was a retail analyst and um
[54:32] you know everybody kind of thinks
[54:33] running one of these big chains is easy
[54:36] and there's a lot into it and it's like
[54:38] well it's really easy to start an
[54:41] American retailer in any category cuz
[54:44] America's so big it's worth over $50
[54:46] billion almost any category.
[54:48] >> Yeah. All you have to be able to do is
[54:51] have a fleet of a thousand stores in 50
[54:54] different states that have very
[54:55] different climates, consumer
[54:57] preferences.
[54:59] You need to have them stocked with the
[55:01] right products at the right time for
[55:03] that region at the right prices. They
[55:06] need to be staffed by friendly and
[55:07] knowledgeable employees who don't steal
[55:09] from you
[55:09] >> who turn over at 100% a year.
[55:11] >> Turn over at least 100% a year. The
[55:13] stores need to be clean and well lit.
[55:15] And if you can do that,
[55:18] presto, $50 billion dollars. Yeah.
[55:20] >> And like in the history of American
[55:22] business, like you can I mean it's more
[55:24] than one hand, but you don't have to go
[55:27] through many.
[55:28] >> Yeah.
[55:28] >> It's really hard to do. And
[55:31] >> having that abstraction layer, having it
[55:35] work, having it seamless is, I think,
[55:38] way harder to do than people think. And
[55:41] I do think what something I think is
[55:42] very interesting about cursor, I'd love
[55:44] your opinion on this is like everybody
[55:47] else in the lab space, you know, had
[55:50] this like
[55:53] we're creating a digital deity, you
[55:54] know, and AGI and ASI like we're
[55:58] >> and the Curser guys were just like we
[56:00] want to make great product.
[56:02] >> Yes. Exactly. in a in a strange way o of
[56:05] everybody at the frontier. Um probably
[56:08] Kerser and it was the most product
[56:11] focused.
[56:11] >> Yes.
[56:12] >> Yeah. I'd say in you know now they're
[56:13] part of SpaceX but that suits Elon and
[56:16] his mindset really really well.
[56:18] >> Yeah.
[56:19] >> Let's make it an engineering problem.
[56:21] You know create the model factory and
[56:23] then we need to have a really good
[56:25] product.
[56:26] >> Yeah.
[56:26] >> You know the you know the the the Tesla
[56:29] cars they're amazing. I mean it's I
[56:30] don't I don't know if you drive one but
[56:32] it drives
[56:33] >> everywhere. Yeah. Yeah. But like the
[56:34] what cursor figured out is
[56:37] >> they're they had I would say a similar
[56:40] instate vision as what those others guys
[56:42] had.
[56:43] >> It was just a different path to get
[56:44] there and it's sort of like a practical
[56:45] meet the customer with what with where
[56:47] they are meet the technology where it
[56:48] is.
[56:49] >> Um and I think you know they'll sort of
[56:51] they have already demonstrated that they
[56:52] kind of led their way up into autonomy
[56:54] from from that starting point.
[56:56] um coding is unique compared to
[56:59] everything else in knowledge work. This
[57:01] this would be like in support of the
[57:02] point that Microsoft is in a good
[57:03] position
[57:04] >> because it is verifiable and perfectly
[57:07] documented and like nothing else in
[57:08] enterprise
[57:09] >> is verifiable and perfectly documented
[57:11] and so it will be messy like that that
[57:14] leads you to a good you know bullcase
[57:16] for something like Microsoft that
[57:17] abstraction layer
[57:18] >> if they execute but it's really really
[57:21] hard to make it really simple for
[57:24] >> oh you know click my co-pilot link to
[57:26] all my stuff train a model yeah
[57:29] >> on our data
[57:31] convince me that you're not going to
[57:32] share it with anyone else and then put
[57:34] it behind a router that's seamless for
[57:35] me and continuously upgrade that open
[57:38] source model.
[57:39] >> Yeah. It's not just some middleware like
[57:40] it's very hard to do. Yeah. And and by
[57:42] the way, they're going to compete
[57:43] they're going to be competing with not
[57:45] only the labs to be that abstraction
[57:47] layer
[57:48] >> but data bricks. So like
[57:51] >> Palunteer um the inference the inference
[57:54] providers um the application companies
[57:57] right so like Harvey has done an
[57:58] incredible job of this and you know like
[58:01] legal has sort of in take off and um and
[58:04] and I think they can see the future of
[58:06] how to be that abstraction layer um and
[58:08] do the work um but like legal is also
[58:11] unique because it's very documented and
[58:13] it's somewhat verifiable tax we'll see
[58:15] that we see see things like that but
[58:17] like the the one and a half billion the
[58:18] really appealing brought by is going to
[58:20] be very messy to go get.
