Back to storiesAIGlobal Highlights

OpenAI taps TensorWave for 457MW of AMD capacity in Nvidia hedge

By Editorial Desk2 min read
OpenAI taps TensorWave for 457MW of AMD capacity in Nvidia hedge
TensorWave will supply 457 megawatts across US sites, just 7.6% of OpenAI's 6GW AMD deal, as OpenAI moves to loosen its near-total reliance on Nvidia chips.

OpenAI has named TensorWave as the first neocloud provider under its 6-gigawatt AMD compute deal, handing the company 457 megawatts across multiple US data centre sites. That is only 7.6% of the total capacity in the October agreement, but it is TensorWave's largest deployment ever, and a company spokesperson said it will need to roughly triple its staff to support it.

The stakes are about dependence. OpenAI's models run almost entirely on Nvidia GPUs, a source familiar with its compute infrastructure said, leaving it exposed to any disruption at the chipmaker. Rival Anthropic spreads workloads across Google TPUs, AMD GPUs, Amazon Trainium and Nvidia GPUs, its S-1 filing showed, and has built a "proprietary orchestration layer" to route work to the cheapest chip.

OpenAI is trying to replicate that. It is designing its own chip architecture, Jalapeño, with Broadcom, due to scale up production in 2027, and is reportedly planning to buy Google TPU chips. The first gigawatt from the AMD deal is set to come online in the second half of this year; TensorWave's capacity is expected in the first half of 2027, with full deployment by early 2028.

A TensorWave spokesperson said the timing was set by this week's Supercomputing 2026 conference, described as "one of the last major compute conferences of the year", rather than by turmoil at Nvidia. The announcement lands as Nvidia holds its first earnings call without Jensen Huang, who posted on X that he would be "honored" to return as CEO. The board told CNBC it is not considering bringing him back and is working with search firms on a replacement, while an oversight committee monitors compliance with US export controls ahead of Huang's closed-door House hearing on Dec. 1.

What it means for Pakistan: Pakistan does not sit inside this chip supply chain, but the direction of travel matters for its plans to host data centres and court AI investment: the compute that would serve Pakistani cloud and AI users is being locked up years ahead through deals measured in gigawatts. That leaves Pakistani buyers and any state-backed AI or cloud projects relying on whatever capacity global providers allocate them, at prices set by the same handful of chipmakers and neoclouds. No Pakistani figure, official or project has been linked to these agreements.

Related topics & keywords

  • #ai
  • #amd
  • #nvidia
  • #data centres
  • #compute
  • #openai

Related Stories

View all news →