Thinking Machines Lab–Boost Run

Thinking Machines Lab agreed in May 2026 to pay Boost Run about $471.7M to rent 5,000 Nvidia B300 GPUs, with storage and CPU servers, for an initial 36 months — orders it cannot cancel, with fees owed whether or not it uses the servers.

A deal between Thinking Machines Lab, Boost Run

Why it matters

Thinking Machines Lab took on more of the risk of a late AI-compute build in its deal with Boost Run than Cohere did in its own deal. Thinking Machines Lab's orders are non-cancelable once placed, and fees start when Boost Run says the servers are ready, regardless of use. Cohere owes nothing for a rack until it accepts it.

Tap a figure for the line in its source.

Thinking Machines Lab gives Boost Run

  • under two order forms for an initial 36 months — non-cancelable once placed, fees non-refundable and owed regardless of actual usage, accruing from Boost Run's notice that the servers are ready

Boost Run gives Thinking Machines Lab

  • Nvidia B300s across Boost Run's data centers, with shared network storage and CPU node services
  • uptime per server each month, or fee credits of 10% to 100% of that server's month — credits against later invoices, never cash
  • the only ways out for Thinking Machines: a material breach Boost Run leaves uncured for 30 days, or, as for either side, a force-majeure outage of more than 60 days

May 28, 2026 disclosed in Boost Run's 8-K, which names Thinking Machinesupdated Nov 19, 2026signed May 21 2026 disclosed by the parties

What happened

In May 2026, Thinking Machines Lab signed a service agreement and two orders with Boost Run. Thinking Machines Lab is the AI lab founded by Mira Murati, OpenAI's former chief technology officer. Boost Run is a small Nasdaq-listed company in Illinois that rents out servers of Nvidia GPUs. The orders cover 5,000 of Nvidia's B300 GPUs, with shared storage and CPU servers, for an initial 36 months. The total is about $471.7 million, or roughly $157 million a year. Once placed, the orders cannot be cancelled. Thinking Machines Lab owes fees from the day Boost Run says the servers are ready, whether or not it uses them. It can walk away only if Boost Run leaves a serious breach unfixed for 30 days or an outage beyond either side's control lasts more than 60 days. Boost Run disclosed the deal a week later in a filing that names Thinking Machines Lab. In October, Boost Run said it had delivered the first phase of a multi-phase project to an unnamed AI company under agreements signed in May. On our reading, that refers to the Thinking Machines Lab deal. The price per GPU-hour and any up-front payment are in the order forms. Those forms were not filed.

Sources 10 documents — 6 SEC filings · 2 company releases · 1 research note · 1 other document

Research trail Found 135 documents and read 131 of them while checking this deal and related ones.

  1. SEC · SEC filingsec.gov
    States 5 figures: $471.7M contract value · GPUs · non cancelable · exits · parties
  2. SEC · SEC filingsec.gov
    States 3 figures: fees accrue · term and renewal · 99.5% sla
  3. SEC · SEC filingsec.gov
    States 2 figures: $39.87M seller scale · Q3 silence
  4. company releaseboostrun.gcs-web.com
    States first phase
  5. company releasethinkingmachines.ai
    States lab other compute
  6. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  7. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  8. SEC filingsec.gov
    Cited by the record; no single figure is quoted from it.
  9. research noteccir.io
    Cited by the record; no single figure is quoted from it.
  10. other documentgpurentals.ai
    Cited by the record; no single figure is quoted from it.

The record's own summary of its sourcing: Boost Run 8-K Item 1.01 (filed May 28 2026) and EX-10.1, the MSA; Boost Run Q2 2026 10-Q and Q3 2026 10-Q (filed Nov 12 2026); Boost Run release (Oct 26 2026); Thinking Machines Lab announcement (Mar 10 2026); CCIR (2026) — read 2026-11 · as of Nov 2026

The evidence 12 sourced figures · 1 dispute

The record's headline Thinking Machines Lab–Boost Run: ~$471.7M for 5,000 Nvidia B300 GPUs over an initial 36 months (May 21 2026) — two order forms under one MSA, non-cancelable once placed, fees non-refundable, owed 'regardless of actual usage' and running from Boost Run's ready-for-use notice; a first phase delivered by October, on the monitor's matching of an unnamed release

