Skip to content

Business

TensorWave signs six-year deal to supply OpenAI with 457 MW of AMD-powered infrastructure

Edited and reviewed by AI Ping Editorial Team

3-line summary

  1. TensorWave will provide dedicated capacity on AMD hardware for OpenAI at multiple U.S. data center sites under a six-year deal.
  2. The first capacity is expected in the first half of 2027, with full deployment targeted for early 2028.
  3. The 457 MW figure refers to IT power capacity. TensorWave did not disclose the site locations or GPU count.
BeginnerBeginners
Rows of black server racks in a fictional teal-lit data center
AI-generated conceptual image: rows of AI server racks in a fictional data hall · AI Ping

TensorWave announced November 16 that it had signed a six-year contract to provide OpenAI with data center infrastructure powered by AMD GPUs. The deal covers 457 megawatts of IT power capacity across multiple U.S. sites. TensorWave expects the first capacity to come online in the first half of 2027 and the full deployment to be in place by early 2028.

The contracted infrastructure is dedicated to OpenAI. TensorWave says it is building the systems for model training and inference, the work of running models after training. The announcement gives OpenAI an additional named infrastructure provider and a schedule for that deployment. It leaves the size of the first phase and the arrival dates at individual sites open.

What the 457 MW measures

TensorWave’s release describes the figure as IT capacity. The U.S. Department of Energy’s data center metering guide treats servers, storage and networking gear as IT electrical loads. Total facility power also accounts for equipment that supports the IT load, including building cooling, power distribution and lighting. What a facility draws from the grid depends on those supporting systems and how much of the equipment is in use.

Megawatts describe power. Megawatt-hours measure the energy used over time, which depends on the level and duration of the load. TensorWave’s 457 MW is a combined planned capacity across sites. The announcement contains no facility-by-facility breakdown or expected usage, so it gives no basis for calculating how much energy an individual site would consume.

A power distribution cabinet and a server rack linked by a feeder
AI-generated conceptual image: fictional power distribution equipment supplying a server rack · AI Ping

How the AMD equipment fits together

The facilities will use AMD Instinct MI455X accelerators in Helios racks. AMD’s reference design combines 72 GPUs with server processors and networking equipment in one rack. Its MI455X specifications list 432 GB of HBM4 memory per accelerator and direct liquid cooling. AMD gives the rack up to 31 terabytes of high-bandwidth memory. That memory gives the accelerators fast access to model data during training and inference. The published configuration describes the hardware design. TensorWave gave no figures for the number of racks or individual accelerators under the OpenAI contract.

TensorWave plans to connect the racks over high-speed Ethernet. Large training jobs divide work among multiple accelerators, so the links that carry data between systems affect how quickly the job runs. The power available to the equipment measures a different part of the system. TensorWave had already announced adoption of the MI455X-based Helios design for its own cloud in July. The new contract names a customer and a scale for a dedicated deployment.

Stacked computing trays and blue coolant tubes inside a fictional server rack
AI-generated conceptual image: computing trays and cooling lines arranged within a fictional rack · AI Ping

Where it fits in the existing AMD partnership

OpenAI and AMD announced a 6-gigawatt agreement across several generations of AMD hardware in October 2025. The companies said at the time that the first 1-gigawatt deployment, using MI450-series equipment, was scheduled to begin in the second half of 2026. In earnings remarks on November 3, 2026, AMD said it had started shipping the MI455X in volume in the third quarter. It named OpenAI among customers that had brought their first Helios racks online.

AMD’s filing for the 2025 deal says OpenAI’s binding commitment for the initial gigawatt could be fulfilled directly, through an affiliate or through authorized third-party buyers. A related stock warrant carries separate vesting conditions tied to delivery and purchase milestones. TensorWave’s new announcement describes its role operating and supplying the dedicated data center capacity.

OpenAI compute chief technology officer Uday Ruddarraju described the TensorWave deal as an extension of work under the existing partnership in TensorWave’s release. The companies gave no breakdown of which phase of the 6-gigawatt target the 457 MW would count toward. TensorWave also left the individual sites and contract price undisclosed.

Sources