Gavin Baker on Anthropic's S-1, GPU Prices, and SpaceX

Summary

• Baker reads Anthropic’s S-1 as an upside surprise with substantial enterprise headroom. He points to $10 billion burned, plus encouraging profitability, cash generation and unit economics. About half the Fortune 500 has adopted Anthropic, yet it represents just 0.1% of their expenses—suggesting diffusion is “maybe earlier than I thought we were.”
• Anthropic’s disclosed customer mix challenges the concentrated-spending bear case. Baker cites only 250 accounts spending over $10 million, while a host highlights 11,000–12,000 spending over $100,000. Baker sees a “very diversified spending base,” while allowing that Ramp’s concentration data might still describe AI spending overall.
• Open-weight infrastructure companies are outgrowing OpenAI and Anthropic, according to Baker. Fireworks, Baseten, Together, Modal and Cognition underpin his argument that demand is broadening beyond the major labs. Atreides led Fireworks’ latest round, where growth meant “the multiple was effectively going down” daily. Growth beyond Cursor is particularly notable as SpaceX insources Cursor’s compute.
• GPU pricing signals tightness through 2027 despite the announced supply pipeline. Baker reports daily rental-price increases during October and expensive forward capacity: portfolio companies seeking another 4,000 Blackwells for September 2027 face sharply higher quotes than for their May–June 2026 clusters. “We’re not in backwardation anymore.”
• Compute returns increasingly depend on execution and retaining pricing upside. Baker praises CoreWeave’s operations but says substantial capacity was contracted far below today’s market. A host argues that more spot exposure and prepayments can produce paybacks “well under a year.” Running clusters resembles “a Formula 1 team”; Baker says a new Jensen product with preassembled racks could shorten deployment and reduce working capital tied up in idle GPUs.
• Training demand creates a potential inference-supply risk within Baker’s bullish framework. A bearish argument he considers compelling is that Anthropic dedicates 74% of its compute to training. That suggests room for more inference revenue, but if training needs decline, capacity could enter the inference market.
• IPOs and repriced cloud contracts could reduce the build-out’s dependence on external financing. SpaceX and Anthropic becoming public and fully unlocked could restore mutual funds’ capacity for private investments. Higher cloud pricing could also lift operating cash flow enough to fund more of 2027’s infrastructure internally; Baker does not know whether external financing has peaked, but says, “I think we’re getting close.”

Deep dive

1. Anthropic’s S-1 pairs enormous scale with low enterprise penetration

• Baker calls the S-1 “pretty clearly an upside surprise”: $10 billion burned, with encouraging profitability, cash generation and unit economics. About half the Fortune 500 has adopted Anthropic, yet it represents 0.1% of their expenses. Enterprise diffusion may be “earlier than I thought we were.” Baker also cites roughly $67 billion annualized if the reported third quarter is annualized.
• A host argues that concentrated AI spending is natural among early obsessives, citing Ramp data suggesting 1% of companies account for 90% of spending. Baker disagrees about Anthropic’s customer mix: only 250 accounts spend over $10 million, while a host cites 11,000–12,000 spending over $100,000. Baker calls the S-1 more authoritative and says it shows a “very diversified spending base,” while allowing that Ramp might still describe AI overall.
• Baker’s broader demand signal is open-weight infrastructure: Fireworks, Baseten, Together, Modal and Cognition are growing faster than either OpenAI or Anthropic. He says that was true during OpenAI’s summer acceleration and remains true amid Anthropic’s October reacceleration. Baker says Anthropic lost share in summer and is now gaining it. He doubts it would leave the October ARR blank in the S-1 without expecting a strong number, while hoping the market moves beyond monthly ARR figures.
• A host says private-market investors on the West Coast already know these companies are at scale and growing, while the news is less familiar on the East Coast. Atreides led Fireworks’ latest round, and Baker says growth made pricing difficult: “The multiple was effectively going down” every day. He finds inference-cloud growth beyond Cursor especially striking because SpaceX has been insourcing Cursor’s compute. His assessment is that the ecosystem has accelerated and diversified.

2. Grokbot wins Baker through ease of use and information routing

• A host asks whether Cursor’s product and distribution can help Grokbot despite xAI models not stunning on a particular benchmark. Baker answers as a user: Grokbot is “the most impactful AI product” for him and his firm, and he has about 17 Grokbots. He had kludged together a setup with Claude Code and says that may have been user error.
• Baker describes investing as “making sure the right information gets to the right person at the right time,” and says Grokbot is very good at that. A host adds the Sunday-night research example: an agent can spend a million tokens investigating a market-size question without asking an employee to work for six hours.
• Personal agents have not displaced his assistant Susan. “I guess I have executive assistant privilege,” Baker says; he has downloaded Instinct and played with Muse, but says Grokbot has been transformative for his professional use cases.