[58:22] >> Yeah. Although I do always think and um
[58:24] you know I think probably in their heart
[58:26] of hearts Harvey and Lora think oh if we
[58:29] solve this
[58:30] >> we could be that abstraction layer for
[58:31] everyone.
[58:32] >> I think probably in their heart of
[58:34] hearts cognition thinks something like
[58:35] that too.
[58:36] >> I think everybody thinks and by the way
[58:38] there's like massive validation of the
[58:39] category because Kirkland Ellis said
[58:42] >> we're going to spend 500 million bucks
[58:43] to build this ourselves. Like first of
[58:45] all you know like good luck that's going
[58:48] to be very hard. Yes.
[58:49] >> Um, but that actually tells you that the
[58:52] pie is really big, right? Huge.
[58:53] >> Yeah, it's massive.
[58:54] >> And that's it's and you know, just um
[58:56] and I'm sure they have a very smart head
[58:58] of a head of AI, but it's not like a
[59:00] $500 million onetime build. That model
[59:04] has to be continuously updated,
[59:05] switching out the base model. Then all
[59:07] of that has to happen transparently. But
[59:10] I think you're going to have this huge
[59:11] collision between,
[59:13] you know, products like Fireworks Nexus,
[59:15] these legal agents, coding agents, big
[59:19] companies like Microsoft,
[59:21] >> Data Bricks,
[59:21] >> Data Bricks, Snowflake coming up,
[59:23] >> you know, for sure. Um, you know,
[59:26] Salesforce, I think, is going to, you
[59:28] know, Salesforce and Workday and all
[59:30] these companies. This is like
[59:31] everybody's going to go after it. It's
[59:33] just going to come down to who executes
[59:34] the best and
[59:37] >> and this is just you know who has the
[59:39] lowest costs.
[59:40] >> Yes, exactly.
[59:40] >> But it's going to be very hard I think
[59:42] over time unless you're re if you're not
[59:45] vert vertically integrated you have to
[59:47] be so good to emerge as that abstraction
[59:50] layer.
[59:50] >> Yeah. Yeah. To be the lowcost provider
[59:52] very hard
[59:53] >> because yeah we you're just simply not
[59:55] going to be the lowcost provider if
[59:56] you're not vertically integrated if you
[59:58] don't own your own compute over the very
[60:00] long long term. Um and you know it's
[60:04] that's another reason like I um you know
[60:06] I increasingly look at these
[60:07] hyperscalers on EV to net PP&E.
[60:10] >> Yes.
[60:10] >> Because net PP& is compute and that is
[60:13] just what the market thinks you're going
[60:14] to monetize your fleet of compute at and
[60:17] you can kind of look at them and there's
[60:18] some pretty obvious inefficiencies too.
[60:20] >> Yeah. Yeah. Yeah.
[60:21] >> Yeah. Kind of an AI version of price to
[60:23] book.
[60:23] >> Yeah. I like the price to book. Okay.
[60:27] >> Um so okay you you mentioned Jensen. you
[60:29] know, I I'd share your sentiment like
[60:30] he's like carrying this industry
[60:31] forward. Like tell me your thoughts on
[60:33] Nvidia.
[60:35] >> So, um
[60:37] I think he's in a very very good
[60:40] position and his strategy of being
[60:42] vertically integrated but horizont
[60:44] horizontally open and it's like okay
[60:47] like let's just say um
[60:50] you know let let's say there's some
[60:52] accelerator that emerges that is really
[60:53] really really really good. almost
[60:56] certainly it will be better if it can
[60:58] plug into and this is why like I know
[61:00] you have an accelerator investment my
[61:03] number one thing is if you're a
[61:05] semiconductor CEO the only thing you
[61:08] should ever say is thank you Jensen
[61:11] thank you for creating this opportunity
[61:13] thank you how can we work with you we
[61:16] want to enable you sure we're going to
[61:18] compete with you on the edges
[61:20] >> but you know my rule of thumb for
[61:21] accelerators every 1% share today is
[61:23] probably worth a hundred billion Yes.
