General Compute's Nvidia challenge rests on a chip loan of up to $400 million
The start up can borrow up to four times its Series A, secured against the hardware it buys. Its pitch is that chips win through deployed capacity, which needs a lender.
The loan is central to General Compute's financing. General Compute, a San Francisco start up that rents AI inference capacity, announced a $100 million Series A on 27 October. It can also borrow up to $400 million from the lender Upper90 to buy chips.
The loan ceiling is four times the round and funds the challenge to Nvidia. The company says so itself: financing comes before engineering.
The equity is $115 million in total, the $15 million seed from May plus the Series A. Blackbird and Abstract co-led the round, with Fuse Venture Partners, Village Global, Bold Capital, QED Investors and 645 Ventures among the investors. The loan ceiling is about 3.5 times that equity. A ceiling is what the company may borrow, not money it has raised.
A customer contract worth $100 million in annual recurring revenue paved the way for the round, TechCrunch reported on 27 October. The post money value is $800 million, against $60 million after the May seed.
General Compute has raised 115 million dollars of equity, 15 million in a May 2026 seed and 100 million in a Series A on 27 October. Its Upper90 debt facility allows borrowing of up to 400 million dollars, about 3.5 times the equity, and the first draw is the Cerebras agreement.
View data
| Capital item and status | Millions of dollars |
|---|---|
| Seed round, raised May 2026 | 15 |
| Series A, raised 27 October 2026 | 100 |
| Upper90 debt facility, ceiling secured July 2026 | 400 |
What the financing buys
The money buys hardware. General Compute buys systems from Cerebras, SambaNova and Positron AI, installs them in rented data centre space and runs them. The release says the cloud also offers GPUs alongside the new chips.
Customers pick one of two products. One is a guaranteed tokens per second service level, a promised speed of output. The other is dedicated bare metal servers.
The loan is not new. It dates from July and was tied to a SambaNova purchase. Inference means running a trained AI model to answer requests, and collateral means the chips secure the loan, so a lender can claim them if the borrower defaults.
The company's post on its Cerebras agreement explains the logic. "Chips do not win markets on architectural merit," it reads. "They win on deployed capacity, and deployed capacity is a financing problem before it is an engineering problem."
The fastest growing AI infrastructure companies are, it says, "extensions of NVIDIA's balance sheet". They get Nvidia equity, Nvidia GPUs as collateral and Nvidia purchase guarantees. A rival chip has no such channel, and General Compute says it was built to be one. The loan provides that financing channel, giving the lender control over access to it.
The lender and the resale risk
Upper90 is the lender taking that bet. TechCrunch reported in July that the loan might be the first to take inference specific chips as collateral. Chief executive Billy Libby says his firm financed Crusoe's GPU purchases in 2021, which he believes was the first loan against advanced chips. He says Upper90 was "compensated for the risk" then.
Traditional lenders avoided those deals because GPUs can lose value quickly. CoreWeave later turned chip backed loans into a business model, and the financing became common. GPUs are now comparatively well understood, TechCrunch wrote, so Upper90 is turning to inference. "We think open source models are going to be important," Libby said.
Resale value is central to the borrower's case. General Compute's chief executive, Finn Puklowski, told TechCrunch in July: "there's not too many buyers for them." The lender's collateral is only as good as the market for non-Nvidia silicon, so resale value is the number to watch.
Positron is the second hardware commitment. In a 20 October post, General Compute said it will buy Positron's Atlas systems, available from the first quarter of 2027. It has also committed to the next generation Titan systems, built on Positron's Asimov chip.
Positron's chief executive, Mitesh Agrawal, described the pairing as "debt financing for purpose-built inference silicon and a cloud offering that matches the hardware to the workload". The split is by stage. GPUs process the prompt, and memory focused accelerators generate the tokens.
SambaNova is the oldest relationship. General Compute built its May seed round around SambaNova silicon, according to TechCrunch. Its SN50 inference chips are power efficient and need no water cooling, so they install faster than GPUs in more kinds of data centre. SambaNova's chief executive, Rodrigo Liang, says "the real question is whether you can make money" selling speed.
The operating bill sits on top of the chip bill. The release says General Compute deploys the hardware into colocation space and operates it: cooling, networking, energy and software. Cerebras's chief technology officer, Sean Lie, says inference performance "depends on how the full system is deployed and operated".
Abstract's David Kwon shows how fast this came together. "When we first met, they were pitching an idea," he said. "Within months they had a big partnership with Cerebras and a debt facility of up to $400 million."
Power, speed and the launch deadline
Power is the next limit. Blackbird's James Palmer says: "A year or two ago, the bottleneck was silicon. Now it's power." General Compute has more than 50 megawatts of capacity in its pipeline. Palmer adds that "both power and silicon supply chains are slow to scale".
Speed gives buyers a reason to accept a new vendor. The company's own table assumes a coding agent that makes 400 model calls of about 300 tokens each. At 100 tokens per second each call takes 3 seconds, so the task takes 20 minutes. At 2,000 tokens per second each call takes 0.15 seconds, and the task takes one minute.
The table is the company's assumption, not a customer result. But a buyer who values that gap does not need General Compute to match Nvidia's scale, only to deliver the promised speed.
The Cerebras post calls the multi year agreement the largest single hardware commitment the company has made. Its first workload is agentic coding. Capacity opens to customers in the first quarter of 2027, as do Positron's Atlas systems.
That quarter ends on 31 March 2027 and is the first test of the claims. Paying coding workloads on Cerebras and Positron systems by then would show that the hardware securing the loan is earning revenue.