Anthropic Signs 20-Year, 191MW Riot Compute Lease in $9.1B Deal

Riot Platforms signed a 20-year, 191 MW data center lease at its Rockdale, Texas campus on August 10, 2026 — a deal expected to generate about $9.1 billion in contract revenue through June 2048, with Bloomberg reporting that the frontier lab behind it is Anthropic, maker of the Claude models. Two five-year extension options could lift total potential value to roughly $16.1 billion, placing this among the largest single-tenant AI compute commitments of the year.

The deal structure, broken down

Riot Platforms (NASDAQ: RIOT), one of the world's largest Bitcoin miners, disclosed the Data Center Lease and Services Agreement alongside its second-quarter 2026 earnings. The counterparty is described in the press release and SEC 8-K filing only as "one of the world's leading frontier AI labs." Bloomberg subsequently identified Anthropic, citing people familiar with the transaction. Neither company has confirmed the tenant's identity publicly — the contract is confirmed, while the tenant attribution rests on Bloomberg's reporting — but the market has clearly accepted the attribution: Riot shares jumped roughly 25% in after-hours trading.

The financing structure matters as much as the headline number. Riot secured a $573 million interim financing facility from Morgan Stanley to fund early development costs while it finalizes an investment-grade credit backstop — the kind of financial engineering that only works when the tenant carries investment-grade credit, itself a signal about who the tenant is. For scale, 191 MW is enough electricity to power roughly 143,000 homes.

Why frontier labs are locking compute for two decades

This lease is the second anchor tenant Riot has secured in under seven months. AMD signed its first Rockdale lease on January 16, 2026 — an initial 25 MW of critical IT capacity delivered on time and on budget by May 2026, with an expansion bringing AMD's contracted total to 50 MW and an option to grow toward 200 MW. Combined, Riot now has 241 MW under signed leases representing roughly $9.8 billion of contracted revenue with "two of the most important companies in the AI ecosystem," in CEO Jason Les's words.

Anthropic's broader capacity portfolio shows the pattern is strategic, not opportunistic. The lab already has commitments across Amazon Web Services, Google Cloud, and Microsoft Azure, plus deals with Volta Infra and xAI, and it signed a separate July 22, 2026 partnership with AMD to deploy up to 2 GW of Instinct MI450-series GPUs. The Riot lease adds a second long-dated, single-site anchor. Three structural reasons explain why frontier labs are converting cash into decades of guaranteed electricity and floor space:

1. Interconnection queues are the real bottleneck. New transmission interconnection requests now wait years in most U.S. grids. Rockdale already has fully approved, energized interconnection — the single most valuable asset a compute buyer can acquire in 2026. This is the same logic driving the sovereign AI infrastructure buildout we covered earlier, where nations treat compute like energy grids.

2. Financing economics reward long-dated tenants. An investment-grade tenant lets developers like Riot pre-finance multi-hundred-megawatt builds with bank facilities, converting future capacity into today's construction capital.

3. Roadmap insurance. The only way a lab can promise a 2029-2030 model generation is to own the power that generation will require. Training runs for frontier-scale models now cost hundreds of millions of dollars in energy alone, and inference for 24x7 agentic traffic is consuming an ever-larger share of that envelope — an economics shift we explored when inference spending overtook training for the first time.

The Bitcoin-miner-to-AI-tenant migration is industry-wide: Core Scientific, IREN, Applied Digital, TeraWulf, and Hut 8 are all pursuing similar HPC conversions, driven in part by activist investors who argued that U.S. power portfolios command far more value leased to computing tenants than mining coins. Riot's own balance sheet tells the story: it ended June with more than $1.2 billion in liquid assets — $548.9 million in cash plus 11,380 bitcoin worth about $666 million — and it has been selling bitcoin inventory to fund the equity portion of its data center buildout. For a broader look at how hardware dollars are flowing across the sector, see our breakdown of Intel's $15 billion AI compute bet.

