Can one company be another’s shareholder, compute supplier, primary distributor AND competitor? In the AI ecosystem, Microsoft is all four to Anthropic.

Microsoft is a shareholder in Anthropic, having committed up to $5 billion in November 2025. It sells Anthropic cloud capacity that reabsorbs some of that capital, backed by a $30 billion Azure agreement . It distributes Claude to enterprise customers through Azure. And it competes directly through the seven in-house MAI models it launched in June 2026 .

When we map the various commercial agreements across the AI landscape, nearly every large company in this ecosystem holds several positions at once. The industry’s chipmakers, cloud platforms and model providers have become one another’s investors, suppliers, sales channels and competitors, frequently playing multiple roles simultaneously.

Chart 1: Multi-Vector Partnerships

Chipmakers and cloud hyperscalers increasingly occupy multiple overlapping economic positions across the leading AI labs. Amazon, for instance, holds all three available positions across both OpenAI and Anthropic: equity stakes, multi-year compute contracts and distribution partnerships.

For the frontier labs, these relationships deliver scale: growth capital, specialized compute and instant distribution into thousands of enterprise clients the cloud providers spent decades winning. For the hyperscalers, holding all three positions removes the need to predict where the profits land:

  • If frontier models sustain their pricing power and capture the industry’s profit: The hyperscaler wins through its equity stake, sharing directly in upside of the lab’s valuation growth.
  • If models commoditize into low-margin utilities: The hyperscaler still wins through its compute contracts, converting the lab’s ongoing infrastructure spend into guaranteed, long-term cloud revenue.
  • If value concentrates at the application layer: The hyperscaler wins through its distribution channel, owning the enterprise customer relationship and controlling AI-infused solutions delivery.

Chart 2: Interlocking Grid of Mutual Dependency

No single alliance exists in a vacuum. Mapped across the ecosystem, these commercial agreements form an interlocking grid of mutual dependency: competitors cross-supply, cross-distribute and co-fund one another at scale.

The majority of companies in this matrix depend on at least four counterparties.

One clear pattern is frontier lab multi-sourcing. Anthropic is running Claude across Nvidia GPUs, Google TPUs and AWS Trainium . Sourcing across three infrastructure providers prevents hardware lock-in and creates leverage in future capacity negotiations.

Across the grid, the underlying corporate strategy is universal: diversify dependencies upstream to protect supply lines, while aggregating choices downstream to preserve customer access and pricing power.

Chart 3: The AI Infrastructure Value Chain

Traced through the AI infrastructure value chain, equity and project finance dollars convert almost immediately into contracted compute capacity.

For frontier AI labs, compute is by far the largest cost, ahead of talent, training runs and go-to-market expenses.

One clear pattern is the pairing of an equity investment with a purchase obligation that’s typically of far greater size. For example, on April 20 of this year, Amazon announced a further $5 billion investment in Anthropic. At the same time, Anthropic committed more than $100 billion to Amazon Web Services over ten years, and it has signed comparable agreements with Microsoft and Google .

The capital buys contracted capacity, not owned infrastructure. A lab that built data centers would own an asset it could sell or borrow against, which is the approach Meta has taken. A lab that pre-commits to purchasing compute capacity owns nothing and owes for years.

On the supply side, the benefit of the relationship appears twice. Amazon recently has seen more direct impact from Anthropic’s rising valuation than from selling compute: $53.4 billion of Amazon’s $62.6 billion in second-quarter net income was a paper gain “primarily from our investments in Anthropic,” versus a total of $16.6 billion of operating income for AWS, which sells the compute.

The asymmetry is this: the capacity obligations are fixed in size and dated in years. The revenue that covers them, from customer subscriptions to agent fees, is contracted in months.

Chart 4: AI Labs Partnerships Beyond AI Infrastructure

Beyond the AI infrastructure ecosystem, AI labs have built extensive networks of partnerships and investments across the broader economy. This chart shows OpenAI’s and Anthropic’s acquisitions, investments and partnerships by sector.

OpenAI has used acquisitions to build consumer products and applications around its models. It has acquired companies across consumer hardware , developer tools , product analytics , health records and media .

Anthropic’s acquisitions have been concentrated in developer tools and software infrastructure, including Bun and Stainless . Its expansion into other enterprise sectors has relied more heavily on its partner network .

Both companies moved directly into enterprise deployment in 2026. OpenAI launched its Deployment Company in May, with more than $4 billion committed by a TPG-led consortium, and agreed to acquire Tomoro, adding approximately 150 forward deployed engineers . Anthropic anchored a $1.5 billion services venture built on Fractional AI , renamed Ode with Anthropic in July, which then acquired Casper Studios in August.

Partnerships extend distribution; acquisitions determine what the model provider owns directly. Both labs are moving closer to the enterprise customer, where software and implementation services make the model harder to replace.

Chart 5: Who Keeps the Margin?

The question underlying every one of these partnerships is who keeps the margin. It resolves into three contests, each decided by what it costs a customer to replace one supplier with another. Where replacement is straightforward, competition pushes price toward the cost of delivery; where it is not, the supplier keeps the difference.

1. Compute providers versus compute buyers. Compute providers design and sell the accelerators that train and run AI models: NVIDIA and AMD via merchant chips, and Microsoft, Google, Amazon and Meta via custom silicon. The buyers are the cloud platforms that assemble them into data centers, and the AI labs that buy directly or rent capacity through those clouds. Providers hold the pricing power today for two reasons: first, demand has run ahead of supply; and second, the labs typically write their training and inference systems against NVIDIA’s CUDA layer, so moving those systems to AWS Trainium or Google TPUs requires additional (expensive) code conversion, compilation and performance testing. NVIDIA’s gross margin was 75.0% in the quarter ending July 2026, against 72.4% a year earlier.

2. Cloud platforms versus AI model providers. Enterprises can buy AI models directly from cloud platforms: Amazon’s Bedrock , Microsoft’s Foundry and Google’s Gemini Enterprise Agent Platform each sell competing models alongside their own. Most enterprises buy this way today rather than contracting directly with a model provider; their data already sits on the cloud platform, and one vendor means one contract, one bill and one security review. That gives the cloud platform the stronger position: it carries several interchangeable models, so switching between them requires only a setting change rather than a vendor swap, and it already owns the customer billing, security and compliance relationship. AI model providers are responding by building their own direct sales channels . Where models remain interchangeable, the cloud platform keeps the profit; where one becomes uniquely embedded in a company’s operations, the AI model provider does.

3. AI model providers versus enterprise customers. AI model providers sell access to their models, priced by the token; The buyers are companies putting those models inside their own products and operations. They buy AI as an input, so they judge it on price and performance, and treat dependence on one supplier as a risk. The buyers hold the stronger position today. For most tasks the leading AI models are close substitutes, and routing software directs each request to the lowest-cost model capable of completing it. Spending moves quickly as a result: OpenAI’s share of enterprise API spending fell from 50% in 2023 to 27% in 2025, while Anthropic’s reached 40% . AI models are responding by building bespoke agentic systems and embedding their models in company workflows, raising the cost of replacement. Buyers also absorb a second cost: every deployment teaches the AI model provider how that industry works, and both labs are moving into the application layer their customers occupy.

The strongest position belongs to companies with the greatest number of ways to earn and the fewest outcomes on which their returns depend. That is why a single company becomes shareholder, supplier, distributor and competitor to several rival model providers at once. Holding four positions across competing AI labs is a hedge that pays whichever way the market resolves.

The views expressed are the author's own and do not necessarily reflect the official policy or position of any of her affiliations. This article is for informational purposes only and is not investment advice, research, or a recommendation with respect to any security.