Nvidia / 6 October 2026

Cohere Commits $525.6 Million to Nvidia Cloud Capacity

Cohere has signed an approximately US$525.6 million agreement with Boost Run for dedicated GPU cloud infrastructure, with the provider identifying Nvidia GB300 NVL72 systems as the computing platform. The five-year service arrangement was signed on 30 September and disclosed in an October SEC filing and company announcement. Initial infrastructure acceptance is expected in the second quarter of 2027. The filing makes clear that the commitment depends on delivery and acceptance requirements, with a portion payable in advance. Nvidia’s connection is the hardware underlying the service, while the contract value belongs to Boost Run’s relationship with Cohere. The deal provides a named customer commitment for Blackwell capacity, with specific protections against late delivery.

The acceptance provisions are central to the contract’s economic meaning. Each rack’s service term begins when Cohere accepts it, rather than on a single date applied automatically to all the planned infrastructure. The agreement also prevents monthly fees being invoiced for a rack before acceptance or deemed acceptance under its terms. That ties the start of recurring billing to delivery of usable capacity. A customer reserving scarce infrastructure can therefore support the provider’s investment without immediately paying the full recurring charge for equipment still being installed. The provider must manage the interval between committing capital to that equipment and earning the service payments that repay the investment over time.

Cohere has additional protection around a July 2027 deadline. If a specified minimum amount of infrastructure has not been accepted by 15 July, it can terminate and recover prepaid amounts; if the minimum has been accepted, it can terminate the order for infrastructure that remains unaccepted and receive the associated refund. Boost Run’s parent guarantees the relevant refund obligation, limited to prepayments received and not applied against fees. The filing does not disclose the prepayment amount or the minimum rack threshold. Those missing commercial details prevent an estimate of the precise cash at risk, but the disclosed structure establishes that the customer’s upfront funding comes with meaningful delivery conditions.

Boost Run chief executive Andrew Karos called the agreement “an important step in efficiently deploying our financing into contracted revenue.” The provider says the contract lifts its total contracted value above US$2.6 billion, a cumulative measure of service commitments across agreements rather than annual sales. Its offering combines GPU capacity with networking, storage and CPU services. The filed agreement also addresses the customer’s ability to use the infrastructure within its own services, including conditions around direct resale. That matters because the commercial user of a cloud platform may itself serve many downstream organisations. Operating the equipment, managing access and supplying the surrounding services are part of what the customer is purchasing beyond the Nvidia processors.

The agreement illustrates how customer commitments can help organise an infrastructure buildout without transferring all execution risk to the customer. Prepayment supplies some funding before the full service period, while acceptance milestones and refund rights preserve leverage over delivery. Boost Run must still obtain and operate the required equipment, and the parent guarantee creates a potential cash obligation if the conditions for refunds arise. The company’s announcement does not disclose the price paid to Nvidia or the number of GPUs involved. Its US$525.6 million headline therefore cannot be treated as a chip order of the same value, even though the commitment supports a concrete use for Nvidia’s rack-scale Blackwell systems.