Size
~$471.7M combined over an initial 36 months, two order forms (8-K) · 5,000 NVIDIA B300 GPUs with shared network storage and CPU node services · non-cancelable once placed, fees non-refundable and owed regardless of usage · ≈ $157M a year spread evenly (the monitor's arithmetic) · no price, prepayment, schedule, site or MW filed
When
May 21 2026 (MSA and two order forms signed) → May 28 2026 (8-K, naming Thinking Machines) → Oct 26 2026 (Boost Run: first phase of a multi-phase project delivered under agreements signed in May, customer unnamed) → each order's initial 36 months, from start dates no filing gives
Status
signed May 21 2026; 8-K filed May 28. Boost Run's Oct 26 2026 release says the first phase of a multi-phase project was delivered 'under agreements signed in May 2026' to a customer it does not name — on the monitor's matching, this one (first-phase); how many of the 5,000 GPUs that phase holds, and when the rest follow, no document on file says. Boost Run's Q3 2026 10-Q (filed Nov 12 2026) has been read: it does not name Thinking Machines or mention the first phase (q3-silence)
Structure
a Boost Run Service Agreement (the MSA, 'Version 1.5.1') between Boost Run Inc. and Thinking Machines Lab Inc., and two Order Forms executed concurrently, May 21 2026. The MSA is Boost Run's standard master agreement for its platform, under which 'Orders may have varying term durations, ranging from hours to multiple years'; the orders carry this deal's commercial terms — 5,000 B300 GPUs with shared storage and CPU nodes, an initial 36 months, ~$471.7M combined. Each order is non-cancelable once placed and its fees non-refundable, owed regardless of usage and accruing from Boost Run's notice that the services are ready; orders renew automatically for the shorter of their term or a year unless either side gives notice. Only the MSA is filed (EX-10.1, redacted); the order forms — and with them the price, the delivery schedule and any prepayment — are not
Programme
Boost Run lab contracts
Parties
Thinking Machines Lab customer — Thinking Machines Lab Inc.; signed the MSA and two order forms, non-cancelable once placed, fees owed regardless of use from Boost Run's ready-for-use notice; may end them for Boost Run's uncured material breach (30 days) or, as either side may, a force-majeure outage of more than 60 days. No statement of its own on the deal is on file · Boost Run provider — Boost Run Inc., the listed parent, signed directly (Cohere's contract is with its subsidiary Boost Run LLC); supplies bare-metal B300 servers with storage and CPU nodes across its data center facilities; disclosed the deal, naming the customer, in its 8-K
Record
disclosed / primary source · as of Nov 2026

Flows 2 as the record states them

  • Thinking Machines Lab → Boost Runcompute-spend~$471.7M combined under two order forms for an initial term of 36 months (8-K) — non-cancelable once placed, fees non-refundable and owed 'regardless of actual usage', accruing from Boost Run's ready-for-use notice (MSA); ≈ $157M a year spread evenly (the monitor's arithmetic; delivery is phased); the price per GPU-hour and any prepayment are in the unfiled order forms
  • Boost Run → Thinking Machines Labcompute5,000 NVIDIA B300 GPUs 'across the Company’s data center facilities', with shared network storage and CPU node services — bare-metal servers on Boost Run's platform; no site, megawatts or server count filed

