$10 billion. Six years.

On August 4, Bloomberg dropped a bombshell: Anthropic has signed a $10 billion compute procurement agreement with cloud infrastructure startup Volta Infra Holdings, with a contract term of six years. That works out to roughly $1.67 billion per year — exceeding OpenAI's entire 2024 revenue. And Volta, the company landing this $10 billion order? It was just founded earlier this year. It's six, maybe seven months old. The bubble tea shop that opened downstairs in your neighborhood has probably been around longer.

That's not even the wildest part. The data center is being built in Tydal, a small town in central Norway near the Arctic Circle, powered by hydroelectric dams; it runs on NVIDIA's next-generation Vera Rubin chips, which haven't even entered mass production yet; and the actual operator is Bitdeer Technologies, a Nasdaq-listed Bitcoin mining company. NVIDIA supplies chips, Norway supplies cheap electricity, a mining company supplies the facility, and Anthropic supplies the cash — four players that seem to have nothing to do with each other, assembled into the most surreal compute consortium in AI industry history.

The real significance of this deal isn't the dollar amount. It's that it marks a fundamental shift in AI compute procurement logic: leading AI companies are systematically bypassing the big three cloud providers — AWS, Azure, GCP — and going straight to a new breed of compute suppliers built on the triad of "NVIDIA-backed middlemen + cheap electricity + ready-made infrastructure."

Deconstructing Volta: NVIDIA + Norwegian Hydro + Bitcoin Mines

To understand how counterintuitive this deal is, you have to break down what Volta actually is.

Volta isn't a cloud service provider in the traditional sense. It doesn't have AWS's global data center network, Azure's deeply entrenched enterprise customer base, or even a properly functioning website you can visit. Its core assets are three things: NVIDIA's priority supply access, Norway's cheap electricity, and Bitdeer's existing mining facilities.

First, NVIDIA's backing. Multiple sources confirm Volta is an "NVIDIA-supported" cloud infrastructure company. What does that mean? It means Volta isn't getting GPUs that anyone can queue up to buy — it's getting priority access to NVIDIA's next-generation Vera Rubin architecture chips. Vera Rubin is Blackwell's successor, expected to enter mass production in 2026 with performance doubling again. While other companies are still fighting over Blackwell allocations, Volta has a packaged supply of pre-release next-gen chips locked in.

Second, cheap hydro power in Tydal, Norway. Tydal sits in central Norway, surrounded by large hydroelectric dams. Nordic electricity prices consistently hover around $0.04–0.06/kWh — roughly one-third to one-fifth of Chinese industrial electricity rates. Training a frontier LLM, electricity accounts for 30–40% of operating costs. Building the data center in Norway saves billions of yuan in power costs alone over six years. On top of that, hydro is 100% clean energy, satisfying tech companies' carbon neutrality pledges. Two birds with one stone.

Third, Bitdeer's ready-made mining facilities. Bitdeer is a Nasdaq-listed Bitcoin mining company founded by Jihan Wu. Volta's data center in Tydal directly reuses Bitdeer's existing mining infrastructure — the server halls, power supply systems, cooling, and network access are all already in place. That slashes at least 18 months off the time from groundbreaking to production. For Anthropic, time is more valuable than money: getting compute 18 months late could mean falling irreversibly behind in the model race.

The planned data center capacity is 133 megawatts — equivalent to the residential electricity consumption of a mid-sized city. All 133MW dedicated to AI inference and training, placing it firmly in the first tier of AI-specific data centers.

One sentence to sum up Volta's business model: NVIDIA supplies chips, Norway supplies cheap power, Bitdeer supplies the facility, Anthropic supplies the cash. Everyone gets what they want; all the risk sits with Anthropic.

Four Deals in Six Months: Anthropic's Compute Anxiety

Volta isn't Anthropic's first compute gamble — it's the fourth in a six-month compute-buying spree.

Earlier this year, Anthropic struck a deal with SpaceX to leverage Elon Musk's data center resources; in Q1, it signed a large-scale chip procurement agreement with AMD, seeking compute backup beyond NVIDIA; in Q2, it partnered with Akamai to use their edge computing and CDN networks for distributed inference; in August, it dropped $10 billion on Volta, building (co-building) a dedicated AI data center in Norway.

Why is Anthropic in such a rush? Three reasons.

First, explosive growth in Claude's user base. Claude passed 100 million daily active users earlier this year, with enterprise customer counts up 400% year-over-year. Every new user, every API call, every task executed in Work or Codex mode consumes real GPU compute behind the scenes. Insufficient compute means slower service, request queuing, and user churn — in a competitive environment surrounded by ChatGPT, Gemini, and DeepSeek, even a 30-minute outage can cause irreversible user attrition.

Second, insatiable compute hunger from next-generation model training. Industry consensus: training a frontier model at the GPT-5/Claude-4 level costs between $1–5 billion per run. And it's not a one-time training job — continuous iteration, RLHF, safety alignment, multimodal expansion, every round burns GPU hours. The Vera Rubin chip is designed precisely for the next generation of trillion-parameter models; locking in chip supply and accompanying data centers ahead of time is essentially locking in a ticket to the next model race.

