Frontier AI labs monetize compute at an estimated $100 billion annualized run-rate revenue per gigawatt of inference capacity, yet persistent physical shortages of power, wafers, and copper make underbuilding far more likely than overcapacity through 2028. (5:00)
Data center infrastructure economics remain exceptionally strong, with 1 GW builds achieving 9-to-10-month contract paybacks and drawing substantial private debt financing from Blackstone, KKR, and Apollo against NVIDIA GPU assets. (10:00)
Long-term data center constraints may push balance-of-plant economics into orbit, where reusable SpaceX Starship launches could lower 1 GW deployment costs below terrestrial balance-of-plant equivalents by 2027–2028. (35:00)
Enterprise software architectures are migrating toward hybrid multi-model routing, using frontier closed models for reasoning and planning alongside fine-tuned open-source models for private, lower-cost execution. (50:00)
Detailed notes
## Background
Published by David George and Gavin Baker on RSS and web platforms on August 31, 2026, this episode features a16z's David George sitting down with Gavin Baker to examine the state of the AI boom. According to the show description, the discussion explores why demand for intelligence may be drastically underestimated, the non-zero-sum value capture across frontier labs and NVIDIA, compute payback economics, and the debate between overbuilding bubbles versus structural shortages.
## TL;DR
Gavin Baker and David George argue that the 21st century is fundamentally shaped by Elon Musk and Jensen Huang, with severe physical supply constraints—such as power, wafers, and copper—limiting AI rather than a lack of real demand. The discussion outlines how frontier labs and cloud builders achieve sub-one-year paybacks self-funded by massive operating cash flows, creating an expanding 'and' ecosystem rather than a winner-take-all monopoly. Looking ahead through 2028, the conversation covers the necessity of framing data centers as working-class engines to avoid political backlash, the emergence of orbital compute powered by SpaceX Starship, enterprise hybrid multi-model routing, and NVIDIA's durable, multi-chip architectural and financing moat.
## Atmosphere & style
The dialogue carries a rapid-fire, intensely analytical energy reminiscent of an insider briefing between two elite technology investors digging through capital expenditure models. Gavin Baker speaks with sweeping conviction and granular technical fluency, jumping effortlessly from balance-of-plant electrical math and private credit syndication to futuristic visions of orbital data centers and asteroid extraction. David George serves as a pragmatic, sharp sounding board, continuously testing the bull case against real LP anxiety and enterprise procurement realities to create a deeply compelling, techno-optimist tour de force.
## Bulletin
- `05:00` Frontier AI labs monetize compute at an estimated $100 billion annualized run-rate revenue per gigawatt of inference capacity, yet persistent physical shortages of power, wafers, and copper make underbuilding far more likely than overcapacity through 2028.
- `10:00` Data center infrastructure economics remain exceptionally strong, with 1 GW builds achieving 9-to-10-month contract paybacks and drawing substantial private debt financing from Blackstone, KKR, and Apollo against NVIDIA GPU assets.
- `35:00` Long-term data center constraints may push balance-of-plant economics into orbit, where reusable SpaceX Starship launches could lower 1 GW deployment costs below terrestrial balance-of-plant equivalents by 2027–2028.
- `50:00` Enterprise software architectures are migrating toward hybrid multi-model routing, using frontier closed models for reasoning and planning alongside fine-tuned open-source models for private, lower-cost execution.
## Key takeaways
- `00:00` The 21st-century technological shift is defined primarily by Elon Musk and Jensen Huang, with current platform risk driven by physical supply shortages rather than demand deterioration.
- `03:00` AI value capture is non-zero-sum, allowing frontier labs, open-source models, neoclouds, application layers, and NVIDIA to generate substantial value simultaneously.
- `05:00` Frontier labs will generate massive operating cash flows but negative free cash flow as all incremental capital is re-invested into GPUs, custom XPUs, and power capacity.
- `15:00` Current paying AI adoption rests on fewer than 10 million heavy users out of 1.5 billion global knowledge workers, leaving vast headroom for token demand growth.
- `20:00` Unlike historical debt-fueled speculative bubbles, hyperscalers are self-funding the vast majority of current AI infrastructure from operating cash flows.
- `22:50` Data centers serve as massive economic drivers for working-class trades and small-town municipal tax bases, making local economic benefits a far more effective public pitch than existential risk or China competition.
- `28:45` Supply chain allocations indicate zero uncommitted compute capacity through 2028, pointing to persistent intelligence shortages rather than an infrastructure glut.
- `35:00` Rapid drops in reusable Starship launch costs could make orbital compute balance-of-plant substantially cheaper than terrestrial data center buildouts constrained by grid and labor inflation.
- `40:00` xAI maintains a hedged compute strategy: internal usage absorbs capacity if first-party apps surge, while excess capacity can be sold into severe market shortages if overbuilt.
- `50:00` Enterprises are adopting multi-model architectures to avoid vendor lock-in and mitigate the data security risks of sharing proprietary business context with frontier labs.
- `57:30` To achieve enduring low-cost status, enterprise AI application and infra providers will ultimately need vertical integration and direct ownership of their compute layer.
- `01:00:00` NVIDIA's defensibility stems not merely from single accelerators, but from a tightly integrated eight-to-nine-chip system stack spanning switches, CPUs, DPUs, and interconnect fabrics.
- `01:02:44` Residual Value Guarantees provided by NVIDIA enable private equity debt financing that significantly lowers the cost of capital compared to rival ASIC deployments.
- `01:10:00` Rapidly diverging open-source model architectures favor general-purpose GPU clusters over prematurely specialized ASICs that risk obsolescence.
## Detailed walkthrough
### `00:00` The Era of Elon and Jensen, and Searching for Bearish Data
Reflecting on the scale of transformation underway, Gavin Baker frames the current technological inflection point in historic proportions: “When the history of the 21st century is written, you know, there's like the Victorian age, I think this will be like the age of Elon and Jensen, because they are fundamentally altering the fabric of human society and civilization.” While historical paradigm shifts inevitably spur overinvestment cycles and bubbles, the defining tension today is whether market skepticism overlooks an unprecedented physical and technological supply constraint. In response to political and social pushback surrounding massive power buildouts, Gavin emphasizes the domestic upside: “We are re-industrializing America and it's awesome.” Looking further out on the horizon of compute limits, he envisions physical expansion beyond Earth, noting that “an increasing fraction of the world's compute is going to be in orbit” and that asteroid mining will become an inevitable reality.
Returning to immediate market conditions, David George asks Gavin about his recent travels across the West Coast over the summer, where Gavin actively hunted for credible bearish signals to balance his view. Gavin reports coming up empty-handed: “My standard question is, can you tell me one quantitative data point in your business that's getting worse? Just one. That's my standard question. And it's at least in July and August, I haven't been able to find a single person.” While Anthropic appeared slightly muted—likely owing to a quiet period ahead of corporate milestones—the rest of the ecosystem surged throughout July and August. OpenAI accelerated, open source picked up even more velocity, and xAI’s Grok exhibited dramatic acceleration following the rollout of Grok 2 (GrokBot).
The divergence between operational reality and public markets is striking. Gavin points out that public equities have "fallen out of bed" over the prior two months, reminding listeners that “you can drown crossing a river that's on average two feet deep.” Although broad indices hide the damage, pure-play AI equities have suffered steep drawdowns despite underlying business fundamentals accelerating across the board.
- Historical precedent shows revolutionary technological shifts typically induce market bubbles, leading to overvaluation and eventual overbuilds, yet current constraints indicate compute shortages rather than excess.
- High-level index benchmarks mask severe drawdowns within key publicly traded AI beneficiaries.
- Operational acceleration in mid-2024 spanned proprietary frontier labs (OpenAI, Grok), open-source ecosystems, and enterprise pipelines.
---
`03:00` Non-Zero-Sum AI: The "And" World vs. The "Or" World
David references a podcast appearance by Eric Vishria with Patrick O'Shaughnessy proposing that the artificial intelligence landscape is inherently non-zero-sum. Rather than a solitary winner capturing all economic rent, value is distributed across the entire tech stack: proprietary frontier labs (Anthropic, OpenAI), vertically integrated platforms (SpaceX/xAI, Meta), hyperscalers deploying custom silicon (Google selling TPUs), open-source developers, specialized neoclouds, inference platforms, and well-executed end-user enterprise applications.
Addressing the noise around Anthropic, Gavin and David observe that the lab appears to have reset its revenue definitions and cleaned up accounting figures to produce apples-to-apples comparisons with OpenAI. The competitive dynamics between frontier labs resemble a high-stakes cat-and-mouse game. Frontier labs sit on advanced model checkpoints, strategically timing releases to overshadow rivals—such as Anthropic waiting for an OpenAI release (like Project Astra) only to publish Claude 3.5 Sonnet or Opus updates hours afterward.
```mermaid
flowchart TD
accTitle: Value Capture Across the Non-Zero-Sum AI Stack
accDescr: Diagram illustrating how value accrues across infrastructure, open source, frontier labs, and applications without a single winner-take-all outcome.
David notes that institutional limited partner (LP) discussions are invariably gripped by pessimism: “Every LP conversation that we have starts with like, how's this all going to go wrong?” LPs constantly search for failure points, wondering whether open-source models will destroy proprietary labs or whether foundational model commoditization will wipe out software moats. David rejects that premise outright: “I'm like, this is all wrong. This is not an or thing. It's an and thing, right? This is an and thing. Frontier is going to work really well. N minus one models are going to work really well. Open source is going to work really well. There's going to be a bunch of application companies that work really well. The clouds are probably going to be fine.” Gavin points out that this structural expansion leaves one primary beneficiary anchoring the foundation: “The last 26 years have taught me not to bet against Jensen.”
- Anthropic’s perceived lull reflects accounting true-ups, pre-IPO quiet period disciplines, and calculated pacing of model checkpoint releases.
- Rather than an "either/or" competitive displacement, value creation behaves as an "and" phenomenon spanning frontier labs, N-1 checkpoints, open source, and neocloud providers.
- NVIDIA remains structurally positioned in the center, capturing massive downstream demand across both open and closed model ecosystems.
---
`05:00` Frontier Lab Economics and the Training vs. Inference Tradeoff
Gavin and David explore the corporate cultures and capitalization strategies of frontier developers as they approach the public markets. Anthropic has increasingly emphasized a mission-driven approach, even screening candidates in culture interviews by asking how they would feel if their equity went to zero to weed out mercenaries. Gavin observes the practical flaw in this posture: “That's great. We want missionaries, but we also want people to make money. And at the end of the day, you can't afford the compute you want for your mission if you go—if the equity goes to zero.” While Anthropic has evolved into an "accidental enterprise company" purely as a byproduct of funding its mission, OpenAI and xAI operate with explicitly commercial orientations.
The conversation pivots to the unprecedented capital allocation mechanics facing these labs. Gavin illustrates the volatile tradeoff between serving revenue-generating inference and pursuing research breakthroughs in training:
> “Let's just say they have 10 gigs of power and they're allocating eight to inference. And let's just say they're monetizing that inference at whatever, 60 billion a year. So $480 billion a year in revenue. Which is like on a revenue payback basis would be like a one-year payback on a revenue basis, not a gross profit basis... Let's say they have a big research breakthrough and they decide, wow, it is to our long-term advantage to go from eight gigs allocated to inference to two gigs allocated to training to eight gigs on training. And then your revenue just went from 480 to 120 and, I think your annualized revenue, and I actually think they would do that.”
Public market equity investors accustomed to the smooth, steady margins of legacy internet giants like Google and Meta will face extreme whiplash. In traditional tech, infrastructure costs and revenue growth were decoupled; platforms did not need to switch off active user monetization to train their next algorithms. Frontier AI labs hold direct levers over their near-term revenue depending on which model checkpoints they publish, where they price along the Pareto frontier, and how many gigawatts of compute they divert away from inference back into training clusters. Consequently, while these labs will generate staggering operating cash flows, Gavin anticipates they will not produce positive free cash flow anytime soon, opting instead to funnel every dollar of operational cash flow and incremental profit back into acquiring GPUs, custom XPUs, and securing data center power capacity.
- Anthropic and OpenAI are estimated by industry observers to monetize compute at roughly $100 billion in annualized run-rate revenue per gigawatt of allocated inference capacity.
- Unlike traditional SaaS or internet platforms with decoupled infrastructure, frontier labs face zero-sum power allocations between active revenue generation (inference) and long-term capability jumps (training).
- Capital expenditures are projected to outpace operating cash flows for the foreseeable future, preventing near-term free cash flow generation as labs reinvest all available capital into compute infrastructure.
`10:00` Compute Allocation and the Economics of Sub-One-Year Paybacks
David George and Gavin Baker break down where capacity actually goes inside frontier AI labs: a vast portion of their power is tied up running their own heavily subsidized first-party products rather than pure pre-training. Gavin frames the breakdown across a typical large cluster footprint: roughly 8 gigawatts might go directly toward inference on first-party products, 2 gigawatts are dedicated to internal research, and another 2 gigawatts go to foundational pre-training (with post-training inference layered on top). Given the unshakable conviction around scaling laws, none of the frontier players are prioritizing immediate free cash flow generation. David points out how temporary hesitations have backfired: Satya Nadella appeared to blink around the time of the Davos interview where he defended Microsoft's $80 billion capex commitment, and Dario Amodei publicly warned that while losing share is bad, overspending into bankruptcy is worse. Anthropic took a more conservative stance, while OpenAI and xAI (Colossus) went all-in, rapidly winning back market initiative.
Turning to cloud infrastructure providers like Nebius and CoreWeave, Gavin highlights the unprecedented speed of capital payback. A new gigawatt build costing roughly $50 billion can often secure 50% to 60% in upfront customer prepayments, immediately reducing net outlays to around $25 billion to $30 billion. On contract pricing smoothed around $2 to $3, payback sits around 9 to 10 months; on spot markets earning $5 to $8, or inside xAI's rapidly deployed mega-clusters, capital returns inside of a single year. Furthermore, premier financial giants are willingly funding these deployments at attractive rates.
```mermaid
flowchart TD
accTitle: Frontier Lab Compute Allocation and Payback Cycle
accDescr: Flowchart illustrating the split of compute capacity across applications and the capital payback cycle for infrastructure providers.
subgraph Allocation["Compute Deployment Split"]
A["Total Lab Capacity"] --> B["8 GW: First-Party Inference"]
A --> C["2 GW: Internal Research"]
A --> D["2 GW: Pre-Training & Post-Training"]
end
subgraph Economics["Financing & Payback Dynamics"]
E["$50B / GW Initial Capex"] --> F["50%–60% Upfront Prepayments"]
F --> G["$20B–$25B Net Outlay"]
G --> H["Spot Monetization ($5–$8/hr)"]
H --> I["Payback: Sub-1 Year"]
end
```
- Upfront customer prepayments on new gigawatt buildouts typically absorb 50% to 60% of total capital expenditure upfront.
