Figure AI's $3.5 Billion GPU Bet: Why a Robot Maker Is Spending More on Computing Than It Has Ever Raised
Figure AI has just made a bet that reshapes how we should think about the humanoid robot business. The company announced a $3.5 billion partnership with Nscale to rent up to 100,000 graphics processing units (GPUs) on Nvidia's Vera Rubin platform, with deployment beginning in the second half of 2027 at a facility in Barstow, Texas. The commitment is remarkable for one reason: it exceeds Figure's entire lifetime fundraising by $1.6 billion.
Why Is Figure Spending This Much on Computing Power?
Figure's founder Brett Adcock explained the logic plainly: "Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute." The company has raised approximately $1.9 billion since its founding, with major rounds in February 2024 at a $2.6 billion valuation and September 2025 at a $39 billion valuation from investors including Nvidia, Microsoft, the OpenAI Startup Fund, Jeff Bezos, Brookfield, and Intel Capital. Yet the GPU rental agreement alone will consume a stream of payments over years, starting before the company has shipped robots in volume.
Brett Adcock
"Our AI model, Helix, becomes more capable the same way every learned system does: with more data and compute," stated Brett Adcock, Figure's founder.
Brett Adcock, Founder at Figure AI
As of April, Figure had deployed more than 350 of its Figure 03 robots with customers. At BMW's Spartanburg plant, an earlier model helped produce over 30,000 X3 vehicles in ten months, with the newer version now handling logistics sequencing. Two years of fundraising at rising valuations have produced a few hundred robots in the field, yet Figure is now committing to a compute bill that dwarfs the company's balance sheet.
How Does This Deal Actually Work?
The structure reveals how modern AI infrastructure gets financed. Figure will pay for GPU capacity as it consumes it, with payments beginning in late 2027. Nscale, the two-year-old compute provider, is taking an equity stake in Figure alongside the deal, effectively betting that the robot maker will succeed and continue paying its rent. The arrangement allows Figure to pre-commit its next several funding rounds to its GPU supplier, while the supplier gains a financial incentive to ensure those rounds happen.
Nscale itself is a young company with an ambitious growth plan. It signed a $45 billion six-year deal with Anthropic two weeks before the Figure announcement and supplies Microsoft under contracts covering roughly 200,000 GB300 systems. The company has told investors its contracted leases now total about $103 billion. Nscale is seeking $3.5 billion in pre-IPO financing ahead of a potential listing, with funding coming from $1.5 billion in convertible notes led by Third Point at a $30 billion valuation cap, plus about $2 billion from Nvidia.
The financing chain reveals how Nvidia is anchoring the entire ecosystem. Nvidia sells the Vera Rubin racks, invests in the landlord that buys them, and already holds a stake in the tenant that will rent them. This mirrors how turbine makers have financed power projects for decades, standing behind both builder and buyer so a plant gets financed before it earns anything.
What Does This Mean for the Humanoid Robot Industry?
The Figure commitment exposes a fundamental split in the humanoid robotics category. Three weeks before Figure's announcement, Unitree, the world's largest humanoid maker, listed in Shanghai at about a $9 billion valuation after bids exceeded 8,000 times the shares on offer. Unitree sells robots at gross margins near 60 percent, earned about $235 million in revenue last year, and made a profit doing it. Its IPO raised about $904 million.
Figure's single compute commitment is worth nearly 40 percent of Unitree's entire company valuation. Figure has agreed to spend roughly four times what Unitree raised in its IPO on GPUs alone. This comparison shows where the money in the category goes: Unitree's business is building and selling machines at a margin, while Figure's spending signals that its business is training the AI model that runs the machine, a business that consumes capital at data-center scale before a single robot ships in volume.
Steps to Understanding the Humanoid Robot Financing Model
- Hardware vs. AI Split: Unitree operates as a traditional hardware manufacturer with 60 percent gross margins, while Figure treats AI model training as its core business, requiring massive compute infrastructure investment before volume production.
- Capital Intensity: Figure's $3.5 billion GPU commitment exceeds its lifetime fundraising, while Unitree's entire IPO raised $904 million, showing how differently the two companies finance growth.
- Supplier Alignment: Nscale takes equity in Figure alongside the compute deal, creating financial incentives for both parties to succeed and ensuring the infrastructure gets built before revenue arrives.
- Ecosystem Financing: Nvidia invests in both the landlord and the tenant, mirroring traditional power project financing where equipment makers anchor both sides of a deal to ensure capacity gets built.
What Risks Could Derail This Plan?
The timeline is tight and the chain is long. Vera Rubin systems are not broadly available until 2027, and the Barstow facility has to be built out. Nscale must complete its IPO listing, and Figure must keep raising money to pay rent that starts before its robots ship in volume. A slip anywhere in that sequence pushes revenue into a later year for everyone above it.
Figure's next funding round will show whether investors will finance a rent bill nearly twice the size of the company's lifetime capital. The deployment count, which stood at 350 robots in April, will show whether the compute is producing machines that customers actually pay for. The humanoid business now has a rent bill, and it comes due in Texas before the robots ship in volume.