Jane Street commits $13 billion to Crusoe for GPUs
The trading firm's second enormous compute deal in a month, plus a forecasting model that beats everything and can't be used commercially.
Jane Street signed a five-year cloud contract with Crusoe worth about $13 billion, its second huge GPU commitment in weeks. Bloomberg reported the deal, which gives the trading firm clusters of GPUs and supporting infrastructure for AI training and inference through Crusoe's cloud platform, and Reuters confirmed the report. Crusoe is the highest-profile cloud customer Jane Street has landed, and Bloomberg says the contract helped draw interest to Crusoe's ongoing raise — the company was reported in July to be discussing around $3 billion at roughly a $30 billion valuation. Crusoe and Jane Street did not immediately respond to Reuters' requests for comment.
The pattern is the point. Jane Street led Fluidstack's $1.5 billion round at an $18 billion-plus valuation earlier this month — we covered it in Fluidstack triples to an $18B valuation in six weeks on Jane Street's bet — and now it has committed $13 billion of contracted spend to one of Fluidstack's direct competitors. A market maker buying capacity from multiple neoclouds at once is not picking a winner; it is hedging a compute shortage it expects to persist for the full life of these contracts. And the shape of the commitment says something about what Jane Street thinks it is buying: five years of training and inference capacity is not a research budget, it is an operating assumption that model-driven trading gets more compute-hungry, not less.
For the neoclouds this is the best signal available and the worst one at the same time. Long-dated, credit-worthy demand is exactly what you need to underwrite a data center build, and a marquee name on the contract makes the next fundraise easier. But selling five years of capacity today means having nothing left to sell at next year's prices — and every neocloud that signs a deal like this is betting that GPU scarcity holds long enough to justify locking in the rate.
Google released TimesFM-3, a 330M-parameter time series foundation model that reads many variables at once — under a license that bars commercial use. Most forecasting models look at one series in isolation; TimesFM-3 is natively pre-trained for the multivariate case, so a forecast for ice cream sales can draw on related product lines, foot traffic, and known future events like a scheduled promotion. It alternates two kinds of attention — one causal and backward-looking within a single series, one sideways across all series at a given moment — and it decodes the whole forecast horizon in a single pass instead of patch by patch, which cuts latency and stops errors compounding along the way. Google says it takes the top average rank on Gift-Eval, FEV-Bench and the TIME leaderboard in both point and probabilistic forecasting, with TimesFM-2.5 — state of the art when it launched in September 2025 — now at the bottom of those same tables.
The license is the strategic part. The weights are on Hugging Face under a non-commercial license while the code stays Apache-2.0, and the paid path runs through BigQuery's AI.FORECAST. Google is not hiding the research; it is reserving the product. For anyone doing demand planning, capacity forecasting or observability work, that means the best open forecasting model is available to evaluate and off-limits to deploy — the same squeeze that turned open-weights LLMs into a licensing argument, now arriving in a field that had been genuinely open.
What to watch: whether other neoclouds land Jane Street-scale contracts this quarter, and whether anyone ships a commercially licensed multivariate forecaster that closes the gap.
Is selling five years of GPU capacity forward a sign of confidence in demand, or a sign these companies think prices peak now? Tell us in the comments.