- Contract payback cycles calculate out to roughly 9 to 10 months, whereas aggressive spot monetization produces even faster paybacks.
- Major private equity and asset management players (Blackstone, KKR, and Apollo) are actively providing low-cost debt financing secured by GPU assets because useful economic lives continue to expand alongside escalating token ROI.
> “In my career as an investor, there haven't been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks.” — Gavin Baker
---
`15:00` Demand-Side Diffusion, Token Budgets, and the Grok Moment
Addressing concerns about an infrastructure bubble, David and Gavin examine the demand side, arguing that monetization to date rests on an astonishingly tiny fraction of the global workforce. David suspects that the heavy, paying user base delivering tens of billions in annualized revenue across the industry is not even 30 million people, but likely sub-10 million developers and super-users. Within venture-backed startups and enterprises, internal usage displays a severe power law: median software engineers use a modest amount, while elite engineering talent burns 10x to 100x more compute. Across the Andreessen Horowitz portfolio, tech-forward and AI-native startups regularly spend high single digits up to 10% or more of their total human payroll budget purely on model tokens each month, compared to legacy enterprises like banks sitting around 1%.
Gavin shares that Atreides Management saw its internal token consumption explode 100x between March and August. Following the rollout of Grok Enterprise, adding just two intensive users threatened to drive consumption up another 10x to 20x within a single month. Reflecting on user adoption, Gavin compares the generational shift to watching his parents adopt the iPhone and iPad: despite his own deliberate efforts to stay at the cutting edge at age 50, younger 23-year-old engineers operate natively in these tools. While Claude Code and Codex represented a major breakthrough by moving developers from writing 20% to over 90% of their codebase with AI, building bespoke podcast summarizers or custom trackers still required hours of tinkering. Grok, by contrast, generated superior, fully automated workflows in just 7 to 12 seconds. David emphasizes that this marks the critical transition from passive, reactive knowledge tools to proactive agentic systems that analyze the day's intelligence across bot fleets and directly recommend daily strategic actions.
- Enterprise token spend follows an extreme power-law distribution, with leading engineering teams allocating roughly 10% or more relative to payroll against 1% at traditional financial institutions.
- Out of an estimated 1.5 billion global knowledge workers, active heavy usage remains concentrated in an elite group likely smaller than 10 million individuals.
- Atreides expanded its token run rate 100-fold over a five-month span, driven by rapid automated synthesis and domain-specific bots.
> “So at Atreides, our internal token consumption has gone up 100x from the month of March. March through August, 100x our token spend.” — Gavin Baker
> “And now you have a Grok bot that says, 'What are the recommended actions based on everything the other bots have learned today? What recommendations do you have for me today?'” — David George
`20:00` Endless Token Consumption, Historical Bubbles, and Physical Bottlenecks
David George begins by describing the current state of personal AI usage: tools like Grok, Codex, and portfolio company Town operate in the background, analyzing workflows to propose actionable automations. While power users sit on the bleeding edge of adopting autonomous agent workflows, George notes that this marks only the beginning of a massive demand wave: once general users click "yes" to automate their entire workflow, the result will be an endless consumption of tokens.
```mermaid
flowchart LR
accTitle: AI Adoption and Supply-Side Constraints
accDescr: Diagram illustrating how rising user demand collides with physical and regulatory constraints to cause an intelligence shortage.
subgraph Demand ["Escalating Demand"]
A["Power Users<br>(Millions)"] --> B["Mass Adoption<br>(500M+ Users)"]
B --> C["Endless Token<br>Consumption"]
end
subgraph Bottlenecks ["Physical & Macro Bottlenecks"]
D["Watts & Wafers"]
E["Raw Inputs<br>(Copper & Power)"]
F["Capital & Regulation<br>(Higher Real Rates)"]
end
subgraph Outcome ["Market Dynamic"]
C --> G["Global Compute Shortage"]
Bottlenecks --> G
G --> H["Massive Undersupply<br>& Rising Intelligence Prices"]
end
```
Gavin Baker tempers this excitement by contextualizing it within market history. Drawing a distinction with early financial manias—noting an earlier podcast appearance where he erroneously linked the South Sea Bubble to the invention of marine chronometers for longitude when it was actually more of a speculative mania akin to tulip mania—Baker stresses that genuine transformative platforms invariably experience bubbles:
- Deeply transformative technologies (railroads, steel mills, automobiles, radio, television, personal computers, and the internet) inevitably spark speculative market run-ups.
- Asset overvaluation leads to physical overbuilding.
- The critical danger arises when buildouts are financed with debt rather than operating cash flows, because “debt-funded build-outs, they demand immediate ROI, not an ROI in two years.”
- Fortunately, the bulk of today's AI infrastructure investment is still being covered directly out of hyperscaler operating cash flows.
Beyond capital structure, Baker emphasizes that physical bottlenecks are acting as natural governors on the speed of the buildout. Pointing to discussions with Jamin Ball ("Jazz") and Patrick O'Shaughnessy, Baker cites "watts and wafers" as foundational limits that cap raw industrial capacity. The copper thesis highlights this scaling problem: if a few million current power users have already plunged the world into an acute global compute shortage, supporting 500 million users will demand a staggering number of new copper mines and power generation plants. These real-world constraints, combined with rising real interest rates and heavy-handed regulation, are actively slowing the rollout—a friction Baker ultimately views as healthy for society.
`22:50` The AI Public Relations Crisis and the Working-Class Case for Data Centers
Baker shifts his focus to the worsening political and regulatory climate in America, voicing frustration with how frontier lab leadership discusses artificial intelligence. Referencing an exchange on X with Dario Amodei and Sholto Douglas of Anthropic, Baker critiques the AI safety community's doom-and-gloom rhetoric. When Dario defended his public output as balanced because he had published one optimistic essay and one pessimistic essay, Baker countered that offering an existential, terrifying message 50% of the time—echoing figures like Eliezer Yudkowsky who claim everyone will die if advanced AI is built—poisons the public perception. Instead of abstract promises about curing cancer in hypothetical futures, the industry must produce concrete medical breakthroughs and tell a grounded, factual story. Baker highlights his favorite biblical proverb, arguing that “the truth shall set you free,” but insists that “the only group that can tell the AI industry's truth is the AI industry.”
Baker and George argue that the actual narrative of data centers is exceptionally positive for blue-collar Americans:
- Data center development represents one of the greatest economic engines for working-class citizens, creating lucrative career paths for electricians, plumbers, and HVAC technicians that out-earn NPV-negative four-year college degrees.
- Facilities that use behind-the-meter power generation transform municipal budgets, routinely multiplying small-town tax bases by 10x rather than merely doubling them.
- Environmental criticisms often rely on outdated claims: natural gas serves as a relatively clean transition fuel, and allegations regarding excessive freshwater consumption have been thoroughly debunked.
`25:00` Re-Industrializing America and Learning from Meta's Small-Business Playbook
George points out that geopolitical arguments for AI infrastructure—such as the urgent imperative to stay ahead of China—fail to resonate with everyday citizens because foreign invasion seems distant and abstract across the Pacific Ocean. Voters are instead focused on immediate economic concerns: local affordability and whether new infrastructure will improve their quality of life. Baker cites Loudoun County, Virginia, to demonstrate that high-density compute infrastructure and local prosperity go hand-in-hand:
| Region | Characteristic | Economic Outcome |
|:--- |:--- |:--- |
| **Loudoun County, VA** | Highest density of data centers in the United States | Highest per-capita income county/ZIP code in the U.S., driven by massive local tax receipts |
| **Rust Belt / Left-Behind Towns** | Former manufacturing and steel centers hollowed out by plant closures | Economic revitalization driven by high-paying infrastructure construction, maintenance, and power generation |
| **United States (Macro)** | High domestic natural gas abundance ($2–$3/MMBtu) vs. Europe/Asia ($20–$25/MMBtu) | Unprecedented industrial energy cost advantage fueling domestic manufacturing |
Baker warns that hostile rhetoric against domestic data centers is being actively fueled and laundered through platforms like TikTok via Chinese Communist Party-aligned influence operations. The reality on the ground is that cheap American natural gas combined with data center demand is actively re-industrializing the nation—a long-sought, bipartisan economic objective.
To correct this public narrative, Baker recommends that the leading companies in the ecosystem (including Meta, Google, OpenAI, Anthropic, SpaceX, NVIDIA, AMD, and Broadcom) emulate Sheryl Sandberg’s early quarterly earnings strategy at Meta. During Meta's initial public quarters, Sandberg consistently highlighted 10 to 15 specific, relatable small businesses—such as a Des Moines cake bakery run by two single mothers that expanded into 15 locations and 50 employees—to humanize the commercial value of their ad platform. AI labs and semiconductor providers must similarly share tangible human case studies instead of leaning solely on abstract national defense arguments.
`28:45` The Risk of Underbuilding Through 2028
George and Baker conclude by contrasting public panic over an AI speculative bubble with the harsh supply realities facing the sector. While market commentators consistently worry about overbuilding and excess capacity, every credible supply-chain forecast through 2028 shows virtually zero unallocated compute capacity.
Given mounting regulatory pushback, grid delays, and political friction, scheduled data center deliveries are far more likely to face severe delays than cancellation from lack of demand. As a result, the primary macro risk facing the market through 2028 is not an oversupply of compute, but a structural undersupply. In an environment where demand from hundreds of millions of users continues to outpace available supply, Baker and George agree that the ecosystem will likely face significant upward price increases for access to intelligence.
`30:00` Compute Scarcity, the Cost of Open Source, and the "Age of Elon and Jensen"
David George and Gavin Baker explore the counterintuitive dynamics of AI pricing and compute capacity. David brings up a thesis posed by Dwarkesh Patel that token costs might actually increase tenfold rather than collapse, running contrary to popular belief. Gavin agrees that in a supply-constrained environment with explosive demand, price spikes are entirely plausible because frontier models deliver immense user surplus, justifying premium pricing over cheaper models.
Gavin points out that if data center development is blocked by regulatory pushback or "de-growthers," the unintended consequence will be severe compute inequality where only the wealthiest individuals and tech giants can afford access. David notes that delivering cheap consumer services usually depends on advertising models that take years to monetize, making near-term scarcity even more painful.
Addressing open-source AI, Gavin pushes back on the common misconception that open weights equal free inference. It takes the same amount of physical compute to run an open-source token as a frontier token on an equivalent-sized model, and open weights can actually be less efficient on a per-task basis. Furthermore, licensing terms like Moonshot AI's Kimi stipulate a 30% revenue share.
Reflecting on the infrastructure giants driving this transition, Gavin praises Jensen Huang and Elon Musk as civilizational figures reshaping society:
> “I think, like, you know, kind of when the, when the history of the 21st century is written, you know, there was like the Victorian age. I think this will be like the age of Elon and Jensen.” — Gavin Baker
David agrees, emphasizing how SpaceX and Starlink unlocked consumer surplus by delivering satellite internet to underserved global regions where terrestrial economics never penciled out, establishing space-based infrastructure as a proven model.
---
`35:00` Orbital Compute Architecture and Terrestrial Cost Realities
Gavin and David dismantle common misconceptions about orbital data centers. Rather than resembling the "Death Star" or massive terrestrial warehouses floating in space, an orbital compute unit is roughly the size of an airplane: a central rack of chips flanked by solar array "wings" and a radiator kept in perpetual shadow within a sun-synchronous orbit.
```mermaid
flowchart LR
accTitle: Orbital Compute Unit Architecture
accDescr: Layout of an orbital compute satellite showing solar wings, central compute rack, and an unlit radiator in sun-synchronous orbit.
subgraph Orbit["Sun-Synchronous Orbit Architecture"]
Solar["Solar Wings<br/>(Continuous Power)"]
Rack["Compute Rack<br/>(e.g., NVL72 / Rubin)"]
Radiator["Thermal Radiator<br/>(Kept in Shadow)"]
end
Solar --> Rack
Rack --> Radiator
```
Gavin recounts debating a physics PhD investor friend who insisted orbital compute was physically impossible—until he met directly with SpaceX engineers and recanted. Gavin argues that outside skeptics, regardless of credentials, have rarely spent more than a few hours on the problem, whereas thousands of elite SpaceX engineers have treated thermal dissipation and orbital mechanics as a solved engineering challenge that is in many ways simpler than Starlink satellites with complex phased-array antennas.
The two investors compare the capital expenditure of building data centers on Earth versus launching them into space:
Gavin highlights that terrestrial balance-of-plant costs ($15 billion per gigawatt) will suffer severe inflation due to rising electrician wages, copper shortages, and optical cabling bottlenecks:
> “So what you have to compare it to is the cost of launch. And with Starship reusability, that goes to under a billion. So the economics just instantly flip.” — Gavin Baker
While speed-of-light latency limitations mean training clusters will remain grounded, Gavin predicts orbit will become critical swing capacity for inference. He references Elon Musk's timeline of co-designing a NVIDIA Rubin rack with Jensen Huang slated for launch in Q4 2027 (or 2028 allowing for delays). David notes that between orbital compute and Starlink Mobile tapping an $800–$900 billion wireless market alongside satellite broadband, SpaceX is addressing a nearly $2 trillion global market.
`40:00` xAI's "Heads You Win, Tails You Win" Compute Hedge and Global Starbase Expansion
David George and Gavin Baker look beyond theoretical orbital compute to examine the immediate terrestrial momentum behind xAI, Grok, GrokBot, and Cursor. Baker highlights how advertising telemetry on X is expanding and predicts that xAI will likely bundle services—combining Starlink, GrokBot, and X advertising packages in much the same way Google originally scaled its cloud business by bundling it with search ads. George notes that xAI finds itself in an enviable strategic position because its bets pay off either way:
> “I actually like the AI position that they're in because it's like heads you win, tails you win in the sense that their first party business is growing very fast and they caught up to the frontier like very quickly.”
