One deal, end to end

The walkthrough, in plain English

The situation. Aurora Robotics is a Series A company in São Paulo training a vision-action foundation model for warehouse robots. The training run is the single largest line item on its budget for the next two quarters. The team has priced it out: eight NVIDIA H200 GPUs running continuously for six months. That is 8 × 24 × 365 ÷ 2 = 35,040 GPU·hours. At a representative spot price of $4.50 per GPU·hour, the raw compute bill is $157,680.

That number is deceptively small. The problem is not the headline cost — it is when the cash leaves. To pay it, Aurora either dips into runway it raised to spend on engineers, or it raises an earlier, more dilutive round specifically to hand the proceeds to a cloud provider. Either way the most illiquid asset in the company — its equity — is being spent to rent the most commoditized one — GPU time. That is the arbitrage Compute for Equity exists to clear.

Step 1 — Price the compute

The Oracle quotes the deal: 35,040 GPU·h of H200 capacity, sourced from a partner datacenter with abundant, cheap energy. Rather than invoice Aurora $157,680 in cash, the datacenter agrees to invest that compute. The instrument is a Compute-SAFE.

Step 2 — Structure the Compute-SAFE

The provider commits a $1,000,000 Compute-SAFE with a six-month draw window — Aurora pulls compute as the training run demands it, up to the cap. The instrument carries a $25M valuation cap and a 20% discount. Note the headroom: the committed envelope ($1M) is larger than the priced training run ($157,680), so the same instrument can absorb a follow-on run or an inference burst without re-papering the deal.

Step 3 — Draw the compute, under SLA

Over six months Aurora draws against the SAFE as it trains. Every GPU·hour is metered. The protective term that matters: if average uptime in any month dips below 95%, that month's draw rolls forward — Aurora never gives up equity for compute that did not run. The provider, not the startup, carries the operational risk.

Step 4 — Convert at the next round

Nothing converts until Aurora raises a priced round. When it does, the Compute-SAFE converts at the better of the cap and the discounted round price. That single rule is why the founder can bound the cost before signing:

• Bear ($20M pre). The 20% discount sets the price at $16M — below the $25M cap — so the discount governs. $1M ÷ $16M = 6.25% dilution. The downside case is the most expensive, because a soft round makes the discount bite.
• Base ($30M pre). Discounted price is $24M, just under the $25M cap, so the discount still governs. $1M ÷ $24M = 4.17% dilution.
• Bull ($80M pre). Discounted price would be $64M, far above the cap, so the cap governs. $1M ÷ $25M = 4.0% dilution. The better Aurora does, the cheaper the compute looks in hindsight.

What the provider gets — the other side of the trade

A Compute-SAFE only exists because both sides win. The datacenter that invests the compute is not doing Aurora a favour; it is making an investment it could not otherwise access. Its GPUs are a depreciating asset that earns a thin, commoditized rental margin when sold as raw hours. By accepting equity instead of cash for a slice of that capacity, the provider converts a low-margin rental into an equity position in a vetted, venture-backed AI company — the kind of upside that is normally reserved for the startup's cash investors. The provider is, in effect, underwriting Aurora with hardware it already owns. The SLA-suspension clause keeps that underwriting honest: the provider is paid in equity only for compute it actually delivers to spec, which aligns its incentive to keep the cluster healthy with the startup's need for reliable training. And because the instrument is standardized and custodied under ADGM rules, the position is — eventually — transferable on a secondary market, so the provider is not locked in until Aurora's exit. Cheap energy in, equity upside out, with the operational risk priced where it belongs. That symmetry is what makes the deal repeatable rather than a one-off favour.

Why a founder takes this deal

For somewhere between 4.0% and 6.25% of the company — deferred until a priced round and bounded on both ends — Aurora gets a six-month H200 training run, preserves roughly four months of cash runway, and pushes GPU price and uptime risk onto the party best able to manage it. The cash alternative spends real money today and still dilutes, just one round earlier and without the cap. The Compute-SAFE turns a cash expense into a bounded, deferred equity cost aligned with the company's own success. That is the whole thesis, in one deal.