Gavin Baker on Anthropic's S-1 surprise, compute market going into contango, and why AI diffusion is still early

Gavin Baker on Anthropic, compute contango, and why enterprise AI diffusion is earlier than expected

Gavin Baker, partner at Atreides Management, makes a straightforward case: the Anthropic S-1 was an upside surprise, enterprise AI adoption is earlier than most appreciated, and the compute market has quietly moved into contango.

Anthropic S-1

Baker's read on the filing is unambiguously positive. The company has burned only $10 billion to reach $6.7 billion annualized revenue, with what he describes as encouraging unit economics and cash generation. The customer base is more diversified than skeptics expected — roughly 11,000 to 12,000 accounts spending over $100,000, and only 250 accounts spending more than $10 million, which undercuts the narrative that a handful of hyperscalers are carrying the whole number.

What catches him off guard is the penetration figure. Roughly half the Fortune 500 has adopted Anthropic, yet AI spending represents just 0.1% of those companies' expenses. Baker reads that not as a ceiling but as an indication that diffusion is earlier than he thought — "maybe earlier than I thought we were."

He also notes that Anthropic left a blank space for October ARR in the filing, and argues the company wouldn't have done that unless it knew the number would be strong.

“I do think the Anthropic S-1 was pretty clearly an upside surprise. The fact that they've only burned $10,000,000,000 — the profitability, the cash generation, the unit economics, they're all very encouraging. Whatever the number is, 67 billion annualized... To be that big with such low penetration into the Fortune 500 is crazy. October was a strange month because the price to rent essentially any kind of GPU has gone up every day — we're not in backwardation anymore.”

Compute contango

The more consequential signal Baker flags is what's happening in the forward market for compute. The spot price for GPU capacity has risen every day through October. More importantly, Carmen Lee, founder of SiliconData (an Atreides portfolio company), told GTC that the compute market is now in contango — future delivery costs more than spot, even with the widely-discussed supply additions coming online.

Atreides portfolio companies trying to secure clusters for mid to late 2027 are finding pricing dramatically higher than they expected, despite full awareness of incoming supply. Baker argues this is not legible to public markets even though the evidence is publicly available — Sarah Fryer (OpenAI CFO) has said OpenAI could sell its compute commitments at a 3x to 5x premium today, and that any long-term contract signed before May 2026 is deeply in the money.

CoreWeave is the cautionary example: excellent operators who locked in long-term offtake agreements at prices now well below market. Baker says CoreWeave is aware of the problem and is actively trying to shift toward more spot exposure. SpaceX's approach — shorter contracts, willingness to sell into the spot market, large prepayments — is closer to the right model for the current environment, and Baker thinks it implies paybacks well under a year.

Operator quality as a constraint

Baker frames GPU cluster operation as roughly analogous to running a Formula One team — a high-execution physical challenge that many new entrants underestimate. He names CoreWeave, Crusoe, Nebius, and SpaceX as companies with proven track records. Many others with land and power commitments lack the operational depth to energize clusters quickly, and the tightening market is functioning as a forcing mechanism to concentrate capacity with competent operators.

Nvidia's DSX Pod announcement at GTC Berlin — a pre-assembled, seven-rack unit designed to shorten energization timelines — matters here. If energization time drops from the current three-to-six months, hundreds of billions in assets currently sitting idle start generating revenue faster, which Baker views as a meaningful working-capital improvement for the industry.

Open-weight inference clouds

Baker argues the inference cloud layer is growing faster than either OpenAI or Anthropic. He cites Fireworks AI, Base Ten, Together, Modal, and Cognition as companies at scale that are predominantly open-weight-first and all accelerating. Atreides led Fireworks' last round, and Baker says pricing was difficult because revenue was growing fast enough that "yesterday's price is not today's price" by the time the deal closed.

Financing the 2027 build-out

Baker expects the next wave of capital to come partly from an unexpected source: American mutual funds. Firms like Capital Research, T. Rowe Price, Franklin Templeton, and Goldman Asset Management are currently constrained by liquidity rules on the percentage of illiquid assets they can hold. When Anthropic and OpenAI go public, those positions get unlocked, freeing up buying power that redeploys into the next generation of private AI companies.

Gulf sovereign wealth funds — particularly MGX and Mubadala in the UAE — have already been active. Atreides and Valor co-led the Crusoe round with Mubadala. But Baker's most concrete near-term claim is that the build-out itself may become increasingly self-financing: as hyperscaler cloud contracts reprice to market and spot exposure increases, operating cash flow accelerates, and he thinks the industry may be approaching peak reliance on external financing.

The bear case he takes seriously

Baker singles out one bearish argument worth engaging: Anthropic is dedicating 74% of its compute to training. That's bullish if it means inference capacity is being held back and revenue could be much higher. It's bearish if anything reduces the compute required for training, because a large volume of supply could suddenly hit the inference market. He credits an LBO analyst on X with surfacing the point and says it's the kind of negative signal he actively looks for.

His overall posture is that he called Anthropic losing market share in the summer and is now calling a reacceleration in October. He's not forecasting smooth sailing — he's tracking the data and updating.