A signed slice of a much larger promise
The AMD–OpenAI partnership announced in October 2025 was a headline number: 6 gigawatts of AMD GPUs over multiple years and generations. Headline numbers do not run models. Somebody has to build the sites, install the racks and operate them. On November 16, 2026, TensorWave put its name to a specific slice of that work. TensorWave describes itself as the first publicly announced neocloud supporting the AMD–OpenAI agreement.
Under a six-year agreement, TensorWave will provide OpenAI with 457 megawatts of IT capacity across multiple US data center sites. The hardware is AMD's Instinct MI455X, in systems purpose-built for AMD's Helios rack-scale design, linked by high-performance Ethernet. TensorWave will operate it as dedicated, single-tenant capacity, reserved for OpenAI rather than shared with other customers. TensorWave describes itself as an AMD Helios launch partner and says this is its largest deployment to date.
Put the 457 megawatts against the partnership it belongs to and the scale changes shape. ALFW's arithmetic puts it at about 7.6% of the 6 gigawatts AMD and OpenAI announced. That does not make it small. It makes it an early named piece of the program, and the release does not establish the allocation or delivery status of the remainder.
Announced is not the same as available
The release establishes no delivered capacity under this agreement. TensorWave expects capacity to begin coming online in the first half of 2027, with the full deployment expected by early 2028. Those are the company's expectations, stated in its own release, not completed milestones. The release states a six-year agreement and separately gives expected deployment dates; it does not specify when the term begins, and nothing in it is relief for anyone short of compute this quarter.
The release also leaves out the two details that would let a reader price the deal. It discloses no dollar value, and it names no sites, only "multiple U.S. data center sites". TensorWave says it has secured more than 2 gigawatts of long-term data center capacity overall. Secured capacity, like announced capacity, is a claim about what a company controls or intends to build. It is not evidence that 457 megawatts of finished, powered, networked racks exist.
Why the AMD part is the actual story
OpenAI's infrastructure has mostly meant Nvidia hardware, its own Stargate sites, and large cloud partners. This agreement is a test of whether that pattern widens. TensorWave says it operates one of the largest AMD-exclusive AI clouds, and it presents this deployment as proof that AMD infrastructure can run at the scale frontier labs require. AMD's Vamsi Bopanna is quoted in the release calling TensorWave "a strong partner in scaling AMD Instinct infrastructure". OpenAI's CTO for Compute, Uday Ruddarraju, is quoted welcoming the agreement as a way to advance the previously announced AMD partnership.
Read those quotes for what they are. Both are statements inside TensorWave's own announcement. They confirm the parties are aligned publicly. They do not confirm performance, cost per token, or how much of OpenAI's real workload will end up on MI455X systems once the capacity arrives. The release supplies no performance or cost-per-token results.
What to watch
Delivery can be assessed against the announced schedule, with site and commissioning disclosures providing additional evidence. Three things to track: whether TensorWave names sites and construction starts; whether the H1 2027 start date holds; and whether other neoclouds are named for further slices of the 6 gigawatts, which would show the AMD program naming more of its build-out. Until then, count the megawatts when they switch on.
Sources
- [1] TensorWave announcement, Business Wire, November 16, 2026 (primary release; all figures, quotes and timeline above; read first-hand by ALFW)Read source
- [2] AMD–OpenAI strategic partnership announcement, October 2025 (6 GW context) (label only; no URL recorded)
- [3] Data Center Dynamics, "OpenAI signs on for 457MW AI capacity with TensorWave", November 16, 2026 (secondary; not relied on for core figures)Read source