Gavin Baker’s AI Investment Case: Adoption Depth, Compute Contracts, and the Cash to Build

1. Guest Background

Gavin Baker, managing partner of Atreides Management, joins John Coogan and Jordi Hays on TBPN to discuss the investment economics surrounding AI. The conversation moves from the Anthropic S-1 described by the speakers to enterprise adoption, open-weight businesses, GPU availability, and financing. Its central question is how expanding use becomes demand for infrastructure—and whether selling that infrastructure can generate enough cash to support further construction.

Baker brings experience across more than technology investing. He says he ran a diversified fund for nine years, invested in energy and natural resources, and previously worked at Fidelity. That background helps explain his attention to electricity, physical execution, and capital allocation alongside software revenue. It does not establish equal expertise or privileged access in every adjacent market.

Two disclosed investments are particularly relevant. Atreides led Fireworks’ latest round and invested in Silica Data, whose compute-market observations Baker cites. These relationships provide context for his access to fundraising and procurement information, while also creating an economic interest in parts of the argument. His testimony should therefore be read as an informed investor’s account, with distinctions preserved between reported disclosures, portfolio anecdotes, and forecasts.

2. What the Episode Covers

Baker starts with an unexpectedly favorable reading of the Anthropic S-1 discussed on the show. He highlights approximately $10 billion of cumulative cash burn and describes cash generation and unit economics as encouraging. The interview does not walk through the financial statements, so those comments should not be expanded into a claim that the company has achieved overall net profitability. He also mentions roughly $67 billion when annualizing the reported third quarter. His separate reference to current ARR does not make those two measures interchangeable, nor does either establish completed annual revenue.

The more consequential observation concerns adoption depth. Baker cites approximately half the Fortune 500 using Anthropic while related spending amounts to about 0.1% of their expenses. He sees substantial room for expansion within existing customers. The hosts discuss organizations still figuring out deployment, suggesting that buying access and embedding AI in everyday work are different stages. The expense denominator is not explained, and the hosts move between Fortune 500 and S&P 500 references, limiting how precisely the population can be described.

A disagreement over concentration sharpens that distinction. A host invokes Ramp data suggesting that a small share of companies accounts for most AI spending. Baker responds with Anthropic account tiers: roughly 250 accounts spending more than $10 million and 11,000–12,000 spending more than $100,000. He finds that customer base more diversified than expected, while allowing that the wider AI market may look different. An economy-wide spending distribution and one vendor’s customer distribution can coexist; account thresholds alone do not provide a complete revenue-concentration calculation.

Baker also argues that growth extends beyond the leading proprietary model companies. He names Fireworks, Base Ten, Together, Modal, and Cognition as open-weight-oriented businesses growing faster than OpenAI or Anthropic, without presenting comparable growth tables. His Fireworks financing story illustrates the speed he has in mind: revenue rose so quickly during negotiations that the implied valuation multiple kept falling. It is an investor’s transaction example rather than a standardized industry comparison.

His product example is more personal. Baker calls Grokbot the most consequential AI product yet for himself and his firm, reports using about 17 Grokbots, and contrasts its simplicity with workflows he assembled using Claude Code. He identifies the benefit as routing the right information to the right person at the right time. A host adds the example of starting research late on a Sunday without disturbing employees. Neither supplies measured time savings or investment returns.

3. Core Views: Reasoning, Examples, and Limits

Baker’s answer to concerns about incoming GPU supply begins with prices and procurement. He reports rising rental prices during October and cites Silica Data founder Carmen Lee’s statement that compute had entered contango: forward access was expensive as well as spot capacity being tight. The concrete example is a portfolio company receiving a 4,000-Blackwell cluster in May or June 2026, then encountering substantially higher quotes for another cluster of the same size in September 2027. His reasoning is that buyers and sellers already know additional supply is planned, yet those procurement channels still price future access at a premium. This challenges an effortless-oversupply thesis, but does not establish a universal price curve. The episode does not normalize geography, service terms, GPU configurations, or delivery conditions, and a quote is not proof of future realized demand.

A second distinction separates operational competence from contract economics. Baker praises CoreWeave’s reliable clusters while arguing that low-priced, long-term fixed contracts restrict its participation in higher market prices. In his account, owning useful hardware does not automatically capture scarcity rents: the supplier may already have promised its output away. He prefers greater spot exposure and large customer prepayments in the prevailing market, suggesting paybacks well below a year. That is a conditional investor assessment, not a demonstrated return available to every operator. No full cost model accompanies it, and spot exposure also transmits falling prices to the seller more quickly.