Resolved figures 12 each with the verbatim line from its document

  • contract-value$471.7M
    ~$471.7M combined across the two order forms signed with the MSA on May 21 2026, over an initial term of 36 months — Boost Run's own figure, in the 8-K it filed a week later (Item 1.01, filed, not furnished). It is a bundle: the orders cover the 5,000 B300 GPUs 'along with related shared network storage and CPU node services' (gpus), and the 8-K does not split the value. Spread evenly it is ≈ $157.2M a year (the monitor's arithmetic, 471.7 ÷ 3); delivery is phased (first-phase), so it is not a calendar-year revenue figure. The order forms, which set prices and any prepayment ('The prepayment amount, if any, is set forth in the applicable Order', MSA §3.d), are not filed — only the MSA is (EX-10.1)
    “Pursuant to the Orders, the Company will provide GPU server rental services for an initial term of 36 months, with a combined total contract value of approximately $471.7 million.”
    Boost Run 8-K Item 1.01 (filed May 28 2026) · sec.gov
  • gpus5,000
    5,000 NVIDIA B300 GPUs, rented as 'NVIDIA B300 GPU servers' on Boost Run's platform, 'where bare metal server provisioning takes place' (8-K), across its data center facilities — no site, megawatts or server count is filed. The MSA itself names neither the customer nor the chip: both are in the 8-K and in the unfiled order forms
    “The Orders provide for the deployment of a total of 5,000 NVIDIA B300 GPUs across the Company’s data center facilities, along with related shared network storage and CPU node services.”
    Boost Run 8-K Item 1.01 (filed May 28 2026) · sec.gov
  • non-cancelable
    the commitment binds when an order is placed, not when servers arrive: each order is non-cancelable for its term and its fees non-refundable. The 8-K adds that Thinking Machines 'is obligated to pay all fees for the full duration of the initial term or any extended term, regardless of actual usage, except as expressly set forth in the MSA or the applicable Order' (the MSA's §10.a says the same), and that 'Upon any termination, all outstanding fees become immediately due and payable' — which the filed MSA words more narrowly: on termination 'any undisputed outstanding Fees owed to Boost Run at the time of termination will become due and payable in accordance with Section 3' (§10.f)
    “Once an Order is placed on the Boost Run platform, it is non-cancelable for the specified term duration, and any fees associated with that Order are non-refundable.”
    Boost Run 8-K Item 1.01 (filed May 28 2026) · sec.gov
  • fees-accrue
    when the meter starts: at Boost Run's notice that the servers are ready for use — not at Thinking Machines' acceptance or first use — though they are payable only after acceptance ('Customer shall pay all such accrued Fees following acceptance'), and Boost Run may invoice them 'Once acceptance has occurred or the Acceptance Period has expired without written notice of non-conformance'. Acceptance is deemed on written acceptance, on production use, or when a seven-business-day testing period passes without a notice of material non-conformance (MSA, acceptance testing). Under Cohere's agreement no fee is payable on a rack before it is accepted (boostrun-cohere fees-on-acceptance)
    “For clarity, Fees shall accrue from the date Boost Run notifies Customer that the Services are ready for use, regardless of whether Customer has commenced production use, and Customer shall pay all such accrued Fees following acceptance.”
    Boost Run Service Agreement v1.5.1 — Boost Run 8-K EX-10.1 (filed May 28 2026) · sec.gov
  • exits
    Thinking Machines' ways out, as filed: for cause, on Boost Run's material breach left uncured 30 days after notice (8-K; the MSA's §10.d runs both ways, and counts Thinking Machines' failure to pay undisputed fees within 10 days of their due date as a material breach); and, in the MSA but not in the 8-K's summary, either party may end the agreement 'without further liability, effective immediately upon written notice, where a Force Majeure Event causes the interruption of Services lasting more than sixty (60) consecutive days' (§10.e). The MSA provides no termination for convenience
    “The Customer may terminate the MSA, or an applicable Order, for cause if the Company materially breaches its obligations and fails to cure such breach within thirty (30) days of receiving written notice.”
    Boost Run 8-K Item 1.01 (filed May 28 2026); MSA §10.d–e (EX-10.1) · sec.gov
  • term-and-renewal
    36 months is the INITIAL term (contract-value). After it, the MSA renews each order automatically, a year at a time, unless either side gives notice — 90 days before expiry for a term of 24 months or longer (§10.c) — or the order sets a different renewal; Thinking Machines may instead extend an order by three months to three years on three months' notice (§10.b). The MSA commences May 21 2026 and runs to the end date in the applicable order (8-K). Cohere's agreement, by contrast, 'shall not automatically renew under any circumstances' (boostrun-cohere exits)
    “Unless either Party provides written notice of non-renewal at least the applicable Notice Period before the expiration of the then-current term, the applicable Order will automatically renew for successive renewal terms equal to the shorter of (i) the length of the then-current term or (ii) one (1) year, unless a different renewal term is specified in the applicable Order.”
    Boost Run Service Agreement v1.5.1 §10.b — Boost Run 8-K EX-10.1 (filed May 28 2026) · sec.gov
  • sla99.5%
    99.5% availability per server, measured monthly; below it Boost Run credits 10% of that server's monthly fees (95.00–99.49%), 25% (90.00–94.99%), 50% (85.00–89.99%) or 100% (below 85%), capped at the month's fees for that server and issued only as credits against future invoices, with 'no cash value' (MSA Exhibit F)
    “Boost Run will provide the Services with a SLA Metric of 99.5% and above to Customer.”
    Boost Run Service Agreement v1.5.1, Exhibit F — Boost Run 8-K EX-10.1 (filed May 28 2026) · sec.gov
  • parties
    Boost Run Inc., the listed parent itself — not a subsidiary, as in Cohere's agreement — and Thinking Machines Lab Inc., which the 8-K names; 'Concurrently with the MSA, the Company and the Customer executed two Order Forms'. The 8-K adds that neither Boost Run nor its affiliates has 'any material relationship with the Customer, other than in respect of the Agreements'. Filed under Item 1.01 on May 28 2026, a week after signing; the MSA is EX-10.1, 'by and between Boost Run Inc. and Thinking Machines Lab Inc.' (exhibit index), though its own text names no customer. Boost Run's later registration statements list it as with 'Thinking Machines Labs, Inc.'
    “On May 21, 2026, Boost Run Inc. (the “Company”) entered into a Boost Run Service Agreement (the “MSA”) with Thinking Machines Lab Inc. (the “Customer”), pursuant to which the Company agreed to provide high-performance managed GPU compute and cloud infrastructure services to the Customer.”
    Boost Run 8-K Item 1.01 (filed May 28 2026) · sec.gov
  • first-phase
    the one word on delivery: Boost Run's Oct 26 2026 release says it has delivered the first phase of a multi-phase project 'to a leading AI research and product company under agreements signed in May 2026', agreements that 'provide for the deployment of NVIDIA B300 GPUs across the Company’s data center facilities, along with related network storage and CPU node services'. The release names no customer; it is read here as this deal — the same month, chips and wording as the 8-K — a match the release does not make. How many of the 5,000 GPUs that phase holds, and when the rest follow, it does not say. An unfiled company release
    “Boost Run has also delivered the first project phase of a multi-phase project to a leading AI research and product company under agreements signed in May 2026.”
    Boost Run release, Oct 26 2026 (company PDF) · boostrun.gcs-web.com
  • seller-scale$39.87M
    what the contract is to Boost Run: its whole revenue was $39.87M in the quarter to Sep 30 2026 (Q3 10-Q) and $31.14M in the quarter Thinking Machines signed (Q2 10-Q), while this one contract, spread evenly, is ≈ $39M a quarter (the monitor's arithmetic, 471.7 ÷ 12) once fully delivered; its revenue for all of 2025 was $26.89M (S-1; boostrun entity). The Q3 10-Q puts the quarter's $35.59M rise in leasing revenue 'primarily' down to 'the execution of certain significant long-term customer contracts', naming none; Boost Run's filings on file give no 2026 revenue by customer
    “Total revenue was $39.87 million for the three months ended September 30, 2026, an increase of 492% compared to $6.74 million for the three months ended September 30, 2025.”
    Boost Run Q3 2026 10-Q — MD&A, Financial Summary · sec.gov
  • q3-silence
    what Boost Run's Q3 2026 10-Q (filed Nov 12 2026) does not say: it names neither Thinking Machines nor Cohere, gives no customer's name, and does not mention the first phase delivered or any acceptance, so nothing in it confirms or adds to this contract's size, price, start or delivery. Its commitments note describes only an unnamed June 8 2026 order of 240 GPU servers for three years and 'multiple long-term customer agreements' entered in September 2026 'with multi-year terms extending through 2032' — September, not May, so not read as this deal
    “During September 2026, the Company entered into multiple long-term customer agreements with multi-year terms extending through 2032.”
    Boost Run Q3 2026 10-Q — Note 18, Commitments and Contingencies · sec.gov
  • lab-other-compute
    Boost Run is one of Thinking Machines' suppliers, not its main one on any document's account: two months earlier the lab announced a partnership with Nvidia to deploy at least a gigawatt of Vera Rubin systems, 'targeted for early next year', and a month earlier Google Cloud announced a new agreement to supply it A4X Max VMs with Nvidia GB300 GPUs (Apr 22 2026). Neither has a record yet (the thinkingmachines entity). 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

