Working Knowledge
Search⌘K

The Board · Archived snapshot

Follow the roles, then the contracts.

Most infrastructure headlines involve three overlapping roles: model developers, compute providers, and chip or fabrication companies. The organizations and contracts change; the role map is the durable part.

Snapshot · July 13, 2026

They train and publish model families

The model developers

They train and publish model families

OpenAI

GPT and Codex; long-running strategic and infrastructure relationship with Microsoft.

[1][2]

Google DeepMind

Gemini model family inside Google.

[3]

Anthropic

Claude; uses capacity from several providers and describes AWS as its primary cloud and training partner.

[4]

Meta

Llama open-weight model family, distributed under Meta's licenses.

[5]

SpaceX / xAI

Grok model family; SpaceX announced that it acquired xAI on 2 February 2026.

[6]

Mistral, Qwen, and DeepSeek

Important European and Chinese model developers with open-weight releases alongside hosted products.

[7][8][9]

They finance, build, and rent infrastructure

The compute providers

They finance, build, and rent infrastructure

Microsoft Azure

Cloud and AI infrastructure provider with a strategic OpenAI relationship.

[10][1]

Amazon Web Services

Cloud capacity plus custom Trainium accelerators; Anthropic describes AWS as its primary cloud and training partner.

[11][4]

Google Cloud

Cloud TPU capacity and Google's vertically integrated model and infrastructure stack.

[12][3]

SpaceX Colossus

SpaceX agreed to provide Anthropic all compute capacity at the Colossus 1 data center in a deal announced 6 May 2026.

[4]

Oracle and CoreWeave

Additional GPU and AI infrastructure capacity outside the three largest cloud platforms.

[13][14]

They design chips, software, and fabrication

The silicon stack

They design chips, software, and fabrication

NVIDIA

GPU accelerators plus the CUDA software platform used by a large share of AI workloads.

[15]

AMD

Accelerators and the ROCm open software stack for GPU computing.

[16]

Google TPU, AWS Trainium, Microsoft Maia

Custom accelerators designed by cloud providers for parts of their own AI workloads.

[12][11][17]

TSMC

A major contract manufacturer for advanced high-performance chips; fabrication is a different role from chip design.

[18]

How value moves

Four recurring flows.

Roles overlap. Customer, supplier, investor, and competitor can describe the same pair of companies at different layers.

01

Chip designers and fabricators supply accelerators and manufacturing capacity to cloud and model companies.

[15][16][18]
02

Cloud and specialist infrastructure providers package that hardware into capacity that model developers can use.

[10][11][12][14]
03

Infrastructure relationships may also include investments, preferred-provider agreements, or custom-chip commitments, so customer and financier can be the same company.

[1][4]
04

The same model developer may spread training and inference across several hardware and cloud platforms rather than rely on one stack.

[4]

Evidence

Sources for the July 13, 2026 snapshot

  1. 1
    Microsoft and OpenAI extend partnership

    Microsoft · Announcement · checked 2026-07-13

  2. 2
    Codex is now generally available

    OpenAI · Announcement · checked 2026-07-13

  3. 3
    Gemini models

    Google DeepMind · Documentation · checked 2026-07-13

  4. 4
    Higher usage limits for Claude and a compute deal with SpaceX

    Anthropic · Announcement · checked 2026-07-13

  5. 5
    Llama

    Meta · Documentation · checked 2026-07-13

  6. 6
    xAI joins SpaceX

    xAI / SpaceX · Announcement · checked 2026-07-13

  7. 7
    Mistral AI models

    Mistral AI · Documentation · checked 2026-07-13

  8. 8
    Qwen model repository

    Alibaba Qwen · Documentation · checked 2026-07-13

  9. 9
    DeepSeek model repository

    DeepSeek · Documentation · checked 2026-07-13

  10. 10
    Azure AI infrastructure

    Microsoft Azure · Documentation · checked 2026-07-13

  11. 11
    AWS Trainium

    Amazon Web Services · Documentation · checked 2026-07-13

  12. 12
    Cloud TPU

    Google Cloud · Documentation · checked 2026-07-13

  13. 13
    OCI AI infrastructure

    Oracle · Documentation · checked 2026-07-13

  14. 14
    CoreWeave cloud platform

    CoreWeave · Documentation · checked 2026-07-13

  15. 15
    CUDA platform

    NVIDIA · Documentation · checked 2026-07-13

  16. 16
    ROCm software

    AMD · Documentation · checked 2026-07-13

  17. 17
    Microsoft Azure Maia AI accelerator

    Microsoft · Announcement · checked 2026-07-13

  18. 18
    High-performance computing platform

    TSMC · Documentation · checked 2026-07-13

Return to the current Board →