The same caution applies to Baker’s retelling of Sarah Friar’s statement that earlier compute commitments could be resold at three to five times their price. It reinforces his interpretation that older contracts have gained economic value. However, the interview supplies neither an executed resale nor the contractual rights and obligations that would determine whether such value could actually be realized. A favorable mark on committed capacity and cash received from transferring it are different things.

Physical delivery provides another constraint. Baker compares operating a large GPU cluster to running a Formula One team: coordinated execution matters, and possession of land, power, and an offtake agreement is insufficient. He says some operators need three to six months after hardware delivery to energize their GPUs. His interest in the seven-rack preassembled DSX Pod follows from that delay. If it shortens commissioning, less capital sits in equipment that produces no revenue. The economic mechanism is clear, but the product’s realized improvement is not demonstrated in the episode. Announced hardware capacity should therefore remain distinct from stable, usable compute.

Baker then connects pricing and execution to financing. He argues that public listings and subsequent unlocking of SpaceX and Anthropic holdings could free mutual funds’ capacity for private investments under illiquid-asset constraints. Separately, he expects cloud contract repricing, greater spot exposure, and revenue growth to lift operating cash flow and fund more of the 2027 buildout than observers anticipate. These are separate channels: reallocating investment capacity is not the same as generating operating cash. His forecast requires the expected commercial improvement to become available cash, and he explicitly stops short of saying external financing has already peaked.

The strongest counterargument comes from Baker himself. Calling himself a responsible optimist, he describes searching for negative facts and highlights a bearish reading of a claim that Anthropic assigns 74% of compute to training. If training needs decline, substantial resources might enter inference, weakening the scarcity thesis. The percentage remains a speaker-cited figure, and the transfer is a scenario rather than an observed event. Nevertheless, it identifies a missing variable in simple supply forecasts: usable inference capacity can increase through reassignment of existing equipment, not only through new construction.

4. Learning and Application

For enterprise AI research, the interview suggests maintaining separate measures of customer reach, spending depth, account size, and revenue concentration. Record what each statistic measures before drawing a growth conclusion. A large customer count can coexist with limited deployment, while numerous substantial accounts need not establish a low share of revenue from the largest customers. Keep annualized quarterly revenue separate from ARR and completed financial-year revenue, retaining the calculation and source. This preserves Baker’s adoption argument without giving conversational estimates more precision than the evidence supports.

For compute procurement or infrastructure evaluation, build a delivery-to-cash comparison. Match hardware, delivery windows, and service conditions before comparing prices. Then examine fixed versus adjustable pricing, prepayment requirements, and the expected date of productive operation. Prepayments may relieve a supplier’s initial funding burden, but require buyers to commit cash sooner and assess delivery risk. Spot exposure offers participation in price increases while reducing revenue certainty if prices fall. Any proposed rapid payback should show its assumptions about equipment cost, power, operating expense, utilization, and commissioning delays. Hardware arrival should not silently become the first revenue date.

Stress tests should target the links in the optimistic argument. Consider training capacity moving into inference, forward quotes weakening, or clusters taking longer to energize. For each case, examine achievable rental revenue, utilization, and the resulting funding gap. Likewise, newly available mutual-fund investment capacity should not be booked as capital that a particular operator will receive. Baker’s stated willingness to change his mind becomes more useful when translated into explicit observations that would reduce confidence before they occur.

Agent adoption can begin with one information-routing task rather than a target number of agents. An illustrative pilot would specify a research input, an intended recipient, and a delivery deadline, then assess accuracy, timeliness, and required human review. This is an application derived from the interview’s mechanism, not a workflow Baker says he already operates. His own product preferences also depend on existing support: he acknowledges having an exceptional executive assistant, shaping his interest in personal-assistant tools. Teams with different staffing and responsibilities should test their own bottlenecks rather than copy his ranking.

There is also a useful separation between research breadth and portfolio commitment. Baker says his firm follows energy, industrial infrastructure, and natural resources without making them a large allocation, preferring areas where it believes it has an advantage. Adjacent sectors can supply important signals without automatically becoming investments. Following a connection to AI is the start of an inquiry; deciding whether the available evidence supports committing resources is a separate judgment.

Finally, possible social benefits should not substitute for evidence about timing or financing. When a host asks whether drug revenues may arrive too late to support current infrastructure costs, Baker offers an unnamed old-drug repurposing story. The episode identifies no drug, disease, or supporting study. That anecdote cannot establish efficacy, commercialization speed, or cash available for construction. Keep such opportunities as questions for further verification, rather than using an uncheckable success story to close a funding model.

Source

• Original episode: Gavin Baker on Anthropic's S-1, GPU Prices, and SpaceX