The Two Economies of AI

AMD reported $11.5 billion in quarterly revenue last week — a record, up 50% from a year ago, driven by data-center AI demand. Its shares fell 7% the same day. ^[1] That is not a glitch. It is the model layer’s condition compressed into a single earnings report, and it keeps repeating. Alphabet posted its first-ever negative free cash flow — minus $5.9 billion — with 2026 capital-expenditure guidance reaching $205 billion. ^[2] Meta’s free cash flow collapsed 91%. AMD’s own free cash flow dropped 20% even as it tripled capital spending to $808 million. ^[1] The companies building and running the most advanced AI models are generating more revenue than ever and watching the cash drain out the bottom. Meanwhile, the models themselves are getting cheaper for everyone else. Chinese open-weight models, led by DeepSeek, have surpassed American models in downloads on Hugging Face. ^[3] SAP’s CFO, Dominik Asam, put the enterprise view plainly.

It requires much more excruciating assurance levels. — Dominik Asam

Uber kept its AI spending broadly flat this year not by using less AI but by switching to cheaper models and optimized defaults — its cost per token fell even as adoption rose. ^[4] The model layer is caught in a double squeeze: the infrastructure costs are staggering, and the product is becoming a commodity. The deployment layer tells the other half of the story. SAP’s cloud backlog grew 26% to €22.9 billion. ^[5] Palantir’s US commercial revenue rose 121%, driven by its AI platform that connects models to real-world operational data. ^[6] ServiceNow projects subscription revenue doubling from $15.7 billion to $30 billion by 2030 by embedding AI agents into the workflow systems enterprises already use. ^[7] Cloudflare launched an open-source AI workspace this week that integrates models directly with a company’s internal systems. CEO Matthew Prince made the case for integration over siloed AI.

For AI to truly transform an enterprise, it can't live in a silo or behind a developer bottleneck. — Matthew Prince

The competitive bottleneck has moved. It is no longer whether the model is smart enough. It is whether you can plug it into a company’s actual data, its governed workflows, its compliance requirements. SAP CTO Philipp Herzig named the shift in a single sentence this spring.

Enterprise AI doesn't stall because the models aren't good enough; it stalls because the data isn't ready for AI agents. — Philipp Herzig

Snowflake’s SVP Christian Kleinerman made the same point from the data-platform side.

Customers want AI that works directly on their governed data, not in isolated systems. — Christian Kleinerman

SAP is freezing hiring and travel to fund a $3 billion-plus investment in data-integration companies — Dremio, Prior Labs, Reltio — because it sees the integration layer as existential, not optional. ^[8][9] European IT services firms are riding the same wave: Capgemini raised its growth target after a 9.2% bookings increase, Sopra Steria upgraded its outlook, and OVHcloud reported 20.2% public-cloud revenue growth — all driven by the technical complexity of wiring AI into fragmented corporate databases and strict governance frameworks. ^[5] Beneath this split sits a financial machinery that makes the model layer’s position more precarious. The capital flowing into AI infrastructure travels in a circle: Nvidia invests $100 billion in OpenAI; OpenAI buys compute from Oracle and CoreWeave, which buy Nvidia hardware; OpenAI partners with AMD for chips in exchange for a 10% stake. ^[10][11] The commitments are enormous — a $300 billion Oracle compute deal, a $250 billion Microsoft Azure contract, a $350-to-$500 billion Broadcom custom-accelerator deal for 10 gigawatts of compute — and they require cash the labs do not generate from operations. ^[12] Sam Altman acknowledged the overexuberance directly.

Is AI the most important thing to happen in a very long time? My opinion is also yes — Sam Altman

He was simultaneously planning $1.4 trillion in data-center spending over eight years. ^[13] The demand is real. Micron’s high-bandwidth memory supply is sold out through 2028, with non-cancellable contracts extending to 2030. ^[14] OpenAI forecasts $100 billion in annual recurring revenue by 2027, running 15% ahead of plan, and projects $200 billion by 2030. ^[15][12] But the revenue is increasingly coming from deployment channels — enterprise subscriptions, a ChatGPT advertising pilot that hit $100 million annualized in six weeks, Agents-as-a-Service — rather than from pricing power on model intelligence itself. ^[16] This is the context in which OpenAI and Anthropic both filed confidential S-1 IPO registrations in June. ^[17] They are racing to public markets to fund compute commitments that outstrip their operating cash flow, and they are doing so at the moment their own models are becoming the commodity input someone else builds the premium product on. Nasdaq and Russell implemented fast-entry index rules in May that shortened IPO-to-index inclusion to as few as five trading sessions, a mechanism that will channel index-fund inflows into these offerings within days of their debut. ^[18] Truist found that high-profile tech IPOs suffer an average 55% first-year drawdown. The two most important AI labs are about to find out whether public-market investors see the same value the private ones did. ^[18]