The data center and energy market context

This deal does not happen in a vacuum; it is the product of a systemic shortage. Data center demand in 2026 is being driven less by the number of servers being shipped and more by the number of megawatts a single AI cluster can absorb. A full 191 MW single-tenant build-to-suit is at the frontier of what any private developer has signed, and it only works because Rockdale's interconnection is already approved — a status that can take four to six years for new grid connections in constrained regions. Riot's broader portfolio compounds the advantage: more than 1,100 acres and 1.7 GW of power capacity across its two Texas facilities, with the Corsicana campus (a 1 GW facility) under a separate letter of intent with an unnamed tenant.

The financing also reveals how the AI infrastructure market now prices certainty. Morgan Stanley's $573 million interim facility bridges construction while Riot finalizes an investment-grade credit backstop. In practical terms, the tenant's credit quality is what lets a former Bitcoin miner finance a billion-dollar buildout at reasonable rates — the same reason hyperscalers and frontier labs have become the preferred counterparties for every data center developer in North America. What miners lose in Bitcoin revenue upside — Riot's mining revenue fell to $113.7 million from $140.9 million year over year — they gain in contracted, recurring NOI: Riot projects $7.3 billion to $8.2 billion in cumulative net operating income over the base term, or $365 million to $411 million per year. That is the difference between speculating on a volatile commodity and collecting utility-grade rent for twenty years.

What it means for inference cost and the Claude roadmap

Run the unit economics and the deal starts to look like a pricing hedge. Roughly $9.1 billion over 20 years is about $455 million per year in average base-term revenue — a sum equal to 65% of Riot's annualized Q2 revenue. Per megawatt, that's about $2.4 million per MW-year, or roughly $200,000 per MW-month, for built-to-suit Tier 3 space and power. For an enterprise workload that needs, say, 4 MW of power for a training cluster, that is roughly $9.6 million per year of fixed data center cost before a single GPU is purchased. Long-dated locks like this are exactly how Anthropic can underwrite stable, predictable API pricing across a 20-year horizon — capacity it controls, at costs it fixed in 2026, is capacity it does not have to buy at spot rates during a future supply crunch.

For the model roadmap, the signal is bullish. Committing 191 MW in Texas plus 2 GW of AMD Instinct GPUs tells you Anthropic is planning model generations years beyond Claude Opus 5 and Claude Fable 5, and that it expects demand for Claude inference to keep compounding. At the same time, Riot's own quarterly data center revenue reached $23.2 million — operating lease plus tenant fit-out services — showing the recurring-revenue math is starting to work, even as the miner posted a $237.2 million net loss for the quarter while it funds the conversion.

Why This Matters for Developers

If you build on Claude — or on any frontier API — this deal is a supply-chain signal you should read carefully. In our production deployment at SaaSNext, where we run agentic pipelines that burn millions of tokens a day across Claude and other models, token price volatility and capacity throttling are the two operational risks that actually keep us up at night. A 20-year compute lock like this one is Anthropic's explicit answer to both: it removes the capacity ceiling that forces labs to throttle during demand spikes, and it stabilizes the cost base that underwrites API pricing. When a frontier lab owns its power for two decades, the enterprise cost of AI inference gets more predictable, not less.

Practically, developers should update their capacity planning model. Treat frontier API uptime and pricing as variables that improve as labs like Anthropic close compute deals, and watch the delivery milestones — the December 2027 first block of 96 MW and the June 2028 full deployment — as markers of when Claude capacity will expand materially. Also consider the competitive side: every MW Anthropic locks is capacity that a smaller lab cannot rent, which reinforces the moat around frontier-tier models. For engineering teams the takeaway is threefold: (1) reassess long-term API contracts with the assumption that frontier capacity is only growing; (2) bake spot-price volatility hedges into your cost model rather than assuming every quarter will match the last; and (3) treat multi-year compute commitments at the frontier labs as a macro signal that the inference price curve is heading down, not up. Keep pace with the capacity deals and pricing moves on our latest AI news hub.

Timeline

Sources

• Riot Platforms official press release — Q2 2026 results and strategic highlights
• Bloomberg — Anthropic Strikes $9 Billion Deal with Cloud-Computing Firm Riot
• The Block — Riot-Anthropic $9.1B lease coverage
• Data Center Dynamics — Riot Platforms agrees 191MW, 20-year lease agreement with Anthropic worth $9.1bn
• Anthropic official newsroom

By Deepak Bagada, CEO at SaaSNext & Principal AI Architect.

Last verified: August 11 2026.