General Compute raises $100M to rent out inference chips that aren't Nvidia's

Covers roughly 5am ET Tuesday, Oct 27 through early Wednesday, Oct 28. Tuesday's money went to the plumbing of AI: a cloud built on non-Nvidia inference chips, a data-analytics company rebuilt for agents, and what Imperial College calls a record UK robotics seed. Fifteen venture rounds in the window cleared our checks, plus two Oct 26 rounds the last Daily Brief missed (marked "catch-up"). Every round below is confirmed by the company, its lead investor or both. Figures companies report about themselves are labeled company-reported.

General Compute, a San Francisco inference cloud, raised a $100M Series A co-led by Blackbird and Abstract. Fuse Venture Partners and Village Global joined, along with Bold Capital, QED Investors, 645 Ventures, Carbide Ventures, Uncorrelated Fund, Endeavor Catalyst, Icehouse Ventures, MatterScale, Carya, Plug and Play and angels (General Compute · release via WebWire · TechCrunch). The model is simple to state and hard to run: the company buys GPUs as well as specialist inference chips made by Cerebras, SambaNova and Positron AI, installs them in colocation space, operates them, and sells the output either as dedicated bare-metal servers or under a guaranteed tokens-per-second service level. Software routes each workload to the chip that suits it. The equity sits on top of a previously announced debt facility of up to $400M from Upper90. First major deployments are due to go live in the first quarter of 2027, and the company says more than 50 MW of capacity is in its pipeline (company-reported). Finn Puklowski (CEO) and Jason Goodison (CTO) founded it. The company did not state a valuation. Why it matters for investors: Specialist inference chips need someone to finance, rack and sell them, and General Compute wants to be that channel. Positron's CEO singled out that pairing of debt financing for inference silicon with a cloud that matches hardware to workloads. The economics rest on three things the round alone doesn't prove: utilization once capacity goes live in 2027, the cost of the debt behind the hardware, and the roadmaps of its chip suppliers. Diligence questions: Do customer contracts run as long as the loans? Who absorbs the loss if a supplier's chip falls behind or its maker struggles? How much of the 50 MW pipeline already has signed power and signed demand?

Historical lens. Debt-financed compute already has a public template. CoreWeave's IPO registration statement, filed March 3, 2025, says the company began in September 2017 as The Atlantic Crypto Corporation LLC, took the CoreWeave name in December 2019, and ended its mining business, which mined mainly Ethereum, on September 30, 2022, citing changes to Ethereum's reward policy (CoreWeave S-1). In July 2023 a CoreWeave subsidiary signed a delayed-draw term loan of up to $2.3 billion to pay for part of its computing equipment, secured by that subsidiary's assets and a pledge of its equity. The filing puts the average rate on that borrowing at 14.11% at the end of 2024. A second facility of up to $7.6 billion followed in May 2024. Growth came with concentration: Microsoft supplied 62% of CoreWeave's 2024 revenue. Our interpretation, not established fact: General Compute is applying a similar structure, venture equity on top and chip financing underneath, but with Cerebras, SambaNova and Positron hardware rather than Nvidia's. What lenders charge against that less familiar collateral could set its pace as much as demand does.

Hex, the data-analytics company founded in 2019, closed a $100M financing at a $1.2B valuation led by Peak XV and 137 Ventures, with Amplify, Avra, a16z, Redpoint, Sequoia and Zero Prime participating. Hex did not name a round stage (Hex). The money came with a product reset. CEO Barry McCardel writes that after its best-ever quarter the company started over on a clean-sheet version of the product, which more than 2,500 customers use (company-reported). The result, Hex 4.0, treats every analysis as plain code that teams own, and adds context and controls meant to keep work trustworthy whether it happens inside Hex or in outside agents such as Claude, Codex or Cursor. McCardel says agents wrote all of the code for the rewrite, guided by Hex's engineers (company-reported). He writes, half-jokingly, that early this year investors expected most software to die. Why it matters for investors: This is a useful counterpoint to the "AI eats SaaS" trade, and also a warning. A six-year-old company with more than 2,500 customers concluded that adding AI to an architecture set in 2020–21 was heading for diminishing returns, and rebuilt instead. Investors in pre-2023 software companies should ask the same question of their own holdings: does the data model work for agents, or only for people clicking through screens? The new risks at Hex are execution risks: moving 2,500 customers onto a rewritten platform without churn, and earning a $1.2B price with a product customers have only just started to see. Falling cost of rewrites, if agent-written code is as productive as Hex says, cuts both ways: it helps incumbents rebuild, and it helps new entrants copy them.

Machenta, a London-based Imperial College London spinout built on work from the university's Robot Learning Lab, raised a $40M seed co-led by Creandum and Union Square Ventures (Creandum · Imperial College London · Tech.eu). Imperial calls it the largest seed round ever raised by a UK robotics startup. CEO Edward Johns, an Imperial professor whose field is robot learning, founded the company with his PhD student Vitalis Vosylius and colleagues from the lab. Their foundation model is pre-trained only on simulated data and is meant to let a robot perform a new task after one demonstration, with no further training (company-reported). Creandum writes that the lab's research included a robot picking up 1,000 everyday tasks, each from one demonstration, in less than a day, and that its partners tested the system with their own props during diligence. The money goes to scaling the model and growing the team at a new office and lab in London's King's Cross. Why it matters for investors: Robot-learning startups are splitting into two camps: those buying ever more real-world demonstration data, through teleoperation or filmed chores, and those betting simulation can replace most of it. Machenta is a pure simulation bet, so its costs scale with compute rather than with data collection. At $40M, this seed prices a research team and a technical thesis, with no customers disclosed. Diligence questions: How well does one-shot learning hold up on cluttered, real production lines? Is the business model a licensed model, software per robot or full deployments? Which robot makers will integrate it first?

Round notes

Not investment advice. The AI Cap Table is published for information only and is not investment, legal, or tax advice. Nothing in this newsletter is an offer or solicitation to buy or sell any security. Funding data comes from public sources we cite and verify, but errors are possible; do your own due diligence and consult a qualified adviser before making any investment decision.