Chronologie

A logician and a neurophysiologist show networks of simplified neurons can compute logic — the founding abstraction.

'Computing Machinery and Intelligence' reframes 'can machines think?' as a testable question.

McCarthy, Minsky, Shannon and colleagues coin 'artificial intelligence' and set its agenda.

The first trainable neural network hardware; the Navy predicts thinking machines. The hype-and-winter cycle begins.

A proof of single-layer limits helps freeze neural-network funding for a decade — the first AI winter's seed.

Rumelhart, Hinton and Williams show multi-layer networks can learn internal representations.

Convolutional networks deployed on US mail — deep learning's first industrial success.

Brute-force search takes chess; the win is narrow but the symbolism is not.

NVIDIA opens its GPUs to general computation — the platform deep learning will land on.

Fei-Fei Li's 14-million-image dataset bets that data is the bottleneck.

Two consumer GPUs halve the error rate. Deep learning's era — and NVIDIA's second act — begins.

Goodfellow's adversarial networks and Google's sequence-to-sequence models open generation and translation.

Altman, Musk, Sutskever, Brockman and others launch a nonprofit lab with $1B pledged.

Deep RL takes Go a decade early; move 37 enters folklore. 280M people watch.

NVIDIA builds matrix math into silicon — the first AI-specific GPU feature.

Eight Googlers publish the transformer. Everything on this site descends from it.

OpenAI and Google show pre-training + fine-tuning beats task-specific models.

OpenAI stages the release citing misuse — the first mainstream fight about publishing model weights.

175B parameters; in-context learning; Kaplan's curves say more compute keeps paying.

Ampere becomes the workhorse of the first scaling wave.

The first mass-market LLM product — code completion from Codex.

Protein structure solved to experimental accuracy; 200M structures follow.

Hopper's Transformer Engine and FP8 are built for the models to come.

DeepMind: most models are undertrained on data. Every training plan changes.

Image generation on consumer GPUs; the fine-tune ecosystem and the copyright wars begin.

October 7 rules cut China off from A100/H100-class chips and advanced fab tools — compute becomes statecraft.

A 'research preview' reaches 100M users in two months — the fastest product adoption ever recorded, and the starting gun for everything in this database's Trends section.

Capability jumps again; technical details go dark; Microsoft commits ~$10B more.

The first chipmaker to a $1T market value, on data-center guidance that stuns Wall Street.

Meta makes open weights a strategy, not a leak.

A800/H800 workarounds banned; performance-density rules close the loopholes.

OpenAI's board crisis ends with Altman back and the board remade — governance meets capability in public.

Two-die B200 and the GB200 NVL72 make the rack the unit of computing — and liquid cooling mandatory.

Test-time compute becomes the new scaling axis.

Memphis build resets industry assumptions about how fast a frontier cluster can rise.

Hinton (Physics) and Hassabis/Jumper (Chemistry) — the academy ratifies the era.

1.742 exaFLOPS on AMD MI300A — the public sector's biggest disclosed machine.

AWS commits its own silicon to Anthropic at hundred-thousand-chip scale.

OpenAI, SoftBank, Oracle and MGX unveil an up-to-$500B US buildout; Abilene is site one.

An open reasoning model at disclosed bargain cost; NVIDIA falls ~17% in a day (~$589B) — then the Jevons argument wins the year.

The GPU cloud lists at ~$23B — the market's first pure-play AI infrastructure stock.

Google's inference flagship: 9,216-chip pods for the 'age of inference'.

AMD's CDNA4 reaches Blackwell class; OpenAI appears on stage.

First company to the mark — ahead of every national stock market but a handful.

Frontier reasoning from Colossus; a trillion-parameter open MoE from Beijing.

A router unifies fast and thinking modes for ChatGPT's ~700M weekly users.

First open OpenAI language models since GPT-2 — the open-weights race is now three-way.

Circular or catalytic? The vendor-financing debate defines the autumn.

First shut US reactor ever returned to service — the nuclear revival's proof point.

Google Cloud deal worth tens of billions; a gigawatt of TPU capacity slated for 2026.

GTC DC: 'half a trillion in bookings' visibility through 2026.

Trained on TPUs end to end — the strongest non-NVIDIA statement yet.

Three frontier releases in three weeks (Gemini 3, GPT-5.1, Opus 4.5) — cadence itself becomes the story.

Structured records on this site reflect public information as of mid-January 2026; the News section stays live. See Methodology.