[61:25] >> So there's no need to go head on with
[61:28] Nvidia.
[61:28] >> Yeah.
[61:29] >> Um just pick a niche, get your 1%. Make
[61:33] sure that you know
[61:34] >> is very big.
[61:35] >> He has he has nine chips. Yeah.
[61:37] >> Um you know he's got he's got multiple
[61:39] flavors of accelerators. He's got CPUs.
[61:42] >> He's got you know Ethernet switches. He
[61:44] has two kinds of GPUs. You know he's got
[61:47] you know we've gone from just um scale
[61:49] out networking being a thing. We have
[61:50] scale up scale out scale across now
[61:52] scale in.
[61:53] >> Yeah. So just try to find a way to plug
[61:55] into his ecosystem.
[61:56] >> By the way, this is not foreign. Like
[61:58] his biggest customers all have competing
[62:01] products with various of those nine
[62:03] chips.
[62:04] >> Yeah. And just try to find a way to plug
[62:06] in, but just be nice to him. Be nice. Be
[62:11] nice. It's all personal. Yeah. You know,
[62:13] and it's just like sometimes like, you
[62:15] know, you hear some of these and it's
[62:17] like, have you ever seen game tape of
[62:20] the Chicago Bulls when Jordan was is,
[62:23] you know, it's game 50 of the season.
[62:25] >> Yeah.
[62:26] >> And he's a little bored.
[62:27] >> Yeah.
[62:28] >> And the Bulls are down cuz, you know,
[62:29] they're up eight games. You know,
[62:30] they're up eight games over the number
[62:32] two person in their conference.
[62:34] >> And he's a little bored. And then
[62:36] somebody
[62:36] >> somebody talks
[62:37] >> Somebody who's you who's who's kind of
[62:39] young decides, I'm going to talk to
[62:41] him because we're beating him. And then
[62:42] he just looks
[62:43] >> and it's like
[62:44] >> and it's like
[62:44] >> it's the best. Those are my favorite.
[62:46] >> It's amazing. Yeah. Yeah. You We've all
[62:47] seen, you know, whatever the last dance.
[62:50] >> Just don't do that.
[62:51] >> Yeah. Exactly.
[62:52] >> You know, just just like, "Hey, Michael.
[62:54] Man, I'm so happy to be on the court
[62:56] with you." Like that's that's to that's
[62:59] that's the move. But the reason it's
[63:00] particularly important is because
[63:03] Jensen's data centers are financable.
[63:05] >> Yes. And it goes back to that point like
[63:08] let's say it's $50 billion
[63:12] um for an Nvidia data center you need a
[63:16] $15 billion equity check.
[63:18] >> Yeah.
[63:19] >> Okay. You can finance the other 35
[63:21] billion.
[63:22] >> Yeah.
[63:22] >> And it's not circular financing. I have
[63:24] a lot of respect for the people I have
[63:25] met from Blackstone and KKR and Apollo.
[63:28] Yeah.
[63:28] >> And they're underwriting each of those.
[63:30] >> Yeah. and they finance it. And then
[63:34] there's a residual value guarantee,
[63:36] which as long as that residual val value
[63:39] guarantee is less than the gross profit
[63:40] dollars he's getting from selling the
[63:42] chips into that data center,
[63:44] >> it's like essentially it's super NPV
[63:47] positive with very little risk for him.
[63:50] >> Um, and then he, you know, he gets a
[63:52] revenue share. So if you're um and his
[63:56] data centers are the most financable.
[64:00] >> Yes.
[64:00] In I like let's just say a good case for
[64:04] probably TPUs are the second most
[64:06] financable.
[64:07] >> It probably takes I don't know double
[64:10] the equity check at least. Yeah.
[64:11] >> And then the rates on the rest of it are
[64:14] higher.
[64:15] >> Yeah. Exactly.
[64:15] >> And so cost of capital is a huge
[64:19] advantage and that's why you just want
[64:21] to be part of his ecosystem. And you can
[64:24] see he's he's he has all these chips.
[64:27] He's acquiring land power and shell
[64:29] companies now matchmaking them with
[64:31] offtake agreements. I think one reason
[64:33] he's doing these RVGs is if he doesn't
[64:35] do them, it's kind of an anthropic and
[64:37] open AI dominated world because they can
[64:39] pay the most for compute. He can
[64:42] effectively help other people
[64:44] >> compete with anthropic and open AI.
