Why it matters
Thinking Machines agreed to pay $65M a year for Crusoe to run the lab's open models on a dedicated B200 cluster under a service-level agreement. The price covers both chips and operations. With no public contract term or GPU count, the annual rate leaves both total spending and the size of the cluster unknown.
Tap a figure for the line in its source.
Thinking Machines Lab gives Crusoe
- for Crusoe Managed Inference under an annual agreement — no term, total, minimum commitment or exit terms stated
Crusoe gives Thinking Machines Lab
- a dedicated cluster of Nvidia HGX B200 systems on InfiniBand that Crusoe runs and supports as an SLA-backed endpoint — no GPU count, megawatts or site stated
- production serving of the lab's Inkling models, GLM 5.2 and 5.3 and its own fine-tuned variants; batch inference for synthetic data is a stated next step, not yet in scope
What happened
In September 2026, Thinking Machines Lab agreed to pay Crusoe $65 million a year to run the lab's AI models for its users. Mira Murati, OpenAI's former chief technology officer, founded the AI lab. Crusoe is best known for building the Stargate data center in Abilene and also runs a cloud business. It dedicates a cluster of Nvidia B200 GPUs to the lab and operates it with its own engineers. A service-level agreement commits Crusoe to a specified level of performance. Crusoe runs the lab's open models and fine-tuned versions of those models. One of the open models is the lab's own Inkling, which anyone can download. The lab was already using the service on the day the deal was announced. The announcement gives no contract length, GPU count or location. Crusoe said the lab joins customers who have brought this business past $100 million a year in contracted revenue. Crusoe made the same claim six days earlier, so the lab's share of that revenue is undisclosed.
Sources 18 documents — 9 company releases · 3 press reports · 3 other documents · 1 social post · 1 SEC filing · 1 market-data page
Research trail Found 70 documents and read 68 of them while checking this deal and related ones.
- States 5 figures: $65M annual value · deployment · workloads · next step · $100M seller scale
- States 2 figures: lab other compute · nvidia ties
- States in production
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
- Cited by the record; no single figure is quoted from it.
The record's own summary of its sourcing: Crusoe release (Sep 23 2026; newsroom and GlobeNewswire), LinkedIn post (Sep 23 2026), Series F release (Sep 17 2026), Managed Inference launch release (Nov 20 2025), price list and Self-Serve Deployments post (Jul 14 2026), blog (Oct 21 2026); Thinking Machines Lab announcement (Mar 10 2026); Dow Jones via Tiger Brokers (Sep 17 2026); The Next Web (Sep 23 2026); The Inference (Sep 26 2026); SAVRN; The Information via Dealroom (Sep 23 2026) — read 2026-11 · as of Oct 2026
The evidence 8 sourced figures · 3 disputes
The record's headline Thinking Machines Lab–Crusoe: $65M a year for Crusoe to serve the lab's open models on a dedicated Nvidia HGX B200 deployment it runs (Sep 23 2026) — inference bought as an SLA-backed managed service, not GPUs by the hour; no term, GPU count, site or price stated, and whether the lab is most of Crusoe's 'past $100 million' inference book or an addition to it, no document says
- Size
- $65M a year (Crusoe's release) · Crusoe Managed Inference on a dedicated Tailored Deployment of NVIDIA HGX B200 systems with Quantum-2 InfiniBand · no term, total, minimum, GPU count, megawatts, site or SLA level stated · ~770–879 B200s is arithmetic on list prices, not a disclosure (contested)
- When
- Sep 23 2026 (announced, in production; no signing date given) → Oct 21 2026 (Crusoe names the lab among teams 'already building on Crusoe Cloud') → a term no document states
- Status
- announced Sep 23 2026 and live: Crusoe's post the same day says the lab 'is running its production workloads on Crusoe Managed Inference' (in-production); batch inference for synthetic data generation is a stated next step (next-step). No document on file gives a signing date
- Structure
- an annual agreement for Crusoe Managed Inference, delivered as a Tailored Deployment — the bespoke tier of Crusoe's inference service, in which its engineers optimize and run a dedicated, benchmarked, SLA-backed endpoint for the customer's models, beside Serverless Inference (priced per token) and Self-Serve Deployments (priced per GPU-hour). Crusoe's price list publishes no Tailored Deployments price. The agreement itself is not public: its term, any minimum commitment, its termination rights and the SLA's level are not disclosed; Crusoe's standard online terms are not it
