Inference providers
Companies that run models on GPUs or custom chips and sell the output by the token.
Who is making money
The companies earning or raising real money in this part of AI. Suppliers are marked: they sell to the builders rather than to end users.
Makes the GPUs that most AI inference runs on, and paid about $20B in late 2025 to license Groq's inference chip technology.
$89.0B data centre revenue in the quarter to July 2026 (fiscal Q2 2027), out of $96.2B total
Open profileDesigns its own inference chips and runs a cloud that serves open models quickly.
$3.5B in a 2026 round that followed its Nvidia licensing deal
Open profileMakes very large AI chips and now sells hosted model access as well as the hardware itself.
$510M in 2025, with $87.9M net income
Open profileRents Nvidia GPU cloud capacity to AI labs and companies, and sells GPU servers.
About $760M annualised in 2025 (Sacra estimate)
Open profilefast open-model hosting
Company says more than $1B annual recurring revenue (Jul 2026). Sacra estimates about $800M annualised by May 2026
Open profileopen-model hosting
About $1B annualised (Sacra estimate, Feb 2026)
Open profileHosts the main library of open AI models and datasets, and sells inference, storage and subscriptions to companies.
Annualised revenue above $150M, from one unofficial report. The company does not disclose revenue
Open profilepay-as-you-go GPUs
About $300M annualised (CEO, 2026, reported)
Open profileLets developers run AI models through an API. Now part of Cloudflare, feeding its Workers AI platform.
Bought by Cloudflare for $57.4M cash consideration, completed 1 Dec 2025
Open profileruns production models for companies
About $600M annualised run-rate, reported in mid 2026 (not audited)
Open profileHosts open AI models through an API at low prices and is building its own data centre capacity.
Company says annual recurring revenue above $100M (2026). Not audited
Open profileServes image, video and audio models through an API for app builders.
Reported to have doubled from $200M to $400M annualised in about five months (one outlet, 2026)
Open profileMakes its own AI chips and sells inference cloud access and on-premise systems to large organisations.
About $11B after a round of about $1B (July 2026)
Open profileWhere the money flows
Apps and companies pay per token or per GPU second. The hosts pay chip makers and data centres.
licence for Groq's inference technology, with Groq's founder and president joining Nvidia
w.mediacompute capacity deal covering 750 MW over three years
sec.govacquired Replicate, with some payment held back
sec.govtook part in the Series D round
orrick.com
The openings
Gaps still open in this area, from ones a small business could take to ones only giants can play.
- Small business Manage AI running costs and vendor choice for clients
- Startup money Specialist hosting: in-country, compliant or ultra-fast
- Giants only Inference chips or large GPU clouds
Prices per token keep falling and the big clouds and labs sell direct, so independent hosts risk thin margins.
Sources: noqta.tn, dealroom.co
The map updates three times a week.
241 companies across 25 areas, 62 money flows and 89 openings so far. New research lands every Tuesday, Thursday and Saturday.
Back to the full mapFree daily email
Get the map updates in your inbox
A short read every morning on what is moving in AI, including new areas of the map as they go live. Free.
You are in
Your first AI Pulse lands tomorrow morning. Keep an eye on your inbox.