Third, compute is becoming the "oil reserve" of the AI industry. NVIDIA production is limited, TSMC CoWoS packaging capacity is a bottleneck, data center build cycles are 18–24 months, and prime cheap-power locations are genuinely scarce — the number of places worldwide suitable for large AI data centers (cheap power + cool climate + political stability + network connectivity) is actually quite small. Don't sign now, and in six months you might not even have a place in line.

The deeper reason: Anthropic is preparing for a fall 2026 IPO. Locking in long-term, stable, cost-controlled compute supply before going public not only tells investors a good "infrastructure independence" story but also hedges against compute price volatility over the next several years.

Bypassing Big Cloud: The New Compute Procurement Paradigm

The most noteworthy signal in this deal isn't the $10 billion figure itself — it's that Anthropic chose Volta instead of AWS, Azure, or GCP.

For the past three years, the relationship between AI companies and cloud providers was a "deep binding" model: OpenAI tied to Microsoft, Anthropic to Google and Amazon, xAI to Oracle. Cloud providers put up capital, compute, and distribution channels; AI companies provided models and technology; both sides signed 5–10 year exclusive cooperation agreements. It was a "marriage" — each side got what it needed, but each also constrained the other.

Now things are changing. Anthropic is simultaneously signing with six vendors — Google, Amazon, SpaceX, AMD, Akamai, Volta — and increasingly favoring non-traditional cloud providers. Three structural shifts underlie this:

First, cloud vendor GPU markups are too high. GPU instances on AWS and Azure typically cost 3–4x bare-metal costs, because virtualization, networking, storage, operations, and margins are layered on top. For a customer like Anthropic spending hundreds of millions annually on compute, bypassing cloud providers and going straight to bare-metal suppliers saves 30–50% on compute costs. Over six years, that's billions of dollars in savings.

Second, cloud vendors themselves are building AI models. Google has Gemini, Amazon has Nova, Microsoft has the Copilot family. When your cloud provider is also your direct competitor, handing them your core compute lifeline is essentially showing your cards to your opponent. Anthropic can't help worrying that Google might "favor" its own Gemini when allocating TPU/GPU resources.

Third, the compute supply ecosystem is diversifying. CoreWeave has gone public, proving the "AI-specific cloud" model viable; Nebius is expanding rapidly in Europe; mining companies pivoting to compute supply (Bitdeer, Foundry Digital) have ready power and facility resources; the Middle East and Nordics are aggressively courting AI data center deployments. Compute supply is no longer dictated by the big three clouds; buyers have more leverage and more options.

SpaceX is doing something similar. Another piece of news the same day revealed that Musk delivered SpaceX's first earnings report, and Starlink's AI compute revenue surged 247% in Q2 — reselling GPU compute turned out to be more profitable than launching rockets and satellites. When "outsiders" like SpaceX and Bitdeer enter the AI compute market, the traditional cloud providers' moat is eroding.

Risks and Endgame: What a Six-Year Contract Bets On

Of course, this $10 billion gamble isn't without risks.

The biggest risk is time. Six years is an eternity. Six years ago (2020), the AI industry was still debating whether GPT-3 could write coherent sentences; now we're talking about AGI and Agents. If new computing architectures that don't require massive GPU fleets emerge within three years — purpose-built ASICs, photonic computing, quantum computing, even neuromorphic chips — today's GPU data centers could become expensive scrap metal. While Anthropic locks in six years of compute, it also locks itself out of flexibility to switch technology trajectories.

The second risk is Volta's execution capability. A months-old company building and operating a 133MW hyperscale data center in Norway is enormously complex. NVIDIA providing chips is one thing; completing construction on time, maintaining stable power, and running operations without major incidents for six years is another. Bitdeer has experience operating large mining facilities, but mining farms have far lower requirements for stability and network quality than AI data centers.

The third risk is AI demand growth falling short of expectations. Before the dot-com bubble burst in 2000, telecom companies were signing similarly insane long-term contracts — leasing undersea fiber, building data centers, buying servers — and when the bubble popped, 90% of those contracts became worthless. If AI application adoption is slower than expected, or if more efficient algorithms dramatically reduce compute requirements (as DeepSeek proved small models can deliver strong results), a $10 billion long-term contract becomes massive stranded cost.

But none of these risks will stop Anthropic, or the industry as a whole. Because a more fundamental truth is: AI's second half isn't about algorithms — it's about electricity and chips. Whoever locks in enough cheap power and GPU supply survives the next model race. This is no longer a war for algorithm engineers; it's a war for infrastructure and capital.

With $10 billion, Anthropic is telling the world: it intends to stay at the table for the next six years. As for whether Volta, this "infant company," can handle the delivery pressure of a $10 billion contract, we'll know in six years — but most likely we won't have to wait that long. When the first compute capacity comes online in 2027, we'll find out who won this bet.

See you tomorrow.