Gavin Baker: Why AI Demand Is Outrunning Compute Supply

Gavin Baker, a venture investor, shares findings from conversations with AI leaders conducted over summer, noting that when asked to identify a single quantitative metric getting worse in their business, no one could provide one. This observation leads to a discussion of the accelerating AI landscape across OpenAI, Anthropic, open-source models, and Grok. Baker and George explore the thesis that unlike traditional winner-take-all markets, the AI ecosystem may be positive-sum, where frontier models, N-minus-one models, open-source variants, and application layers can all succeed simultaneously.

The conversation shifts to economics fundamentals. They discuss how inference-heavy allocations (e.g., 8 gigawatts of inference to 2 gigawatts of training) can generate $480 billion in annual revenue with one-year paybacks on a revenue basis. Frontier labs like OpenAI and Anthropic face a strategic trade-off: they could reallocate compute from inference to training to chase capability improvements, but this would slash revenue by 75%, creating pressure that could impact public company valuations. The speakers suggest this revenue control and allocation flexibility represents a unique characteristic of these businesses.

On the supply-demand dynamics, they argue the market is massively supply-constrained. Current monetization occurs across only ~10-30 million heavy-paying users (mostly developers), while 1.5 billion knowledge workers remain largely untouched. Enterprise adoption shows power-law distributions where top engineers spend 10-100x more tokens than median engineers. Atatrade's internal consumption grew 100x from March to August, with Grokbot potentially driving another 10-20x increase monthly—yet this represents genuine productive use, not wasteful spending.

They examine whether historical tech bubbles apply to AI. Every major technology shift (automobiles, railroads, internet) produced overbuilding funded by debt, which demands immediate ROI. However, today's compute build-out is largely funded from operating cash flow, which is more forgiving on timing. More importantly, supply constraints fundamentally differ from past cycles: with supply still tightly constrained and demand concentrated among a small fraction of potential users, the traditional oversupply dynamics may not emerge.

Data center development receives substantial discussion. Contrary to opposition narratives, the speakers argue data centers represent the best development for working-class Americans, revitalizing dying small towns through tax revenue that increases 10x, not doubles. Loudoun County, Virginia—the highest per-capita income county in America—has the highest density of data centers, demonstrating proof of concept. Behind-the-meter power generation and improved environmental practices address historical concerns. They argue the AI industry has failed to communicate these tangible benefits effectively.

Orbital compute and SpaceX's role feature prominently. Baker explains that orbital data centers aren't massive space structures but airplane-sized racks with solar panels and radiators in sun-synchronous orbits. Physics constraints are solvable engineering problems, not fundamental barriers, according to SpaceX's assessment. With Starship reusability, launch costs drop from $35 billion to under $1 billion per gig, making the cost structure competitive with terrestrial data centers while eliminating labor-intensive cooling. A Rubin rack co-designed by Elon and Jensen is scheduled for Q4 2027 launch (possibly delayed to 2028). Even if orbital compute remains supplementary, it addresses mid-single-digit billions in marginal supply.

Asteroid mining emerges as a long-term SpaceX application. Psyche and other asteroids contain more precious metals than Earth's crust. With Starship reusability and Optimus robots, capturing asteroids into stable geosynchronous orbits and mining them becomes economically viable. This ties into Jeff Bezos's vision of Earth becoming zoned residential with heavy industry in space.

The speakers discuss frontier model dynamics as the labs prepare for IPOs. Anthropic rebased accounting metrics and conducted quiet-period testing before expected reacceleration. A competitive dynamic exists where labs game each other's releases—Anthropic waiting for OpenAI's Astra before releasing Fable upgrades. Public markets will force labs to balance mission alignment with shareholder returns; massive revenue cuts to fund research may prove difficult once equity holders expect returns.

They present a multi-model future dominated not by single winners but by ensembles. No single model excels at everything; a Pareto curve exists across performance dimensions. Microsoft's bet against frontier model dominance positions it well if an ensemble approach emerges. Specialized open-source models, funded by chip companies like NVIDIA or Google monetizing TPU sales, could approach frontier capabilities. Companies like Fireworks, with their Nexus router product, embody this vision: users select frontier models for complex tasks while open-source handles execution at lower cost, all behind transparent routing.

Data privacy and corporate strategy intersect here. Companies won't share enterprise context with frontier labs due to IP and regulatory concerns. Instead, they'll fine-tune open-source models on proprietary data, maintaining ownership while controlling costs and capabilities. Legal AI (Harvey, LexGPT) demonstrates this pattern—Kirkland & Ellis announced $500 million to build internal systems, validating the market size while showing the difficulty of continuous model updates and transparent optimization.

Microsoft, Databricks, Snowflake, Salesforce, and Workday will compete to become the abstraction layer for enterprises, routing between frontier models and open-source. The winner will be determined by execution quality and, critically, cost structure. Vertical integration (compute ownership) becomes essential for low-cost provision over time. The battle resembles other infrastructure layers historically.

NVIDIA's dominant position receives extensive analysis. Jensen Huang commands ~70-80% of accelerator supply and has locked up complementary supply chains: DRAM, memory, photonics, fabrication capacity. His strategy of being "vertically integrated but horizontally open" allows him to work with all competitors while extracting value through multiple channels: direct chip sales, residual value guarantees (RVGs) on financed data centers, revenue sharing, and warrants. His ability to finance data centers at lower cost than competitors through partnerships with Blackstone, KKR, and Apollo provides additional advantage. Competing accelerators should integrate with NVIDIA's ecosystem rather than fight head-on; each 1% market share is worth ~$100 billion.

XAI's decision to not build competing silicon but work with NVIDIA represents strategic wisdom compared to competitors who announced proprietary chips. Cerebras, custom chips for labs, and other accelerators face long development cycles and execution risk. When chips fail (or lack product-market fit), recovery requires hundreds of millions to billions in additional funding. NVIDIA's advantage isn't just current technology but the ecosystem, financing, and supply chain flexibility to evolve at scale.

Regulation and real rates pose headwinds. Rising real rates increase capital costs for compute buildout. Regulatory uncertainty around data centers, environmental concerns, and AI development creates friction. The speakers argue the industry must communicate truthfully about benefits rather than pursuing abstract narratives about staying ahead of China. Concrete examples—cured cancer, job creation, town revitalization—resonate better with the public than geopolitical arguments.