I asked Excel to clean a messy revenue table this morning. It worked, the way it has a hundred times before. What I did not know until I read the reporting over coffee is that the model answering me was probably not the one I assumed. Microsoft has started quietly routing tens of thousands of prompts a week inside Excel and Outlook to its own in-house models, branded MAI, in place of the OpenAI and Anthropic engines that used to sit behind those features.

The company is not coy about why. In June its AI chief, Mustafa Suleyman, said the plain part out loud: “We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost.” Microsoft had been on track to spend around 500 million US dollars a year renting Anthropic’s models. The target it has set for that bill is zero.

When the company that owns the distribution starts making the model, the model stops being the scarce, priced thing.

Here is the part that matters for anyone building on top of AI. The single most valuable customer the frontier labs have is now making a good-enough version of their product and dropping it into software that hundreds of millions of people already open every morning. When the company that owns the distribution starts making the model, the model stops being the scarce, priced thing. It becomes a component.

Watch what that does to price. Microsoft’s MAI-Thinking-1, shown at its Build conference in June, is pitched as matching Anthropic’s Opus 4.6 on a common coding benchmark at a lower cost. You do not have to trust the benchmark to read the move. Once a competent model ships inside the tool at the platform’s own marginal cost, the price a lab can charge for that same capability, inside that tool, trends toward zero. The labs keep the frontier. They lose the pricing power on everything the frontier has already made ordinary.

Southeast Asia has been living this ahead of the memo. OCBC in Singapore now runs more than thirty internal tools on open-weight models: Google’s Gemma for summaries, Alibaba’s Qwen for code, DeepSeek for reading the market, across six markets from Hong Kong to Vietnam. The Chinese open families price access somewhere between a fifth and a thirtieth of the American frontier. A regional bank worked out, quietly and early, that most of what a business does with a language model does not need the smartest one, and that the base model is a commodity you rent cheap or host yourself.

So the real question for a founder or an operator concerns what you own that a platform cannot bundle. A business that is a thin layer over a model capability, a summariser, a formatter, a chatbot with a nicer login, can be absorbed the week that capability becomes a checkbox in the suite or a line in the bank’s own IT roadmap. What survives is the part no platform can package: the workflow your customers have wired into their day, the proprietary data only you hold, the relationship where they call you, by name, when it breaks.1

A capability you cannot tell apart is a capability nobody can keep charging a premium for.

Back in the spreadsheet, the cleaned table looked exactly as it always had. I could not tell which model did the work, and that was the whole point. A capability you cannot tell apart is a capability nobody can keep charging a premium for.

Footnotes

  1. The uncomfortable version for the labs: their biggest customers are also their most capable competitors, and every invoice teaches the customer precisely what the thing would be worth to build in-house. ↩