Vendors & EquipmentAmdOpenaiTensorwaveAI Infrastructure
OpenAI signs for 457MW of dedicated AMD-based AI capacity
OpenAI’s six-year TensorWave agreement covers 457MW of dedicated AMD Instinct MI455X capacity, with operations due in 2027 and full deployment by early 2028.
Vendors & EquipmentWhy it matters
- OpenAI signed a six-year, 457MW capacity contract with TensorWave.
- Operations are expected in H1 2027, with full deployment by early 2028.
- AMD says one Helios rack provides up to 2.9 exaflops FP4, 1.4 exaflops FP8, 31TB HBM4 and 1.7PB/s of memory bandwidth.
- AMD claims 15% more peak FP4 performance and 50% more HBM capacity than Nvidia’s Vera Rubin NVL72 specifications.
- TensorWave raised $350 million in June 2026 and claims 2GW of long-term US data-center capacity.
The story
By Elena Vasquez
OpenAI has signed a six-year agreement with AMD-focused neocloud TensorWave for 457MW of dedicated AI compute capacity, the provider’s largest deployment contract to date.
The contract covers multiple data-center locations, with IT capacity powered by AMD Instinct MI455X GPUs in Helios rack-scale systems. TensorWave will operate the systems as dedicated, single-tenant capacity for OpenAI. Operations are expected to begin in the first half of 2027, and the deployment should reach full status by early 2028.
What will OpenAI receive?
AMD launched the Helios racks in July. They combine MI455X accelerators with 6th-generation AMD EPYC “Venice” server CPUs, AMD ROCm software and AMD Pensando networking within one rack-scale architecture.
AMD’s published specifications for a single Helios rack list:
- Up to 2.9 exaflops of peak FP4 performance.
- Up to 1.4 exaflops of peak FP8 performance.
- 31TB of HBM4 memory.
- 1.7 petabytes per second of memory bandwidth.
How do AMD’s rack claims compare?
AMD says Helios delivers 15% more peak FP4 performance than Nvidia’s published specifications for the Vera Rubin NVL72 rack. It also claims 50% more high-bandwidth memory capacity, six percent more HBM bandwidth and 50% more scale-out bandwidth.
The vendor further claims up to 30% more tokens per dollar. These are AMD comparisons rather than customer operating results from the OpenAI deployment. The commercial test is whether TensorWave can install and run the contracted capacity on schedule.
Piotr Tomasik, TensorWave’s president and co-founder, said: “AI infrastructure is entering a new phase, where the ability to deliver hundreds of megawatts of advanced compute reliably and at speed will separate the companies that can support frontier AI from those that cannot.”
He continued: “TensorWave was built to prove that AMD infrastructure could operate at the scale and reliability that the world’s most demanding AI companies require. This deployment takes that work to an entirely new level, with TensorWave helping deliver a meaningful portion of the infrastructure behind OpenAI and AMD’s next generation of AI compute.”
Uday Ruddarraju, CTO for compute at OpenAI, said: “As demand for AI continues to grow, expanding access to reliable, large-scale compute is critical to our ability to build and deliver increasingly capable models.”
“We’re excited to advance our previously announced partnership with AMD through this agreement with TensorWave, which will help us continue scaling AI to serve more people and organizations around the world.”
For OpenAI, the agreement advances a partnership it had previously announced with AMD. TensorWave remains responsible for providing and operating the capacity; the 2027 start means the announcement sets a future delivery test rather than reporting present output from the installation.
What is behind TensorWave’s expansion?
TensorWave secured $350 million in a Series B funding round in June 2026, which the company said would enable it to expand its AI infrastructure footprint. It also claims to have secured 2GW of long-term data-center capacity across the United States.
One part of that expansion involves TecFusions, a known customer of TensorWave. In January 2026, the two companies signed an additional 20MW capacity agreement for data centers in Pennsylvania and Arizona.
That customer relationship now sits alongside the end of a separate arrangement under which TecFusions would have leased capacity at Fermi’s Project Matador. The two companies officially went their separate ways earlier this month.
The OpenAI award gives TensorWave a named customer for its largest deployment while the company builds out a claimed 2GW US footprint. It also concentrates commercial execution on AMD’s newest rack platform and a 2027 start date that has yet to become an operating result.
The next concrete checkpoint comes in the first half of 2027, when operations are due to start. Full delivery by early 2028 will show whether AMD’s rack specifications and TensorWave’s expansion plan can support the contracted 457MW workload.
Also reported
Source: datacenterdynamics.com
More from Elena Vasquez
Show full bio
Market editor covering companies, deals and strategy in Carrier business, 5G and 6G, fiber and broadband at Telecom Gazette.
523 articles