How AI Compounds

Progress looked sudden because the tributaries were already moving.

Modern AI is not one invention. It is the convergence of mathematical ideas, software practice, silicon, capital, data center scale, and distribution. This page is the primer: how the slope steepened, why the flywheel matters, and why investors should care.

2006AWS launches

Anyone can rent a server by the hour. Compute becomes a utility.

2007CUDA

Nvidia makes the GPU programmable. The bedrock under every model that comes next.

2015Google TPU v1 in production

Google quietly puts the first AI-specific chip into its datacenters. Hyperscalers will all build their own silicon.

2016AWS launches GPU cloud

Anyone can rent an NVIDIA GPU by the hour. AI startups stop needing their own datacenters.

2016First DGX-1 to OpenAI

Jensen Huang hand-delivers Nvidia's first AI supercomputer to a small lab in San Francisco.

2019TSMC begins EUV production

ASML's $200M lithography machines start printing the chips that train modern models.

Sep 2022Hopper / H100

Nvidia's transformer engine enters production. Training scale starts to look like an industrial product cycle.

May 2023DGX GH200

Nvidia links 256 Grace Hopper superchips into one giant memory system. Frontier training becomes a cluster architecture problem.

Sep 2023Amazon backs Anthropic

AWS becomes Anthropic's primary cloud and points Claude training toward Trainium.

Dec 2023TPU v5p

Google exposes the in-house accelerator stack behind Gemini. The hyperscalers are no longer just renting Nvidia.

Mar 2024Blackwell / NVL72

Nvidia stops selling chips as parts and starts selling rack-scale AI factories.

Jul 2024AI capex arms race

Hyperscalers and neoclouds race to turn power, land, memory, networking, and GPUs into model capacity.

Oct 2024EPYC 9005

AMD launches Turin. AI clusters still need host CPUs for orchestration, data movement, and inference plumbing.

Dec 2024Broadcom crosses $1T

Custom AI accelerators and Ethernet make the ASIC supplier a trillion-dollar company.

Dec 2024Trillium TPU GA

Google's sixth-generation TPU becomes a cloud product, turning internal model infrastructure into rented capacity.

Feb 2025Blackwell ramp

Nvidia calls Blackwell the fastest product ramp in company history as data center revenue keeps compounding.

Apr 2025Ironwood TPU

Google unveils its first TPU designed specifically for the age of inference.

Sep 2025OpenAI-CRWV tranche

OpenAI adds a third CoreWeave tranche, taking cumulative commitments to roughly $22B. The neocloud lane becomes contracted capacity.

Oct 2025Gigawatts become the unit

OpenAI's Nvidia, AMD, and Broadcom commitments and AWS Project Rainier move frontier capacity into gigawatts and Trainium chips.

Nov 2025Ironwood goes commercial

Google's TPUv7 Ironwood moves from announcement to commercial pull-through, with Anthropic as the demand signal.

Dec 2025Broadcom AI revenue doubles

Custom accelerators and Ethernet AI switches become a second shovel-maker revenue curve.

Jan 2026Nvidia backs CoreWeave

Nvidia takes equity and adds a compute backstop. Neocloud capital structure becomes part of the accelerator supply chain.

Feb 2026$194B data center year

Nvidia reports fiscal 2026 data center revenue of $193.7B. AI demand is no longer a forecast; it is booked revenue.

Mar 2026Vera Rubin

The hardware roadmap turns into rack-scale systems for pretraining, post-training, test-time compute, and agentic inference.

Apr 2026Google Cloud +63%

Google Cloud Q1 revenue jumps 63% YoY to $20B and backlog nearly doubles to $462B. TPU demand is becoming booked cloud revenue.

May 2026CoreWeave backlog hits $99B

CoreWeave Q1 backlog rises to $99.4B, capex guide moves to $31B-$35B, and 2027 exit ARR is mostly contracted.

