Microsoft bets on Maia chips: why did major companies’ payments to NVIDIA grow faster than their spending?

Executive summary

On 12 October, Microsoft announced the third generation of its AI chip, Maia 300, and said it would expand the Maia family from 2027 to facilities providing several gigawatts, making it the main system it relies on for inference: running models after training to answer user requests. Headlines presented the announcement as a direct threat to NVIDIA, the largest supplier of these chips. This research assesses the announcement against a common benchmark, then asks a question whose answer can be tested with numbers: have companies that make their own chips paid NVIDIA less so far?

The answer in NVIDIA’s own disclosures is no. In the quarter ended 26 July, revenue from its hyperscale category, comprising public cloud providers and the largest consumer internet companies, reached 48.7 billion dollars, up 102% from a year earlier. Over roughly the same period, combined capital expenditure (Capital Expenditure, Capex), meaning spending on assets such as buildings and equipment, at Microsoft, Alphabet, Amazon and Meta grew 87%, from 88.25 to 165.05 billion dollars in the quarter, even though all four companies had operated their own chip programmes throughout the year. NVIDIA’s revenue from this category therefore grew faster than the four companies’ spending, with no measurable effect yet from substituting their own chips for NVIDIA’s.

The real threat concerns the period after 2027 and is confined to inference: Maia is not used to train models, as Microsoft relies on NVIDIA and AMD for training, does not rent it to external customers, and no one outside the company has tested it yet. Analysts at Evercore ISI estimate that Maia will account for around 10% of Microsoft’s AI chip spending by Q1 2027. Even so, the scale is substantial: using the ratio NVIDIA itself cited for one of its sites, each gigawatt corresponds to around 35 to 47 billion dollars in its revenue per chip generation.