[64:46] >> Yeah. In the same way that he stood up
[64:47] the neo clouds in the first place. Yeah.
[64:49] >> It's just democratizing compute which is
[64:51] good for the world. Again, I think he's
[64:52] a patriotic American. His his interests
[64:54] are aligned with that though with with
[64:55] with the patriotic American ones, right?
[64:58] Fragmentation, right?
[64:59] >> Fragmentation, no dominant AI. Exactly.
[65:02] Which is which is really good because
[65:03] he's like a he is a ruthless competitor.
[65:06] And it's awesome that his incentives
[65:09] around fragmentation of AI,
[65:10] fragmentation of models, and you know,
[65:13] fragmentation of power um are completely
[65:16] aligned with what's good for America.
[65:18] And just going back to open source, just
[65:20] like I I just can't take it that people
[65:24] think that Jensen is like the world's
[65:27] biggest advocate for open source and
[65:29] it's somehow the a giant risk to his
[65:32] business.
[65:33] >> Yeah, exactly. No, it's great for his
[65:34] business. It's great for his business.
[65:35] >> It's amazing for his business because it
[65:37] means that instead of, you know, having
[65:39] a 90% margin on top of a token made with
[65:42] an Nvidia GPU,
[65:44] >> maybe it's a 40% margin. So more of
[65:47] those tokens are going to be consumed
[65:48] which means you need more compute.
[65:50] >> Yeah, exactly.
[65:51] >> Um
[65:52] >> in a supply constrained world
[65:53] >> in a supply constrained world and you
[65:55] know let's just what percentage of the
[65:57] world's supply has he locked up?
[66:00] >> 70 80 somewhere in there. And then um
[66:04] >> you're talking about fab capacity.
[66:06] >> All of it. All of it. You know it's just
[66:07] because he's saw this coming before
[66:09] everybody else.
[66:10] >> Yeah. And all the system supply chain.
[66:12] >> Yeah. He's got he's got the fab
[66:14] capacity. Yeah. locked up. He's got DRAM
[66:17] capacity locked up. He's got NAND
[66:19] capacity. He's got laser capacity. He
[66:22] has capacitor capacity. He has, you
[66:24] know, what you need to make the racks.
[66:27] And it's just like he, you know, he used
[66:28] to say, if I go back,
[66:32] >> you know, 15 years, he'd say, "Listen,
[66:33] I'm making a two or three billion dollar
[66:35] bet every two years, and I'm moving
[66:38] really, really fast."
[66:39] >> Yeah. Now he's making these multiundred
[66:42] billion dollar bets, bringing the supply
[66:44] chain alongside him. He's bringing the
[66:47] financing alongside him by kind of
[66:49] standardizing it, making it easy for the
[66:51] very smart people at Blackstone, KKR and
[66:53] Apollo and Goldman Sachs and Morgan
[66:54] Stanley, JP Morgan to finance
[66:57] >> and like that is hard to compete with.
[67:00] >> Yeah. And you know it is um
[67:04] we um my firm trades we have a pretty
[67:07] big portfolio private portfolio
[67:08] companies uh that are semiconductors
[67:11] and it's just um you know Elon said a
[67:15] lot of people are going to learn a hard
[67:16] lesson in hardware and like I will just
[67:19] say I've learned a lot of hard lessons
[67:20] in semiconductor investing like you can
[67:23] you can bet on the best team and you
[67:26] tape the chip out you feel great okay
[67:28] we've taped it out and it and that's
[67:30] happening happening faster than ever
[67:30] right now.
[67:31] >> Yeah, it's happening faster than ever.
[67:32] You feel great about it and we're
[67:34] getting really good with the emulation
[67:36] and the simulations and you feel great
[67:38] about it
[67:40] and then um you know you'll experience
[67:42] this the chip comes back from the lab
[67:44] everybody you get a facetime from the
[67:45] CEO they plug it in.
[67:47] >> Yeah. you know, and like and then
[67:50] sometimes it doesn't work, you know,
[67:53] it's just like
[67:54] >> Yeah, this famously happened with
[67:55] Cerebrus twice, right? Like, and they've
[67:57] powered through and like they've done
[67:58] great.
[67:59] >> Well, I don't I think the chip I think
[68:00] each Cerebrris chip worked, it just
[68:04] struggled to find product market fit.
[68:06] >> Yeah. Yeah. Fair.
