Your Landlord Is Also Your Rival

Anthropic, the company that makes Claude, has committed $200 billion to Google Cloud over five years. Google makes Gemini, a direct competitor to Claude. The commitment represents 40% of Google's total revenue backlog. ^[1] That is not ownership. It is tenancy with extra steps. The full ledger is larger. Anthropic has also committed $100 billion to Amazon, which runs its own Bedrock models; $30 billion to Microsoft, which sells Copilot; and is negotiating a $10 billion compute lease with Meta, which builds Llama. ^[2][3][4] That is over $340 billion in compute commitments to four companies that all run rival AI models. OpenAI, the other major lab that owns no data centers, has committed $100 billion to Amazon and joined a $500 billion joint venture with Oracle to build data centers under the Stargate banner. ^[2][5] The AI industry has split into infrastructure landlords and tenants, and the tenants pay rent to their own competitors. The compounding mechanism is straightforward. Every dollar a tenant lab pays for compute becomes revenue for a company that builds a competing model. That revenue can be reinvested — in chips, in research, in the next generation of the rival model. The landlord collects the rent before the tenant's own model earns anything. Combined hyperscaler capital expenditure is projected at roughly $570 billion in 2026, with Amazon at $200 billion, Alphabet at $180 to $190 billion, and Microsoft at $190 billion — Meta's $125 to $145 billion sits on top. ^[6] Amazon's custom Trainium chips alone carry $225 billion in revenue commitments. ^[6] The global data center capacity deficit sits at 12 gigawatts — only 8.9 GW operational against 21.1 GW of demand — and cloud service backlogs have reached roughly $2 trillion. The bottleneck is not money. It is power, labor, cooling, and electrical components. The landlords control the supply. The tenants are trying to escape. The results so far are not encouraging. OpenAI's Stargate initiative has paused its UK project and abandoned expansion in Abilene, Texas, facing regulatory hurdles, energy constraints, and community opposition. ^[7] Community opposition alone blocked or delayed roughly $130 billion in data center projects in the first quarter of 2026. ^[7] OpenAI is now negotiating a $500 billion, 10-gigawatt data center on federal land in Ohio — a 20-year lease with a first phase targeted for 2028 — but the Abilene and UK retreats suggest the path from negotiation to concrete is not straight. ^[7] Anthropic's situation is even more revealing. Its IPO filing discloses over a dozen preliminary agreements to lease and manage its own data centers in the United States, exceeding one gigawatt of capacity. ^[8] But to secure those leases, Anthropic needs Google to provide a financial guarantee — and Google is co-designing the server chips. ^[8] Anthropic is trying to move toward managing its own infrastructure, and it still needs its largest competitor to co-sign the lease. The company was driven to this scramble by what it called an "infrastructure emergency": an 80-fold increase in revenue and usage to $30 billion annualized that forced it to lease SpaceX's Colossus 1 supercomputer — 220,000 GPUs drawing 300 megawatts — just to keep Claude running. ^[9] In March, Anthropic had to reduce Claude session limits during peak hours, with users exhausting five-hour allowances in minutes. ^[10] Then, in July, Meta crossed the line from the other direction. On July 1, the company launched Meta Compute, a service selling raw GPU compute and access to its proprietary models to external customers — directly competing with AWS, Azure, Google Cloud, and the neocloud providers CoreWeave and Nebius. CoreWeave and Nebius shares fell roughly 17% on the announcement; Meta's rose 10%.

It's definitely on the table. — Meta

Meta is spending $125 to $145 billion on infrastructure this year and hired AWS veteran Dave Brown to run the compute business. ^[11] Its $27 billion Hyperion data center in Louisiana is backed by state-fast-tracked approval for seven gas-fired power plants generating 7,400 megawatts, plus 2,500 megawatts of renewables — one of the largest single-site AI infrastructure projects in the world. ^[12] But the line does not run clean. Meta is also a customer of Amazon, which supplies it with tens of millions of Graviton processor cores. ^[2] Meta and Microsoft both rent additional capacity from CoreWeave and Nebius — CoreWeave carries a $100 billion revenue backlog, and Nebius grew 684% year over year. ^[13] Meta is still building models: after Llama 4 underperformed, the company hired Alexandr Wang as Chief AI Officer and invested $14.3 billion for a 49% stake in Scale AI. ^[14] Meta Compute is not a pivot away from model-building. It is a second front. The pattern extends beyond the hyperscalers. Nvidia launched a revenue-sharing financing model in July — DSX AI factories — that lets AI startups access GPUs and token credits without upfront capital in exchange for a share of revenue. ^[15] Nvidia is no longer just a chip supplier; it is becoming a compute landlord too. At the enterprise level, companies facing soaring token costs are shifting to smaller models and building in-house GPU compute. PNC CEO Bill Demchak said his bank would build its own GPU compute rather than remain dependent on external token providers. ^[16] The result is a form of double competition. The companies that own the infrastructure layer — Google, Amazon, Microsoft, and now Meta — compete against their rivals twice. Once in the model market, where Gemini faces Claude, where Bedrock faces GPT, where Llama faces both. And once by collecting rent from the labs trying to beat them, rent that compounds into the next round of infrastructure and the next generation of models. The tenants know the trap they are in. They are spending hundreds of billions to get out of it. So far, the landlords hold the keys.