```mermaid
flowchart TD
accTitle: xAI Dual-Track Compute Strategy
accDescr: Diagram showing how xAI wins whether first-party applications dominate or compute capacity is sold as cloud infrastructure.
subgraph Inputs["Aggressive Compute Investments"]
C["Massive GPU / Cluster Buildout"]
end
subgraph ScenarioA["Scenario A: First-Party Dominance"]
FP["Rapid First-Party Scaling<br>(Grok, GrokBot, X Platform)"]
BD["Bundling Engine<br>(Starlink + Grok + X Ads)"]
FP --> BD
end
subgraph ScenarioB["Scenario B: Overbuilt Capacity"]
EXT["External Cloud Infrastructure"]
REV["Monetize via Compute Scarcity"]
EXT --> REV
end
C -->|"Application Success"| FP
C -->|"Excess Compute"| EXT
```
If xAI overbuilds capacity beyond what is needed for internal training and inference, the extreme broader shortage of compute allows them to sell the excess at premium rates as cloud infrastructure. Baker contrasts this with an OpenAI or Anthropic maximalist view, pointing out that even if frontier labs develop internal silicon, they will not own reusable Starships or multi-pad spaceports. As terrestrial power and cooling costs inflate while Starship launch costs deflate, space-based infrastructure becomes a viable long-term hedge. The conversation shifts enthusiastically to SpaceX's infrastructure expansion, including Starbase Louisiana and future global spaceports across the Middle East, the least bureaucratic nations in Europe, and East Asia (Japan or South Korea), targeting roughly 5,000 launches per year. Baker distinguishes SpaceX's rapid turnaround—engineering launchpads to catch, re-stack, refuel, and re-launch Starships more than twice a day—from crude recovery experiments, such as a Chinese test using a wire-catch rig originally brainstormed on the SpaceX subreddit.
- **Enterprise Bundling Potential**: Packaging Starlink connectivity, GrokBot enterprise access, and X ad spend mirrors early hyperscaler adoption playbooks.
- **Starship Launch Cadence**: SpaceX is designing pads for an operational tempo exceeding two launches per day per pad.
- **Global Spaceport Footprint**: Projected spaceport development spans domestic hubs like Louisiana to international locations across Europe, the Middle East, and Asia.
---
`44:00` The $20 Trillion Company: Asteroid Mining, Orbital Industry, and Mars 2030s
George recalls an investor gathering where participants debated what business would become the world's first $10 trillion company. Baker’s response at the time cut through the speculation:
> “I have no idea, but I know which one's going to be the first $20 trillion company.”
When pressed on SpaceX's most futuristic economic driver over a ten-year horizon, Baker points to asteroid mining. Targeting celestial bodies like Asteroid 16 Psyche—which contains more gold, silver, and platinum than the entirety of Earth's crust—SpaceX could utilize Starship and lunar waystations to shepherd captured asteroids into stable geosynchronous orbits above isolated, uninhabited U.S.-controlled atolls in the Pacific. From there, autonomous Optimus humanoid robots would handle processing and extraction, dropping refined materials down to Earth virtually free of transit costs. Baker links this vision to Jeff Bezos’s longstanding thesis:
> “Jeff Bezos said something very interesting. He said, I think in the future... Earth is going to be zoned residential.”
Under this framework, heavily polluting resource extraction and heavy manufacturing migrate off-planet into the vacuum of space, eliminating industrial emissions while utilizing limitless spatial volume. Looking roughly eight years out, Baker predicts a fleet of modified Starships (Mars Colonial Transporters) touching down on Mars. Using cargo bays equipped like Pez dispensers, Starships will deploy Optimus robots carrying American flags, solar panel arrays, industrial batteries, compute racks, and Mars-orbiting Starlink satellites. The initial phase will stream high-resolution 4K feeds of robotic construction directly back to Earth before human crews arrive to establish permanent habitation.
---
`47:30` Microsoft’s Dilemma, Model Ensembles, and NVIDIA-Funded Open Source
The discussion turns to Big Tech’s positioning, comparing Apple's conservative foot-dragging with Microsoft’s enterprise-facing AI posture. Baker notes that Microsoft attempted to develop an in-house frontier foundation model and came up empty-handed:
> “Satya said, we're going to have our own models that are very competitive. I think he said that 18 months ago. They don't have their own models that are competitive.”
Despite that shortfall, Baker believes the macroeconomic landscape is turning in Microsoft’s favor because the frontier is shifting from monolithic black-box models toward multi-model ensembles along a Pareto efficiency curve. Major Global 1000 enterprises will rarely rely on a single model for every task; instead, they will route workflows through specialized architectures and top-tier open-source weights.
```mermaid
flowchart LR
accTitle: AI Supply Chain Dynamics Between Labs and Chipmakers
accDescr: Flowchart illustrating the competitive incentives pushing frontier labs into custom silicon and chipmakers into funding open-source frontier models.
subgraph FrontierLabs["Frontier AI Labs"]
ASIC["Design Proprietary ASICs"]
end
This dynamic sparks cross-industry friction: as frontier labs design proprietary ASICs to bypass hardware dependencies, chipmakers face strong incentives to retaliate by funding open-source model ecosystems directly. Baker observes that NVIDIA could effortlessly bankroll a $50 billion to $100 billion training run, citing tactical investments and acquisitions like Poolside to push open weights to frontier parity. Meanwhile, Google may be pivoting its compute monetization by renting external TPU capacity to generate massive cash flows while staying within striking distance of the frontier. Baker argues that commodity open-source models reaching near-frontier performance is the ultimate boon for Microsoft and horizontal application software providers, who can integrate cheap, world-class intelligence directly into their software suites without paying exorbitant rents to closed frontier labs.
`50:00` The Rise of Hybrid Multi-Model Architectures
Enterprises are increasingly moving toward multi-model and hybrid architectures rather than depending on a single closed provider. Gavin Baker explains that organizations can now take an open-source base model—noting that Nemotron historically functioned as a solid pre-trained model without heavy post-training—and customize it with reinforcement learning (RL) and supervised fine-tuning (SFT) on their own proprietary data. Sharing sensitive enterprise context directly with a frontier lab can be dangerous, especially given shifts in Zero Data Retention (ZDR) policies. David George agrees: “And sharing that with a frontier lab, that may be hazardous for your financial health.” Gavin emphasizes that by owning the post-trained model, companies protect their IP and control their capabilities and operating costs.
A --> B
B --> C
B --> D
B --> E
C --> F
D --> F
E --> F
```
Real-world systems are already converging on routed setups. Gavin points to Grok (and setups utilizing mixes of models like Gemini 3.7 Flash, Grok, and Claude Opus behind an intelligent router), while David points out that systems can use the most capable frontier model for high-level planning while routing execution to cheaper models. Gavin notes this shifting competitive landscape creates a much friendlier environment for players like Microsoft compared to a world governed by a duopoly. Meta has mounted an impressive comeback into frontier relevance, while Google’s Gemini has suffered narrative and capability swings. Gavin highlights Fireworks AI's Nexus product as a prime example of this emerging model:
- Fireworks AI’s CEO Lin Qiao was among the first to articulate this setup: letting an enterprise bring an open-source model, running RL on internal data (for clients like Goldman Sachs, Morgan Stanley, JPMorgan, Fidelity, or a16z), and placing it behind a seamless router alongside frontier options.
- Tech leaders like Satya Nadella (referring to "specialized intelligence") and Alex Karp at Palantir have adopted their own variations of this multi-model paradigm.
`54:15` The Battle to Become the Ultimate Abstraction Layer
David and Gavin debate who will capture the coveted position of being the default abstraction layer between users and intelligence. Gavin uses an analogy from his days as a retail analyst to demonstrate that concepts sounding simple on paper are brutal in execution. In retail, anyone can theoretically build a $50 billion business: all you need is a thousand stores across 50 states with different climates and tastes, stocked with the right products at the right prices, staffed by friendly employees who don't steal from you and turn over at 100% per year, kept clean and well-lit. Yet in American history, only a handful of retail chains have pulled it off.
Building a corporate AI abstraction layer presents a similar challenge. Gavin praises Cursor’s pragmatism compared to the quasi-religious tone of traditional labs: “And the Cursor guys were just like, we want to make great product.” David concurs, pointing out that Cursor shared a long-term vision of autonomy with the labs but chose an engineering-first path, meeting users where they are rather than preaching about digital deities. Coding emerged as the primary beachhead because it is verifiable and perfectly documented—unlike the vast majority of messy enterprise knowledge work.
- Coding provides clean synthetic environments, giving tooling providers a clear path to early product-market fit and measurable autonomy.
- The messy reality of unorganized enterprise data makes providing a turnkey, trusted abstraction layer far more complex than simple middleware.
`57:30` Competitive Collision: Vertical Integration and Cost Advantage
A massive collision is brewing across the software stack. Enterprise software giants, inference providers, and vertical agent companies are all attempting to capture the same territory. Specialized vertical applications like Harvey in legal are demonstrating how domain-specific abstraction layers take root, benefiting from legal work being highly documented and somewhat verifiable (with tax expected to follow).
Validation of the category's immense size is visible in corporate spend: law firm Kirkland & Ellis committed $500 million to build internal AI infrastructure. Gavin observes that while such an effort is daunting—because an AI stack is not a one-time build, but requires continuous model swaps, routing, and post-training—it underscores the massive TAM. Looking ahead, players across the entire ecosystem will collide:
- **Contenders**: Middleware/inference platforms (Fireworks Nexus), vertical agents (Harvey, Cognition), data platforms (Databricks, Snowflake), and enterprise SaaS incumbents (Microsoft, Palantir, Salesforce, Workday).
- **The Deciding Moat**: Long-term winners will need rock-bottom operating costs. Gavin asserts: “Because, yeah, you're just simply not going to be the low-cost provider if you're not vertically integrated if you don't own your own compute over the very long term.”
`01:00:00` Valuing Hyperscalers and Jensen Huang’s Dominance
David George and Gavin Baker open by examining how to properly value hyperscalers, proposing an enterprise value (EV) to net PP&E (property, plant, and equipment) metric. Gavin points out that net PP&E represents compute capacity, and looking at EV relative to net PP&E effectively prices what the market expects a company to earn on its fleet of hardware—acting like an AI-native version of price-to-book value. Gavin highlights how this exposes glaring valuation inefficiencies across the public hyperscalers.
The conversation naturally shifts to NVIDIA's CEO, Jensen Huang. Gavin asserts that Jensen is effectively carrying the entire industry forward through a strategy of vertical integration paired with horizontal openness. For any aspiring semiconductor startup or accelerator company, Gavin offers blunt strategic advice: do not attempt to fight NVIDIA head-on. With every 1% of the accelerator market likely worth $100 billion, niche enablement is vastly superior to suicidal competition. NVIDIA commands an ecosystem encompassing at least nine distinct chips—including diverse accelerator flavors, CPUs, Ethernet switches, two kinds of DPUs, and architectures spanning scale-up, scale-out, scale-across, and now scale-in networking.
```mermaid
flowchart TD
accTitle: The NVIDIA Compute Ecosystem and Strategy
accDescr: Visualizing NVIDIA's multi-chip architecture and why hardware startups must integrate rather than compete directly.
subgraph Fabric ["Networking Topologies"]
E["Scale-Up"]
F["Scale-Out"]
G["Scale-Across"]
H["Scale-In"]
end
subgraph Market ["Market Strategy for New Entrants"]
I["Avoid Head-on Battle with Jensen"]
J["Capture 1% Niche Share (~$100B Value)"]
K["Integrate Horizontally into the Ecosystem"]
end
Hardware --> Fabric
Fabric --> Market
```
Gavin draws an analogy to the 1990s Chicago Bulls to illustrate why semiconductor founders should never antagonize Jensen Huang publicly. He recalls regular-season games around game 50 where the Bulls were up eight games in the standings and Michael Jordan was visibly bored—until an inexperienced opposing player started talking trash. Jordan’s eyes would light up, and he would ruthlessly dismantle them. “My number one thing is if you're a semiconductor CEO, the only thing you should ever say is thank you, Jensen,” Gavin argues, adding that the winning playbook is to express gratitude and find ways to co-exist along the edges rather than waking the giant.
- Gavin's rule of thumb: Every 1% market share in AI accelerators represents roughly $100 billion in enterprise value.
- NVIDIA's silicon suite encompasses accelerators, CPUs, Ethernet switches, and two types of DPUs across scale-up, scale-out, scale-across, and scale-in fabrics.
---
`01:02:44` The Financeability of NVIDIA Data Centers and Residual Value Guarantees
Gavin explains that NVIDIA's greatest structural moat is not merely chip performance, but financeability and the cost of capital. Building a hypothetical $50 billion NVIDIA-powered data center requires roughly a $15 billion equity check, with the remaining $35 billion easily financed through top-tier debt providers. Institutional giants like Blackstone, KKR, and Apollo are actively underwriting these massive debt packages because they trust NVIDIA hardware liquidity and residual value.
Crucial to this structure are Residual Value Guarantees (RVGs). When Jensen provides an RVG, the guaranteed downside is significantly lower than the gross profit dollars NVIDIA captures upfront from selling the silicon into the facility. This creates an overwhelmingly net-present-value (NPV) positive transaction with negligible downside for NVIDIA, on top of which the company often secures an ongoing revenue share. In contrast, alternative custom silicon like Google's TPU is substantially harder to debt-finance; Gavin estimates a TPU cluster requires at least double the equity check along with noticeably higher borrowing rates.
Furthermore, Gavin notes that Jensen is actively assembling land, power, and shell companies, matchmaking them with compute offtake agreements. By underwriting these RVGs, Jensen purposefully props up alternatives and neoclouds to ensure the industry is not monopolized by OpenAI and Anthropic. “And it's not circular financing. I have a lot of respect for the people I have met from Blackstone and KKR and Apollo,” Gavin emphasizes, showing how disciplined private credit enables NVIDIA's massive deployment machine.
- Financing structure for a $50B NVIDIA cluster: approximately $15B in equity, with $35B raised as debt backed by private credit (Blackstone, KKR, Apollo).
- TPU-based clusters require at least double the equity contribution and endure higher interest rates due to lower secondary-market liquidity.
- Jensen's matchmaking engine pairs real estate, power infrastructure, and shell companies with compute offtake agreements to democratize compute access.
---
`01:05:00` Open Source, Supply Chain Dominance, and the Brutal Reality of Hardware
Gavin pushes back against the common bear narrative that open-source models threaten NVIDIA's business model. On the contrary, Jensen is one of open source's fiercest champions because lower software margins drive token consumption to explosive levels. If a proprietary model charges a 90% software gross margin per token and an open-source model drops that margin to 40%, aggregate token demand explodes—requiring dramatically more NVIDIA compute in a supply-constrained environment.
