The stack eats the model

Vertical integration is becoming the new AI moat.
The chart says the quiet part out loud: nobody wants to be “just” an application, “just” a model, or “just” a cloud provider anymore. Google is closest to full-stack control (with app, model, cloud, and silicon) and that’s why I said many times in this newsletter that is probably going to be the most vauaed company by the end of the decade…or much sooner.

Microsoft has distribution and Azure, but weaker silicon. AWS has infrastructure and chips, but weaker model/app momentum. OpenAI and Anthropic own the model layer, but depend heavily on other people’s infrastructure and chips. That dependency is now strategic vulnerability. In AI, margin will migrate to whoever controls the bottleneck and the bottleneck is not always the model. Sometimes it is compute, sometimes it is distribution, sometimes it is proprietary workflow data.

The SpaceX–Cursor logic is strange, but strategically coherent.
A $60bn stock deal for Cursor sounds insane in normal corporate finance language. But in AI-stack language, it makes sense. SpaceX has compute ambition through Colossus, distribution through the Musk ecosystem, and now, potentially, a high-value software workflow… coding. Cursor has product usage, developer behavior, interaction data, and one of the most important agentic AI use cases. SpaceX is not buying “a coding app”. The WSJ piece says SpaceX agreed to buy Cursor parent Anysphere for $60bn in stock, days after its historic IPO, specifically to strengthen its AI coding power. The IPO gave SpaceX an expensive currency, and now it is using that currency to buy missing layers of the stack.

The future will look less like competition between companies and more like competition between empires.
OpenAI vs Anthropic is too narrow. The real race is Google-stack vs Microsoft/OpenAI-stack vs Amazon-stack vs SpaceXAI-stack vs Meta-stack. Each empire wants applications, models, infrastructure, chips, data, and distribution. The interesting prediction is that “best model” will matter less over time than “best system”. Google may win in consumer AI because it controls search, Android, YouTube, Gmail, Maps, TPU, and Gemini. Microsoft may win enterprise because it controls Office, Windows, GitHub, Azure, and corporate identity. SpaceXAI may become the weirdest one, but with satellites, data centers, energy, compute, Grok, X distribution, and now Cursor-like developer workflow data. It may look absurd from the outside, but so did Amazon owning cloud infrastructure before cloud became the profit pool.

OpenAI will be forced to integrate downward. It cannot remain only a model company renting the future from Microsoft, Nvidia, and energy suppliers. Expect more infrastructure deals, chip partnerships, data-center announcements, maybe even acquisitions in enterprise workflow software.

Anthropic will stay more focused, but that focus becomes a trade-off. It may win regulated enterprise trust, but it risks becoming the “responsible model provider” inside someone else’s stack. That is a good business. It may not be the best public-market story.

Cursor-like products will become the most valuable AI applications because they produce proprietary behavioral data. The next frontier is observing how humans solve problems, where they hesitate, what they accept, what they reject, and how agents can take over multi-step work.

SpaceXAI will use its inflated stock as acquisition currency. That is the underrated point. A meme-like valuation is also a weapon. If the market gives you expensive paper, you can buy real assets with it.

The AI IPOs will be priced less like software companies and more like geopolitical infrastructure plays. Investors will not ask only “what is revenue growth?”, but they will ask “who owns the stack, who controls the bottleneck, and who can finance the next $100bn of compute without blinking?”