[68:06] >> For the first two generations, the chip
[68:08] worked. It just didn't have product. And
[68:10] they've done great with it. Yes.
[68:12] >> Yeah. But there's a different thing
[68:13] between you, you plug it in, doesn't
[68:14] work at all.
[68:15] >> And it doesn't work at all. Exactly. And
[68:17] then it's like if it doesn't work at
[68:19] all, you might be back to the drawing
[68:22] board and hey, we need another, you
[68:24] know, hundreds of millions of dollars,
[68:27] billion dollars, and we've we've learned
[68:29] our lesson. It's going to work the next
[68:31] time two years from now.
[68:32] >> Yeah. Assuming you can finance it. Yeah.
[68:34] >> As Yeah. Assuming you can get financing.
[68:36] So, it's um you know, semiconductors are
[68:40] hard. Like the real world is hard. Like
[68:43] hardware is hard. and what he is doing
[68:47] at the scale he is doing at and the
[68:49] speed and bringing all of this alongside
[68:52] him cuz you know the land and the power
[68:54] has to come.
[68:54] >> Yeah.
[68:55] >> You know the entire supply chain has to
[68:57] come the financing has to come.
[68:59] >> And so given that he's you know 70 80%
[69:04] whatever we want to say you just want to
[69:06] plug into that ecosystem.
[69:07] >> Yeah. Part of why Elon made the decision
[69:10] he made right. Yeah.
[69:11] >> Yeah. which I also think was like a very
[69:14] high elo move.
[69:15] >> Yeah, totally.
[69:16] >> So,
[69:17] you've had everybody else try and build
[69:19] their own ASIC.
[69:20] >> Yeah,
[69:21] >> they've gotten up on stage. Sometimes
[69:23] they say negative things about, you
[69:25] know, Jensen or Nvidia or take shots.
[69:28] >> Um you I did think it was pretty smart.
[69:31] You know, the jalapeno team last night
[69:33] and we should give credit where credit
[69:34] is due. Jalapeno is the I would say the
[69:39] first good ASIC other than TPU or
[69:42] tranium I have seen from internal
[69:44] >> in a in a what seems to be a pretty
[69:45] short amount of time.
[69:46] >> Pretty short amount of time. It's
[69:48] impressive. We should give credit where
[69:49] credit is due.
[69:50] >> They do have a good team working.
[69:51] >> They have a good team. Yeah.
[69:53] >> Um so they had a really good team. I
[69:55] think they had a lot of advantages and I
[69:57] do think
[69:58] >> if you are a lab and you have the model
[69:59] and you see the direction of research
[70:01] that's a big advantage for designing
[70:03] your own chip. But then you go back to
[70:04] Nvidia and they work with everyone.
[70:07] >> Yes.
[70:07] >> And everybody, you know, keeps thinking
[70:09] it's going to really standardize. And if
[70:11] you look at the three big, you know,
[70:13] Chinese open source models, Deepseek,
[70:15] Kimmy, Quinn, they're kind of all um
[70:19] evolving in very different ways.
[70:22] >> Yeah.
[70:23] >> And they can, you know, they can all run
[70:24] on, you know, a more general purpose
[70:26] chip, um, a GPU, but you're going to
[70:29] need, if you want to specialize,
[70:31] >> Yeah. You're going to need general
[70:32] purposes at a minimum for the types of
[70:33] evolution you see from that. Yeah.
[70:35] >> So, um like I think he's I'm very happy
[70:40] his incentives as a CEO are perfectly
[70:43] aligned with what's good for America.
[70:45] >> Yes.
[70:45] >> Um
[70:47] so I just make sure your semiconductor
[70:51] guys do not talk trash about Michael
[70:54] Jordan.
[70:54] >> Be nice to be nice to MJ. Be nice to MJ.
[70:57] Yeah. Exactly.
[70:57] >> Yeah. And then it's like, you know,
[70:58] sometimes it's like, you know, you tug
[71:00] on Superman's cape and you get
[71:01] confident.
[71:02] >> Yeah.
[71:02] >> You know, you get confident and you
[71:04] start to talk a little bit of trash.
[71:06] Well, you know, Superman sometimes he
[71:07] just flies away like that's what
[71:09] happened to the TPU team.
[71:10] >> Yeah. You know, and you know, Jalapeno,
[71:13] they're tugging on Superman's cape a
[71:15] little bit.
