Partnerships were the densest signal category in AI during 2026 — denser than product launches or funding. The deals draw a dependency graph worth reading carefully, because every partnership is also an exposure. Here's the map in four tiers.
Tier 1: Labs ↔ Clouds (The Capacity Deals)
- OpenAI ↔ Microsoft, amended: Microsoft stays the primary cloud partner, but its IP license became non-exclusive and OpenAI may now serve products across any cloud — the year's most consequential contract change.
- OpenAI ↔ AWS: a $100 billion, 8-year expansion; ~2 GW of Trainium capacity. OpenAI buying its former rival-cloud's custom silicon marks the end of the exclusive-cloud era.
- OpenAI ↔ Oracle: AI campuses underway at two Texas sites plus New Mexico, Wisconsin, and Michigan.
- Anthropic ↔ Amazon: up to 5 GW secured for Claude; nearly 1 GW of Trainium2/3 online by end-2026.
- Anthropic ↔ Google + Broadcom: multiple gigawatts of next-generation TPU capacity starting 2027 — a lab dual-sourcing custom silicon from two hyperscaler ecosystems simultaneously.
- NVIDIA ↔ SSI: long-term partnership plus investment, with Vera Rubin access expected to raise SSI's compute an order of magnitude — the vendor-investor circularity in its purest form.
The read: exclusivity is dead; every lab now runs a multi-cloud, multi-silicon portfolio, and every cloud runs a multi-lab one.
Tier 2: Clouds/Platforms ↔ Chipmakers
- Meta ↔ AMD: up to 6 GW of Instinct GPUs, explicitly framed as diversification.
- Meta ↔ Broadcom: extended through 2029; co-developing the industry's first 2nm AI accelerator.
- Microsoft ↔ AMD: Helios racks deployed on Azure for frontier inference.
- Google ↔ Broadcom: long-term TPU development and supply, with component supply assurance through 2031.
- NVIDIA ↔ SK Group: a $500B+ initiative including a 2 GW Vera Rubin "AI factory."
- NVIDIA ↔ Lumentum: $2 billion investment plus multibillion purchase commitments — locking optical supply the way TSMC capacity was locked a cycle earlier.
The read: the second-source era arrived. Every major buyer now holds at least one non-NVIDIA silicon lane, and suppliers are being tied down two and three levels deep in the bill of materials.
Tier 3: Model Distribution Alliances
- Microsoft ↔ Hugging Face: open-weight models on Foundry Managed Compute, weights pre-staged in Azure with Microsoft running enterprise serving — open models with hyperscaler operations.
- xAI ↔ Amazon Bedrock: Grok 4.3 generally available with a 1M context — a lab whose owner runs a rival social platform distributing through AWS anyway.
- Perplexity ↔ Samsung: API deployment across hundreds of millions of devices; Perplexity also reports 6 of the "MAG7" as platform users, while running inference on CoreWeave.
The read: distribution now flows through everyone else's channels; the marketplace layer (Bedrock, Foundry, Vertex) is where enterprise model selection actually happens.
Tier 4: Enterprise and Industrial Alliances
- Mistral ↔ Airbus, BMW, ASML: an AI stack for industrial engineering — the European sovereignty play made concrete.
- Accenture ↔ Microsoft: Copilot rolling out to ~743,000 people, the largest enterprise deployment to date.
- FOX ↔ AWS: preferred AI cloud for media workflows.
- Sony ↔ TSMC: MOU toward a Japan JV for next-gen image sensors — AI demand reshaping adjacent supply chains.
Reading the Graph: Three Risks
- Concentration disclosed: Microsoft's ~45% of commercial RPO from OpenAI; CoreWeave's anchor-heavy backlog (Meta up to ~$21B). The graph has single points of failure and they're documented.
- Circular financing: vendors investing in customers who buy their products (NVIDIA↔SSI and adjacent structures) flatters demand signals.
- Geopolitical couplings: TSMC Arizona's $20B capital injection, AMD's export-license constraints on MI308, sovereign builds from Japan to France — the map is being redrawn along jurisdictional lines as much as commercial ones.
