Figure has disclosed just over $1.6bn of equity across two rounds, a $675m Series B in 2024 whose investors included Microsoft, NVIDIA, Amazon's Industrial Innovation Fund and Bezos Expeditions [5], and a Series C exceeding $1bn at a $39bn valuation last September led by Parkway Venture Capital [4]. The floor on this compute commitment is therefore roughly twice everything the company has raised in public view [2]. The largest obligation on Figure's books is owed to a counterparty that is also now a shareholder [2].
Divide the floor by the reserved fleet and you get about $35,000 a card; divide the ceiling and you get $60,000 [3]. Those are numbers in the neighbourhood of buying the hardware outright, which suggests what is actually being sold: a reserved slice of a campus that does not exist yet, funded in part by the roughly $3bn of debt Nscale closed weeks before the deal for GPU sites in Texas and North Carolina [7]. The contract term is undisclosed, so the annual rate cannot be computed, and the per-card figure is a bound rather than a price.
On Nscale's side the arithmetic is about the listing. Figure's floor is about 6.9% of the contracted book Nscale has shown prospective investors, and the ceiling about 11.8% [1], which is a large share to have written with a counterparty whose shares you hold. Two terms decide how it reads, and neither is public: whether the strategic investment was cash at the $39bn mark or credit against the contract, and whether the $3.5bn is take-or-pay [2]. This structure, an infrastructure provider investing in the customer buying its compute, has drawn criticism on Wall Street as a way to make demand look bigger than it is [12]. Jensen Huang, whose chips sit underneath, calls it a flywheel: training on Vera Rubin through Nscale's cloud, validating in Isaac Sim, deploying on GPUs inside the robots [13].
The demand looks real. Figure's Index programme is ingesting 30 minutes of human training video every second from paid creators using a phone app [9], which is 1,800 times real time, or roughly 43,200 hours of footage a day [5], and Brett Adcock's stated logic is that Helix gets more capable with more data and compute [11]. Figure also says it has lacked the infrastructure to run sustained large-scale training at pace until now [19]. The claim that robot training runs can exceed large language model training appears without a number attached [15], so the case that robot foundation models are their own capex category rests on contract sizes rather than measured training scale.
Read against the field, the contract sizes are the pattern. Skild AI raised $1.4bn in January above a $14bn valuation [16], Apptronik's Series A extension pushed its round past $935m [17], and Physical Intelligence was reported in March to be discussing about $1bn above $11bn, still unconfirmed months later [18]; the sector total implies roughly $4.8bn in 2025 and about $3.8bn of net addition this year [4]. My read is that Nscale bought queue priority in reverse, paying with equity to lock a 2027 campus tenant, and Figure paid with cap-table room because it is short of compute, not cash [20]. That is wrong if Nscale discloses a cash purchase at the $39bn mark alongside an unconditional commitment, in which case this is one large customer at market rates. It is also wrong if Figure starts paying from robot revenue, and the source material shows no customer buying the output yet [21].
