Cerebras Systems just committed to a data center in Mikkeli, Finland that an independent study pegged at between €1.0 billion and €1.7 billion in regional investment, and the company is currently losing money on nearly every dollar it brings in. That tension is the whole story of AI infrastructure right now: the orders are real, the losses are real, and the race is to outrun one with the other.

The Mikkeli campus, announced September 1 in partnership with Compute Nordic Finland, will eventually reach 165 megawatts of capacity. Construction on the first 50 megawatts is already underway, with expansion steps to 80 megawatts and then the full 165 planned in stages. Each service order runs seven years. That is a long leash for a company that went public on the Nasdaq under the ticker CBRS and is still burning through cash at a significant rate.

The order book that justifies the bet

The rationale is sitting in Cerebras's own numbers. When the company reported second-quarter results on August 12, it disclosed $25.4 billion in customer obligations not yet delivered. Cloud revenue for the quarter ended June 30 rose 281% from a year earlier. The company also raised its full-year outlook on every core metric.

Those are not small numbers. $25.4 billion in backlog dwarfs a single quarter's sales by a factor that makes the Finnish campus look less like speculation and more like a construction crew trying to catch up with a queue.

Cerebras also has a supply-chain edge that matters right now. It avoids two components that are scarce across the entire AI hardware industry: a type of high-bandwidth memory (called HBM) and a specialized chip-packaging process. By sidestepping those bottlenecks, it can scale production while competitors wait in line. Core gross margin reached 41% in the second quarter, up roughly nine percentage points from a year earlier, suggesting the underlying economics improve as volume grows.

The losses that complicate it

But the financials carry a serious asterisk. On standard accounting terms, second-quarter gross margin was 14% and the operating margin was negative 265%. Cerebras promotes a tighter measure that strips out stock compensation, warrant costs, and pass-through data center expenses, and even that "core" figure showed an operating margin of negative 16%.

It also matters which revenue number you use. Core revenue for the quarter was $209.9 million; the standard accounting figure was $180.1 million. The gap reflects how the company structures its reporting, and readers following the stock need to know which version they are reading.

The outlook for the third quarter does not improve things. Cerebras guided core operating margin to between negative 25% and negative 23%, worse than what it just delivered, even as core revenue guidance of $214 million to $216 million represents only a modest step forward.

Mikkeli adds a specific kind of pressure. Compute Nordic Finland handles development, operations, and the customer interface for the campus. That means Cerebras is effectively buying capacity from a partner on seven-year commitments. The deal only works if the customers already in that $25.4 billion backlog stick around, pay on schedule, and ideally bring friends.

What this actually represents

The broader pattern here is not unique to Cerebras. Across AI infrastructure, companies are making commitments at a scale and speed that would look reckless in almost any other context. The bet is that demand will grow fast enough to convert today's losses into tomorrow's margins before the cash runs out or the orders slow down.

Cerebras has more supporting evidence for that bet than most. A 281% revenue jump and $25 billion in signed obligations are not nothing. The Finland campus sends waste heat back to the local community and runs on closed-loop water cooling, which suggests at least some discipline in how the physical infrastructure is being built.

But a company guiding to deeper operating losses next quarter, while committing to a multi-billion-euro campus abroad, is making a very specific wager: that speed of execution matters more right now than profitability, and that the window to capture AI infrastructure share is short enough to justify the cost. Whether that logic holds depends entirely on what happens to that backlog.