3. Compute scarcity extends into 2027, but operators capture different economics

• Baker says October rental prices for essentially every kind of GPU have risen daily. Silicon Data, in which Atreides is an investor and which Baker calls the most authoritative source, says compute is now in contango: forward capacity is expensive too. Portfolio companies with 4,000 Blackwells for May or June 2026 are shocked by the cost of adding another 4,000 in September 2027. Despite supply coming online, Baker says it is hard for them to find compute well into 2027: “we’re not in backwardation anymore.”
• Baker takes Sarah Friar’s comments literally: long-term compute commitments signed before May 2026 are “way in the money.” A contract buying compute for perhaps $10 or $15 a watt could be resold at a large premium; Friar described a “3, 4, 5X premium.”
• Baker says CoreWeave runs GPUs well but locked in substantial compute under long-term contracts now far below market. A host explains that CoreWeave sold at fixed prices and argues that more spot exposure and fewer long-term offtake agreements could improve the economics. The host says selling into the spot market could mean paybacks well under a year and large prepayments.
• The host compares running a GPU cluster to running a Formula 1 team, requiring experienced operators and coordinated execution. Baker agrees that the key is bringing on GPUs, energizing them and running them well. He names CoreWeave, Crusoe, Nebius, SpaceX and the big hyperscalers as companies with a track record of execution, while saying the future remains uncertain.
• Baker points to a new Jensen product: seven racks, preassembled, that you “just literally plug in and it works.” He says some operators otherwise take three or six months to energize GPUs; at this scale, six months could leave hundreds of billions of dollars in assets generating no revenue. Shortening deployment would take working capital out of the market.

4. Responsible optimism includes a training-driven supply risk

• Baker says he treats X as an exercise in “cultivated equanimity,” searching criticism for useful bearish arguments. One he finds compelling is that Anthropic dedicates 74% of its compute to training. That could indicate room for more inference revenue, but if training needs decline, substantial capacity could enter the inference market. “I’m not on X to be right”; he says he is trying to learn and get to the right answer.
• A host raises the timing concern in biology: infrastructure costs arrive now, while drug revenues may take years. Baker says “we’ll see how long it takes,” and points to Montana’s new-drug program and the possibility of using existing compounds for new diseases. He says AI has helped identify a drug that has existed for 50 or 60 years and can effectively cure a rare disease that devastates 10,000 children a year. A friend did the work, but Baker says it is the friend’s story to tell.
• Baker says AI’s public-relations effort is “-5 out of 100” and the industry needs to tell more stories. His public case includes lives saved, benefits for blue-collar Americans and small towns, reindustrialization, and increased electrical generation: “You cannot reindustrialize without power.” He recounts a skeptic demanding data centers in America’s wealthiest county, then points to Loudoun County, which Baker says is the wealthiest and has the highest per-capita number of data centers.

5. Commodity research is broader than Atreides’ portfolio exposure

• Asked how far he follows energy infrastructure and commodities, Baker says he pays close attention. He ran a diversified fund for nine years and previously invested in energy and natural resources. He has not made those areas a large part of the portfolio, but says investors need to cast a wide research net and focus capital where the firm has a competitive advantage.
• Baker says gold has long had a place in portfolio construction. He finds it amusing how strongly some gold bulls react to asteroid mining, calling it a challenge to their worldview; a host compares the possibility of technological disruption to lab-grown diamonds.
• His concrete private-market exposure is a company with a chemical process for refining rare earths, rather than acquiring mines. “The problem with rare earths isn’t that they’re rare”; refining them is dirty. Baker also credits DARPA’s efforts and says investors never know where a weak but important signal will come from.

6. IPOs and cloud cash flow could loosen the capital constraint

• Baker argues that SpaceX and Anthropic IPOs can unlock another pool of private capital. American mutual funds face limits on illiquid holdings; when those companies are public and fully unlocked, funds regain room to buy private assets. Baker names Capital Research, T. Rowe Price, MSAM, GSAM and Franklin Templeton as existing participants.
• Sovereign funding is already substantial, particularly in the GCC and UAE. Baker singles out MGX and Mubadala, calling the UAE “extremely front-footed” on AI; Atreides and Valor co-led Crusoe with Mubadala.
• Baker expects aggregate hyperscaler and neocloud revenue growth to keep accelerating as underpriced contracts reset and more compute gets spot exposure. Revenue is dropping through at high margins, lifting operating cash flow and potentially financing more of the 2027 build-out than people expect. His hedge matters: “I don’t know if we’ve seen the peak of external financing,” but “I think we’re getting close.”
• A host suggests public ownership could moderate polarized AI sentiment as direct holders grow from perhaps 50,000 people to millions. Baker favors direct ownership over a foundation whose board or bureaucrats allocate the money. He says every child in America will have a SpaceX share and encourages more companies and wealthy people, including Michael Dell and Brad Kirshner, to support broad participation: “Give everybody a stake in capitalism.”