2012AlexNet wins ImageNet

Hinton's team crushes every image-recognition method by 10 points. Trained on two gaming GPUs.

2014Google acquires DeepMind

Google buys a London AI lab for ~$500M. The team that will build AlphaGo two years later.

2015OpenAI founded

Altman, Sutskever, Brockman, Musk launch a non-profit aimed at general AI.

2016AlphaGo beats Lee Sedol

DeepMind's AI wins Go 4–1 against the world champion. Watched live by 200 million.

2017Attention Is All You Need

Eight Google researchers publish the architecture under every modern LLM.

2020GPT-3

175-billion parameters. Writes code from English. The first commercially useful model.

Nov 2022ChatGPT

A free chat box. 100 million users in two months — the fastest-adopted consumer product ever.

Mar 2023GPT-4

The chat toy becomes a benchmark machine. Reasoning, coding, and multimodal ambition move into the product lane.

Jul 2023Llama 2

Open-weight models become commercially usable. The frontier starts leaking into every developer's laptop.

Dec 2023Gemini and Mixtral

Google enters with Gemini 1.0; Mistral proves open MoE models can matter. The field stops being one-company theater.

Mar 2024Claude 3

Anthropic pushes the benchmark race into a real multi-frontier market.

May 2024GPT-4o

Multimodal interaction becomes the consumer default, not a research demo.

Jun 2024Claude 3.5 Sonnet

Coding and artifacts turn the model from answer box into work surface.

Sep 2024o1 reasoning

Test-time compute becomes a product primitive. The model can spend more compute to think harder.

Oct 2024Computer use

Claude learns to look at screens, move a cursor, click, and type. Agents start needing full-stack inference.

Dec 2024Gemini 2.0

Google frames the next phase around agents, tool use, and AI that can take actions.

Jan 2025DeepSeek R1

An open reasoning model shocks the market. Efficiency improvements widen demand instead of ending the hardware cycle.

Feb 2025Claude Code

The terminal becomes an agent surface. Models start editing files, running tests, and doing multi-step software work.

Mar 2025Gemini 2.5 Pro

Google's thinking model resets the benchmark board and pushes long-context reasoning into the mainstream race.

Apr 2025o3 and o4-mini

Reasoning models get tool access and move closer to agentic workflows inside ChatGPT.

May 2025Claude 4

Frontier models are now sold on coding, agents, and sustained multi-hour work.

Aug 2025GPT-5

OpenAI unifies fast answers, reasoning, coding, and agentic API work under one flagship model.

Sep 2025Anthropic ARR hits $7B

SemiAnalysis estimates Anthropic external ARR crosses roughly $7B as Claude Code's viral growth starts showing up in dollars.

Oct 2025Anthropic books 1M TPUs

Claude's next phase pulls roughly one million Google TPUs into view. Software demand starts reserving silicon like a strategic asset.

Nov 2025Claude Opus 4.5

Anthropic raises the agentic-coding bar and holds the daily-driver position for serious software work.

Dec 2025Cursor crosses $100M ARR

Coding-agent monetization steepens. The software layer starts proving it can turn tokens into enterprise revenue quickly.

Jan 2026OpenAI-Cerebras latency tier

OpenAI signs a 750 MW Cerebras deal; GPT-5.3-Codex-Spark makes low-latency inference its own product lane.

Feb 2026Claude Code goes industrial

Claude Code authors roughly 4% of public GitHub commits while Anthropic hits a $14B run-rate and raises at a $380B valuation.

Mar 2026Open weights compress the cycle

GLM, Qwen, Kimi, and Composer push agentic-coding capability forward in roughly 13-week cycles at lower inference cost.

Apr 2026Opus 4.7 / GPT-5.5

Claude Opus 4.7 and GPT-5.5 reset the coding-agent bar as model-lab ARR moves from roughly $9B at year-end toward $30B-$44B+ in 2026.