- Parties
- Thinking Machines Lab customer — Thinking Machines Lab; pays $65M a year for Crusoe to serve its production models, without managing its own inference stack. Speaks in the release only through its ML infrastructure lead, Myle Ott; none of its own pages on file mentions Crusoe · Crusoe provider — runs and supports the dedicated HGX B200 cluster as a Tailored Deployment of Crusoe Managed Inference on Crusoe Cloud; issued the release (its newsroom and GlobeNewswire)
- Record
- disclosed / primary source · as of Oct 2026
Flows 2 as the record states them
- Thinking Machines Lab → Crusoecompute-spend$65M a year under 'a $65 million annual agreement' (Crusoe's release, Sep 23 2026) — no term, total, minimum commitment or termination terms stated, so no total is drawn; it pays for a managed service (GPUs, Crusoe's engineering, an SLA), not GPU-hours
- Crusoe → Thinking Machines LabcomputeCrusoe Managed Inference on a dedicated Tailored Deployment of NVIDIA HGX B200 systems with Quantum-2 InfiniBand, run and supported by Crusoe as an SLA-backed endpoint — no GPU count, megawatts, SLA level or site stated
Resolved figures 8 each with the verbatim line from its document
- annual-value$65M$65M a year — 'a $65 million annual agreement', Crusoe's own figure in the joint release it issued on Sep 23 2026 (its newsroom; the GlobeNewswire copy, 11:07 ET, has the same words). Neither company files with the SEC, so nothing filed or audited stands behind it; Thinking Machines speaks in the release only through a quote that gives no figure. The release states no term, total contract value, minimum commitment or termination terms, so $65M a year is neither a total nor a floor on one. It pays for a managed service — the GPUs, Crusoe's engineers running them and an SLA — not for GPU-hours (contested deployment-size)
“Crusoe, the industry's first vertically integrated AI infrastructure provider, and AI innovator Thinking Machines Lab today announced a $65 million annual agreement to power inference for the lab's open models on Crusoe Cloud.”
Crusoe release (Sep 23 2026) · crusoe.ai - deploymenta dedicated Tailored Deployment of NVIDIA HGX B200 systems connected with NVIDIA Quantum-2 InfiniBand (workloads), which Crusoe runs and supports, so the lab does not manage its own inference stack. Tailored Deployments are the bespoke tier of Crusoe Managed Inference — 'Work directly with Crusoe's Engineering team for the highest level of optimization and a dedicated, benchmarked, SLA-backed endpoint' (Crusoe's Self-Serve Deployments post, Jul 14 2026) — and Crusoe's price list gives them no price, unlike its Serverless token rates and Self-Serve GPU-hour rates. The SLA's level, the GPU count, the megawatts and the site are not stated (contested deployment-size, site)
“Crusoe will run and support the cluster directly as a Tailored Deployment – a dedicated, benchmarked, SLA-backed endpoint – giving Thinking Machines Lab the price-performance and headroom to scale without managing its own inference stack.”
Crusoe release (Sep 23 2026) · crusoe.ai - workloadsinference, not training: production serving of the lab's Inkling models (Inkling is 'Our Open-Weights Model', the lab's post of Jul 15 2026), GLM 5.2 and 5.3 (open models developed by Z.ai, per The Next Web) and the lab's own fine-tuned variants. One summary of The Information's report on the deal (AI Market Watch) calls it 'dedicated GPU capacity for model training'; the release says 'to power inference', and Dealroom's summary of the same report says 'inference not training' — not adopted. The lab's training compute is elsewhere (lab-other-compute)
“Through Crusoe Managed Inference, Thinking Machines Lab will serve a range of production workloads – including its Inkling models, GLM 5.2 and 5.3, and its own fine-tuned variants – on a dedicated Tailored Deployment of NVIDIA HGX B200 systems connected with NVIDIA Quantum-2 InfiniBand networking, engineered for high-throughput, cost-efficient serving at scale.”
Crusoe release (Sep 23 2026) · crusoe.ai - in-productionlive when announced: Crusoe's company post on LinkedIn the same day says the lab is running its production workloads on Managed Inference, 'on a dedicated, SLA-backed Tailored Deployment we run and support'; the lab's ML infrastructure lead, quoted in the release, says the service 'took us from evaluation to production quickly'. On Oct 21 2026 Crusoe again named Thinking Machines Lab among 'teams like Thinking Machines Lab, Ollama, Base44, Fireworks, Roboflow, and SpAItial' already building on Crusoe Cloud (its blog). No document on file gives a signing date
“Thinking Machines Lab is running its production workloads on Crusoe Managed Inference.”
