Product1 publisher3 min readPublished
The Information now puts the raise at $5bn to $6bn on at least $40bn pre-money, with Nvidia writing roughly half the cheque, which makes this a compute-supply arrangement for anyone buying Tinker cluster time.
The Product Desk · Product desk
Compiled by The Product DeskSomething wrong?How this is made
The person with a problem here is whoever pushed a fine-tuning run onto Tinker last quarter and now owns the vendor-risk paragraph. A limited partner at Andreessen Horowitz, by contrast, was briefed on the round and has a spreadsheet for this sort of thing.
The money into the lab arrives at one place. Inkling, the 975-billion-parameter open-weight model shipped in July, earns nothing on the weights [8]. Tinker, generally available since December 2025, rents GPU clusters for training and fine-tuning: the customer keeps its data and algorithms behind an API, Thinking Machines runs the hardware and bills for the time [9]. Revenue is metered silicon. Nvidia would put up roughly half the round, near $2.5bn to $3bn [3].
The arithmetic is small enough to do in the margin. A raise at the top of the range puts post-money near $46bn [19], so $3bn is about 6.5 percent of the company, and $2.5bn into a $45bn post-money is about 5.6 percent [21], before counting whatever Nvidia already holds from the $2bn seed at $12bn post-money [17] and from its undisclosed March investment [7]. Accel, in from the seed, is in talks to lead [2]. Everyone else in the syndicate is being asked for the other $2.5bn to $3bn [22], against a partnership to deploy at least a gigawatt of Vera Rubin systems with first deployment targeted for early 2027 [7]. RuntimeWire made the point plainly under a heading saying a billion dollars does not buy a gigawatt [24].
What the reporting does not say is that the equity is conditioned on hardware purchases, and neither party disclosed the March cheque [7]. So the circularity shows up in the shape of the deal, not in any disclosed term. Nvidia describes its role as backing American open-source AI developers [11]; two days later it turned up as a backer of Nscale's pre-IPO financing [12].
The price is harder to hold. Stephanie Palazzolo puts annualised revenue in the hundreds of millions, which RuntimeWire notes is self-reported through people close to the business and unaudited [13], while TechCrunch's source said more than $100m [14]. That is why the multiple is a band of roughly 80 to 400 times revenue [15], which implies revenue somewhere between $500m and $100m [23]. Amir Efrati called the price lower than the company wanted but still high for a lab this young, especially as a multiple of revenue [16]. The company had asked $50bn or more late last year and did not close it, partly because people were leaving, co-founder Andrew Tulloch among them [18][20].
For a buyer, two questions fit on one card. Does the same counterparty appear on your vendor's cap table and in its bill of materials. Is the revenue figure you are pricing continuity against audited by anyone. Here both answers are unhelpful, and the mitigation is portability rather than reassurance: Tinker supports outside model families as well as its own [10], a capability worth exercising rather than just admiring. Running one real job on someone else's cluster costs a few hours of compute, and it is a cheaper way to learn your switching cost than finding out during an allocation squeeze.
Ranked by verification strength, evidence, and original report placement.
The Information's current version of the story puts Thinking Machines' raise at $5bn to $6bn, on a pre-money valuation of at least $40bn.
Accel, an investor since the seed round, is in talks to lead the new round.
Nvidia would supply roughly half the capital in the round, somewhere near $2.5bn to $3bn.
Against $40bn pre-money, the revenue multiple lands somewhere between roughly 80 and 400 times, depending which end of 'hundreds of millions' is accurate.
The Information published on 3 September and then updated the story; the early version put the raise at more than $1bn, and the new version is five to six times that and names Nvidia.
TechCrunch reported the story on the afternoon of 3 September, before the update; RuntimeWire carried the same figure that day; Tech Funding News built a full analysis on the smaller figure on 4 September, a day after the update landed.
Publishers with included, body-backed reporting in this cluster.
1 article · September 6, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
build
1.5% of Hugging Face repos take 99.2% of downloads, and the ceiling is Chinese1 publisher
product
Nvidia now carries $99bn of its own customers on its balance sheet1 publisher
build
Thinking Machines will pay $50,000 in its own credits to see if Inkling breaks under fine-tuning1 publisher
build
SGLang's one-GPU Qwen3.8-27B recipe is the useful half of the release1 publisher
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One relay, a backer's own briefing at the end of it
Every dollar figure here comes down a single chain: The Information's updated report, sourced to Andreessen Horowitz telling its limited partners about a company it already owns, retold by The Next Web. Efrati and Palazzolo are named, which is more than most financing rumours offer, but nobody at Thinking Machines, Nvidia or Accel is quoted, no filing exists, and the revenue that anchors the multiple has never been audited.
Product in market, demand undisclosed
There is real machinery behind the round. Tinker has been selling cluster time since December 2025, Inkling's weights are freely available, and a gigawatt of Vera Rubin capacity is nominally booked for early 2027. What no one has published is how much Tinker anybody is actually buying; the hundreds-of-millions annualised figure is the only demand signal in the story and it comes from the company's side of the table.
Raise understated in circulation, price still ahead of proof
The error in circulation ran the unusual direction: everyone who copied the pre-update $1bn understated the raise by a factor of five or six, so the news was quieter than the facts. Pricing tilts the other way. $40bn pre-money against unaudited revenue somewhere between $100m and $500m works out at 80x to 400x, and roughly half the cheque comes from the vendor whose hardware that revenue runs on, which makes the valuation less of an independent verdict than it looks.
Both named parties talking up their own book
Andreessen Horowitz led the $12bn seed and briefed its own limited partners on the new price, which is the only reason the number reached print. Nvidia's half of the round funds a buyer of Vera Rubin systems and it surfaced again two days later inside Nscale's pre-IPO financing, so the same capital keeps appearing on both sides of the same sale while being described as support for American open-source developers.
Direction firm, arithmetic provisional
Two points hold up under scrutiny: the round is in talks at a level far above what most write-ups reported, and Nvidia is meant to cover about half of it. The parts a price depends on stay soft. The revenue denominator is unaudited, Accel's lead is unconfirmed at a level its peers declined ten months ago, and the pre-money-versus-post-money distinction alone moves the headline number by $6bn.