NVIDIA reinforces this dynamic by locking down an estimated 70% to 80% of the planetary supply chain needed to build AI infrastructure. Gavin details how Jensen locked up not only TSMC fabrication capacity, but also DRAM, NAND, high-frequency lasers, specialized capacitors, and rack assembly components. Fifteen years ago, Jensen was making bold $2 billion to $3 billion bets every two years; today, he makes multi-hundred-billion-dollar bets that drag the global supply chain and Wall Street's capital markets alongside him.
Reflecting on his own investments at Atreides, Gavin shares the painful lessons of venture semiconductor investing. Even when backing an elite team that runs flawless emulations and tapes out a chip quickly, physical silicon often fails when it returns from the fab. He references Elon Musk's observation about hardware humility and contrasts it with startups like Cerebras, whose chips worked technically but struggled through two generations to find product-market fit before executing successfully. Having a chip that completely fails to function can set a company back hundreds of millions of dollars and two years—if the startup can even survive the financing drought.
Gavin praises Elon Musk's decision to lean heavily into NVIDIA hardware as a high-ELO strategic play, rather than risking unproven custom silicon. While acknowledging that specialized internal efforts like the Jalapeno ASIC show impressive progress for in-house teams who understand frontier model research trajectories, Gavin concludes that NVIDIA remains unmatched because it designs for the entire world's workloads simultaneously. “Elon said a lot of people are going to learn a hard lesson in hardware. And like, I would just say I've learned a lot of hard lessons in semiconductor investing,” Gavin remarks, underlining why NVIDIA's operational scale and hardware execution remain nearly impossible to duplicate.
- NVIDIA controls an estimated 70% to 80% of critical worldwide compute hardware supply, spanning fabs, DRAM, NAND, optics/lasers, and capacitors.
- Compressed model margins (e.g., drops from 90% to 40% margins) induce higher token volume, which directly translates to higher compute consumption.
- Cerebras chips functioned technically across their initial iterations, but required multiple generations to lock down genuine product-market fit.
`01:10:00` Model Divergence and Respecting "Superman" in Semiconductors
The discussion opens with the misconception that artificial intelligence architectures are rapidly consolidating toward a single, uniform standard. David George and Gavin Baker point out that looking across leading frontier models reveals substantial architectural variation rather than uniformity. In particular, the major open-source Chinese models—DeepSeek, Kimi (Moonshot AI), and Qwen (Alibaba)—are all evolving along noticeably different trajectories. While a general-purpose processor like a GPU can readily support this widening breadth of model evolutions, trying to specialize custom silicon too early creates major vulnerabilities, as general-purpose compute remains essential to keep up with the moving target of AI research.
Turning to NVIDIA and Jensen Huang, Baker remarks on how well the company's executive incentives align with broader national and economic interests:
> “I’m very happy his incentives as a CEO are perfectly aligned with what’s good for America.”
The conversation shifts to an analogy about competitors underestimating Jensen Huang's dominance, comparing him to Michael Jordan and Superman. Baker warns that semiconductor teams often grow overconfident after a single success and make the mistake of trash-talking:
> “You know, sometimes it's like, you know, you tug on Superman's cape and you get confident.”
They observe that Google’s TPU team previously fell into this trap. While newer challenger chips (such as internal ASIC efforts or specialized silicon) have built hardware that is legitimately competitive with a single NVIDIA chip, they are only rivaling one component out of a tightly integrated eight- or nine-chip system stack. Citing Dylan Patel from SemiAnalysis, Baker highlights that Jensen functions essentially as the central banker and Federal Reserve of the artificial intelligence ecosystem. In that light, strategic players like Elon Musk made the right calculation by cooperating closely with NVIDIA rather than attempting a head-on confrontation.
```mermaid
flowchart TD
accTitle: The AI System Advantage
accDescr: Diagram illustrating why competing with NVIDIA at a single chip level falls short of system-level integration.
subgraph Competitors["Challenger Approach"]
A["Custom Accelerator / ASIC"]
end
subgraph NvidiaSystem["NVIDIA Integrated System Stack"]
B["Flagship GPU"]
C["Networking / NVLink / InfiniBand"]
D["Switch Silicon / NVSwitch"]
E["CPUs / Superchips"]
F["System-Level Interconnect & Software (CUDA)"]
B --- C
C --- D
D --- E
E --- F
end
A -.->|"Matches only 1 of 8-9 system components"| B
```
`01:12:15` Inferring Customer Preferences Through Financing Deals and Outro
In a severely supply-constrained semiconductor market, determining organic customer preferences is exceptionally difficult because demand is so high that buyers will take whatever allocation is available. If a company secures TSMC wafer allocation and the necessary high-bandwidth memory (DRAM/HBM) to pair with it, its inventory will sell out regardless of vintage, keeping older chips like the H100 priced at steep premiums.
Baker argues that the clearest signal of authentic market preference is not raw unit sales, but the structure of the financial deals chip vendors must cut to secure adoption:
> “In a world that is so supply chain constrained, it's actually really hard to tell what true customer preferences are.”
To break down these dynamics, Baker outlines a clear hierarchy of deal structures across the industry:
* **Direct Vendor Investment (Overcoming the Cold-Start Problem):** Cloud providers invest directly into the customer to mandate chip usage, seen when Google (TPU) and Amazon (Trainium) invested in Anthropic. As long as the cash invested is lower than the gross profit generated by the compute commitment, the vendor secures an economic return while forcing real-world usage to mature the chip.
* **Residual Value Guarantee (RVG) Financing:** Deals structured with major private credit and private equity firms (such as Blackstone, Apollo, KKR, or Goldman Sachs). If the guaranteed residual value remains below the vendor's gross profit, the chip provider is protected against downside while retaining upside via revenue-sharing arrangements.
* **Warrant Incentives Tied to Pricing:** Contracts where equity warrants are granted to customers alongside fixed compute prices per million tokens. If chip performance improvements outpace stock appreciation, the vendor captures net value.
* **Uncapped Equity Warrants:** Deals that grant outright warrants with no performance hedge, which can result in a deeply negative net present value (NPV) because rising company valuations simply hand away equity value to win basic hardware adoption.
Because NVIDIA commands authentic market pull, it does not need to resort to value-dilutive concessions, and smart capital continues to invest behind its ecosystem.
David George closes the conversation by thanking Gavin Baker for the detailed discussion. The episode concludes with the standard podcast outro, directing listeners to subscribe on Apple Podcasts, Spotify, X (@a16z), and Substack, followed by the formal Andreessen Horowitz compliance disclosure clarifying that the episode is purely informational and does not constitute investment, tax, or legal advice.
## Notable quotes
> `10:00` In my career as an investor, there haven't been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks. — Gavin Baker
> `00:00` When the history of the 21st century is written, you know, there's like the Victorian age, I think this will be like the age of Elon and Jensen, because they are fundamentally altering the fabric of human society and civilization. — Gavin Baker
> `20:00` Debt-funded build-outs, they demand immediate ROI, not an ROI in two years. — Gavin Baker
> `03:00` The last 26 years have taught me not to bet against Jensen. — Gavin Baker
> `01:12:15` In a world that is so supply chain constrained, it's actually really hard to tell what true customer preferences are. — Gavin Baker
> `00:00` An increasing fraction of the world's compute is going to be in orbit. — Gavin Baker
> `22:50` The only group that can tell the AI industry's truth is the AI industry. — Gavin Baker
## Mentioned
David George, Gavin Baker, Elon Musk, Jensen Huang, Eric Vishria, Patrick O'Shaughnessy, Satya Nadella, Dario Amodei, Jamin Ball, Sholto Douglas, Eliezer Yudkowsky, Sheryl Sandberg, Dwarkesh Patel, Jeff Bezos, Lin Qiao, Alex Karp, Dylan Patel, OpenAI, Anthropic, xAI, Grok, Grok 2, GrokBot, Grok Enterprise, Claude 3.5 Sonnet, Claude Opus, Claude Code, Codex, Project Astra, Meta, Llama, Google, TPU, SpaceX, Starlink, Starlink Mobile, Starship, Mars Colonial Transporter, NVIDIA, NVL72, Rubin, AMD, Broadcom, CoreWeave, Nebius, Colossus, Blackstone, KKR, Apollo, Atreides Management, Town, Cursor, Kimi, Moonshot AI, Loudoun County, Virginia, Asteroid 16 Psyche, Optimus, Microsoft, Apple, Poolside, Nemotron, Fireworks AI, Nexus, Goldman Sachs, Morgan Stanley, JPMorgan, Fidelity, a16z, Palantir, Gemini 3.7 Flash, Harvey, Kirkland & Ellis, Cognition, Databricks, Snowflake, Salesforce, Workday, TSMC, Cerebras, Jalapeno ASIC, DeepSeek, Alibaba, Qwen, SemiAnalysis, Amazon, Trainium
## Video description
a16z’s David George sits down with Gavin Baker to unpack the state of the AI boom, why demand for intelligence may still be dramatically underestimated, and why the outcome doesn't necessarily have to be winner-take-all.
David and Gavin explore the possibility that frontier labs, open-source models, applications, clouds, and NVIDIA can all capture significant value as AI adoption expands. They dig into the economics of the infrastructure buildout, why compute investments can have unusually fast payback periods, and what happens when today's relatively small group of heavy AI users expands to hundreds of millions of people.
They also debate the risk of an AI bubble versus an AI shortage, the backlash against data centers, orbital compute, the rise of multi-model architectures, and NVIDIA's position at the center of the AI supply chain. Gavin makes the case that the AI buildout could help reindustrialize America, while David explores whether the bigger near-term risk is not overbuilding, but failing to build enough.
Gavin Baker and David George argue that the artificial intelligence sector is characterized by severe physical and capital undersupply rather than an impending speculative bubble, creating sub-one-year infrastructure paybacks and reinforcing NVIDIA's central market position.
Severe Compute Constraints and Frontier Monetization (5:00)
Despite steep drawdowns in public pure-play AI equities, Baker finds operational metrics across frontier labs accelerating rather than deteriorating. High monetization rates—estimated by market observers at roughly $100 billion in annualized run-rate revenue per gigawatt of allocated inference—mean providers can recover capital quickly. New gigawatt buildouts costing $50 billion frequently secure 50% to 60% in upfront customer prepayments, which cuts net outlays to $20 billion to $25 billion and enables contract payback cycles within 9 to 10 months.
Frontier developers face an acute trade-off between near-term revenue and model capability. Unlike traditional software platforms that operate with decoupled infrastructure, AI labs must choose between allocating limited gigawatts of compute to active inference or diverting capacity back into model training. Baker expects labs to operate without positive free cash flow over the near term, opting instead to funnel all operating cash flows back into purchasing hardware and locking down power capacity.
Physical Governors and American Industrial Benefits (20:00)
Baker and George contend that historical technological transformations inevitably invite speculative capital, but physical constraints—namely available power ('watts') and silicon manufacturing ('wafers')—act as a natural governor against runaway overbuilding. In an environment where heavy internal usage is concentrated among fewer than 10 million core software engineers and power users, widening adoption to the broader global workforce will intensify existing shortages in copper, power generation, and electrical grid capacity, leaving little unallocated compute capacity through 2028.
In response to mounting public and regulatory pushback against data centers, Baker points to Loudoun County, Virginia, as evidence that concentrated compute infrastructure drives local economic prosperity. Data center projects generate high-earning trade jobs for electricians and HVAC technicians while significantly expanding municipal tax bases. Moreover, Baker and George note that the United States holds a distinct competitive edge through cheap domestic natural gas ($2 to $3 per MMBtu versus $20 to $25 in Europe and Asia), serving as an engine to re-industrialize domestic manufacturing.
Terrestrial Inflation and Orbital Compute Economics (35:00)
As terrestrial data center balance-of-plant expenses—such as land, power, and labor—swell to roughly $15 billion per gigawatt, Baker anticipates that orbital infrastructure will become economically competitive for inference workloads. Under a satellite configuration operating in sun-synchronous orbit, IT rack hardware remains comparable to terrestrial equivalents at roughly $35 billion per gigawatt, while balance-of-plant costs are eliminated via solar arrays and vacuum radiative cooling. With reusable launch platforms like SpaceX's Starship targeting launch expenses under $1 billion per gigawatt, total orbital deployment costs could fall below terrestrial equivalents.
Baker highlights Elon Musk's timeline targeting the orbital launch of an NVIDIA Rubin compute rack by late 2027 or 2028. While light-speed latency constraints dictate that complex model training runs will remain grounded on Earth, orbital deployments are positioned to serve as vital swing capacity for global inference demand.
Multi-Model Ensembles and Enterprise Software Layers (50:00)
Enterprises are actively pivoting away from single closed-model dependencies in favor of multi-model routed architectures. By combining fine-tuned open-weight models trained on proprietary corporate data with an orchestration layer that delegates planning to frontier models and routine tasks to cheaper models, enterprises protect internal intellectual property and manage token costs. George and Baker point to tools like Cursor and Fireworks AI's Nexus as early examples of practical abstraction layers meeting developers within verifiable coding workflows.
This shift lowers barriers for application vendors and incumbents such as Microsoft, Palantir, and Salesforce. However, Baker stresses that as vertical agents, data platforms, and inference middleware converge into direct competition, sustained long-term defensibility will require vertical integration and low-cost access to compute.
NVIDIA's Systemic Moat and Capital Advantages (62:44)
NVIDIA maintains an structural advantage that extends far beyond individual chip performance. Baker highlights that the company delivers a full nine-chip system stack spanning CPUs, Ethernet switches, DPUs, and specialized networking fabrics. Controlling an estimated 70% to 80% of critical worldwide hardware components—from TSMC wafers to optics, lasers, and capacitors—NVIDIA enables private credit lenders like Blackstone, KKR, and Apollo to easily finance its clusters with modest equity cushions.
Customer deal structures reveal underlying market preferences amid supply shortages. While proprietary custom silicon like Google's TPU or Amazon's Trainium often requires direct vendor equity investments or equity warrants to encourage enterprise migration, NVIDIA's hardware enjoys robust debt financing supported by high secondary-market liquidity and low-risk Residual Value Guarantees.
The conversation concludes that the primary market risk through 2028 is not an oversupply of compute infrastructure, but rather a persistent physical undersupply that will likely trigger upward price spikes for intelligence access.
[10:00] Gavin Baker: In my career as an investor, there haven't been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub-one-year paybacks.