May 2026Codex doubles in a week

After GPT-5.5, Codex revenue doubles in under seven days. The model-to-agent-to-inference feedback loop is now measured in days.

2022 onward

After ChatGPT, the stream stops looking narrow.

The public saw a chatbot. Underneath it, every layer of the stack started accelerating at once: frontier models, open models, reasoning models, coding agents, GPU systems, custom silicon, data centers, power, memory, and networking. The river widens because each improvement creates more reasons to use AI, and each new use case asks the physical world for more capacity.

AI factories

The model race becomes an infrastructure buildout.

By the Blackwell and Vera Rubin cycle, this is no longer just better chips. It is rack-scale architecture, networking, storage, memory, power, cooling, and capital formation arranged around one question: how much intelligence can be produced, served, and improved per watt, per dollar, and per unit of scarce capacity?

Hardware and software are not two stories. They are the same story, told from opposite banks.
ChatGPT was not the beginning of AI. It was the moment the river became visible from the road.
$194B

Nvidia fiscal 2026 data center revenue, up 68% year over year.

04 layers

Infrastructure, models, applications, and the capital recycling loop between them.

~25 yrs

Length of the compounded research that produces the current era.

The floodgates

After 2022, the river becomes a torrent.

The important part is not that ChatGPT appeared suddenly. It is that the layers beneath it were ready to amplify each other. The left bank is the physical substrate; the right bank is model capability and distribution. From 2022 onward, both banks start throwing off milestones every few months.

Hardware, capital, capacity

Left bank

The physical stack starts moving like a release cadence.

H100, Trainium, TPU, Blackwell, Broadcom ASICs, neocloud finance, and cloud backlog turn AI from a software launch into an industrial buildout.

  1. Sep-Nov 2022: H100 enters production just before ChatGPT turns latent AI demand into a public product shock.
  2. May-Dec 2023: DGX GH200, Amazon's Anthropic/Trainium commitment, and Google's TPU v5p show hyperscale AI is becoming a cluster business.
  3. Mar-Dec 2024: Blackwell/NVL72, capex escalation, EPYC host CPUs, Trillium TPU, and Broadcom's $1T moment widen the hardware story beyond GPUs.
  4. Sep-Dec 2025: CoreWeave commitments, gigawatt deals, Project Rainier, commercial Ironwood, and Broadcom AI revenue make capacity the strategic asset.
  5. Jan-May 2026: Nvidia/CoreWeave finance, Nvidia's $194B data center year, Vera Rubin, Google Cloud backlog, and CoreWeave's $99B backlog show demand converting into contracted infrastructure.

Models, agents, distribution

Right bank

The software current stops being a single channel.

GPT-4, Llama, Gemini, Claude, DeepSeek, coding agents, and model-lab ARR create a many-frontier market.

  1. Nov 2022: ChatGPT makes the capability legible to the public at 100 million users in two months.
  2. Mar-Dec 2023: GPT-4, Llama 2, Gemini 1.0, and Mixtral widen the field beyond one lab.
  3. Mar-Dec 2024: Claude 3, Llama 3, GPT-4o, Claude 3.5 Sonnet, o1, computer use, and Gemini 2.0 compress the release cycle.
  4. Sep-Dec 2025: Anthropic ARR, TPU reservations, Claude Opus 4.5, and Cursor ARR show agents becoming revenue, not just demos.
  5. Jan-May 2026: Cerebras latency, Claude Code commits, Anthropic's run-rate, open-weight catchup, Opus 4.7, GPT-5.5, and Codex doubling in a week turn the flywheel into a same-week feedback loop.

Why it matters

The compounding is the point.

When model quality improves, new uses appear. When new uses appear, usage rises. When usage rises, the world needs more and better infrastructure. AI progress is therefore not just a software story. It is a demand shock into the physical stack.

Investor lens

History becomes a map of pressure.

The important question is where the next turn of the flywheel creates strain: capability strain, capacity strain, or a narrow gate that the market has not yet recognized.