Boost Run has secured a large customer commitment before the infrastructure behind it enters service. Its approximately $525.6 million agreement with Cohere reserves dedicated GPU cloud infrastructure and related services, but the roughly five-year term starts separately for each rack only after Cohere accepts it.
That structure makes delivery the central event, not a footnote to the contract announcement. Initial acceptance is expected in early Q2 2027. A July 15, 2027 deadline gives Cohere termination and refund rights if minimum capacity is not accepted, with additional protections for unfinished portions even if that minimum is reached.
The contract starts with a promise, not completed revenue
Boost Run LLC and Cohere signed the service agreement on September 30, 2026, and Boost Run disclosed it on October 6 through a regulatory filing and company announcement. It was the young provider’s largest disclosed customer commitment at the time.
The planned package includes NVIDIA GB300 NVL72 systems, network storage and CPU-node services. NVL72 links 72 Blackwell-generation GPUs in a high-bandwidth rack-scale design intended for demanding training and inference. That architecture description is not a disclosure of this project’s GPU count.
Each accepted rack begins its own approximately five-year service period, so different parts of the deployment can start at different times. The agreement does not automatically renew; it remains effective until the orders expire or an earlier contractual termination occurs.
The $525.6 million is total contract value, or TCV: payments committed over the agreement’s life, subject to delivery, acceptance and termination conditions. It is not present-period revenue, collected cash or profit. Part is payable in advance, but the disclosed documents do not reveal that amount.
Boost Run says the deal took portfolio-wide TCV above $2.6 billion. This indicates future business under signed arrangements, provided the relevant obligations are fulfilled and agreements remain effective. It does not establish that the corresponding infrastructure is already installed, accepted or generating cash from operations.
July’s deadline creates two different customer remedies
If Cohere has not accepted the specified minimum infrastructure by July 15, 2027, it can terminate the agreement and receive a refund of all prepaid amounts. The public filing does not disclose the minimum threshold, so an outside reader cannot calculate the exact capacity needed to avoid that outcome.
Reaching the minimum does not remove every remedy. Cohere can still terminate the relevant order portion covering infrastructure not accepted by the deadline and recover the associated prepayments. The distinction between an agreement-wide remedy and an unfinished-order remedy is material.
Those protections define consequences, not a prediction that either party expects failure. Boost Run’s parent also guarantees certain refunds linked to unused prepayments. Advance cash can support construction, but the company must preserve its ability to meet service and refund obligations.
Customer acceptance is separate from ordering or installing hardware. Cohere must confirm that the delivered infrastructure meets contractual requirements before a rack’s service term begins. The public documents omit the technical acceptance tests and much of the detailed delivery schedule.
The companies also have not disclosed the final hosting facilities, project megawatts, total GPUs or individual rack timetable. It would be speculative to infer those numbers by dividing contract value by assumed GPU rental prices. Equally, the absence of a location announcement alone does not prove delay.
A much larger backlog meets a smaller operating business
Boost Run began Nasdaq trading in May 2026 after combining with Willow Lane Acquisition Corp. At that point it reported $940 million in contracted customer revenue. The subsequent TCV growth describes an expanding portfolio of obligations to finance, deliver and operate, as well as demand.
Its second-quarter filing supplies a different measure of scale: $31.14 million in quarterly revenue and an operating loss of approximately $12.9 million. These period results cannot be directly compared with one contract’s five-year value, but they show why future commitments and current operations must be kept distinct.
At June 30, 2026, the company reported $120.17 million in unrestricted cash and a $23.58 million working-capital deficit. It explained that customer prepayments contribute to current liabilities because deposits are recorded before services are delivered. Prepayments can bring cash in ahead of revenue recognition without becoming earned profit.
The Cohere arrangement illustrates both the financing benefit and its constraint. Deposits can help fund capacity, while acceptance milestones and refund guarantees keep that funding tied to performance. More signed business can improve visibility for suppliers and lenders while increasing the capital and execution burden.
