So the base commitment is about twice every dollar Figure has ever raised. The stretch goal is roughly three times.

Let's start with what didn't happen. No money moved this week. The first GPUs don't land until the second half of 2027, on hardware — Nvidia's Vera Rubin platform — that is only now beginning to ship in volume . This is a multi-year commitment, not a completed order with a massive wire transfer.

The interesting thing is that Barstow, Texas is not a new campus. It's the Cedarvale facility in Ward County that is owned by Ionic Digital, a former bitcoin miner. Nscale leased the whole site last October in a 10-year, roughly $2 billion triple-net lease. And the site's announced purpose was a different customer's deployment: Microsoft, deploying roughly 104,000 Nvidia GB300 GPUs with deliveries starting this quarter.

Cedarvale currently runs about 234 to 240 megawatts: four 50-megawatt buildings originally built to house roughly 66,000 bitcoin miners.

Now do the math on Figure's allocation. By my count, a hundred thousand Vera Rubin GPUs is about 1,389 NVL72 racks. At a reported 190 to 230 kilowatts per rack, that's 264 to 320 megawatts of compute load. Call it 340 to 420 megawatts with cooling and facility overhead.

That is more than the entire current capacity of the site, on top of an existing Microsoft commitment .

In addition, Microsoft holds an option on a further 700-megawatt phase at Cedarvale starting in late 2027, and the total campus could then theoretically reach 1.2 gigawatts. Note: Microsoft has walked away from a non-binding letter of intent for an Nscale facility before: 1.35 gigawatts at Nscale's West Virginia campus. Google was in talks for the same site and walked too; Anthropic eventually took it , for $45 billion.

Meanwhile, Texas paused its data center interconnection reviews in August after requests in the ERCOT queue reached 474 gigawatts — roughly five times the state's record peak demand, and about 90% of it data centers, per Governor Greg Abbott. Nscale has not disclosed how Cedarvale is powered, and I could find no interconnection filing or county permit record for the expansion this deal implies.

But there's another gap. You don't just need a data center and power and cooling. You also need the things that you need cooling for … the chips themselves.

Morgan Stanley analysts have pegged a single Vera Rubin NVL72 rack at roughly $7.8 million, with more than $2 million of that in memory alone. Other estimates run to $8.8 million as HBM4 prices climbed. Multiply by 1,389 racks and you get $10.8 billion to $12.2 billion in hardware before land, building, power infrastructure … or any margin for Nscale.

So $3.5 billion doesn't buy 100,000 GPUs. It buys somewhere between 32,000 and 45,000 of them. Even the $6 billion maximum potential commitment covers only about half the GPU fleet the announcement describes.

So "$3.5 billion of compute" may describe rental spend over a term rather than capital equipment. Or the 100,000 figure may be a ceiling reached only well past $6 billion. Or Nscale may be getting volume pricing analysts can't see.

It gets more complicated: the Nscale IPO

If you thought that was already pretty complicated, buckle up. Nscale is reportedly targeting a U.S. IPO listing as early as this month, with Goldman Sachs and JPMorgan advising. The company is looking to raise as much as $3 billion at something like a $25 billion valuation. (There is no public filing yet.)

The IPO pitch deck depends on order backlog and future projected revenue, of course. Nscale's contracted revenue was reported at roughly $51 billion in early August. By this week it was reported at $103 billion, roughly doubling after the Anthropic deal. Meanwhile, Nscale's actual recognized revenue was about $33 million for all of 2025, about $37 million in the first quarter of 2026, and just over $100 million in the second.

A person familiar with the discussions told Reuters the $103 billion figure was "illustrative and not intended as formal revenue guidance." Fortune, in June, contrasted CEO Josh Payne's claim that "every dollar is backed by real revenues" against the only public accounts on record: a $24 million loss for the seven months to December 2024.

Figure's $3.5 billion arrived on that pile days before an IPO roadshow.

About $62.5 billion of the $103 billion has a named customer attached: Anthropic, Microsoft, and now Figure. The remaining $40 billion or so belongs to unnamed "AI-native clients."

