Training AI models takes a lot of compute, and compute is in scarce supply.

The claim has two parts: training AI models takes a lot of compute, and compute is scarce. True if evidence from around mid-2026 shows that training frontier AI models needs very large amounts of compute, and that AI compute (GPUs, accelerators, data-centre capacity) was supply-constrained, for example through shortages, allocation limits, long lead times or capacity limits that companies have reported. Mostly true if both parts hold but the scarcity was easing or limited to particular segments. Mixed if only one of the two parts holds. Mostly false if the compute requirement is true but supply was broadly plentiful at the time. False if both parts are contradicted.

Set before the research began.

1. E1 American AI Companies Can’t Get Enough Chips Center for a New American Security (CNAS) · 7 May 2026

   “In 2026, artificial intelligence (AI) chip production has become a binding constraint on the pace of the AI compute buildout. Demand for computing power to train and deploy advanced AI models continues to grow exponentially, outpacing many chip manufacturers’ forecasts.”

2. E2 American AI Companies Can’t Get Enough Chips Center for a New American Security (CNAS) · 7 May 2026

   “The compute used to build and run AI systems has been growing exponentially, driven by continued growth in compute used for training and enhancing new models”

3. E3 The Growing Compute Shortage Apollo · 15 June 2026

   “On-demand GPU capacity is effectively sold out, and even older-generation chips are seeing rental rates increase to levels not seen since 2024.”

4. E4 AWS cloud growth accelerates as AI demand strains capacity Cloud Computing News · 31 July 2026

   “Even at that amount, we will still not have enough capacity to meet all of the demand we have in 2026”

Two CNAS passages from May 2026 (E1, E2) say that the computing power needed to train and deploy advanced AI models was growing exponentially, driven partly by compute used for training new models. That supports the claim that training takes a lot of compute, though neither passage gives a figure. For the scarcity part, E1 says AI chip production had become a binding constraint in 2026. Apollo (E3, June 2026) reports that on-demand GPU capacity was effectively sold out and that rental rates were rising even for older chips. E4, an article on AWS capacity, quotes a statement that 'even at that amount' there would still not be enough capacity to meet all 2026 demand. Together these independent sources support both parts of the claim around mid-2026.