Article 337: General Compute Raised $100M on a $100M Inference Contract While CoreWeave Hit a 12-Month Low
It is a genuinely bad week to raise a neocloud Series A. CoreWeave, the publicly traded landlord of AI compute, is at a twelve-month low. Elon Musk is levering $40 billion of debt to buy capacity while predicting that 200+ neoclouds face "guppy-level death and consolidation".
General Compute raised $100 million anyway. Led by Blackbird Ventures and Abstract Ventures, at an $800 million post-money valuation - up from $60 million after a May seed round.
What they actually built
General Compute made two decisions that cut against the consensus.
First, it bought inference chips instead of GPUs. It struck a $400 million deal for SambaNova silicon - described as the first debt financing of inference chips - and has since added Positron and Cerebras to its supplier list. Everyone else chased GPUs because GPUs are what you can rent out at a known price. Inference accelerators are cheaper per token and useless for training, which makes them a bet on where the money eventually goes.
Second, it did not build data centres. It uses smaller power sources, rents space inside other companies' facilities, and converts ex-bitcoin-mining sites. That caps the upside - it will not out-build a hyperscaler - and removes the constraint that is currently strangling the sector. Puklowski expects 75 megawatts deployed in 2027 and access to roughly 200MW of power, "more than we can probably fund and get hardware for."
The demand side is a single customer worth $100 million in ARR, plus cloud companies that want bare metal with no software layer to resell to developers.
The bear case is honest
Finn Puklowski calls fears of a chip-led collapse "massively overstated" and argues Musk is talking his book to depress GPU prices. That is a coherent read of Musk's incentives and not a refutation of his point.
The real risk is not demand. It is financing: the company has bought chips on debt and now needs equity and debt on an ongoing basis. "Capital is going to be a full-time job for me," Puklowski told TechCrunch. A $100M ARR contract is a wonderful thing to raise against and a terrible thing to lose, and inference-chip supply chains are considerably less liquid than Nvidia's.
Why the contrast matters
CoreWeave rents GPUs to anyone. General Compute rents the cheapest possible tokens to specific inference clouds. The first is a commodity business with a financing structure that assumes GPU resale value stays high. The second is a narrower business that only works if inference - not training - becomes where the money is.
Every forecast says it does. The disagreement is timing, and the neoclouds that financed themselves on the assumption it would happen in 2026 are the ones the sector is now consolidating.
Key Takeaway: The neocloud shakeout is not a demand problem, it is a duration problem - too many companies borrowed against chips to serve a market that arrived slower than their debt schedule. General Compute is a bet that the smart move is to own almost nothing, rent everything, and buy the silicon nobody else wants to underwrite. That is either the shape of the next generation of infrastructure or the most elaborate way to be left holding depreciating hardware. The $800M valuation says investors are willing to find out.