Green Owl

AI

AI just flipped

AI inference spending passes AI training in 2026, Gartner forecasts. Using AI now costs more than building it.

Numbers as of 1 Oct 2026

The full story

Six glass vessels, one per year from 2024 to 2029, show spending on renting AI chips in the cloud per Gartner: 7, 18, 38, 60, 81 and 109 billion dollars. Blue at the bottom is training, teaching the AI. Green on top is inference, the AI answering you. In 2026, marked The flip, inference is 21 billioAn AI is taught once per model, shown as one blue book. It answers all day, shown as a field of green chat bubbles: ChatGPT alone gets more than 2.5 billion questions a day. Every answer uses a chip again.Teaching, the blue part, is Nvidia's turf. A blue book carries the Nvidia logo. 6 million developers use Nvidia's own software, built up over 20 years. Moving to another chip means rewriting that work. Then how do rivals get in?Answering, the green part: Google built a chip made for answering. A green chat bubble carries the Google logo. Anthropic signed up for up to 1 million Google chips, in a deal worth tens of billions of dollars, October 2025. Google's newest chip, Ironwood, is made for answering. Google is not the onAnswering, the green part: AMD got two giant orders. A green chat bubble carries the AMD logo. 6 gigawatts plus 6 gigawatts of AMD chips were ordered by OpenAI in October 2025 and Meta in February 2026, with first chips in late 2026. A gigawatt is the electricity the chips use. So how big does this From 2025 to 2029, answering grows almost twice as fast. Rented AI chips for teaching: 9 billion dollars to about 38 billion, about 4 times. For answering: 9 billion to about 71 billion, about 8 times. The growing part is the open one. Total: 18 billion to 109 billion, about 6 times. Cloud rental onWho wins the answering part? A row of the Nvidia, Google and AMD logos, plus Other. One winner, or many? Follow Green Owl if you want more.

Which AI do you ask the most?

Gartner tracks what companies spend renting AI chips in the cloud. 2025: $18.3B, split about evenly. 2026: $37.5B. Inference (the AI answering you) $20.6B, training (teaching it) $16.9B. 2029: $108.6B, with inference above 65%. Gartner's August update lifts 2026 to $42B, still 55% inference.

Whose chips? Training still runs mostly on Nvidia. Inference is where rivals get in: Google built its Ironwood TPU for inference, and Meta has served live Llama 3.1 405B traffic on AMD's MI300X.

Caveat: this is cloud rental only, not all AI spending. Splits marked ~ are approximate.

Want to see how this looks now?

When enough people ask, we re-check the numbers and redraw it.

Where the numbers come from

Show all 27 sources

Companies in this picture

AMDGoogleNvidia

More like this

+7 cards+7+8 cards+9+8 cards+4+7 cards