Data Center Dynamics reported that Boost Run operated from several TierPoint facilities at the end of 2025 and listed more locations in 2026. It also reported quarterly colocation lease costs of $10.56 million, up 576% year over year. Neither the existing sites nor higher spending establish that Cohere’s future cluster has been installed or accepted; the company has not linked particular leases to the contract.
Cohere is buying optionality as well as NVIDIA capacity
Cohere develops the Command model family and North enterprise platform. Customers concerned with privacy, deployment control, regional hosting and regulation may need dedicated capacity, defined data residency, isolated networking and private deployments alongside model quality.
Sovereign AI adds requirements for a country or institution to control data, infrastructure and governance. Dedicated clusters can support particular hosting arrangements and more predictable scheduling than fluctuating shared accelerator availability. These are reasons for this infrastructure model, not proof that an undisclosed Boost Run location meets every customer requirement.
The NVIDIA commitment is not evidence that Cohere has abandoned AMD. An expanded AMD collaboration announced in September 2025 made Cohere’s models and North available on AMD Instinct infrastructure. The two relationships instead show access to different hardware platforms and partners.
Such diversity can match workloads to performance, memory, availability and geographic needs while reducing reliance on one supplier. It also leaves Cohere with long-term payments after acceptance and Boost Run with an ongoing availability obligation. Supply certainty comes with infrastructure rigidity.
Hardware will evolve during a five-year term. Public documents do not explain upgrades, substitutions or technology transitions; commercially sensitive agreement portions were omitted from the public exhibit under securities rules. Future software readiness and utilization therefore matter as much as obtaining an initial hardware allocation.
Neocloud economics depend on the surrounding system
Boost Run describes a platform combining bare-metal GPU servers, managed Kubernetes, network storage, CPU nodes and private networking. Bare metal provides direct access to physical machines; managed Kubernetes coordinates containerized workloads. Customers need that surrounding scheduling, identity, network and storage stack, not simply isolated accelerators.
Specialized GPU clouds, often called neoclouds, compete through dedicated clusters, engineering support and deployment flexibility. CoreWeave is a prominent public example. A customer can reserve a specialized cluster while retaining a hyperscaler for general services; the contract does not show that every workload is leaving public clouds.
Hyperscalers retain scale, broad services, established procurement and integrated software. A smaller provider must show reliable operations, network performance, support and financial stability. Its risks include hardware procurement, facility availability, capital requirements, competition and customer concentration as individual commitments grow.
The dependencies are linked: equipment arrival, electrical density, cooling, network validation, storage, CPU services and Cohere’s training and inference software must align. A delay in one can hold back the service. Installed machines and increased lease spending cannot substitute for the customer’s acceptance.
Nor does the agreement automatically change Cohere’s products now. Acceptance is planned for 2027, and the companies have not announced a customer-facing launch directly tied to these systems. Enterprise use, private deployments and sovereign projects explain prospective demand, not completed product releases.
The next evidence must close the delivery gap
A confirmed location and commissioning milestones would improve visibility into power, connectivity and facility readiness. Continued silence near the expected acceptance window would leave an important question open, but would still not by itself prove a contractual breach.
Initial acceptance in early Q2 2027 would support the announced schedule. Updates should distinguish delivered, installed and accepted capacity; only the last activates each rack’s term. Slippage toward July 15 would make the minimum threshold and partial-order remedies more important.
Financial conversion is the other test. Revenue, operating cash flow, infrastructure costs and customer-deposit obligations need to be considered alongside TCV, rather than treating a larger backlog as the finished outcome. Cohere’s own NVIDIA deployment announcements and continued AMD expansion could clarify how the reserved capacity fits its broader strategy.
The related knowledge-base article below discusses organizing evidence, not validating this deal. The same distinction remains useful here: company projections, signed commitments and completed milestones are different kinds of information. The decisive development will be accepted racks entering an operating service, not another large headline contract number.
Sources and related context
Service agreement disclosure · Boost Run announcement · Quarterly filing · AMD and Cohere collaboration · Data-center reporting · Knowledge-base context · Original reporting aggregation entry—not a direct publisher link