The six-gigawatt agreement between OpenAI and AMD finally has a first named builder. On Nov 16, TensorWave, an all-AMD cloud provider, said it had signed a six-year agreement to supply OpenAI with 457 megawatts of dedicated IT capacity across several US data center sites, built on AMD Instinct MI455X GPUs inside AMD Helios rack-scale systems. Capacity begins coming online in the first half of 2027, with the full footprint deployed by early 2028.
The capacity is single-tenant, meaning the halls are purpose-built for one customer rather than rented by the hour alongside other workloads. Neither side disclosed the contract value, which is the number that would settle how much of the 457 megawatts is already underwritten and how much depends on OpenAI expanding its own demand.
The structural significance is that this is the first neocloud, meaning a privately funded operator that rents out accelerators rather than running a general-purpose cloud, to be publicly named as an infrastructure provider under the OpenAI-AMD agreement. AMD signed that six-gigawatt framework in October 2025 with a warrant for up to 160 million shares that vests as OpenAI takes delivery of chips. A framework with a warrant attached is a promise; a named operator with megawatts and delivery dates is a schedule, which is what the warrant structure actually rewards.
For TensorWave the deal concentrates risk as much as it validates the business. The company is private, raised a Series B in June 2026, and has secured more than two gigawatts of long-term data center capacity, according to the announcement. Committing 457 megawatts to a single tenant on hardware built exclusively for AMD Helios racks means revenue and debt service both rest on one contract: if OpenAI slows, TensorWave has little room to redirect that capacity, while AMD keeps selling chips elsewhere and OpenAI can buy compute from many vendors. Magnetar, AMD Ventures and Nexus Venture Partners are underwriting that concentration.
Both sides framed the agreement around delivery rather than ambition. Uday Ruddarraju, who leads compute at OpenAI, said that as AI demand grows, securing reliable large-scale compute is the key input for building and shipping more capable models. Piotr Tomasik, co-founder and president of TensorWave, argued that the ability to deliver hundreds of megawatts of advanced capacity reliably and quickly is now what separates infrastructure providers that can support frontier work from those that cannot.
Context matters for why the schedule is being read as an AMD story rather than a cloud story. Nvidia has spent the past week absorbing the resignation of Jensen Huang and the board search that followed, and its shares have fallen sharply; its own quarterly results land after Tuesday close. A competing platform that can point to named sites, an operator and a 2027 delivery date is the kind of proof investors asked AMD for when the OpenAI warrant was announced.
What this changes is the unit of measurement. AI infrastructure contracting is moving from annual purchases and spot capacity toward multi-year, single-tenant leases measured in megawatts, signed before the steel is bolted down. That makes power, shells and grid interconnects the binding constraint rather than chip allocation, and it shifts credit risk onto whichever party owns the building. For OpenAI the trade is flexibility for certainty; for TensorWave it is a signature customer against a balance sheet that now carries the bet.
Comments (0)
Log in to join the discussion
Log InNo comments yet