TensorWave has booked 457 megawatts of dedicated OpenAI capacity for six years, making it the first specialist AI cloud publicly named inside the 6-gigawatt AMD buildout OpenAI signed in October 2025. By TensorWave's own framing, the deal is roughly 8% of that program.
Neither company disclosed the contract value, the data-center sites or the accelerator count. The announcement landed on 16 November, a day before Nvidia's quarterly results, which is where a deal like this gets read as a positioning statement as much as a purchase order.
The 457 is the easy number. The number that decides whether this is a milestone or a press release is 35: TensorWave's installed capacity today, against the more than 450 megawatts it says it will operate by early 2028.
Pole one: the neocloud becomes the unit of frontier capacity
The OpenAI-AMD agreement is unusually specific for an AI infrastructure deal. AMD's October 2025 filing with the SEC describes a warrant for up to 160 million AMD shares at one cent each, vesting in tranches: the first after delivery of the initial gigawatt of MI450-series GPUs, full vesting at 6 gigawatts, with escalating stock-price targets that reach $600 for the final tranche. OpenAI also made a binding commitment to buy that first gigawatt.
That structure turns capacity procurement into an equity story. Every megawatt an operator brings online moves OpenAI closer to the next vesting step and gives AMD a customer it has partly financed. What the arrangement lacked until this week was an operator with a name attached.
TensorWave fits the slot. It was the first public cloud to run AMD's MI300X, added the MI325X and MI355X, and now operates one of the larger all-AMD training fleets in North America. It raised a $350 million Series B at a $1.55 billion valuation in June, co-led by Magnetar and AMD Ventures, and says it holds more than 2 gigawatts of long-term data-center capacity under contract.
457 MW — dedicated, single-tenant IT capacity for OpenAI
6 years — contract term
~8% — TensorWave's stated share of the 6 GW OpenAI-AMD program (457 of 6,000 = 7.6%)
45.7% — share of the opening 1 GW tranche
H1 2027 — first capacity online
Early 2028 — full deployment
Piotr Tomasik, the president and co-founder, framed the agreement as proof that AMD infrastructure can meet frontier-scale reliability. That is the supplier's claim, and it is the one the next fourteen months have to settle.
The arithmetic is worth stating plainly, because it cuts both ways. At 457 of 6,000 megawatts, TensorWave accounts for 7.6% of the program, the "about 8%" its president cites. Against the opening 1-gigawatt tranche, the block is 45.7%. A single supplier carrying that much of the first step is a strong endorsement of the neocloud model, and a strong concentration of delivery risk in one young company.
Dedicated capacity has also become the format frontier labs prefer. Single-tenant sites remove the noisy-neighbour problem of shared cloud, and let the customer specify racks, interconnect, cooling and operations instead of renting a menu. The trade is that the customer inherits a counterparty and a construction schedule, not just a bill.
Pole two: 457 megawatts announced is not 457 megawatts energized
Converge Digest drew the distinction the release did not. The 457 MW is planned IT load. It excludes cooling and electrical losses, and it says nothing about how much utility power the sites will draw or how much capacity arrives in the first phase. A 457 MW IT figure can imply a substantially larger grid interconnection, and interconnections are the part of the build that has been slipping across the industry.
The schedule is the second gap. AMD and OpenAI said the first gigawatt of MI450-series hardware would begin deploying in the second half of 2026. TensorWave's first capacity lands in the first half of 2027, with the full 457 MW only by early 2028. The company's block is nearly half of that opening gigawatt, but it starts later than the opening gigawatt was supposed to.
The supply chain has already moved once. TensorWave had been lined up to lease 222 megawatts at Fermi's Project Matador; the two sides separated earlier in November. Its known third-party footprint includes a 20 MW agreement with TecFusions, signed in January, for sites in Pennsylvania and Arizona. A 2-gigawatt pipeline is not the same asset as 2 gigawatts of signed, powered, permitted space.
As we wrote in November, Torus Foundry is already selling twelve-week inference blocks against grid capacity that exists. That is the cheaper version of this business. TensorWave is taking the harder one: it has to build the megawatts before it can rent them.
Who carries the balance sheet
The agreement quietly shifts the capital burden. OpenAI commits to buy capacity; TensorWave buys the GPUs, leases or builds the sites, pays for the power contracts and carries the utilization risk if OpenAI's demand curve bends. That is why the $350 million Series B and the 2-gigawatt land bank matter more than the headline megawatts. They are the buffer that lets a private company promise hundreds of megawatts to a single customer.
The headcount plan says the same thing in a different unit. TensorWave runs 225 people today and expects more than 1,000 by the end of 2027, against a capacity jump from 35 MW to more than 450 MW. Thirteen times the footprint and roughly four and a half times the staff, on a fourteen-month clock.
The hardware is concrete; the price is not public
On paper, Helios is a credible platform. Each MI455X carries 432 gigabytes of HBM4 running at 19.6 terabytes per second; a Helios rack packs 72 of the chips for 31 terabytes of aggregate memory and up to 2.9 exaflops of FP4 throughput. AMD claims 15% more peak FP4, 50% more HBM capacity and up to 30% more tokens per dollar than the published specifications for Nvidia's Vera Rubin NVL72. Those are vendor comparisons made against a competitor's data sheet, not independent benchmarks.
The commercial case is opaque by design. With no contract value, no site list and no accelerator count, nobody outside the two companies can test the unit economics of 457 MW, or confirm whether the capacity counts toward the 6-gigawatt program at all. TensorWave says the deal builds on the partnership. It does not say whether these megawatts are incremental to the 6 GW or a slice of it.
| Dimension | Reading A: the anchor | Reading B: the promise |
|---|---|---|
| What is signed | Six-year, single-tenant capacity contract | Value, sites and chip count undisclosed |
| What is proven | First named neocloud in the OpenAI-AMD buildout | No capacity online until H1 2027 |
| What is measured | 457 MW of IT load | Utility power and phase-one size unknown |
| What it means for AMD | A named anchor customer beyond the lab | One operator does not make a market |
Terms as disclosed by TensorWave, OpenAI and AMD; caveats from Converge Digest
The most useful external comparison is scale. The Las Vegas Review-Journal put 457 MW at roughly the power draw of 375,000 homes. Airtel Africa's flagship East African data center under construction is planned at 44 MW, about a tenth of this single contract. That is the size of the bet OpenAI is placing on one operator's ability to energize sites on time.
The two questions that decide the argument
A yes on both makes the 8% framing a milestone. A no on either makes it a placeholder.
What falsifies the thesis
Three things will settle this without another press release. Capacity online in the first half of 2027 is the first: slip it, and the 8% framing loses its anchor. The utility interconnection behind the 457 MW of IT load is the second. The third is whether OpenAI's purchases through TensorWave vest the next warrant tranche, a detail that would show the megawatts sit inside the 6 GW count rather than next to it.
Until then, the fair reading is narrow. OpenAI has added a named operator to its AMD plan, and TensorWave has found a customer that can absorb a thirteen-fold capacity build. Both facts are real. Neither is the same as 457 megawatts running.