$100 million in annual recurring revenue. That was the number General Compute carried into its first institutional round — not a projection, a signed customer deal.
On October 27 the San Francisco company closed a $100 million Series A co-led by Blackbird and Abstract, at an $800 million post-money valuation. Five months earlier, after a May seed round, investors priced it at $60 million. That is a 13x step in under half a year, in a quarter when much of the neocloud sector was being written off.
General Compute does not sell GPUs by the hour. It buys inference silicon from Cerebras, SambaNova and Positron AI, drops it into leased colocation space, and sells tokens per second under one contract. The thesis: the next scarce resource in AI is not training compute. It is decode speed.
More than 50 MW of capacity in the pipeline, 75 MW of deployments planned for 2027, and roughly 200 MW of power access expected.
The bet is ASIC-first inference, not Nvidia-first capacity — and the binding constraint has moved from chips to power.
The bet against Nvidia's default
Most neoclouds buy the same accelerator everyone else buys, then compete on price and location. General Compute buys the chips Nvidia's customers usually cannot get financed: wafer-scale systems from Cerebras, memory-first accelerators from Positron AI, and SambaNova's dataflow hardware.
The architecture is split by workload. GPUs handle prefill, the prompt-processing stage that behaves like a burst. Purpose-built inference chips handle decode, the token-by-token generation that dominates latency for agents and coding tools. Customers see one endpoint, one SLA, and a routing layer that picks the silicon per call.
The Cerebras stack is the clearest example of what that buys. A CS-3 system puts model weights in 44 GB of on-wafer SRAM with 21 PB/s of memory bandwidth, which keeps decode off an external memory bus. Two systems fit in a 48U rack at 47 kW average draw and 55 kW rated. Cerebras quotes 1,000+ tokens per second per user on multi-trillion-parameter models.
Positron supplies the memory-heavy tier. Its Atlas systems arrive in Q1 2027, with a committed purchase of the next-generation Titan, built on Positron's Asimov chip. The LPDDR design carries terabytes of memory, aimed at holding larger models and longer context resident instead of streaming weights from HBM.
SambaNova was first. General Compute struck a $400 million deal to buy SambaNova chips using debt, which TechCrunch reported as the first debt financing of inference chips. That deal established the template the other partnerships follow: the cloud operator finances the hardware, the chipmaker gets deployed volume, and the customer avoids a capital purchase.
Why token speed became the product
The company's own benchmarks put a GPU cloud at 110 tokens per second per user and its heterogeneous fleet at 2,000. It also claims 2.8x throughput per rack at the same per-user speed. Those are vendor numbers, published without an independent audit, and they should be read that way.
The press release makes the measurement problem visible. Its headline promises results "up to 40x faster than GPUs"; the feature list four paragraphs later says "up to 20x faster." Both can be true against different baselines, but a buyer cannot tell which one applies to a given workload.
The demand argument is stronger than the benchmark argument. The release cites Goldman Sachs projecting 24x growth in inference demand by 2030, and Google's token demand rising 330x between 2024 and 2026. Agents call models repeatedly, so each task multiplies token volume rather than replacing it.
Everyone in this market is selling speed. The real question is whether you can make money doing it.— Rodrigo Liang, CEO and co-founder, SambaNova
That is the right question. Fast tokens matter only if the cost per delivered token clears the hardware, power and financing bill. General Compute's answer is utilization: route each call to the chip with the best economics for that stage, and keep every rack busy enough to pay for itself.
Power is the real balance sheet
General Compute does not build data centers. It rents space inside operator facilities, connects to smaller power sources, and converts former bitcoin-mining sites into AI halls. That choice keeps capital out of concrete and puts it into accelerators, which depreciate faster but can be deployed in months rather than years.
The capacity numbers trace the same logic. The company says 50 MW or more sits in the pipeline, expects to deploy 75 MW in 2027, and believes it will have access to about 200 MW of power. The gap between 75 MW deployed and 200 MW available is deliberate. Puklowski told TechCrunch the power is "more than we can probably fund and get hardware for."
The $400 million debt facility from Upper90, announced in July, is the bridge. Most of it had not been spent when the equity round closed. That matters because the constraint is no longer silicon availability. As Blackbird's James Palmer put it, both power and silicon supply chains are slow to scale, and finding creative routes through those constraints decides who wins.
The market has already priced a version of this idea. As we wrote in October, Hyperbolic raised $30 million on the bet that GPU capacity itself becomes a traded market. General Compute pushes the same logic one layer down: the tradeable unit is not a GPU-hour, it is a guaranteed token rate on whichever chip delivers it cheapest.
One customer, two financing instruments
The round's trigger was concentration, not diversification. A single large customer contract worth more than $100 million in annual recurring revenue gave investors a revenue line to underwrite. TechCrunch did not name the customer, and General Compute has not disclosed contract terms beyond the ARR figure.
That is a strength and a warning. A $100 million anchor proves someone will pay for the service at scale. It also means the $800 million valuation rests on one relationship renewing, expanding, and surviving the Q1 2027 deployment schedule.
The capital structure deserves equal attention. The $100 million equity round and the $400 million debt facility are separate instruments with different claims. Equity bought the buildout runway; debt bought the chips. Puklowski was blunt about the workload: "capital is going to be a full-time job for me, both on debt and equity."
For the chipmakers, the arrangement solves a distribution problem. Cerebras, Positron and SambaNova all build inference hardware that outperforms GPUs on specific workloads, and all of them need operators who can finance, deploy and support it. Palmer's framing was that they "need distribution partners who can take them to scale, who can navigate the complexity of the physical world."
What Q1 2027 has to prove
Nothing major is live yet. General Compute has run some customer workloads, but its first large-scale chip deployments land in the first quarter of 2027. Cerebras capacity opens then, Atlas arrives then, and the debt-financed hardware starts earning then.
Three tests follow. The first is schedule: every partnership names Q1 2027, which leaves no slack for power interconnection, cooling retrofits or chip delivery slips. The second is utilization: 75 MW of deployed capacity only works if the routing layer keeps expensive accelerators busy across workloads with different latency targets. The third is customer breadth. A second anchor contract of comparable size would turn a single deal into a repeatable model.
The neocloud bear case is not abstract. Elon Musk said more than 200 neoclouds face "guppy-level death and consolidation," Microsoft cancelled a multibillion-dollar deal with Nscale, and CoreWeave traded at a 12-month low. General Compute's counter is structural: it owns no data centers, buys chips others cannot finance, and sells a token rate rather than a server.
Does one anchor customer make a business?
Probability: 55% — the anchor deal proves demand, but the deployment schedule leaves little room for a second full-scale qualification cycle.
🟢 Optimistic scenario (30%)
Consequences: the $800M mark looks cheap, and the debt facility becomes a repeatable template for financing non-Nvidia silicon.
🟡 Base scenario (50%)
Consequences: the model survives, but the valuation compresses toward what a single-contract infrastructure business can support.
🔴 Pessimistic scenario (20%)
Consequences: debt service arrives before revenue, and the equity cushion shrinks faster than the pipeline grows.
The inference cloud is a financing business wearing a hardware costume. General Compute's advantage is that it understood this before it bought a single wafer-scale system.
What remains unproven is whether one $100 million contract is a market signal or a customer with unusual needs. Q1 2027 will answer that with deployed megawatts, not slides.