Crusoe on LinkedIn (Sep 23 2026) · linkedin.com - next-stepa stated intention, not part of the announced agreement: batch inference for synthetic data generation — the companies are 'looking to expand into batch inference for large-scale synthetic data generation' (the release); Crusoe's LinkedIn post the same day: 'Next, we're working with Thinking Machines Lab on batch inference for synthetic data generation'. No size, price or date is given
“The companies are also looking to expand into batch inference for large-scale synthetic data generation, turning inference into a flywheel for research.”
Crusoe release (Sep 23 2026); Crusoe on LinkedIn (Sep 23 2026) · crusoe.ai - seller-scale$100Mwhat the deal sits in: Crusoe Managed Inference, generally available since Nov 20 2025 (its launch release), at more than $100M in contracted ARR — a floor with no ceiling, no definition of ARR and no customer breakdown. Crusoe had claimed the same floor six days earlier ('have already contracted over $100M in ARR', its Series F release, Sep 17 2026), and its CEO told the Wall Street Journal the business went 'to an annual revenue run rate of more than $100 million by the summer' (Dow Jones, Sep 17 2026) — revenue being billed, a different measure from contracted ARR. How much of the floor this deal is, the release does not say (contested share-of-managed-inference). An analyst's count: 'Thinking Machines Lab is the one stated cloud contract with a dollar figure' among Crusoe's customers (SAVRN)
“Thinking Machines Lab joins a roster that has taken Crusoe Managed Inference past $100 million in contracted ARR less than a year from launch.”
Crusoe release (Sep 23 2026) · crusoe.ai - lab-other-computethis is the lab's inference supplier, one of several: its training compute is the Nvidia partnership it announced on Mar 10 2026 — at least a gigawatt of Vera Rubin systems, deployment 'targeted for early next year' — and it rents Nvidia B300 GPUs from Boost Run under orders it cannot cancel (Boost Run's 8-K of May 28 2026; boostrun-tml). Those are other deals, and none of their figures is this one's. The lab's own announcement
“Thinking Machines Lab and NVIDIA announced today a multi-year strategic partnership to deploy at least one gigawatt of next-generation NVIDIA Vera Rubin systems to support Thinking Machines’ frontier model training and platforms delivering customizable AI at scale.”
Thinking Machines Lab announcement (Mar 10 2026) · thinkingmachines.ai - nvidia-tiesNvidia is on both sides of the deal and under it: it invested in Thinking Machines — in its seed round (TechCrunch, Jul 15 2025) and again, in an amount not stated, in March 2026 — is named among the investors in Crusoe's Series F (Crusoe's release, Sep 17 2026), and makes the HGX B200 systems the deployment runs on. No document sizes either investment or ties any of it to this agreement
“NVIDIA has also made a significant investment in Thinking Machines Lab to support the company’s long-term growth.”
Thinking Machines Lab announcement (Mar 10 2026) · thinkingmachines.ai
Contested figures 8 not reconciled — every position side by side, with its source and tier
Scope note what this record leaves to other records
NOT this record (carried elsewhere or not carried): Crusoe's Series F — the release's 'anticipated $3.9 billion' at a $30.9B post-money valuation against the ~$3.11B its Form D showed sold at filing (relayed by SAVRN) — its 6 GW+ of gross contracted capacity, $140B+ of total contracted value and 20x+ bookings growth, its campuses (Abilene, the Microsoft expansion there, Wyoming, Texas), credit facilities and the ~$13B five-year Jane Street contract Bloomberg reported (unconfirmed by either party) → the crusoe entity or not carried here. Thinking Machines' funding, its Nvidia partnership and its Boost Run orders → the thinkingmachines entity and boostrun-tml. Not adopted: B200 rental benchmarks (SAVRN's index, Silicon Data, William Blair's list prices) except as inputs to the contested GPU arithmetic; Crusoe's speed claims (up to 9.9x faster time-to-first-token than vLLM, a vendor benchmark whose own blog reports a 4x median) and third-party speed tests; Crusoe's standard online terms (a liability cap, termination for convenience) — generic terms, not this agreement's; SemiAnalysis's modeled token margins for Crusoe; AI Market Watch's 'dedicated GPU capacity for model training' (workloads); The Inference's 'latest valuation' tags for both companies, which give no source or date.