Of course, Nvidia is involved

Bloomberg reported today that Nscale is seeking $3.5 billion in financing ahead of its IPO: up to $1.5 billion in convertible notes and – ding ding ding – roughly $2 billion from Nvidia. Goldman Sachs is reportedly running it, and the notes convert at a double-digit discount to the IPO price, capped at a $30 billion valuation.

That is one day after Nscale announced it would supply Figure. And the pre-IPO raise is $3.5 billion — the identical figure Figure committed to spend.

So here's the full circuit.

Nvidia invested in Nscale's $1.1 billion Series B, and is now reportedly putting in roughly $2 billion more ahead of the listing. Nscale buys Nvidia GPUs: about 200,000 GB300s contracted for Microsoft, up to 100,000 Vera Rubin chips earmarked for Figure. Nscale takes an equity stake in Figure, size undisclosed, and becomes its preferred compute provider. Figure, which counts Nvidia among its own investors from Series B and Series C, commits to buy Nscale compute. That commitment becomes backlog, the backlog underwrites the IPO and the debt, and the proceeds buy more Nvidia GPUs.

So Nvidia is simultaneously on both ends of the transaction and in the middle of it.

Jensen Huang described the goal of this project in a quote in the Figure announcement : "This is the physical AI flywheel that will accelerate the path from models to robots in the world."

Short seller Jim Chanos put it less warmly in July, on a different Nvidia arrangement: "So we are at the point in the cycle where $NVDA has to provide financing guarantees for roughly 2/3rds of the cost of the chips it is selling to the data center project…?! Lol, ok."

Huang has called the circular-financing characterization ridiculous, but clearly, this is at minimum a very creative way of doing business, and at maximum, balance-sheet-boosting shenanigans.

These circular arrangements are getting more common across the industry, and they're no longer buried in footnotes.

AMD, Nvidia's chief rival in AI chips, issued OpenAI a warrant last October for up to 160 million AMD shares — roughly 10% of the company — alongside a deal to deploy six gigawatts of AMD hardware. The equity runs the other direction in this case, from supplier to customer, but the principle is the same: customer and supplier end up owning pieces of each other.

And CoreWeave, a publicly traded AI cloud provider that does more or less what Nscale does, disclosed in an SEC filin g that Nvidia agreed to buy $6.3 billion of its unsold compute capacity through April 2032. Nvidia is a CoreWeave investor and a CoreWeave customer, and it is now the buyer of last resort for capacity CoreWeave can't sell, which is essentially Nvidia guaranteeing demand for the output of its own chips.

What has to be true

For $3.5 billion of compute to be payable out of operations rather than out of the next funding round, Figure needs a deployed, revenue-generating fleet in the tens of thousands, at industrial pricing it has never publicly disclosed, within about two years of a standing start.

That's a tall order.

Figure's most mature deployment is BMW's Spartanburg plant , where a Figure 02 ran about 1,250 hours over roughly 11 months and moved more than 90,000 sheet-metal parts. (It's worth noting that BMW chose a competitor, Hexagon's AEON, for its Leipzig deployment in February.) Figure also announced a distribution-center agreement with Catalyst Brands in May of this year.

BotQ, Figure's factory, targets a build capacity of just 12,000 robots a year. The company will need to expand quickly: Adcock has talked about building 100,000 robots over four years, and has said that the soon-to-come Figure 04 revision is likely the robot that Figure can scale production with.

Maybe all of this happens. Adcock has built and sold companies before, and he raised $700 million at a $6 billion valuation this May for an entirely separate AI hardware startup, which suggests he knows a thing or two.

But it's also possible that a pre-revenue robotics company and a pre-revenue cloud company agreed to point at each other with large numbers, weeks before one of them goes public.

I've asked Figure and Nscale whether the $3.5 billion is binding or a letter of intent, how much equity Nscale took, and how 100,000 GPUs share a campus with Microsoft's 104,000. I'll update with any answers.

This article was originally published on Forbes.com