Thinking Machines will pay Crusoe $65 million a year to serve models it gives away

The 23 September deal is the first hard price tag on open-weights economics, and Crusoe's own numbers put one early-stage lab at up to two-thirds of a young product line's contracted revenue.

Vincent JiangVincent Jiang · 2 min read
Share
Mira Murati, founder and CEO of Thinking Machines Lab, at the 2026 Met Gala
Mira Murati, founder and chief executive of Thinking Machines Lab, whose lab will pay Crusoe $65 million a year to serve its open-weight models.

On 23 September, Thinking Machines Lab agreed to pay Crusoe $65 million a year to serve its production models on Crusoe Managed Inference: a dedicated, SLA-backed deployment of NVIDIA HGX B200 systems wired with Quantum-2 InfiniBand 12. The named workloads include Inkling, whose weights anyone can download for nothing, alongside GLM 5.2 and 5.3, open models built by Z.ai, and the lab's own fine-tuned variants 13. Open weights are free the way a piano is free; someone still pays to move it.

One lab is most of the product line

Crusoe's own arithmetic is the story. The release says the lab joins a roster that has taken Managed Inference past $100 million in contracted ARR, a company-claimed figure, less than a year from launch 12. One customer at $65 million, against a total described only as "past $100 million," puts Thinking Machines at up to about two-thirds of the contracted book, and Crusoe publishes no ceiling. For scale, Crusoe closed $3.9 billion at a $30.9 billion valuation on 17 September 4, and a $13 billion, five-year Jane Street cloud contract has been reported in its business 4. Managed Inference is the runt of that portfolio, and its anchor tenant is one early-stage lab.

One lab is up to two-thirds of Managed Inference's contracted book

$0M$50M$100MThinking MachinesAll other customers (at least)Total contracted ARR (at least)$100M$65M
Data
Contracted ARR
Thinking Machines$65M
All other customers (at least)$35M
Total contracted ARR (at least)$100M
Crusoe says Managed Inference has passed $100 million in contracted ARR less than a year from launch but publishes no ceiling, so the total and the remainder are floors. One customer, Thinking Machines at $65 million a year, accounts for up to about two-thirds of the book. Figures are company-claimed. Source: Crusoe and Thinking Machines Lab release, 23 September 2026 [1][2].1,2

Managed Inference is the runt of Crusoe's portfolio

$0B$5B$10B$15BJane Street cloud contract (5-yr)September 2026 raiseManaged Inference ARR$0.1B
Data
Reported value
Jane Street cloud contract (5-yr)$13B
September 2026 raise$3.9B
Managed Inference ARR$0.1B
Managed Inference is the runt of Crusoe's portfolio. Its contracted ARR, described only as past $100 million, is a fraction of the reported five-year Jane Street cloud contract and the $3.9 billion Crusoe closed on 17 September at a $30.9 billion valuation. Bases differ: ARR is annual, the Jane Street figure is a five-year total, the raise is a single round. Sources: Crusoe and Thinking Machines Lab release [1][2]; TechCrunch, 17 September 2026 [4].1,2,4

The lab that buys compute and sells it

The payer is built for this bill. Thinking Machines raised a $2 billion seed at a $12 billion valuation in July 2025 and is in unsigned talks for $5 to $6 billion more at a $40 billion pre-money valuation, with Nvidia expected to supply about half 5. Openness is not pure cost: Tinker, the lab's fine-tuning platform, sells GPU time, and annualized revenue in the hundreds of millions is a figure relayed by people close to the company, not audited 5. Myle Ott, the lab's infrastructure lead, said the service took his team "from evaluation to production quickly" and freed money for research 1; the two companies also plan batch inference for synthetic data, a flywheel that grows the invoice as it spins 1. The margin between what Tinker bills and what Crusoe collects is visible to no one outside either company.

Cloud revenue while Texas waits

Two days before the signing, Governor Greg Abbott ordered Texas regulators to stop issuing data-center air and water permits until a grid audit is complete; Crusoe's Longhorn and Goodnight projects sit in that queue, with no firm end date 6. The cloud keeps signing while the ground waits. The next reading is the lab's own gigawatt of Nvidia Vera Rubin capacity, targeted for early 2027 5. Until then, the rent is due.

How this brief was made

01Gathered & sourced275 channels · 928 articles▾

Agents swept 275 channels and ingested 928 articles, then de-duplicated and ranked them for signal.

02Verified & cross-validated6 claims · 22 data feeds▾
03Reviewed & edited1 human editor▾

One editor read the draft against the evidence, tuned the framing, and signed off before it shipped.

Become a contributor

Reporting on the business of AI and want it read? We take pitches from outside contributors who bring primary sources and a number worth arguing about.

Share

Deepdive

AI-generated from this story and its cited sources. Not investment advice.

Reader comments

0 comments

    Sign up

    Get your curated digest

    After email confirmation, you will receive a daily digest of the most relevant news that matter to your portfolio