Hyperscale Data signed a Master Services Agreement on 24 June with an unnamed California-based neocloud, committing 20 megawatts of AI compute at its Michigan campus over ten years with a headline value above $1.2 billion. Options carry the deal to $3 billion. The mining operation the site had been running is winding down.
Hyperscale Data disclosed a Master Services Agreement on 24 June that will convert 20 megawatts of Bitcoin mining capacity at its Michigan campus into AI compute for an unnamed California-based neocloud provider, with a headline value above $1.2 billion over ten years. The customer holds an option to scale to 52 megawatts, and a further right to add 32 megawatts on top of that. If both expansions land, total contract value clears $3 billion — an unusually long-duration compute commitment for a customer whose identity has not been disclosed.
The site is the story. Michigan was purpose-built for ASIC mining under Hyperscale's previous corporate identity, when the company operated as a listed miner with hashrate as its headline metric. That business no longer exists in any meaningful commercial sense. The 24 June disclosure formally puts the mining operation into wind-down; the company will run down remaining hashrate contracts and then reprovision the racks for GPU deployment. Nothing about the physical site changes — the transformer capacity, the substation feed, the cooling infrastructure — but the machines behind them do.
Hyperscale is not the outlier. The public-mining sector has now announced roughly $70 billion in cumulative AI and high-performance-computing contracts, and every one of the six largest listed miners has said publicly it expects AI to make up the majority of revenue by year-end. Iren locked in a five-year, $9.7 billion cloud contract with Microsoft in November 2025. TeraWulf sits on $12.8 billion of contracted HPC revenue. Hut 8 struck a $7 billion data-centre lease with Anthropic and Fluidstack. Core Scientific, whose 2022 bankruptcy was the sector's inflection point for AI-adjacent thinking, is now valued more for its data-centre real estate than for anything hashing bitcoin. Analyst modelling from H.C. Wainwright and B. Riley suggests AI could hit 70 per cent of listed-miner revenue by December, up from about 30 per cent at the start of the year.
The economics are not subtle. A megawatt of Bitcoin mining generates gross margins that fluctuate with hashprice, difficulty, and BTC spot — none of which the operator controls, and all of which are pointed the wrong way after 30 June's difficulty environment and bitcoin's slide below $60,000. The same megawatt sold to an AI customer under a ten-year MSA delivers contracted revenue that clears $60,000 per megawatt per month, with take-or-pay structures that eliminate volume risk entirely. The AI customer takes the electricity price risk; the miner-turned-hoster takes a predictable cash flow that the equity market values at a multiple mining revenue does not command.
Nvidia sits in the backdrop of every one of these deals. The chip company priced a $25 billion investment-grade bond offering on 15 June — its largest ever debt deal, drawing $85 billion of orders across seven tranches out to 2056 — and the proceeds are earmarked for the same AI-infrastructure buildout the neocloud operators need capacity for. A thirty-year Nvidia bond is a bet that the demand for compute lasts three decades. The Hyperscale MSA, a ten-year commitment from a customer that presumably plans to run Nvidia hardware for its full term, is a small piece of the same trade.
What the mining industry has quietly stopped selling is exposure to bitcoin. MARA sold 20,880 bitcoin in Q1 for $1.5 billion to cut $1 billion of convertible debt and refile as a compute company. Public miners as a group sold 32,000 BTC in the first quarter — the largest miner distribution on record. The sector's remaining bitcoin exposure is now concentrated in Strategy, Metaplanet, and a handful of smaller treasuries; the listed miners themselves are increasingly a proxy for AI infrastructure demand with a legacy hashing operation attached.
For Hyperscale specifically, the 24 June disclosure resolves a strategic question that had been open since the company's earnings call in February. The board had signalled it wanted to pivot the Michigan site but had not committed to a customer or a timeline. The MSA does both. The operating question that remains is how quickly the reprovisioning happens — moving from S21 Hydros to H200s is a longer engineering job than the equity market usually credits, and analyst models generally assume six-to-nine months of transition capex before the AI revenue starts landing on the P&L.
The tail risk is customer concentration. A single unnamed neocloud with an option ladder to $3 billion of contracted spend is a very large piece of business for a company Hyperscale's size. If the customer is another entity in the recent wave of AI-infrastructure startups that raised on private markets in late 2025 and early 2026 — the ones whose unit economics rely on continued Nvidia H200 supply and continued hyperscaler demand for outsourced training — the MSA is only as durable as the customer's balance sheet. Hyperscale has not identified the counterparty. That is the disclosure the market will want next.