[0:00] Gavin Baker: When the history of the 21st century is written, you know, there's like the Victorian age, I think this will be like the age of Elon and Jensen, because they are fundamentally altering the fabric of human society and civilization.
[20:00] Gavin Baker: Debt-funded build-outs, they demand immediate ROI, not an ROI in two years.
[3:00] Gavin Baker: The last 26 years have taught me not to bet against Jensen.
[72:15] Gavin Baker: In a world that is so supply chain constrained, it's actually really hard to tell what true customer preferences are.
[0:00] Gavin Baker: An increasing fraction of the world's compute is going to be in orbit.
[22:50] Gavin Baker: The only group that can tell the AI industry's truth is the AI industry.
Original transcript
[0:00] When the history of the 21st century is written, you know, there's like the Victorian age, I think this will be like the age of Ilan and Jensen, because they are fundamentally altering the fabric of human society and civilization. What happens if there's like a massive supply shortage? Every time you've had a real profound new technology, you get a bubble because the markets get really excited and they get ahead of themselves, things get overvalued, that overvaluation leads to an overbuild. One of the things that I think has been is this idea that we need to stay ahead of China. You're opposed to data centers. Well, you know what? It's probably the best thing that has ever happened to working class Americans. We are re-industrializing America and it's awesome. Assume that you're right. There's not a physics reason why this can't work. An increasing fraction of the world's compute is going to be in orbit. This sounds crazy, but asteroid mining is going to be a very real thing. Every LP conversation that we have starts with like, how's this all going to go wrong? and chip companies can all win, and what today's compute economics tell us about the sustainability of the build-out. They also tackle the bubble question head-on. Every major technology shift has produced overinvestment at some point. But with compute already constrained, and AI usage still concentrated among a relatively small number of people, what happens when that demand spreads across the broader economy? From data centers and re-industrialization to orbital compute, open source, and NVIDIA, What happens if AI demand keeps outrunning supply? Gavin, you've been out here hanging out on the West Coast over the summer, and you've been talking about the fact that you're like trying to find someone to give you a bearish case to make your sentiment more negative. Have you found anybody? No, and I ask everyone. My standard question is, can you tell me one quantitative data point in your business that's getting worse? Just one. That's my standard question. And it's at least in July and August, I haven't been able to find a single person. Now, if we're being honest, you know, Anthropic is different. in a quiet period, so maybe they've slowed down a little bit. But I do think the rest of the world has... accelerated. OpenAI is clearly accelerated. Open source, I think, has accelerated more. And then I do think Grok, particularly after GrokBot, has had a pretty dramatic acceleration. And so AI overall, it accelerated in July. It accelerated in August and it can't keep accelerating forever, but it's just kind of wild that public stocks have kind of fallen out of bed over the last two months. And I mean – And you can drown crossing a river that's on average two feet deep. And so there's not a lot of action at the index level. But some of these AI names are in pretty significant drawdowns. And they bounced a little bit in August, but still pretty big drawdowns. And things are broadly accelerating. Yeah. Our friend Eric Vichery did a podcast with Patrick O'Shaughnessy, and he said... Maybe everyone wins. Anthropic wins. OpenAI wins. SpaceX wins. Meta wins. Google wins by selling a lot of TPUs. Open source wins. Neo clouds win. The inference clouds win on top of the Neo clouds. Applications win. Yeah. Maybe not all applications. Applications that I think execute well and navigate this are... But that feels like a very possible scenario to me. And there's so much zero-sum thinking in the world. And by the way, on Anthropic, what is... My hypothesis would be, if you're anthropic, one, I think they probably trued up and cleaned up some accounting. Yes, definitely. You'd rather do that. Yes. So you rebased, and now you're comparable to OpenAI. Yeah, in terms of revenue added, in terms of the definition, and now I think kind of revenue added. Exactly. So you kind of rebased. And then they did their testing the waters. And then I would hypothesize because they've executed well, probably the next disclosure is a reacceleration. And then there's always this kind of funny game between the frontier model companies. They always have more advanced checkpoints. Anthropic is clearly waiting for OpenAI to release Astra. Yes. And then it's like the next day, here's Fable 5.1. Yes, exactly. Magically, it just happened to be available soon.
[5:00] Several hours after Astra. Yeah. So I think they're being thoughtful in heading into this IPO, and everyone is shooting at them. Yes. Everybody's shooting at them, and they're in a quiet period, so they can't really shoot back. So there's a lot of gamesmanship. But I do think, having OpenAI Ananthropic be public companies, it's going to be helpful for the market just because it's You know, it's such a powerful force and a lot of public investors, you hear, oh, Sarah Fryer said this at an all hands meeting and it's on the cover of Wall Street Journal. OK, we're going to put that into our model. Yeah. And it's just I think it'll be better for them to be public. I am a little, you know, Anthropic is now in their culture interviews saying, how would you feel if the equity went to zero? Yeah. Because we're looking for people who are mission aligned. Yeah, mission, not mercenary. Yeah. And that's great. We want missionaries, but we also want people to make money. And at the end of the day, you can't afford the compute you want for your mission if you go if the equity goes to zero. Yeah. Like I'm no expert, but I'm pretty sure on that. And then I do think they are. They're like the accidental enterprise company. Oh, for sure. Oh, yeah. They're kind of like the accidental everything. Enterprise is just a byproduct of the mission, the objective. Yeah. Yeah. Whereas I think OpenAI is a little more commercial and obviously SpaceX is a little more commercial. But all of these companies, let's just say they have 10 gigs of power. and They're allocating money. eight to inference. And let's just say they're monetizing that inference at whatever, 60 billion a year. So $480 billion a year in revenue. Which is like on a revenue payback basis would be like a one-year payback on a revenue basis, not a gross profit basis. Yeah, on a revenue basis. Yeah. And I'm trying to use conservative numbers. People seem to think Anthropic and OpenAI are both monetizing at $100 billion a gigawatt today. Yeah, yeah. Let's say they have a big research breakthrough and they decide, wow, it is to our long-term advantage to go from eight gigs allocated to inference to... two gigs allocated to training to eight gigs on training. And then your revenue just went from 480 to 120 and, I think your annualized revenue, and I actually think they would do that. They would make that decision, yeah. N. This is just something that public markets are going to really have to get used to. Yeah. As you say, OpenAI may be a different animal. And I do think the realities, you know, everybody has these ideals about how they're going to manage their business. Then they go public and the stock is volatile and it really impacts employee morale, recruiting, retention. So I'd be surprised if they did such a dramatic cut. But a lot of the revenue is kind of under their control based on what checkpoint they release. Yeah. where they price along this kind of Pareto curve, and then how much they allocate between training and inference. So it's going to be meta and Google and these kind of internet companies. It was just, it was pretty smooth fundamentally, even if the stocks were volatile. Well, there was no massive trade-off they had to make in terms of the cost or infrastructure to serve revenue side. They were totally separate. 100%. Yeah. It's fascinating. So if you go back to Eric's point of it's all going to work, I describe it differently. I've had this conversation with LPs a lot because every LP conversation that we have, it's probably the same for you, starts with how's this all going to go wrong? And it's like, what's going to crash? And I'm like, oh, are the large models screwed or the labs screwed because of open source? And I'm like, this is all wrong. This is not an or thing. It's an and thing, right? This is an and thing. Frontier is going to work really well. N minus one models are going to work really well. Open source is going to work really well. There's going to be a bunch of application companies that work really well. The clouds are probably going to be fine. They're probably going to work really well. The five lab companies are probably going to do really well. Yeah. And NVIDIA is at the center of all of it. Yes. Yes. They're probably going to do pretty well. Yeah. The last 26 years have taught me not to bet against Jensen. Yeah. He's in a pretty good position here. I want to come back to that. The point that you made about training versus inference is an interesting one. It seems to me like the labs will decide... to take all incremental profits. and probably much more than their profits. and invest them in training for a long period of time. Would you think that's This is very different than the clouds. Because the internet companies in the clouds, they just end up being supply-demand driven, and they generate tons of profit, and they can still grow a certain amount. But they don't have some, maybe with the exception of meta, some big long-term bet that's a multi-year payoff. Yeah, I think it's important to kind of be precise. For sure, I don't think they will generate free cash flow anytime soon. I think they're going to generate a lot of operating cash flow. And then they'll use that to buy a lot of GPUs, XPUs, whatever, whatever I'm going to call them, or... Maybe they subsidize. heavily. We do know that data
[10:00] happening at the labs. What have we? their first-party products. So token consumption of their first-party products. Oh, yeah, yeah, yeah. They're doing all this research, and they're spending a lot on data on compute. And their first-party products are like a heavy subsidy. products today, right? Yeah, so it's 8 gigs of inference, and 2 gigs is for internal research, and then, you know, 2 gigs is actually training. Yeah, exactly. And, you know, including probably the inference that goes into post-training. Yeah, I don't... I think given the belief systems that they all seem to have about scaling laws, the which is, continue to hold. I don't think any of them are going to be that focused on generating free cash flow. And you've seen, right, we saw Satya blink. Yes. And Satya really regrets that, I think. Yeah, yeah. You know, he kind of blinked, I think it was last year. He gave that great interview for Davos, and they asked him about all the capex. And he said, I know I'm good for my $80 billion. Right. And I think they blinked a little. They slowed down. They regret that. And then Dario famously, he went on a podcast and he said, listen, some people are being super irresponsible with their spending. And it's a hard decision because if you don't spend enough, you could lose a lot of share. But if you spend too much, you could go bankrupt. And like those are both bad things. But bankruptcy is worse than losing share. So I'd rather be conservative. And he was conservative. And OpenAI was aggressive. And now OpenAI is back in the game. And SpaceX was aggressive. And SpaceX was aggressive. So, you know, like there are clear high ROIs on those independent of supply demand mismatches that are happening. Like clearly that seems to be the right decision, short term and long term. Yeah, absolutely. I mean, we we calculate, you know, Nebius and Corweave both gave some interesting disclosures, but you can kind of get to a nine to 10 month payback. for Nebius because, you know, okay, you bring on a gig, it costs $50 billion, you can get an upfront payment for 50% to 60% of that for customers. So now, you know, you're talking about $25 or $30 billion, and then you can monetize it if you put it into the spot market at a Spot paybacks are probably... Much faster than nine or ten bucks. Yeah, you've got to assume like a smoothed out level, like two bucks, three bucks. Even with that, it's very high payback. Now you can get like five bucks or eight bucks. And then SpaceX, because they build these really big clusters. And I think a really important point is they bring them on fast. Yes. They have an even faster payback and they can monetize it. I have tried to shift to think of pricing per megawatt rather than per GPU. Yeah, yeah. world is, but like SpaceX, the payback feels well inside of that. Yes. And I just... In my career as an investor... There haven't been that many opportunities where you have companies that could deploy tens, hundreds of billions of dollars and get sub one year paybacks. Yes. And it's kind of crazy. And then. Also, like we should also talk if particularly if you're buying NVIDIA GPUs to a lesser extent TPUs, you can finance these. Yes. And there's a very sophisticated, you know. Yeah, it's a very low cost of capital to finance them today. Yeah. And everybody's, you know, worked up about, you know, circularity. And it's like, well, I don't know. I know a lot of smart people who work at Blackstone and KKR and Apollo. And they're the ones that are financing it. They're the ones who are financing it at a relatively low cost. At a relatively low cost. And I think one reason that's happening is useful lives just keep getting extended. And as these models get better and better and better, and the ROI on token spend goes up, the monetization rate per gigawatt goes up. So... I mean, the true equity payback might be – way inside of a year. Yeah, exactly. Exactly. Yeah. And look, there's a case you could make that the prices actually of all this stuff go up. which could make the supply side... economics even more compelling, right? Like, you know, so, on the supply side, like that's the dynamic today like it just is what it is like there's a ton of data points out there that paybacks are within a year yep um I think it's actually interesting to think about the demand side too, because – The knock would be, well, in all these cycles, you get some overbuild, and then that, you know, destroys the economics of the supply side. The demand side today, like, what do we monetize? Like, the monetization of these companies, which are doing, call it 100%. 80 billion of revenue or something in that direction. is on the back of what, like 30 million actual heavy paying users? Like getting real value, I'm talking about like developers. I might take the under on 30 million, man. Yeah, actually,
[15:00] What we see inside our companies is obviously there's a power law in which companies are spending a lot. on tokens, like old banks are probably spending 1%. Very tech forward companies are spending high single digits. But if you actually look at the sort of the power law of what's happening of the actual engineers in those companies, The highest spending engineers are spending... 10 or sometimes 100x more than the median engineer. And so, yeah, your 30 million is probably way overstated. It might be sub 10. And so there's this question of like, where are we at in diffusion? there's one and a half billion knowledge workers. Like it feels like we're nowhere on the demand side and we're massively supply constrained. And what are, I'm just curious, across the A16Z portfolio, What are your best companies spending on tokens per month relative to human compensation? What rough range. Oh, high single digits, some at 10%, like some of the very AI native ones, like 10% plus. And so, you know, and then old economy companies are spending, the ones that are probably doing a good job, like 1%. So it feels to me like, When I look at the supply-demand characteristics, it's like supply stuff – What? People say, is that sustainable? Well, when you pair it with the demand stuff, It feels, it feels like there could be things that disappoint us in terms of like diffusion into the real economy. But it feels like over a 10-year stretch, like we're nowhere. Yeah, absolutely nowhere. And I just... So at Atreides, our internal token consumption has gone up 100x from the month of March. March through August, 100x our token spend. And we just got access to GrokBot Enterprise. And with two people using it, it looks like token spend might 10 or 20x increase. In a month. Yes. From August. Yes. Like, but it's actually extremely valuable. Like we have some heavy GrokBot users here and, like, It is very productive use. Like this is not like wasteful tokens, but. Yeah, I was in, and listen, like I, I try super hard. You know, I always, when I use AI, I just remember when my parents, like I was trying to get them to shift to an iPhone and an iPad and like, you know, get them used to it and like, you know, They did a good job. I give them loads of credit. But, you know, I'm 50 years old. You know, like, how old are you, David? 42. 42. And you see these like 23-year-old kids and just the way they use AI. They're just fluent and native in it. I just feel like maybe in a way that no matter how hard I try, I will never be. And I'm trying really hard. But, you know, like we got Cloud Code. I tried, you know, I built some stuff, did some cool stuff. Yeah. And in like... I don't know. three minutes of type creating GrokBots, I had much better versions of everything I created. You know, so I went on this Patrick O'Shaughnessy podcast like five months ago, and I said, you know, like, I love having a podcast summarizer. Everybody's like, how'd you do it? I was like, well, just use AI and do it. Yes, pretty soon. It takes 10 seconds, and GrokBot. Yes. It's amazing, and it's so good. Yeah. tracker for topics and stocks. Yeah. And like that, all of those would have taken me, I don't know, hours working with Cloud Code. And they each took seven to 12 seconds with GrokBot. It's better. So to me, GrokBot does feel like another, at least for me, like, kind of chat GPT moment because quad code, like, I could see the data. It was powerful. I did some really cool stuff with it. It was like empowering. And this is neat. you know, like family calendar apps, things like that. Yeah. Um, But, This is just 10 seconds, and it's way better than what I was able to do. Yeah, yeah, the cloud code thing was obviously the shift in coding, and our most sophisticated engineers were doing whatever, 20% of their code with AI to whatever, 90 plus. And so now I think everything you described in what you built, with CloudCode or Codex is still kind of reactive. Yeah. In a way, right? Like it's still, you know, it's like summarizers. Yeah. Preparation. It's all like knowledge enhancing, which is part of your job. But it's not actually doing the work for you. - Yeah, and now you have a Grok bot that says, "What are the recommended actions?" - Yes, exactly. - "Based on everything "the other bots have learned today?" "What recommendations do you have for me today?" And that for sure is like, and it was so easy to build. - I now have it, I'm like horse racing,