Contested figures 3 not reconciled — every position side by side, with its source and tier

price-per-gpu-hour3 positions, side by side
not stated — neither the 8-K nor the filed MSA gives a rate; the order forms that carry it (the MSA's invoices show 'the hourly rate per GPU price') are not filed Boost Run 8-K (filed May 28 2026) and EX-10.1 · sec.gov
~$3.59 per GPU-hour, bundled — the contract value divided out: $471.7M ÷ 5,000 GPUs ÷ 3 years ÷ 8,760 hours, as CCIR's table prints it ('3.59 from contract value; includes storage, CPU') and gpurentals' too. The value includes the storage and CPU nodes, so this is a ceiling on the GPUs' own rate, not a rate arithmetic on the 8-K (CCIR's and gpurentals' tables, 2026)
~$3.40 per GPU-hour, GPU-only — CCIR's estimate after costing the CPU nodes and storage 'at posted cloud rates, at half of list' and removing them; a model of the split, not a figure either party states CCIR, B300 rental prices per GPU-hour: signed contracts since March 2026 (2026) · ccir.io

Scope note what this record leaves to other records

NOT this record (carried elsewhere or not carried): Cohere's contract → boostrun-cohere. Boost Run's other customer orders (none names its customer), its company-wide 'TCV' (a metric that counts LOIs and renewal options), ARR, revenue projections, Dell purchasing agreement, project financing, leases and power, SPAC terms, warrants and founder voting control → the boostrun entity or not carried here; two figures (the founder's ~87% of the vote, the SPV's 1,968,750 earnout shares) describe Boost Run's share structure, not this deal, and are not carried. Thinking Machines' funding (the $2B seed; the reported September 2026 round) and its Nvidia and Google Cloud compute → the thinkingmachines entity. Not adopted: B300 rental benchmarks (Silicon Data's index, gpurentals' listed rates, CCIR's contract trend), hardware-cost models (Wing VC, CCIR) and an analyst's ~$67K capex per GPU (Leki, backed out of a ratio) — none is a term of this contract; gpurentals' table calls this customer 'Not named' while citing the 8-K that names it.

More between Thinking Machines Lab and Boost Run

The record holds no other deal between them.

Everything between Thinking Machines Lab and Boost Run →