What 165 MW of inference capacity buys: a rule for judging Cerebras's Finland bet
Use a scorecard to judge whether a large inference buildout is a product or a power commitment.
You are comparing a power contract to a product promise. Trade-off: megawatts are a physical ceiling, tokens are the product, and the contract term decides who carries the risk.
Run the arithmetic before the opinion: the site's usable MW ÷ the MW per token stream your workload needs = streams; streams × expected tokens/sec = tokens/sec; apply cost per million tokens over the term.
Cerebras and Compute Nordic Finland revealed on September 1, 2026 that the Mikkeli, Finland site would scale to 165 MW of contracted IT capacity, with seven-year service orders securing long-term infrastructure. Construction on the first 50 MW phase was already under way.
Megawatts set the ceiling, not the output
Megawatts cap the number of accelerators you can keep running. The token count comes from the workload's MW per token stream, the expected tokens/sec, and the utilization you can prove.
The scorecard turns the buildout into unit economics
Put the vendor on a scorecard before you judge the buildout.
- Contracted MW: the physical ceiling, not the planned ceiling.
- Expected tokens/sec: the throughput your workload actually needs.
- Cost per million tokens: the number that hits your unit economics.
- Latency target: the response time your product can tolerate.
- Contract term: the period over which you are locked in.
- Demand coverage: the share of capacity already tied to revenue.
Ramboll's Finnish Data Center Market Study, dated September 12, 2025, puts the full 165 MW Mikkeli buildout at an indicative €1.0–1.7 billion regional investment, with 80–250 direct permanent jobs and €0.8–2.5 million in annual property tax revenue.
Cost per million tokens is the gap. If Cerebras will not disclose it, derive it from the service orders, the expected token mix, and the term. If it cannot be derived, the MW figure is only a ceiling. Treat the vendor numbers as product claims, not site-level guarantees: the CS-4 rack-scale system provides up to 30x faster inference than GPUs. Scout delivers more than 2,000 tokens per second, more than 30 times faster than closed models such as ChatGPT or Anthropic. Complex reasoning can be delivered in under a second for deep search, copilots, and analysis. Your latency target must be tied to your p95 or p99 requirement, not to a headline speed.
Demand coverage separates a capacity plan from a capacity bet. If token demand is not contracted, the buyer is financing idle power. Ask for the share of capacity committed to named workloads, expected utilization by quarter, and exit or resize terms.
Latency defines the product experience. A high token rate helps only if the first token arrives when your user expects it. Copilots need a sub-second feel; batch analysis can wait minutes. Require p95 and p99 latency for the exact model, context length, and concurrency level you will run. If the number is only a peak, ask for the sustained service level.
The objection is demand, not construction
A long power commitment can outlive the demand that justified it. Cerebras had more than 600 MW of data center capacity operating or contracted for delivery by the end of 2027, and it projected a more than tenfold rise in manufacturing capacity during 2026. Cerebras had $25.4 billion in remaining performance obligations as of June 30, 2026.
In a July 9, 2026 announcement, Cerebras said its first European data center capacity would come online by late 2026, with total capacity reaching 200 MW by late 2027, and framed the European expansion as a way to place fast AI inference compute nearer to users in Europe. Read the late-2027 total as a roadmap, not a Mikkeli load test. The procurement test is whether Cerebras will show the service orders and the utilization plan behind the scorecard. If it will not, request both before you commit.