[20:00] which is like I have GrokBot doing it, Codex doing it, all the action-taking for it. I want to know, make me better at my job, look at everything I do, give me recommended automations you can do. I have Town doing it as well, which is one of our companies, I'm very good at it. And we're kind of on the bleeding edge of trying to do this stuff. just wait till everyone does this stuff. Yeah. And then, and then when we actually click like, yes, go just automate this. Yeah, it feels like that's sort of endless token assumption. Yeah, but I do, we should acknowledge, like, the history of financial markets, like, dating kind of back to the South Sea bubble, is whenever you get this transformational new technology. I actually went on a podcast, I said I thought the South Sea bubble was connected to the invention of longitude and the ability to sail. Turns out it was not. It was kind of like more of a tulip episode. But every time you've had a real profound new technology, whether it's the automobile, the TV, the radio, internet, the PC, railroads, steel mills, you get a bubble because the markets get really excited. And they get ahead of themselves. Things get overvalued. That overvaluation is, leads to an overbuild. And then particularly if you're funding it with debt, um, and, and even today, a majority of this is still being funded out of operating cashflow, which I think is really helpful. Um, um, you know, debt-funded build-outs, they demand immediate ROI, not an ROI in two years. Yeah, you can't be off in the time. You can't be off in the time, but I'm just more, you know, like I... I talked to Jazz about how Watson wafers, Jazz, I guess, and Patrick, Watson wafers are these fundamental constraints. And just the build-out is so big and we're so early that we are – it's like impacting the raw productive capacity of so many industries, you know? Yeah. Now, you know, everybody in copper, there's like an AI thesis. And like, we're going to have to like, think about it to like fill the – you know, if, if, 10% of what we just talked about comes true. You know, we're in this acute shortage with, I don't know, several million people are driving a crazy global compute shortage. What happens when that's 500 million? And, you know, how many copper mines do we need to build that? to like support this. Yeah, yeah, yeah. It's kind of a wild thought. And so like these fundamental constraints, I think are slowing us down. And I... And I think that's good. I actually think that's good for society. And I would now say rates and regulation, you know, real rates are going up. Yes. And it just is what it is, which makes sense because we're like investing a lot. So it makes sense that real rates are going up. And then regulation, man, it's. It is... I'm kind of shocked at what's happening in America. We're in a really bad place. Yeah. And just, you know, I had this exchange with Sholto from Anthropic and Dario on X last weekend. And, you know, Dario said, hey, I don't think I've been negative. You know, I've written two essays. One was positive. One was negative. So being 50% negative, and particularly when it's like, a terrifying negative like an existential an existential negative everybody might be out of out of a job like that eliezer yukowski guy says if we build it everyone will die and it's like how about if we build it like we're gonna cure cancer we're all gonna live forever i thought one of the best things dario said was like what we need to do is stop talking about curing cancer and actually cure cancer actually cure cancer and actually make breakthroughs like But just somebody like my favorite line in the Bible is the truth shall set you free. Yes. But the only person who can the only group that can tell the AI industry's truth is the AI industry. They need to just start telling the truth. Hey, when we've. okay, you're opposed to data centers. Well, you know what? It's probably the best thing that has ever happened to working class Americans. Yeah, exactly. You know, it's like going to college might be significantly NPV negative now because you can go learn how to be an electrician, a plumber, an HVAC tech, and make ungodly amounts of money. Yeah. So this has been amazing for working class Americans. We now have a lot of data that particularly with behind the meter power generation, It transforms a town. Like tax revenue, it doesn't double. It like 10Xs. And it is revitalizing all of these like dying small towns all over America. And listen, we're getting much better at addressing the environmental stuff. Generally, they use natural gas, which is a pretty clean fuel. The water consumption thing is totally debunked. It's totally debunked. It's nothing. It's nothing. So these are like really, really, really good. having a really positive impact on the world.
[25:00] That's without. even considering things like curing cancer, but somebody needs to tell that story. And I think the problem with it now is like the burden of proof is on not curing cancer, but actually delivering some real tangible everyday American benefits beyond using chat, or Grok to answer your questions or substitute for a search engine. It does feel like we're pretty... Close to that. Yeah, it does. And by the way, like one of the things that I think has been correct, but ineffective is this idea that we need to stay ahead of China. Like, it's like, it is true. Like, I'm very much like, I'm a patriot. Like, I believe that. But it's way too abstract. Yeah. The abstract, for the average American. Nobody's worried about China invading America. Yeah, exactly. I don't, I'm pretty sure the Pacific Ocean is really big. Yeah, like, affordability and like, how is this going to change my life for the better or worse, right? And so, I think there's a pretty... immediate impact you could feel like I my favorite is you know Loudoun County Virginia which is like the highest highest per capita income, zip code in the U.S. or county in the U.S. and it has the highest density of data centers. And they make a tremendous amount of tax revenue from data centers. We should do this everywhere. Yeah, it was actually very funny. Someone very opposed to data centers said, oh, you're for data centers? I'd like to see them put in the highest income zip code and the highest income county. They're like, actually, the highest income zip code in America and the highest income county has the highest per capita concentration of data centers. So we've done that. And it worked out really well. It worked out really well. Yeah, but you know, hey, don't bother me with the details. I'm on to my next talking point. That's good. That's good. And all those talking points, it's tragic. Like, there is an organized, CCP-funded campaign where, I think, again, data centers here in America, like I think a lot of it gets laundered through TikTok. And it's just tragic because the other thing that's happening is this is re-industrializing America. The combination of having the Strait of Formos closed, which is amazing for America. You know, natural gas here is two or three bucks. It's now 25 bucks in Europe and Asia, 20 bucks or whatever it is. And natural gas is an important input to the cost of electricity, which is... an important input to almost all manufacturing processes. And so we have a huge cost advantage for that basic input now. And you have that happening now. And you have this. kind of data center boom happening. We are re-industrializing America, and it's awesome. This is what everyone in both parties has wanted for a long time. Yeah, exactly. Like, bring industry back. Small towns that were left behind by the steel nose closing. Well, data centers are bringing them back. Yeah. But somebody has to tell that truth. I mean, I try to do it on every podcast, but, like, I'm just a dude. And, like, your audience is the tech audience that already believes you're preaching the choir, if you will. But, yeah, the story – I met us probably doing the best job. of telling that story, I would think. Yeah, it seems. You know, and I think one reason, it's really wired into Meta's DNA. So one of the first things they started doing as a public company was, I don't remember if it was on their first earnings call, but Cheryl would run through, Cheryl Sandberg, would run through 10 or 15 very specific small businesses that had started using Meta's advertising products and the impact it had on that business. You know, this cake bakery in Des Moines started, you know, worked with Meta. And, you know, it was two women who were single mothers working by themselves. have 15 locations. They employ 50 people. And this has been amazing for Des Moines, and it's been transformative for the years. For them. And they would just run through that Every time. And I do think the entire AI industry, like I'd love to see, you know, everybody, SpaceX, Anthropic, OpenAI, Google, Meta say, hey, Google. here are real businesses and real Americans and like either name the business or get permission to, if you can, name the American or anonymize it. This is a really positive thing it had on their life. Already very tangible, yeah. Yeah, same, NVIDIA, AMD. Broadcom. All of them. Yeah. Just run through specifics because the truth shall set you free, but only if you tell it. Yeah, exactly. Yeah. Yeah. So it seems more likely than given that fact pattern, if you go back to just the sort of macro situation that we're in. that we underbilled on the supply side. Oh, yeah. For like through 28. And by the way, like there's no capacity available with all the forecast builds that will happen through 28, which are probably now going to be delayed given the political dynamics that we have. So – Everybody's worried about oversupply. I'm like more worried about – Yeah, massively undersupply. Massively undersupply. Yeah. Yeah. Exactly. Which, okay, so then if that's the scenario – like you could see a scenario where you see, you know, big price increases. Oh, yeah. Access the intelligence. Yeah, well, good.
[30:00] It's the opposite direction of where everybody thinks this is going to go. Yeah. Well, Dworkash had a wild point. I forget what it was, but he was positing. I forget. Like the cost of a token could go up 10x or something like that. Yes. Which is crazy, but we do live in a supply-demand world. It's conceivable if the demand goes massively. And by the way, the whole premise of this that's happening so far is that there's a massive amount of consumer or user surplus being generated. so like why do people select the frontier tokens when they could use the cheaper tokens to do most tasks there's many reasons why but like the biggest one is because there's a tremendous amount of surplus even if you're using the frontier tokens yeah absolutely and so yeah what happens if there's like a massive supply shortage well I think that would be the you know kind of funny the consequence of like the these like data center de-growthers may be like real compute inequality where big companies and wealthy people can afford compute. And then, you know, two years from now, they'll be on about that. And it's like, well, that happened because of you. You know that happened because you wouldn let us build data centers Yeah and by the way we seen this right The path to a low product delivered to consumers in a mass market is advertising It takes a long time to build an advertising business, as we've seen with all the consumer internet businesses that we've invested in over the years. And so there may be a disconnect in the period where you can't actually offer that. And that would be a terrible outcome. That would be a terrible outcome for the world. Nobody wants that, so we need to build a lot of data centers. Yeah, exactly. Exactly. Yeah. Like a compute inequality like future. That's that's not a good that's not a good future for anyone, which is another reason open source is so important. And just one of the things, you know, I you know, I had a Grok make me make me like a meme of that like three headed dragon. And one of the heads is like kind of confused about like all of the really like stupid, bearish AI narratives. But people have this idea that open source tokens are free. They're not. And it's like it takes the exact same amount of compute. Yeah. All else equal to make an open source token as a frontier token for a comparably sized model. Now, there's a lot of nuances there, but that's broadly true. It just a question of what are the margins that are charged on top of that And even then the Kimi license something that I don think a lot of people appreciate is the Kimi license stipulates a 30 share of any revenue Yeah, yeah, yeah. So Kimi has taken a 30% cut of all the revenue generated on its – and this is because it's open weights, not open source. Yeah, exactly. Yeah. But it's also extremely token hungry too, right? Oh, yeah. Like, it's, it's, even, we're talking on a token basis, but on a task basis, it's far more inefficient. Absolutely. So it's very costly. Yeah. And I just always, like, Jensen, he's a great patriot, great American. Like, we're so lucky to have, we're lucky to have him and Elon. Like, and I think, like, you know, kind of when the, when the history of the 21st century is written, you know, there was like the Victorian age. I think this will be like the age of Elon and Jensen. Yeah. Because they are fundamentally altering the fabric of human society and civilization with AI, SpaceX, making humanity, multi-planetary, Starlink, bringing low-cost internet access to the poorest communities in the world, which is amazing. Which is something that people don talk about but it like an amazing you talked about consumer surplus That is an amazing surplus There was never going to be an economic case to build internet access in those places because of the cost and the willingness to pay And now you could. Now it's there. And any incremental internet capacity is not going to be built in a traditional sense on it's going to come from space. And so that is a huge unlock. I agree. It's a good thing. But we should all be grateful for them because I do think that they're... they're making the future as exciting and inspiring as possible. Say we are in this supply crunch. It's so funny whenever I talk about SpaceX, and it's obviously near and dear to both our hearts, I say, first of all, the orbital data center stuff, it's not like big buildings in space. It's helpful to actually think of it. It's like the size of an airplane. Yeah, people are picturing like the Death Star. Yeah, exactly. Or the Pentagon floating around in space. said that's not what it is at all yeah it's a it's you know whatever the size of an airplane right rack of 72 whatever chips yeah it's it's like five of us standing together is kind of and the airplane is like the wings you have the solar solar wings yeah and then you keep it in a sun synchronous orbit so you have the radiator that's always in the shadow of
[35:00] rack and That's how you call it. And, It's like I can't, it's very hard for me to engage. You know, there's all these people on X and they're like, I am a physics PhD. And this is impossible. And actually, there's a friend who's another investor who actually is a physics PhD who had many arguments with him. And he's like, I am a PhD and this is impossible. And then he goes to the SpaceX day and, you know, he talks to the SpaceX engineer. He's like, well, I was wrong. And so like. If you. Let's say you're an astrophysics PhD. You're brilliant. You're hanging 100 IQ points on me. Have you thought about this for an hour? Have you thought about it for 10 hours? Have you thought about it for five hours? Because you have 10,000 of the world's smartest engineers at SpaceX who, who've thought about this each for hundreds, if not thousands of hours, N. The sum of that, working with very sophisticated engineering tools... It's a solved problem, and in their minds, it's dramatically simpler and easier than a Starlink satellite because a Starlink has to have the phased rays and move around. I think it's – okay, so assume that you're right. I say it's like physics. There's not a physics... reason why this can't work. Cost-wise, it seems... really imposing, But kind of the history of the Elon companies is the cost curve gets dramatically better. When we first invested in SpaceX, Starlink was not commercially available, and the cost Like we had all these questions about how the economics would proceed over time, but the same on the launch side, the same with the Model 3. Like I just have to think that that will get solved paired with the fact that we're going to have massive undersupply. self-inflicted on Earth, it feels clear to me at a minimum it will be swing capacity. And in the fullness of time, maybe it will be larger. Well, no, it's really simple. If we use 50... And it is the people, the question people should be asking about orbital compute, which is the one SpaceX is focused on, is Starship reusability. Yes. Because the math is like, let's just say it's 50 billion a gig. And let's just say 35 of that is IT. Yep. So that's the same. And maybe it grows a little because it's going into space. The rest is power, cooling, labor, all sorts of things that you don't. need in space because you have the solar panel and the big radiator. And that's $15 billion, and it's probably inflationary here on Earth because labor fundamentally feeds into that. We just talked about what's happening to electrician. Yeah, comp, yeah. Yeah, electrician. Materials are all going to go up. Yeah, all of it. Yeah, we're going to run out of – the copper bowls are focused on copper shortages. Yeah, optics, all of it. Yeah, so that $15 billion is inflationary. And So what you have to compare it to is the cost of launch. And with Starship reusability, that goes to under a billion. So the economics just instantly flip. Now, you're always going to train on Earth. There will always be advantages to having... GPUs right next to each other. Like there are, you know, speed of light limitations are a real thing. Latency matters. So data centers on Earth, they're not going anywhere. I think they're going to continue to be very, very valuable. But an increasing fraction of of the world's compute is going to be in orbit. And, you know, Elon said that he and Jensen have co-designed a Rubin rack, and it's going to launch in the fourth quarter of 27. Yeah. And let's just say... let's just say he's off by two quarters. Yeah. I mean, that's, that's 2028. Yeah. That's still okay. That's pretty soon. You know, as Brad Gerstner says, like nobody's really paying attention to this and it's like kind of happening in plain sight. I mean, it kind of, to me, solves for something, you know, mid-seagullage billions today, which by the way, you know, is like, That's just like, keeping share constant. Yeah, exactly. You know, of like what's happening with coding. Not presuming taking any share on GrokBot. Yeah, from 3 billion. And by the way, man, I would just, I'd probably take the over with GrokBot. Yeah. I bet it's like, changing by the day just based on my own usage and the number of people who are hitting their usage limits and then you're starting to get from you know grokbot like hey we hear servers are overloaded every once in a while and like they have a lot of compute um so it's just like okay you don't want to debate orbital data centers yeah No problem. Well, like Starlink Mobile, they have a pretty clear, credible plan for how that's going to work. And that wireless is, call it another $800-900 billion of revenue that they address. So yeah, your mobile plus your broadband, whatever, let's call it close to $2 trillion of a market.
