Today’s Deep Dive covers TensorWave.

We were lucky to have Jeff Tatarchuk, Co-Founder and Chief Growth Officer of Tensorwave, join us on The Compute 100. Check it out before diving in to the story!

TensorWave, the Las Vegas cloud that runs only AMD GPUs, signed a six-year deal on Nov. 16 to supply OpenAI with 457MW of dedicated capacity on AMD Instinct MI455X GPUs in AMD's Helios racks. Capacity starts coming online in the first half of 2027 (TensorWave). TensorWave has gone all in on the weakening of the CUDA moat being an opportunity for developers to scale with AMD. TensorWave's tight partnership with AMD makes them the first neocloud deploying their hardware at scale.
Why it matters: a frontier lab is now committing hundreds of megawatts to a non-NVIDIA hardware, through a startup rather than a hyperscaler.
What you'll learn: TensorWave's core bets, the Helios offering and more.
Three founders and a monopoly
TensorWave was founded in Las Vegas in December 2023 by Darrick Horton, Jeff Tatarchuk and Piotr Tomasik (TechCrunch).
Horton, the CEO, studied mechanical engineering and physics and worked on fusion plasma physics at Lockheed Martin's Skunk Works. He and Tatarchuk had already built a cloud company together, VMAccel, which grew out of crypto mining.
The starting thesis was simple. "We recognized an unhealthy monopoly at work," Horton said in 2024. TensorWave racked its first MI300X systems in April 2024 and set a goal of 20,000 MI300Xs (TechCrunch).
The funding journey
  • Oct. 2024: a $43M SAFE led by Nexus Venture Partners, with AMD Ventures, Maverick Capital and Translink Capital. It was the largest startup round in Nevada's history. TensorWave was then at about $3M of annual recurring revenue (TechCrunch).
  • May 2025: a $100M Series A co-led by Magnetar and AMD Ventures, to deploy 8,192 MI325X GPUs. The company reported a $5M revenue run rate for 2024 and a $100M target for 2025 (Next Platform).
  • Aug. 2025: an $800M tranche of a delayed-draw term loan arranged with Magnetar, backed by GPU inventory and long-term contracts (Futuriom).
  • June 2026: a $350M Series B at a $1.55B valuation, again co-led by Magnetar and AMD Ventures, with more than 2GW of long-term data center capacity secured, company-reported (Magnetar).
The pattern repeats each round: equity from the chip vendor and a credit-focused fund, then debt secured against the GPUs it buys.
The bet: CUDA is no longer the only on-ramp
For a decade, CUDA was the reason AI ran on Nvidia. Models were written against Nvidia's libraries and kernels, so changing hardware meant rewriting software.
That lock-in has loosened. Most new AI code is written in PyTorch, which sits above the chip. "The lanes are opening up on the highway," Tomasik said in 2025, noting that PyTorch-based code "can run on AMD's chips just as easily" (Futuriom).
AMD's ROCm 7 release in September 2025 shipped with Triton, the open kernel language that compiles one piece of code for either vendor's GPUs. It targets AMD hardware automatically (AMD). Open serving engines such as vLLM and SGLang now run on Instinct GPUs out of the box.
The gap hasn't closed everywhere. Hand-tuned CUDA kernels and Nvidia-only stacks such as TensorRT-LLM still don't port, and CUDA has roughly a decade's head start (The Register). TensorWave's business is to absorb that porting work for customers, which is why it hosted "Beyond CUDA," a developer event co-sponsored by AMD, outside Nvidia's GTC in March 2025.
Helios in brief
Helios puts 72 MI455X GPUs, each with 432GB of HBM4, into one liquid-cooled rack that behaves like a single machine (AMD). A 2T-parameter model at 4 bits per weight is 1TB, so it spans at least three GPUs, and those GPUs exchange data constantly (our arithmetic).
Inside the rack, GPUs read each other's memory over 260TB/s of scale-up bandwidth, about 3.6TB/s per GPU. That is roughly 6x the 0.6TB/s each GPU gets to the rest of the data center (our arithmetic). Nvidia's Vera Rubin NVL72 offers 216TB/s (Nvidia).
The novelty is the wiring. Nvidia builds that fabric with its own NVLink switch chips. Helios runs UALink, an open memory-sharing standard, over Ethernet switch silicon that already ships in volume (UALink Consortium). AMD's previous platform, MI355X, scaled up across only 8 GPUs.
The neocloud path for AMD
AMD's demand is now large enough to need a neocloud partner. OpenAI committed to 6GW of AMD GPUs, starting with 1GW of MI450-series in the second half of 2026 (AMD). Oracle plans a 50,000-GPU MI450 supercluster on Helios racks from the third quarter of 2026 (Oracle).
The hard part is outside the chip: powered sites, liquid cooling and operators who can stand up rack-scale systems on a deadline. That is where AMD has been thin. Most independent capacity was built for Nvidia first.
TensorWave fills that gap. It is the first neocloud named as a provider under the OpenAI–AMD deal, its OpenAI sites are designed specifically for Helios (TensorWave), and AMD Ventures has invested in every round. At 457MW, its contract alone covers about 7.6% of OpenAI's 6GW commitment (our arithmetic).
What to watch
First OpenAI capacity is due in the first half of 2027, with full deployment by early 2028. We're watching what customers sign next, how TensorWave ramps, and the benefits that emerge from having more AMD hardware deployed at volume.