Microsoft Just Started Swapping Anthropic Out of Excel and Outlook. Suleyman Just Set the Ceiling on the Anthropic S-1.

Bloomberg broke it Tuesday: Microsoft is now routing tens of thousands of AI prompts every week inside Excel and Outlook to its own MAI models instead of to OpenAI and Anthropic. The migration is targeted at routine, high volume productivity tasks (draft this email, summarize this thread, write this formula, clean this row), the part of the Copilot surface where token cost multiplied by seat count actually shows up on Microsoft's income statement. MSFT closed up about 2 percent on the news. Anthropic and OpenAI did not comment.

The reason the print matters is not that Microsoft is finally shipping the models it announced 35 days ago at Build. It is what Microsoft's AI CEO said out loud while the S-1 clock is running on the largest paying customer of the closed frontier.

The Quote

Mustafa Suleyman, on the record, to Bloomberg: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost." And separately: "Anthropic is extremely expensive and I think many people are urgently looking for alternatives."

Executives do not say "eliminate" on the record about a supplier they intend to keep. That is a public forward guidance number dressed up as an interview quote. What Microsoft told the market on Tuesday is that the Anthropic revenue line inside Copilot is a decaying asset by design, and the design is running in production right now.

The Math

Two rows in that table do most of the work. Microsoft claims MAI-Thinking-1 (35B active parameters) matches Opus 4.6 on SWE-Bench Pro. Even if you assume Microsoft picked the most flattering benchmark it could find, the reasonable inference is that a model an order of magnitude smaller than the frontier tier is now inside a percentage point of it on agentic coding, and Microsoft is running that model on Microsoft compute at Microsoft margin instead of paying Anthropic's per token rate on Anthropic's. Every 100 prompts moved is a lift on gross margin that shows up next quarter, not next year.

Why This Lands Now

Two windows opened at the same time. The Anthropic confidential S-1 lodged on June 1 at a $965B post-money and a $47B ARR run rate. The Microsoft-OpenAI partnership reset in April moved the revenue share to 20 percent through 2030 and kept the IP license intact through 2032, but it also took away exclusivity and gave both sides a runway to reprice each other. Microsoft used the last month of that runway to build a hedge stack and light it up in production.

The Anthropic pitch to a public market has always been a version of "we are the buyable half of a two lab frontier and the largest hyperscaler ships us through Copilot." Suleyman just said the second half of that sentence is now an internal reduce-to-zero target. The S-1 does not have to disclose that verbatim, but the risk factor writer has to account for it. A CEO of a top-three named customer publicly stating that his own goal is to eliminate the vendor is a fact pattern the SEC will expect to see cited.

The Same Move at Two Layers

Ten days ago the hyperscalers made a $3.5 billion consulting turn, AWS with a Forward Deployed Engineering unit and Microsoft with Microsoft Frontier Co. At the time the read was: hyperscalers want to own the workflow layer around the model, because the model itself is compressing. The MAI swap is the same move one layer down. The workflow layer got a Microsoft badge, and the model layer just got one too.

The consequence for a frontier lab that sells to Microsoft is that both the buyer and the deployment engineer inside its enterprise accounts now report up through Satya Nadella. The lab keeps the API surface. It loses the customer relationship in slow motion.

The Pricing Floor Reads This Way

The inference price floor thesis was that per token cost inside the closed frontier would keep falling as compute contracts unlocked capacity in 2027, and that open weight models would set the floor under it. What the Suleyman quote does is add a third floor: buyer captive silicon and buyer captive models. Microsoft is not competing with Anthropic on price at the API. It is walking away from the API for the workloads it can serve itself, and only paying the Anthropic premium for the workloads it cannot. That is a max price ceiling shaped like a demand collapse, not a competitive discount.

The Copilot billing cycle we wrote up on June 30 was the same story from the seat side: Microsoft stopped absorbing the inference subsidy and let end users feel the per token cost. This week we get the platform side. Microsoft is also refusing to absorb the inference subsidy on its own income statement, and rerouting the traffic to a cheaper silicon path is how the number gets down. Both moves are one policy at two layers.

What Anthropic Can Say

Three defenses the underwriter can offer inside the roadshow. One, that a 35B active model matching a frontier one on a benchmark suite is a self serving Microsoft claim, and independent evaluators will land the delta closer to the Anthropic side. That is fair, and probably partly true, but the market has been discounting frontier tier benchmarks for six months anyway, so the argument does not carry much weight against the Suleyman quote.

Two, that Anthropic's enterprise book is not primarily routed through Copilot. Anthropic has its own direct enterprise API, an AWS Bedrock channel, and the Claude Code surface. Copilot revenue is real but it is not the whole book. That is also true. The problem is that Microsoft is the biggest single named customer inside the closed frontier, and the S-1 has to specifically disclose concentration risk. Losing 20 percent of the Copilot slice is not the same as losing 20 percent of the total, but it is a fact the risk factor page has to name.

Three, that Anthropic will retain the top of the workload mix (agentic coding, long horizon research, high stakes reasoning) even as MAI takes the routine end. That is the strongest defense and it is probably right for at least the next two quarters. It is also the same defense OpenAI is running against the same rebalance and the same argument Anthropic just used against the LongCat and GLM open weight incursion at the bottom of the market. Squeezed on both ends by the same argument, the middle is where the frontier premium has to hold, and it is a smaller middle than it was three months ago.

Our Take

The Suleyman quote is worth more than the migration numbers. Migration is a fact. The quote is a policy. Once the largest hyperscaler CEO has told the market that his goal is to eliminate the Anthropic bill, every enterprise procurement officer on the planet gets to open Q3 negotiations with that quote as a cover sheet, and every frontier lab selling into enterprise has to reprice against a public zero. Anthropic and OpenAI both know how to defend against a competitor. They have not had to defend against a customer who says on record that their pricing power is a temporary condition.

For builders and agents, the practical implication is close to what we have been publishing since April. Route your inference through an abstraction that lets you swap models per call, keep an eye on the open weights floor at LongCat-2.0 and GLM 5.2, and expect the closed frontier premium to keep compressing on routine workloads faster than the market currently prices. Reserve the frontier tier for the workloads that actually need it. Everyone else in the stack is now doing the same math Microsoft just published.

Three signposts in the next 90 days. One, whether the Anthropic S-1 amendment names Microsoft as a customer whose spend is expected to decline. If yes, the roadshow is honest and the mark shaves 5 to 10 percent. If no, the underwriters are betting the quote gets walked back. Two, whether Microsoft publishes any Copilot revenue split by underlying model on the July earnings call. Microsoft has never done this before. A first disclosure would be a direct signal to the market. Three, whether OpenAI answers with its own hyperscaler diversification move (a bigger Oracle deal, an AWS expansion, deeper Bedrock integration) that reduces its own Microsoft channel exposure before OpenAI's own IPO window in September.

Anthropic filed the S-1 on the assumption that revenue keeps doubling. Microsoft just put a public ceiling under the biggest slice of that revenue. The doubling curve survives the ceiling only if the workloads Microsoft cannot serve itself keep growing faster than the workloads it can. That is a different pitch than the one the bankers started with 36 days ago, and every S-1 amendment from here through the roadshow is going to have to answer it.