Industries / Technology / Software and SaaS / AI and machine learning
Supply of computing power
This is how AI firms get the chips and machines they train and run models on: the chips sold for AI, the software that ties firms to them, renting by the hour, large purchase deals and the rules on who may buy.
7
Milestones achieved, from 2007 to 2024.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2007
Graphics chips for any program
On 26 February 2007 NVIDIA released a public beta of CUDA, a software kit and C compiler for running general calculations on its graphics chips (GPUs). NVIDIA's Andy Keane said it made the GPU's parallel computing power "accessible to virtually any developer".
Credit
NVIDIA
2010
GPUs rented by the hour
On 15 November 2010 Amazon Web Services began renting Cluster GPU Instances, each with two NVIDIA Tesla M2050 GPUs, at $2.10 an hour. AWS wrote that "any AWS user" could now run GPU computing pay-as-you-go, without buying the hardware.
Credit
Amazon Web Services
2012
Two gaming cards win
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton of the University of Toronto trained an image-recognition network for five to six days on two NVIDIA GTX 580 graphics cards with 3 GB of memory each. It won the ImageNet 2012 contest with a top-5 error of 15.3% (the share of images whose right label was not among its five best guesses), against 26.2% for the next entry. The authors wrote that their results could be improved "simply by waiting for faster GPUs and bigger datasets".
Credit
Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton
2016
A buyer's own chip
In May 2016 Google said it had run its own AI chip, the Tensor Processing Unit (TPU), in its data centers for more than a year, for search results, Street View and the AlphaGo matches. Its 2017 paper, accepted at the ISCA computer architecture conference, measured that first TPU, built for inference (running trained models), as 15 to 30 times faster than the CPUs and GPUs of its day on Google's own workloads. Google later rented TPUs to other firms through its cloud (see the Note "Double the compute").
Credit
Norman Jouppi and colleagues, Google
2018
A shared speed test
MLPerf, started in 2018 by a group that included Google, Baidu, Intel, AMD, Harvard and Stanford, published its first training results on 12 December 2018. Each of its seven tests times how long a system takes to train a set model to a set quality. NVIDIA, Google and Intel submitted results.
LessonA shared test covers only what is entered; NVIDIA skipped the reinforcement-learning test, saying it ran mostly on ordinary processors.
Credit
the MLPerf group, led in its creation by Google
2022
Chips need an export license
On 7 October 2022 the US Commerce Department's Bureau of Industry and Security (BIS) put license requirements on exports of advanced AI chips and the computers that contain them to China. A wider "AI Diffusion" framework of January 2025 was rescinded on 13 May 2025, two days before it was to take effect. In January 2026 BIS moved the Nvidia H200 and lesser chips for China from a presumption of denial to case-by-case review, if the exporter certifies US supply is sufficient and the chips pass independent testing in the United States.
LessonWho may buy a chip can change within months, so a buyer's plans rest on the rules as much as on supply.
Credit
US Department of Commerce, Bureau of Industry and Security
2024
Paper speed versus real speed
In December 2024 SemiAnalysis, a chip research firm, tested AI training on AMD's MI300X against NVIDIA's H100 and H200. On AMD's public software releases, training was "still broken" on most of its tests and needed workarounds, and AMD came within 75% of NVIDIA's performance only with help from several AMD engineering teams. The testers met no NVIDIA-specific bugs. On a basic matrix-multiplication test the MI300X reached about 620 trillion operations a second, against the 1,307 trillion AMD advertised.
LessonA chip's speed on paper means little until its software works; SemiAnalysis titled the report "CUDA Moat Still Alive".
Credit
Dylan Patel, Daniel Nishball and Reyk Knuhtsen, SemiAnalysis
Watch for
Gigawatt-scale chip deals
Deals for AI chips are now sized by the electric power the chips will draw. AMD and OpenAI agreed in October 2025 that OpenAI would deploy up to 6 gigawatts of AMD GPUs, starting with 1 gigawatt of MI450 chips in the second half of 2026; AMD gave OpenAI a warrant for up to 160 million AMD shares that vests as deployment milestones are met. The same month Anthropic said it would use up to one million Google TPUs, with "well over a gigawatt" online in 2026, alongside Amazon's Trainium and NVIDIA's GPUs. In November 2026 AMD said OpenAI, Anthropic, Meta, Microsoft and Oracle now run work on its first Helios racks of MI450 chips.
Achieved whensomeone other than the buyer and the seller confirms the first full gigawatt of either deal running.
Claimed by
AMD and OpenAI, October 2025; Anthropic, October 2025; AMD, November 2026
GPU hours as futures
A futures contract fixes today the price of something bought later, so a buyer can lock in what compute will cost. CME Group and Silicon Data said in August 2026 that they would list futures on Silicon Data's indexes of hourly H100 and B200 rental prices from 5 October 2026, subject to regulatory review. The US Commodity Futures Trading Commission lists the H100 contract as "Approval Pending", with its review extended to 8 January 2027.
Achieved whenthe contracts trade on CME's NYMEX exchange and settle against the index.
Claimed by
CME Group and Silicon Data, August 2026
Open
01
Listed versus rentable GPUs
A public, repeated record of whether GPUs that rental sites list can actually be started, at the listed price, when someone asks for one.
The briefCredit · unclaimed02
Same recipe, three chip families
A public test that runs the same open training recipes, unchanged, on NVIDIA, AMD and Google TPU machines rented by the hour, and reports which ones run on the first try and how fast.
The briefCredit · unclaimed
Open 01 · the brief
Listed versus rentable GPUs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public, repeated record of whether GPUs that rental sites list can actually be started, at the listed price, when someone asks for one.
Why it is still open
Price trackers disagree about eightfold on the hourly price of an H100, because they count different things: one-day spot listings, big-cloud on-demand benchmarks and adjusted index values, Implicator.ai found in July 2026. They record offers, not rentals that went through. SkyPilot's open catalog refreshes listed prices every 7 hours but records no attempts to rent.
Built with what exists
SkyPilot, open software that starts machines on more than 20 clouds from one interface; each provider's own API; a scheduled script (for example GitHub Actions); a public CSV file.
First step
Pick 20 providers from SkyPilot's catalog and record which of them let a new account start an H100 machine without talking to sales.
Done looks like
12 weeks of weekly attempts at 20 providers to start one H100 machine each, recording for every attempt the listed price, whether the machine started within 10 minutes, the price billed and the stated reason when it did not; published with the method.
Open 02 · the brief
Same recipe, three chip families
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test that runs the same open training recipes, unchanged, on NVIDIA, AMD and Google TPU machines rented by the hour, and reports which ones run on the first try and how fast.
Why it is still open
SemiAnalysis's 2024 test needed AMD engineers' help to get AMD within 75% of NVIDIA. Its open InferenceX project now benchmarks NVIDIA, AMD and TPU chips continuously, but for serving finished models, not for training. MLPerf training results are submitted by the vendors themselves.
Built with what exists
training recipes from nanoGPT, TorchTitan and Hugging Face TRL; PyTorch, with PyTorch/XLA for TPUs; each vendor's public container images; rented NVIDIA H100 or B200, AMD MI300X or MI355X, and Google Cloud TPU machines; a script that logs errors and tokens per second.
First step
Run nanoGPT's Shakespeare training example unchanged on one rented NVIDIA machine and one AMD machine, and log what breaks.
Done looks like
20 recipes, each run on all three chip families from the vendors' public images with no code changes, with a published table of first-try success, the fixes needed and tokens per second, repeated after each major software release.