[40:00] Growing. AI... ARR base. Yeah, ARR, you've got the cloud, you know, the sort of cloud business. Yeah, so I don't think, great, you're an orbital compute skeptic, no problem. It doesn't matter. Yeah, exactly. We don't even need to, we can just look at things that are happening today with terrestrial compute, with cursor, with Grok, with GrokBot. By the way, I think XADS are, you know, we have telemetry. They're also growing. You know, I would expect at some point you'll have like a Starlink, GrokBot, X-Advertising Bundle, you know, kind of one of the ways Google built their cloud business is they bundled it with ads and like, hey, we're, you know, maybe you're bundling the ads with AI, but why not do that? Yeah, I actually like the, AI position that they're in because it's like heads you win, tails you win in the sense that their first party business is growing very fast and they, they caught up to the frontier like very quickly. Um, and so they've made the very aggressive compute investments to enable that first party work. Um, and that's the kind of heads you win and like tails you win. Say they overbuilt their capacity for what they need for inference or training. They have a very compelling on the compute side, you know, with like massive scarcity of supply. And so I think that's a really good setup. And there was a bear case that, hey, okay, well, in the open AI and – anthropic maximalist view where they're the only two companies and they're designing their own chips, then like where, what's the room for anyone else? Well, like I don't think they're going to have a reusable starship and multiple spaceports anytime soon. And if the economics of computers such that orbital is where it makes sense increasingly going forward, because starships should be deflationary, you know, terrestrial cooling, you know, power should be inflationary. Well, like even in a world where They fumble the ball with their first-party AI applications. They do still have a massive infrastructure business. Yeah. I'm so fired up about the Starbase Louisiana. Oh, yeah. I can't wait to visit, man. It's so cool. I was reading about it last night. And yeah, it's sort of like they now have the infrastructure for thousands of launches a year. Yeah. coast all over the world. Yeah. Like, you know, at some point you'll probably see one somewhere in the Middle East. You'll see, you know, whatever European country is like the least bureaucratic at the time, you'll see one there. You know, you'll push, I think you'll see probably one in, you know, whether it's Japan, South Korea, who knows? Yeah. Yeah. Yeah. Yeah. Yeah. That's call it, you know, whatever, 5,000 launches a year. Like, that feels very futuristic. Yeah. I mean, it's wild. And I do think a distinction that, you know, SpaceX really tried to kind of hammer home during their IPO is there's a difference between reusability. In China, they did catch kind of a rocket using this. It was actually kind of ironic. It was this kind of jury rigged system of kind of wires that had actually been suggested on the SpaceX subreddit. And like China's clearly paying close attention to the SpaceX subreddit. But that's very different catching that thing from what they're trying to do with Starship where, you know, the booster gets caught with the things and then it gets moved and then the Starship gets caught and then it gets stacked. It gets fueled and just sent right back up. Yeah, two a day, two a day per pad. Those numbers add up pretty fast. And I think they're engineering the pads for more than two a day. Yeah, I think that's a conservative assumption. Yeah. Yeah. What's the, okay, so SpaceX, like, again, you and I have talked a ton about SpaceX. what's like the most futuristic thing that you think about? with SpaceX. like the 10 year killer. Okay. So you and I were at this conference together and there was this whole debate about, um, among a small group of public investors of like, what's going to be the, the first $10 trillion company. And, uh, I think what you said was like, I have no idea, but I know which one's going to be the first $20 trillion company. So like, what's the most futuristic like product or market or technology thing about SpaceX that you can think of? Look, I mean, this sounds crazy, but asteroid mining is going to be a very real thing. We're going to capture, you know, there's asteroid psyche. It has more gold, silver, platinum, you know, every precious metal in it that exists in the Earth's crust. at some point, particularly with Starship, you will be, you know, and we may need that lunar base to make this happen. You'll be able to capture these asteroids. You'll bring them
[45:00] to a stable kind of geosynchronous orbit over some, you know, American-owned atoll in the middle of the Pacific. You know, no humans within whatever 50 miles. You'll, you know, you can imagine like optimist robots, you know. Doing the work, yeah. Yeah, doing the work. And then, you know, delivery to Earth is free, and for sure some of it's going to burn up. But I think that's going to happen. Jeff Bezos said something very interesting. He said, I think in the future... Earth is going to be zoned residential. And, you know, somebody asked him, this was like 15 years ago, what do you mean by that? He's like, all heavy industry will take place in outer space. And then this addresses the pollution concerns, it addresses everything. You know, people always get like really worried about, oh, you know, will we still be able to see the stars? And it's just like, I think it's hard for like the human mind to understand, and it's hard for the human mind to understand, how big space is. Yeah. How big outer space is. Yeah. Yeah. You don't have to worry so much about emissions up there. Yeah. Yeah. So I think that is, um, That's probably the most futuristic thing. But in terms of an economic application, but it does... I mean... I do think in the next few years you're going to have a fleet of starships. Land on Mars next few years. I mean, I don't know. Let's just say at the outside, this is eight years away. Yeah. They're going to land on Mars. We're going to have like, you know, our little ramps going to come out of the Pez dispenser and it's going to be a modified starship, the Mars Colonial Transporter. And it's going to be wild. You're going to have Optimus robots holding American flags, like walk down. And then, you know, they're going to pull out a bunch of solar panels and batteries and racks of compute. And they're going to set all of that up. They'll be dropping Starlinks, you know, and maybe the orbital mechanics don't allow this. I think they will figure out a way to have capacity. So just think how crazy it is to watch the views from Pathfinder or whatever these different Mars rovers are. And 4K video through Optimus robots all over Mars. Yeah. After that, there will be humans. You can inhabit it, yeah. Yeah. Yeah, that is crazy to think about. And that's going to be an amazing moment for America. Yeah, of course. I mean, think about the moon landing. This is a little bit bigger, yeah. Yeah. Yeah. So that seems cool. That's a good one. That's a good one. There's not a lot of chatter about that one out there. Yeah, but I think it's highly likely to happen. Yeah. Yeah. So you mentioned Microsoft. Yeah. And the bet that they made, which is like a little bit of, you know, like Apple's the extreme kind of bet against the future kind of bet they made. And like Microsoft is kind of a gradient of that. Yeah. Like what's your what's your outlook for Bitcoin? their decisions. Well, I do think the world has gotten a lot friendlier for their strategy. You know, they clearly tried to make a frontier model. They failed. Yeah. Satya said, we're going to have our own models that are very competitive. I think he said that 18 months ago. They don't have their own models that are competitive. competitive. But what you're seeing with, um, I think... The future is an ensemble of models. You know, there's a Pareto curve. No one model is going to be the best at everything. And I think the future for certainly, you know, kind of the global, you know, 1,000 biggest companies is you're going to take whatever the best open source model is. I think probably in the very near future, that's going to be an NVIDIA model. The labs making ASICs create very interesting incentives for each to get into each other's business. Incentives for Jensen. Oh, in a world where open source wins, who funds the training? Well, the chip companies could fund the training. Yeah. It's trivial to do a $50 to $100 billion training run for Jensen. And maybe soon, I do wonder if this is kind of Google's super long-term play. They seem to maybe have opted out of the frontier race for now. We're going to monetize our compute at high rates. And we're going to... sell TPUs externally, but that generates so much cash flow and open source is getting closer and closer and closer to the frontier. And it just may be the winner is ultimately just who has kind of the most cash flow to fund these big training runs. But I do think you're going to see American open source led by NVIDIA get really close to the frontier. Like they paid that poolside acquisition was made for a reason. They're doing a lot of smart things, but that is really good for Microsoft and at some level, almost every application software company.
[50:00] Because what you can do now is you can take a base model, and Nematron to date has not had a lot of post-training. It's kind of been a good pre-trained model. You can do with what you want. So if you take a really good pre-trained base model – And then instead of sharing your own kind of enterprise context, that's truly your IP, that's truly the value of, of your company is like the context embedded in all of your data. And sharing that with a frontier lab, that may be hazardous for your financial health. Yeah, certainly with the shift in the ZDR policy, yes. Yes. And so you take a really capable open source model and you do a lot of RL and supervised fine tuning on your own data. So you own it and it's your model. And then if intelligence is like a super important input into your business, Thank you. own and control your intelligence, its capabilities, its cost. And then what we've seen is, From a lot of companies and, you know, GrokBot, my understanding is, you know, I think it's Gemini 3.7 Flash, Grok 4.6 and some Opus. Yep. and what you, and behind a router. Yeah. And you will, um, um, And I'm sure Elon is very focused on having it all grow as soon as possible. Yeah, of course. But I think what you'll see these companies do is they'll have their own model. on their data, and it will work with one or two other frontier models. Not necessarily, but just checking each other. It'll be kind of transparent to the user. Yeah, you can use the most frontier for planning and then have execution run by everything else at slower cost. Yeah, absolutely. And so I think that feels like a very likely future to me. And that is a much Microsoft-friendlier future than one in which there's just only two dominant frontier models. it's like there's going to be at least three with Grok. I do think you've got to give Meta a lot of credit. They've done a great job. Yeah, and I mean, they were out of the game, and they got back in the game. And it's just kind of amazing. Who could have imagined a year ago, you know, when it was like Gemini was a syndact? Yeah, exactly. That this is the scenario that we're in. That Gemini wouldn't even be in the conversation. Yeah, they're not in the conversation. And Muse and Meta would be significantly ahead of them from a capability perspective. chess ever played and you know people you know some people have made bad moves they've made good moves you seem some people come out of the game others come back in um but a future where that future where it's a you know i don't know if we're gonna call it multi-model a hybrid model yeah i don't know what terminology the world is gonna settle on but i think that's the future yeah and I'm... actually surprised. I think the best broad instantiation of that today, outside of GrokBot, outside of Cursor, outside of, you know, like Harvey's done some cool things. Yeah, yeah, yeah, Harvey's done great things with that. Yeah. done it is actually just the fireworks nexus product yeah yeah you can yeah you can Choose your Frontier model. Let us take whatever open source model you want, RL it for you, for your data, for Goldman Sachs, for Morgan Stanley, for JP Morgan, for Fidelity, for A16Z. You have all your own data. You control your intelligence. And we make it transparent behind a router. Yeah. I think that is like a very plausible future. And that's clearly what Lynn from Fireworks, she was the first one to say it. And then Alex Karp and Satya, they both kind of. Yeah, they've taken their own version of it. Yeah. But, you know, Satya, as I say, is specialized intelligence. Like, I think it's very plausible. But this stuff is really hard to do. Like, that sounds easy. I was, it sounds easy to describe, like, the way I describe it to people is, like, who gets to be the abstraction layer to the organization and the users of intelligence? It's, like, the most, whatever, vied after space or position that you could imagine in business, like, in the history of business. Yeah, for sure. Right? I think it's, like, the answer is and, again. Yeah, yes, and, for sure. It's, yeah, who's the arbiter of intelligence for global enterprises? and probably consumers, I was a retail analyst and, um, Everybody kind of thinks running one of these big chains is easy, and there's a lot into it. And it's like, well, I'm not. It's really easy. I started an American retailer. in any category, because America is so big, it's worth over 50 billion dollars. Almost any category. All you have to be able to do is have a fleet of a thousand stores in 50 different states that have very different climates, consumer preferences, restaurants. You need to have them stocked with the right products at the right time for that region at the right prices. They need to be staffed.
