Update: Microsoft and Meta Slash Anthropic Claude Spending Ahead of IPO

Full analysis

What's new since we last covered this: Major customers cutting spending ahead of IPO filing.

Microsoft and Meta both cut what they spend running Anthropic's Claude, and the timing lands right as Anthropic is reported to be heading for an IPO. Microsoft trimmed a projected $1 billion internal Claude bill by about a third and told engineers to ration it, leaning on in-house models and OpenAI's GPT-4.5 for a lot of queries. Meta cut Claude Code usage roughly in half, from around 60,000 engineers to 30,000, after layoffs and the rollout of its own tool, Metacode. Meta was on pace to spend several billion a year with Anthropic in June. It's now at about $105 million a month, or $1.26 billion annualized.

Here's what sets the stakes. Anthropic's own filings show just two customers once made up a quarter of its 2025 revenue, and only about 100 customers spend more than $10 million a year. When your revenue sits on that few shoulders, a shrug from any one of them moves the whole number. This is a story about whether the business selling the model can hold its biggest buyers, and what that tells every company renting a frontier model today.

The Skeptic. "Slashing" is doing a lot of narrative lifting. Meta is still handing Anthropic $1.26 billion a year after the cut. Microsoft's separate $2 billion Copilot forecast with Anthropic reportedly survives untouched. That is not customer flight, it is a hyped adoption peak settling down to a still-enormous number. Most software companies would trade a limb for one customer at nine-figure ARR. And the timing smells. Negative press about customer concentration, dropped right before an IPO, is the oldest short-side setup there is. The Information got fed this by someone with a position. Read it as a leak, not a trend.

The Compute Pragmatist. The money didn't vanish, it moved. Meta's inference spend didn't drop from "several billion" to $1.26 billion because engineers stopped coding. It migrated onto Meta's own GPU clusters running Llama-class models through Metacode. Microsoft did the same, redirecting queries to in-house models it already pays to run. This is the hyperscaler playbook in the open: rent the frontier API long enough to learn the workload, then pull it back onto metal you own. Anthropic's inference revenue is structurally exposed to every customer big enough to build their own. The ones with GPUs and research teams all eventually cross that line. The moat was never the model. It was owning the hardware the model runs on, and Anthropic owns none of it.

The Builder. Sixty thousand engineers on Claude Code, then thirty thousand in a few months. That is a lock-in failure, plain and simple. Meta shipped a replacement and recaptured half its base almost overnight, which means Claude Code never got sticky enough to survive a budget review. Microsoft telling engineers to ration mid-sprint is the fastest way to kill a dev tool. Nothing poisons adoption like a usage meter ticking while you're trying to ship. If you're building product features on a single model's API, your enterprise buyers are running the same build-versus-buy math on you right now, on every workflow you made convenient. The abstraction layer that lets you swap models without a rewrite isn't a nice-to-have anymore.

The Researcher. The substitution pattern is the finding. Microsoft fell back to GPT-4.5 and internal models. Meta built Metacode. Both decided that for a large share of day-to-day coding queries, cheaper and good-enough beats best-in-class. That says frontier capability is commoditizing inside big companies faster than the leaderboards suggest, because the marginal query that a 60,000-person engineering org runs is mostly routine. You don't need Opus-grade reasoning to autocomplete a function. The real question is the elasticity of substitution: once an internal tool clears "good enough" for the bulk of traffic, how much premium work is actually left for the frontier vendor? At Meta and Microsoft the answer appears to be: a lot less than the peak bill implied.

The Enterprise Buyer. A CTO reads this and sees concentration risk, but the opposite kind from Anthropic's. If two Anthropic customers were a quarter of its revenue, those customers had enormous leverage and knew it. The rest of us don't. What I'm signing a Claude contract for is capability I can't reproduce cheaply, and the moment my own team can stand up something that handles 70% of the volume, I'm renegotiating or walking on that 70%. The lesson isn't to avoid Anthropic. It's to never let a single vendor's tool become the critical dependency across tens of thousands of seats without a priced-out exit. Meta had one in Metacode. Microsoft had one in its own models and OpenAI. That optionality is what turned a "several billion" bill into a negotiation.

Where they disagree. The Skeptic says this is normalization after a sugar high, and $1.26 billion from one customer proves Anthropic is fine. The Compute Pragmatist says that misses the direction of travel: every dollar that moved to internal infra is a dollar that doesn't come back, and the biggest customers are exactly the ones who can keep repatriating. Both are looking at the same $105 million a month and seeing opposite futures, a floor versus a way station.

The second split is about what the substitution means. The Researcher reads Microsoft falling back to GPT-4.5 as evidence that routine coding has commoditized. The Builder reads it as a tooling failure Anthropic could have prevented with stickier integration. One blames the model's commoditization, the other blames the product's shallow roots.

What it actually hinges on. One belief: can Anthropic grow new large accounts faster than its biggest customers repatriate workloads onto their own infrastructure? The hyperscalers that can self-serve (Microsoft, Meta, Amazon, Google) are precisely the accounts a model vendor most wants and least controls, because they have the GPUs and the research staff to replace you. Anthropic's defense is visible in the same news window: it's giving startups a free year of Claude Team and $1,000 in credits, which is a move to broaden the base away from a handful of whales and toward many smaller customers who can't build their own model. The council leans toward the Compute Pragmatist. The direction is repatriation by anyone large enough to try it, and the free-startup push is Anthropic conceding exactly that by hunting for revenue that won't walk.

What to verify before reading this as either disaster or nothing: whether Anthropic's next disclosed revenue mix shows the top-customer concentration falling because the base broadened, or because the whales shrank. Those are two very different outcomes hiding inside the same growth number.

Prediction: In Anthropic's first IPO-related financial disclosure (S-1 or equivalent prospectus) expected in 2027, the revenue share from its single largest customer will be lower than the roughly 15% implied by the "two customers at a quarter of 2025 revenue" figure.

Confidence: Medium. Repatriation at the whales is real, but IPO timing and new-account growth are both moving targets.

Why: Microsoft and Meta, the kind of large accounts that drive Anthropic's concentration, are both actively pulling coding workloads onto their own models, and the reported dollar cuts are large enough to shrink the top line of any single-customer share. At the same time Anthropic is spending on free-startup acquisition in this exact window, which broadens the denominator and dilutes any one customer's percentage. Both forces push the top-customer share down, so the number falling is the more likely outcome. It would stay flat or rise only if a huge new single customer signed and immediately dominated the mix, which is the less likely path given that the biggest buyers are the ones defecting, not arriving.

Revisit by 2027-12-31: We're right if Anthropic's IPO prospectus shows its largest single customer accounting for less than 15% of the most recent reported period's revenue. We're wrong if that largest customer accounts for 15% or more.

One caveat the Skeptic earns: a lower concentration number is good for Anthropic's IPO story, so if the filing shows it, expect that framing to be front and center. The question underneath is whether the base broadened or the whales shrank, and the prospectus will be written to make you read it the flattering way.