[71:15] >> Yeah. We'll see.
[71:16] >> We'll see. And it is kind of amazing
[71:18] that like
[71:19] >> Jalapeno did something that none of the
[71:22] big
[71:23] >> like I this is as competitive of a chip
[71:26] as I have seen. Yeah.
[71:27] >> But again, it's just competitive with
[71:29] one of his eight or nine chips.
[71:31] >> Yeah. One of his nine. Yeah, of course.
[71:34] >> They'll continue to work closely
[71:34] together. Yes.
[71:35] >> Yeah. They'll continue to work closely
[71:36] together. So, it's like, hey, that's
[71:38] great. You did the one thing. Well, to
[71:40] actually be competitive with him at the
[71:41] system level, you need another eight
[71:43] chips.
[71:43] >> Yeah. Exactly.
[71:44] >> Yeah.
[71:45] >> Um and he is at and you know, Dylan at
[71:49] some analysis talks about how he's the
[71:52] bank of AI. He's like he's the central
[71:53] bank of AI. He's the Federal Reserve of
[71:55] AI. Yeah.
[71:56] >> And so I actually think it was really
[71:57] smart for Elon instead of like
[72:01] >> competing, you know, with somebody who
[72:03] is
[72:04] >> fully aligned.
[72:04] >> Yeah. Fully aligned.
[72:06] >> Mhm.
[72:06] >> And I think that history is going to
[72:08] judge that to be a wise decision. In a
[72:10] world that is so supply chain
[72:11] constrained, it's actually really hard
[72:13] to tell what true customer preferences
[72:15] are, right?
[72:16] >> Because like you come out,
[72:17] >> Yeah. they'll take anything. Yeah. This
[72:18] is this is how you know that like very
[72:19] old whatever the price is held up of
[72:21] H100 is very high.
[72:23] >> Yeah. Yeah. And if you have a TSM
[72:25] allocation, you're going to be sold out.
[72:27] Yes.
[72:27] >> Particularly if you can get the DRAM to
[72:29] pair with it. You're going to be sold
[72:31] out.
[72:31] >> So, it's actually kind of hard to infer
[72:34] true customer preferences. And I
[72:37] actually think one of the best ways you
[72:38] can like see true customer preferences
[72:41] is the kind of deals they cut with chip
[72:43] companies. So, broadly speaking, you
[72:46] know, the first deal is where the chip
[72:48] company invests
[72:48] >> Yep.
[72:49] >> in a customer. And you saw TPU and
[72:52] Tranium, Amazon and Google do that with
[72:53] Anthropic. Yep. And that was to their im
[72:55] immense advantage because it really
[72:57] helped their businesses, I think, helped
[72:58] those chips really level up because you
[73:00] kind of need to use a chip. There's a
[73:01] cold start problem.
[73:03] >> And um
[73:05] and in that scenario, as long as the
[73:07] dollars you invest are less than the
[73:09] gross profit, you can't lose money. And
[73:11] then there's a scenario where you do the
[73:13] RVG,
[73:14] Blackstone finances it or whoever,
[73:16] Blackstone, Apollo, KKR, Goldman Sachs
[73:18] finances it. Um, and as long as that RVG
[73:21] is actually less than your gross profit,
[73:23] you can't lose money and you have upside
[73:24] probably through a revenue share on top
[73:26] of it,
[73:27] >> then there are deals where you give
[73:30] warrants away, but they're tied to um
[73:33] like a fixed price per million tokens.
[73:35] And as long as the performance of your
[73:37] chip kind of outruns the performance of
[73:39] your stock,
[73:40] >> you're going to do good in that
[73:42] situation. If you just give warrants
[73:44] away, it could be negative NPV because
[73:46] the better the does the more value
[73:49] that's captured by the person. Yeah.
[73:50] >> Yeah. And so you can kind of look at
[73:52] that hierarchy of deals and like infer
[73:55] something about true customer
[73:56] preferences.
[73:57] >> Yes. That's interesting.
[73:58] >> Yeah.
[73:58] >> So Nvidia does pretty good deals.
[74:00] >> Uh like Yeah. I mean there's a reason
[74:03] that people I consider smart are
[74:06] investing in their deals.
[74:07] >> Yeah, I see it. Gavin, thank you. Fun.
[74:09] Always fun to hang with you.
[74:10] >> Thanks, David. This was great, man.
