Compute, underwritten
Notes · Compute · 2026-08-09Six months from incorporation to a ten-billion-dollar order book. The interesting part is not the size. It is the financing structure — and what the market did with it.
Headline terms as announced, 2026-08-04. Three layers, two long-dated contracts, one break clause.
On 2026-08-04, Volta came out of stealth with $300M raised across seed and Series A at a $2.4B post-money — a16z and Altimeter co-leading, NVIDIA, Michael Dell’s family office, Azora and Matter Venture Partners alongside — and a $10B, six-year compute partnership with “a leading AI lab.” Bloomberg named the lab as Anthropic the same day. Neither side has confirmed it.
The founders are infrastructure investors, not chip people. Ricard Boada and Sofia Gumuzio ran AI-infrastructure investing at Brookfield; the cloud software came by acquisition — Genesis Cloud, bought this spring, a Munich GPU cloud that has served some 20,000 users since 2018. Read the org chart and the product is not a datacenter. The product is a capital structure with a datacenter attached.
The structure is worth drawing precisely. Volta signed a 16-year colocation lease with Bitdeer — the NASDAQ-listed bitcoin miner — at Tydal, Norway: 121 MW of IT load, roughly 133 MW gross, four data halls, about $4.7B of contracted payments averaging ~$202/kW-month (3% annual escalators), and an eight-year extension option that would take the total to roughly $8B over 24 years. First halls are targeted for 2026-12-31, the balance by 2027-03-31. Bitdeer supplies the shell, the hydropower and the operations; Volta supplies what goes inside — NVIDIA’s Vera Rubin generation, Dell systems — and the cloud layer on top. Roughly $1.3B of letters of credit, arranged by J.P. Morgan and one other global bank, stand behind the lease.
The arithmetic, taking the announcement at face value: $10B over six years across 121 MW is ≈$13.8M per IT-megawatt-year coming in; the lease costs ≈$2.4M going out. The spread — ≈$11.4M per MW-year, ≈$1.37B a year — is what services the actual capital good: GPU systems that run $30–40M per megawatt at this generation, call it $4–5B for the site. Six-year contracted revenue of roughly 2–2.5× the systems bill is familiar neocloud math. What is less familiar is the sequence. Volta signs the contract first, then finances the hardware against the contracted cashflow — the $5B programme with Azora, project equity plus senior debt, exists to do exactly that, non-dilutively, site after site. “Compute has become a new infrastructure asset class” is the CEO’s line, and it is the entire thesis: the pitch is not better silicon or better software. It is that compute contracts can be financed the way toll roads are.
One honesty clause belongs next to that arithmetic: the $10B is described by Volta as a commitment over the life of the partnership, of which Tydal is the first project. If later sites fold into the same envelope, the per-megawatt figure dilutes accordingly. The announcement attaches the number to this deployment; the announcement is also the only place the number lives.
What the market did with all this is the more instructive print. Bitdeer had disclosed the lease’s existence on June 29 without financial terms; the stock fell about 7% that day. On August 4, with the terms and the tenant’s reported identity public, it opened the pre-market up 23% — and faded to roughly flat by the close, a ~$2.8B company at ~$11.5. Between the open and the close, the market found the two numbers that matter: a no-fee termination right at year ten cuts the firmly-committed value to ≈$2.7B by Benchmark’s arithmetic, and $1.3B of letters of credit covers 28% of the nominal $4.7B. Sell-side moved anyway — Needham took its target to $22 from $19. A year ago this tape bought the headline and asked later. It is reading clauses now — which is, for the asset class Volta wants to create, the beginning of actual underwriting.
The open leg of the structure is duration. Revenue is contracted for six years; the lease runs sixteen. If compute stays scarce through 2032, renewal solves the mismatch. If it does not, the year-ten break is Volta’s escape hatch — and the same clause with the sign flipped is Bitdeer’s residual risk. Add the standard list: one customer, one site, suppliers next to the cap table (NVIDIA invests and sells; Dell supplies, while Michael Dell’s family office invests), a chip generation still ramping, and a first-light deadline five months out. The December 31 date is the rare thing in this story that is simply checkable.
We filed the power argument in May— the binding constraint on compute has moved to electricity and the years it takes to build. Tydal is that argument wearing a term sheet: the scarce input is a powered shell on Norwegian hydro, the miner owns it, and an AI lab is reported willing to commit $10B through a six-month-old intermediary to reach it. The lab pays for certainty, the miner converts megawatts into contracted rent, and the intermediary’s entire margin is the price of assembling certainty out of parts. That margin is contracted, not yet earned — first light is five months out. Whether it is durable is the question the Azora programme’s first term sheet will answer.
A ten-billion-dollar headline is narrative. A year-ten break clause is data. We read the clauses, and we will be watching two dates: December 31, and the day either party says the customer’s name on the record. Not a position; a structure worth filing.
Sources.Lease terms, rates, phasing and the letters of credit are from Bitdeer’s announcement and Form 6-K of 2026-08-04; the raise, valuation, investor list, the $10B partnership, the Azora programme and the Genesis Cloud acquisition are from Volta’s launch announcement of the same day; the tenant’s identity is as reported by Bloomberg and is unconfirmed by either party; the ≈$2.7B firm-commitment arithmetic is Benchmark’s, the $22 target Needham’s, both as reported; session prices are the tape. Per-megawatt figures are our own arithmetic on those disclosures.