Models are commodities, platforms are the prize

Everyone is building models. The winners are building platforms.

This week made the pattern hard to ignore. GeekWire reported that Microsoft's 365 Copilot Researcher now runs a draft-then-critique workflow pairing GPT and Claude, scoring 13.8% higher on the DRACO deep-research benchmark than either model alone. Microsoft doesn't care which model is best. It cares that both models run inside Microsoft.

That framing (the model as a replaceable component, the platform as the durable layer) showed up in three more stories the same week. Meta is reportedly considering licensing Google's Gemini after internal tests showed its next-gen "Avocado" model lagging behind GPT, Gemini, and Claude in reasoning and agentic tasks. The company that spent billions positioning itself as the open-weight alternative is now shopping for someone else's proprietary model. Apple, meanwhile, is building an Extensions marketplace that will let users plug Claude, Gemini, and other chatbots into a reimagined Siri alongside the existing ChatGPT integration. Apple doesn't want to win the model race. It wants to be the App Store for the model race.

The platform play

The common move here is abstraction. Microsoft abstracts models into interchangeable agents inside a workflow. Apple abstracts them into extensions inside an operating system. Meta, unable to compete at the model layer, may end up abstracting its own product surface away from any single model entirely. In each case, the company sitting between the user and the model captures the relationship and the recurring revenue.

This is the classic platform dynamic, and it has consequences for the labs. When Microsoft can swap Claude in for GPT (or GPT in for Claude) based on which scores better on a given task, neither lab has pricing power. They become suppliers competing on benchmarks and margins, while Microsoft owns the customer. The same logic applies to Apple's Extensions marketplace: once users access Claude through Siri, Anthropic's direct relationship with those users weakens.

Which makes Anthropic's reported IPO plans more interesting than a straightforward liquidity event. The company's annualised revenue has hit an estimated $19B, up from $1B fourteen months ago. It recently raised a $30B Series G at a $380B valuation and is now in early talks with Goldman Sachs, JPMorgan, and Morgan Stanley about a potential October listing that could raise over $60B. OpenAI is preparing its own IPO on a similar timeline.

I think the IPO race is partly about locking in platform status before the window closes. Going public gives these labs the capital and credibility to build direct user relationships through products like Claude Code, ChatGPT's integrations, and consumer subscriptions, before the Microsofts and Apples of the world finish turning them into interchangeable backend services. The lab that builds a loyal, direct user base is a platform. The lab that doesn't is a supplier.

The practical question for anyone building on these models: which layer do you want to depend on? If you're building on a single model's API, you're betting that lab maintains its edge and its independence. If you're building on a platform layer (Microsoft's, Apple's, or your own abstraction) you're betting the models stay interchangeable. Both bets carry risk. But the second bet is the one the biggest companies in tech are making right now, and history suggests they're usually right about platform dynamics.

The frontier labs have about eighteen months to prove they're more than suppliers. The IPO clock is ticking, and the platform layers are already being poured.

---