OpenAI Signs Six-Year Deal for 457 MW in AMD GPUs
Contract announced on November 16 secures dedicated capacity using MI455X GPUs in US data centers starting in 2027. Choosing a neocloud rather than a hyperscaler shifts the risk landscape for AI purchasers.

OpenAI has signed a six-year agreement with TensorWave to secure 457 megawatts of IT capacity across multiple data centers in the United States, all utilizing AMD Instinct MI455X GPUs. The announcement was made on Monday, November 16, according to a release distributed by TensorWave and reported by DCD, Unite.ai, and other specialized outlets. The contract value has not been disclosed, but the Las Vegas-based company described this deal as the largest in its history.
The capacity will be set up on Helios racks, AMD's rack-scale architecture that integrates MI455X GPUs, sixth-generation Epyc processors (codename Venice), ROCm software, and Pensando networking. According to the statement, the infrastructure will be dedicated and single-tenant, operated solely for OpenAI. The first phase is expected to become operational in the first half of 2027, with full deployment projected for early 2028.
A Neocloud Instead of a Hyperscaler
The most relevant factor for those procuring infrastructure isn't the size of the deal but the chosen supplier. TensorWave recently raised $350 million in its Series B round, led by Magnetar and AMD Ventures, according to DCD. It is a company built entirely on AMD hardware, having previously been one of the first to operate the MI300X, MI325X, and MI355X accelerators. Committing 457 MW to a company of this scale, instead of consolidating with Microsoft, Oracle, or Google, indicates that OpenAI is diversifying its computing supply chain among smaller, more specialized vendors.
The contract aligns with the partnership between AMD and OpenAI announced in October 2025, which included allocations of up to 6 gigawatts of AMD GPUs. TensorWave did not clarify whether the 457 MW will count against the 6 GW or if it represents additional capacity. The distinction is significant: in the first scenario, the agreement merely details where part of a previously known order will be installed; in the second, it expands OpenAI's exposure to AMD.
The Implications for AMD and Nvidia
For AMD, each contract of this nature serves as proof that ROCm can handle production loads at hundreds of megawatts, a point the company must make against Nvidia’s CUDA platform. If TensorWave replicates such contracts with other leading labs, AMD gains a distribution channel that does not rely on hyperscalers.
The arrangement's weak point is counterpart risk. A neocloud with a short track record takes on the obligation to build and operate nearly half a gigawatt by early 2028, in a market where transformers, turbines, and electrical labor are in high demand. Six-year contracts protect TensorWave's revenue but do not eliminate the risk of delivery delays, which fall on OpenAI.
How This Affects Other Markets
In the United States, the immediate effect relates to the price and availability of power in the states where the sites will be installed, which have yet to be disclosed. The concentration of dedicated capacity for a single client reduces the volume of AMD GPUs available in the spot market for other American companies until 2028.
In the European Union, the contrast is in the model. While OpenAI secures private capacity through a single bilateral agreement, the bloc is still receiving proposals for AI Gigafactories via EuroHPC, with deadlines of December 3, 2026, for the large batch and February 4, 2027, for the medium batch, according to the Commission's bidding portal. A Nordic and Baltic consortium led by Nokia, Nscale, and Telia announced its bid the same day as the TensorWave agreement. The gap between these two timelines outlines where frontier models will be trained over the next three years.
In Brazil, the channel is contractual. Companies consuming OpenAI's API primarily run inference on American infrastructure, which subjects each contract to the international data transfer rules of the LGPD. Additional capacity in the U.S. does not alter this equation but increases dependency on it. For CIOs negotiating multi-year AI contracts, the pertinent question shifts to which physical infrastructure the supplier operates on and whether there is a data residency clause, rather than just which model is in the catalog.
For consultancy boards that resell or integrate OpenAI models, this agreement sends a less-discussed signal: the company's computing layer is becoming more diversified and therefore less predictable in latency and cost by region. Those establishing SLAs with end clients over this layer must read the fine print on availability more carefully than they did in 2025.
Sources
- datacenterdynamics.comhttps://www.datacenterdynamics.com/en/news/openai-signs-on-for-457mw-ai-capacity-with-tensorwave/
- hpcwire.comhttps://www.hpcwire.com/aiwire/2026/11/16/tensorwave-announces-agreement-with-openai-to-deliver-more-than-450mw-of-amd-powered-ai-infrastructure/
- unite.aihttps://www.unite.ai/tensorwave-signs-six-year-openai-deal-for-amd-powered-ai-capacity/
- reviewjournal.comhttps://www.reviewjournal.com/business/vegas-business/innovation/las-vegas-tech-company-strikes-major-deal-with-openai-3915172/
- datacenterdynamics.comhttps://www.datacenterdynamics.com/en/news/amd-based-ai-cloud-tensorwave-secures-350m-series-b-funding/