Compute:Economics
Curated real data

Timeline

The key developments of the AI compute story — model releases, capability shifts, and the deals that moved the market. Demand shocks on the balance page trace back to entries here; major events (E-numbers) can be overlaid on any chart.

E-numbers mark events shown on charts

2021

  1. 2021-06

    GitHub Copilot technical preview

    First mass-market LLM coding assistant; the start of AI-assisted software work as a demand category.

2022

  1. 2022-10

    NVIDIA H100 general availability

    The Hopper generation ships — the unit this platform normalizes to (H100-equivalent). Defined the 2023–25 buildout.

  2. 2022-11E1

    ChatGPT launches

    Fastest product to ~100M users (~2 months). The inference-demand shock that turned AI compute into a scarce commodity.

2023

  1. 2023-01

    Microsoft–OpenAI ~$10B investment

    Anchored the first hyperscaler–lab compute alliance: capital in exchange for exclusive cloud compute.

  2. 2023-03E2

    GPT-4 released

    Frontier capability jump; triggered the H100 shortage era — rental prices peaked above $8/GPU-hr on spot in 2023.

  3. 2023-07

    Llama 2 open weights

    Open-weight frontier-adjacent models spread demand beyond a handful of labs to every enterprise.

2024

  1. 2024-03

    Claude 3 family released

    Multi-lab frontier competition confirmed — no single-vendor demand story; training demand broadens.

  2. 2024-09E3

    OpenAI o1 — test-time compute scaling

    Reasoning models make inference compute scale with problem difficulty: a structural break in the demand curve (inference no longer O(tokens)).

  3. 2024-09E4

    xAI Colossus: 100k H100s in 122 days

    Proved gigawatt-class clusters can be built in months, not years — repriced buildout-speed assumptions across the market.

  4. 2024-10

    Claude computer use (agentic control)

    First frontier-model computer control API: agents operating software directly — a new, long-running inference workload shape.

  5. 2024-11

    Model Context Protocol open-sourced

    Standardized tool connectivity for agents; the plumbing that made agentic workloads deployable at scale.

2025

  1. 2025-01E5

    DeepSeek R1 efficiency shock

    Frontier-class reasoning at claimed ~$6M training cost; NVIDIA fell ~17% (~$600B market cap) in a day — the market's first repricing of compute-efficiency risk.

  2. 2025-01E6

    Stargate: $500B datacenter JV announced

    OpenAI / SoftBank / Oracle / MGX joint venture — the largest single compute-infrastructure commitment announced to date.

  3. 2025-02E7

    Agentic coding goes mainstream

    Claude Code and agent-mode IDEs ship; tokens per developer-task jump orders of magnitude — inference demand decouples from human typing speed.

  4. 2025-03

    CoreWeave IPO

    First pure-play neocloud lists (~$23B initial valuation): public price discovery for GPU-fleet economics.

  5. 2025-05

    OpenAI Codex cloud coding agent

    Delegated, parallel cloud agents normalize long-running autonomous jobs as a workload class.

  6. 2025-08E8

    GPT-5 released

    Frontier release with routed test-time compute by default — reasoning-tier inference becomes the default consumer workload.

  7. 2025-09E9

    OpenAI–NVIDIA $100B / 10 GW "moonshot" pact

    NVIDIA commits up to $100B tied to 10 GW of OpenAI buildout — capital circularity between chipmaker and lab becomes explicit market structure.

  8. 2025-09

    Oracle–OpenAI ~$300B compute contract

    Five-year compute offtake from 2027 (~4.5 GW) — forward-selling compute at unprecedented scale; OpenAI's 2025 deal chain passes $1T.

  9. 2025-10

    OpenAI–AMD 6 GW agreement

    Multi-generation 6 GW Instinct deployment with equity warrants — first credible large-scale second source to NVIDIA.

  10. 2025-11

    Gemini 3 released

    TPU-trained frontier release at scale — custom silicon proves out as a full alternative supply chain for frontier training.

2026

  1. 2026-01

    Agentic engineering & metaharnesses emerge

    Orchestration layers running fleets of agents (plan/verify/synthesize) become standard practice — compute per engineering task rises another step.

  2. 2026-01

    OpenAI–Cerebras ~$10B / 750 MW deal

    Non-GPU accelerator capacity (wafer-scale, 2026–28 delivery) enters the frontier supply mix.