[55:00] by friendly and knowledgeable employees who don't steal from you. Who turn over at 100% a year. Who turn over at least 100% a year. The stores need to be clean and well lit. And if you can do that... Press show, $50 billion. Yeah. And like in the history of American business, like you can, I mean, it's more than one hand, but you don't have to go through many. Yeah. It's really hard to do. Yeah. And. Mm-hmm. Having that abstraction layer, having it work, having it seamless is, I think, way harder to do than people think. And I do think something I think is very interesting about Cursor. I'd love your opinion on this is like everybody else in the lab space, um, you know, had this like We're creating a digital deity, you know, and AGI and ASI. And the Cursor guys were just like, we want to make great product. Yes, exactly. In a strange way, everybody at the frontier, probably Kersher, was the most product-focused. Yes. I'd say now they're part of SpaceX, but that suits Elon and his mindset really, really well. Yeah. Let's make it an engineering problem, create the model factory, and then we need to have a really good product. The Tesla cars, they're amazing. Yeah, yeah, yeah. But what Cursor figured out is they had, I would say, a similar in-state vision as what those other guys had. It was just a different path to get there. And it's sort of like a practical meet the customer with what with where they are and meet the technology where it is. Um, and I think, you know, they'll sort of, they have already demonstrated that they kind of leg their way up into autonomy from, from that starting point. Um, yeah. Coding is unique compared to everything else in knowledge work. This would be in support of the point that Microsoft is in a good position because it is verifiable and perfectly documented. And nothing else in enterprise is verifiable and perfectly documented. And so it will be messy. That leads you to a good bulk case for something like Microsoft, that abstraction layer. If they execute, but it's really, really hard to make it really simple for, train a model on our data, and, convince me that you're not going to share it with anyone else and then put it behind a router that's seamless for me and continuously upgrade that open source model. It's not just some middleware. It's very hard to do. Yeah. And by the way, they're going to be competing with not only the labs to be that abstraction layer, but Databricks. Oh yeah. Like, Palantir, um, the inference providers, um, the application companies, right? So, like, Harvey has done an incredible job of this. Yeah. And, you know, like, legal is sort of in takeoff. Yeah. And I think they can see the future of how to be that abstraction layer and do the work. But, like, legal is also unique because it's very documented. Yeah. And it's somewhat verifiable. We'll see tax. Tax, we'll see that. We'll see things like that. But, like, the $1.5 billion, the really appealing broad pie, is going to be very messy to go get. Yeah, although I do always think, and, um... You know, I think, probably in their heart of hearts. Harvey and Lagora think, oh, if we solve this, We could be that obstruction there for everyone. I think probably in their heart of hearts, cognition thinks something like that too. I think everybody thinks it. And by the way, there's like massive validation of the category because Kirkland and Ellis said, we're going to spend 500 million bucks to build this ourselves. Like, first of all, you know, like, Good luck. That's going to be very hard. But that actually tells you that the pie is really big, right? Huge. Yeah, it's massive. And that's, it's, And, you know, just and I'm sure they have a very smart head of AI, but it's not like a $500 million one-time build. That model has to be continuously updated, switching out the base model. Then all of that has to happen transparently. But I think you're going to have this huge collision between... Products like Fireworks Nexus, these legal agents, coding agents, big companies like Microsoft. Databricks. Databricks. Snowflake coming up. For sure, Salesforce, I think is going to – Salesforce and Workday and all these companies. Everybody's going to go after it. It's just going to come down to who executes the best. And – And this is just, you know, Who has the lowest costs? Yes, exactly. And it's going to be very hard, I think, over time unless you're – if you're not vertically integrated, you have to be so good to emerge as that abstraction layer. Yeah, yeah, yeah. To be the low-cost provider. Yeah. Very hard. Because, yeah, you're just simply not going to be the low-cost provider if you're not vertically integrated if you don't own your own compute over the very long – long term. And, you know, it's another reason like I, you know, I ain't
[60:00] look at these hyperscalers on EV to net PP&E. Yes. Because net PP&E is compute, and that is just what the market thinks you're going to monetize your fleet of compute at. And you can kind of look at them, and there's some pretty obvious inefficiencies to it. Yeah, yeah, yeah. Yeah, kind of an AI version of price to book. Yeah, I like the price to book. Okay, that's good. It's okay, you mentioned Jensen. You know, I share your sentiment. Like, he's, like, carrying this industry forward. Like, tell me your thoughts on Avidia. So, um... I think he's in a very, very good position and his strategy of being vertically integrated but horizontally open. And it's like, okay, let's just say – You know, let's say there's some accelerator that emerges that is really, really, really, really good. Almost certainly it will be better if it can plug into, and this is why, like, I know you have an accelerator investment. 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. Thank you. How can we work with you? We want to enable you. Sure, we're going to compete with you on the edges. Yeah. But, you know, my rule of thumb for accelerators, every 1% share today is probably worth $100 billion. Yes. So there's no need to go head on with NVIDIA. Yeah. Yeah. Just pick a niche, get your 1%, make sure that, you know, he has nine chips. Yeah. You know, he's got multiple flavors of accelerators. He's got CPUs. He's got, you know, Ethernet switches. He has two kinds of DPUs. You know, he's got, you know, we've gone from just scale-out networking being a thing. We have scale-up, scale-out, scale-across, now scale-in. Yeah. So just try to find a way to plug into his ecosystem. By the way, this is not foreign. repeating, products with various of those nine chips. Yeah, and just try to find a way to plug in, but just... Be nice to him. Be nice. Be nice. It's all personal. Yeah. You know, and it's just like sometimes like, you know, you hear some of these and it's like, have you ever seen game tape of the Chicago Bulls when Jordan was, you know, game 50 of the season? Yeah. Yeah. And he's a little bored. Yeah. And the Bulls are down because, you know, they're up eight games. You know, they're up eight games over the number two person in their conference. And he's a little bored. And then somebody. Somebody talks shit. Somebody who's kind of young decides I'm going to talk shit to him because we're beating him. And then he just looks. And it's like. And it's like. It's the best. Yeah. It's amazing. Oh, yeah. We've all seen, you know, whatever the last dance. just don't do that. Yeah, exactly. You know, just, just like, Hey, Michael, man, I'm so happy to be on the court with you. Like that's, That's the move. But the reason it's particularly important is because Jensen's data centers are financeable. Yes. And it goes back to that point. Bye. Let's say it's, $50 billion. For an NVIDIA data center, you need a $15 billion equity check. Yeah. Okay? You can finance the other $35 billion. Yeah. And it's not circular financing. I have a lot of respect for the people I have met from Blackstone and KKR and Apollo. Yeah. And they're underwriting each of those. Yeah. And... They finance it. And then there's a residual value guarantee, which as long as that residual value guarantee is less than the gross profit dollars he's getting from selling the chips into that data center. It's like essentially it's super NPV positive with very little risk for him. And then he gets a revenue share. So if you're... And his data centers are the most Yes. And like, let's just say a good case for probably TPUs are the second most financeable. It probably takes, I don't know, double the equity check at least. And then the rates... on the rest of it are higher. Yeah, exactly. And so cost of capital is a huge advantage. And that's why you just want to be part of his ecosystem. And you can see he has all these chips. He's acquiring land, power, and shell companies now, matchmaking them with offtake agreements. I think one reason he's doing these RVGs is if he doesn't do them, it's kind of an anthropic and open AI-dominated world because they can pay the most for compute. He can effectively help other people compete with anthropic and open AI. Yeah, in the same way that he stood up the neoclods in the first place. Yeah. It's just democratizing compute, which is good for the world. Again, I think he's a patriotic American. American ones, right? Of course, yeah. Fragmentation. Yeah, fragmentation, no dominant AI. Yeah, exactly. Which is really good because he is a ruthless competitor, and it's awesome.
[65:00] that his incentives around fragmentation of AI, fragmentation of models, and, you know, fragmentation of power are completely aligned with what's good for America. And just going back to open source, just like... I just can't take it that people think that Jensen is like the world's biggest advocate for open source. And it's somehow... A giant risk to his business. Yeah, exactly. No, it's great for his business. Yes, it's great for his business. It's amazing for his business because it means that instead of, you know, having a 90% margin on top of a token made with an NVIDIA GPU, maybe it's a 40% margin. So more of those tokens are going to be consumed, which means you need more compute. Yeah, exactly. In a supply constrained world. In a supply constrained world. And what percentage of the world's supply has he locked up? 70%. 80, somewhere in there. And then... You're talking about fab? All of it. All of it. You know, it's just because he saw this coming before everybody else. Yeah, and all this system supply chain. Yeah, he's got the... Fab capacity locked up. He's got DRAM capacity locked up. He's got NAND capacity. He's got laser capacity. He has capacitor capacity. He has, you know, what you need to make the racks. And it's just like, you know, he used to say, if I go back... you know, 15 years, you'd say, listen, I'm making a two or three billion dollar bet. Every two years and I'm moving really, really fast. Now he's making these multi-hundred billion dollar bets, bringing the supply chain alongside him. He's bringing the financing alongside him by kind of standardizing it, making it easy for the very smart people at Blackstone, KKR and Apollo and Goldman Sachs and Morgan Stanley, JP Morgan to finance. And like that is hard to compete with. Yeah. And, you know, it is – We, my firm, Atreides, we have a pretty big portfolio, private portfolio companies. They're semiconductors. And it's just, you know, Elon said a lot of people are going to learn a hard lesson in hardware. And like, I would just say I've learned a lot of hard lessons in semiconductor investing. Like you can bet on the best team and you tape the chip out. You feel great. Okay, we've taped it out. And that's happening faster than ever, right? Yeah, it's happening faster than ever. really good with the emulation and the simulations and you feel great about it, you're And then, you know, you'll experience this. The chip comes back from the lab. Everybody, you get a FaceTime from the CEO. They plug it in, you know, and like, And then sometimes it doesn't work. It's just like. Yeah, this famously happened with Cerebrus twice, right? Yeah. And they've powered through and they've done a great job. I think the chip, I think each Cerebrus chip works. worked. It just... struggled to find product market fit. Yeah, yeah, yeah. For the first two generations. The chip worked. Yeah, yeah, yeah. It just didn't have product market fit. And they've done great with it, yes. Yeah, but there's a different thing between you. you plug it in and it doesn't work at all. Exactly. And then it's like, if it doesn't work at all, You might be back to the drawing board and, hey, we need another... you know, hundreds of millions of dollars, billion dollars. And we've, we learned our lesson. It's going to work the next time, two years from now. Yeah. Assuming you can finance it. Yeah. Yeah. Assuming you can get financing. So it's, um, You know, semiconductors are hard. Like the real world is hard. Like hardware is hard. And... What he is doing at the scale he is doing at and the speed and bringing all of this alongside him, because, you know, the land and power has to come. Yeah. You know, the entire supply chain has to come. The financing has to come. Yeah. And so given that he's, you know. 70, 80 percent, whatever we want to say, You just want to plug into that ecosystem. Yeah, that's part of why Elon made the decision made, right? Yeah, which I also think was like, a very high ELO move. Yeah, totally. So, You've had everybody else try and build their own ASIC. Yeah. They've gotten up on stage. Sometimes they say negative things about, you know, Jensen or NVIDIA or take shots. I did think it was pretty smart. You know, the Jalapeno team last night, and we should give credit where credit is due. Jalapeno is the... I would say the first good ASIC, other than TPU or Tranium IFC from internal. - In what seems to be a pretty short amount of time. - In a pretty short amount of time. It's impressive, we should give credit where credit is due. - They do have a good team working on it. - They have a good team, yeah. So they had a really good team, I think they had a lot of advantages, and I do think if you're a lab and you have the model and you see the direction of research, that's a big advantage for designing your own chip. But then you go back to NVIDIA and they work with everyone.
[70:00] Yes. And everybody, you know, keeps thinking it's going to really standardize. And if you look at the three big, you know, Chinese open source models, DeepSeek, Kimmy, Quinn, they're kind of all. all evolving in very different ways. Yeah, yeah. And they can, you know, they can all run on, you know, more general purpose chip, um, a GPU. But you're going to need, if you want to specialize, yeah, you're going to need general purposes at a minimum for the types of evolution from that. Yeah. Yeah. So, um... Like, I think he's... I'm very happy his incentives as a CEO are perfectly aligned with what's good for America. Yes. So, I just, Make sure your semiconductor guys do not talk trash about Michael Jordan. Be nice to MJ. Yeah, exactly. Yeah, and then it's like, you know, sometimes it's like, you know, you tug on Superman's cape and you get confident. Yeah. You know, you get confident and you start to talk a little bit of trash. You know, Superman, sometimes he just flies away. Yeah. Like that's what happened to the TPU team. Yeah. And, you know, Jalapeno, they're tugging on Superman's cape a little bit. Yeah, we'll see. We'll see. Jalapeno did something that none of the big... like this is as competitive of a chip as IFC. Yeah. But again, it's just competitive with one of his eight or nine chips. Yeah, one of his nine. Yeah, of course. And it's... They'll continue to work closely together. Yeah, they'll continue to work closely together. So it's like, hey, that's great. You did the one thing. Well, to actually be competitive with him at the system level, you did another eight chips. Exactly. Yeah. And he is... And, Dylan at Simi Analysis talks about how he's the bank of AI. He's the central bank of AI. He's the Federal Reserve of AI. And so I actually think it was really smart for... Elon instead of like, Competing? You know, with somebody who is fully aligned. Yeah, fully aligned. Mm hmm. And I think that history is going to judge that to be a wise decision. In a world that is so supply chain constrained, it's actually really hard to tell what true customer preferences are. Right. Because like you come out and they'll take anything. Yeah, this is how you know that like very old, whatever, the price is sold up of H100 is very high. Yeah. And if you have a TSM allocation, you're going to be sold out. Yes. Particularly if you can get the DRAM to pair with it. You're going to be sold out. So it's actually kind of hard to infer true customer preferences. And I actually think one of the best ways you can like see true customer preferences is the kind of deals they cut with chip companies. So broadly speaking, you know. The first deal is where the chip company invests. Yep. And a customer, and you saw TPU and Tranium, Amazon and Google, do that with Anthropic. Yep. And that was to their immense advantage because it really helped their businesses, I think, help those chips really level up. Because you kind of need to use a chip. Yeah, yeah. There's a cold start problem. And... And in that scenario, as long as the dollars you invest are less than the gross profit, you can't lose money. And then there's a scenario where you do the RVG process. Blackstone finances it or whoever, Blackstone, Apollo, KKR, Goldman Sachs finances it. And as long as that RVG is actually less than your growth for profit, you can't lose money. And you have upside probably through a revenue share on top of it. Then there are deals where you give warrants away, but they're tied to like a fixed price per million tokens. And as long as the performance of your chip kind of outruns the performance of your stock, you're going to do good in that situation. If you just give warrants away, it could be negative in PV because the better the stock does, the worse the deal is. The more value that's captured by the person. Yeah. And so you can kind of look at that hierarchy of deals and like infer something about true customer preferences. Interesting. Yeah. So NVIDIA does pretty good deals. Like, yeah. I mean, there's a reason that people I consider smart are investing in their deals. Yeah, I see it. Gavin, thank you. Fun. Always fun to hang out with you. Thanks, David. This is great, man. and Spotify. Follow us on X at A16Z and subscribe to our substack at a16z.substack.com. Thanks again for listening and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures.
[75:00] Thank you.
