On November 16, US AI cloud provider TensorWave announced a six-year agreement with OpenAI. Under the deal, TensorWave will dedicate 457 MW of IT capacity to OpenAI's workloads across several data centers in the US. The news was reported by 36Kr, citing Jiemian.
The hardware is AMD Instinct MI455X, and the sites are being designed around AMD Helios rack-scale systems. The metric that matters here is megawatts, not cards: the contract specifies how much power and cooling the provider commits to deliver for the racks, not the number of accelerators. The timeline is laid out too: first capacity in the first half of 2027, full deployment by early 2028.
The timeline is the most interesting part of this contract. The bet is on hardware that hasn't really been proven in production yet, and on capacity that doesn't physically exist — in 2027–2028. It looks like an attempt to lock in a spot in the AMD queue early, while providers are still learning to assemble Helios racks. Our read: this isn't about today's shortage — it's about the next generation of infrastructure.
The second layer is the deal structure itself. OpenAI isn't buying cards directly; it's leasing dedicated capacity from a third-party cloud company. That makes scaling easier — and makes it easier to walk away from any single vendor. From what we've seen, major AI labs are increasingly spreading themselves across multiple contractors and vendors, because depending on a single supply channel is too expensive — in both money and time.
Who wins: TensorWave gets an anchor customer for six years, and AMD gets a reference deployment next to the biggest player in the game. Who loses: the "NVIDIA-only" position gets weaker as an argument. That said, competition isn't going away, and a second source is always less polished — software, drivers, stack maturity. That's now project risk, not marketing.
The announcement includes no contract value, no GPU count, no power density per rack, and no breakdown by site or state. It's unclear whether 457 MW is the entire contract or a cumulative cap with phases. The source is a recap of the announcement in a Chinese business news feed — no direct quotes from OpenAI, AMD, or TensorWave, and no MI455X specs. We wouldn't take the "2027 → early 2028" timeline at face value: schedules on projects like this tend to slip, and the only real test will be whether the first phase actually ships.
Forecast: by mid-2027, TensorWave or OpenAI will publicly name at least one specific site for this contract — with an address, a state, or a campus name. If that doesn't happen, the first-phase timeline has slipped.
Source: 36Kr
Source: www.36kr.com · post in Telegram