The Investor

Anthropic lost less money here than it lost control. The Information reports that total Microsoft payments to Anthropic are roughly flat, because Copilot pass-through (the money Microsoft forwards when its customers pick Claude) rose enough to cover the internal cut. Derived math puts internal spend near $670M or less, with more than $330M of pass-through growth filling the hole. The starting point is reported two ways: one account says a peak above $1B annualized, another says a planned budget of at least $1B a year. Microsoft's engineers used to choose this revenue. Microsoft's routing chooses it now. Copilot EVP Jacob Andreou says his job is to own quality, and adds: "our job's also to own efficiency for you."

You cannot out-price a free substitute

Microsoft's commercial agreement lets it use OpenAI's GPT-5.6 Sol without paying OpenAI, so inside Copilot Claude does not compete on price at all. It competes only where it is clearly better. Every gain Sol or Microsoft's MAI models make on complex coding moves volume away from Claude.

The pricing is the more interesting puzzle. Copilot charges $30 per seat per month plus usage fees tied to model cost, which lets Microsoft hand routing savings to customers to take share or hold prices and widen its margin. The option is worth more unexercised.

The quieter lever is deterministic scaffolding, meaning fixed, pre-written code in place of model output. A model used to write roughly 100,000 characters of generic code for every AI-built PowerPoint deck. Microsoft now pastes in a pre-written script. Andreou says this improves quality while "reducing the cost a ton."

Uber shows the same shape from the buyer's side. The Pragmatic Engineer reports its token usage keeps rising while its AI costs have been flat since May 2026, after open models and smart routing cut per-token costs by more than 50%. Bloomberg confirmed the trend three weeks later, and also reports startups leaving premium models, which pushes OpenAI and Anthropic toward cheaper enterprise pricing. Usage growth no longer implies vendor revenue growth.

Meta's seats are the cleaner signal

Meta's spring layoffs cut about 10% of its workforce. If they hit Claude Code users proportionally (a generous assumption), they account for about 6,000 of the roughly 30,000 lost seats, leaving something like 80% to staff moving to Meta's internal tools, MetaCode and Muse Code. That is illustrative math. Meta declined to comment. It fits a wider build-it-yourself habit: The Pragmatic Engineer counts 19 in-house agent harnesses across 18 companies, including Stripe's Minions and Google's Agent Smith.

Microsoft plays four roles at once: investor ($5B), compute landlord (a reported $30B Azure commitment), distribution channel and competitor. It can trim a sub-$1B spending line without disturbing the strategic relationship. Add the reported prospectus disclosure that nearly 25% of Anthropic's revenue comes from two unnamed customers, and diligence narrows to one question. If that concentration sits with buyers able to build their own tools, it is churn risk. If it sits with cloud channels reselling Claude, it is distribution dependency.

Where the evidence cuts the other way

Customer demand for Claude inside Copilot is still rising, and The Information says it is unclear whether Microsoft will make a meaningful dent in customers' Claude bills. Spending through Azure and Amazon Bedrock reportedly keeps growing, though that claim is lower-confidence. Anthropic is also winning mid-market work directly: NinjaOne and Druva use its agents to pull Workday data through APIs.

One outlet reporting this, Augment, discloses an Anthropic position and runs a secondary marketplace. Weight its private pricing accordingly.

This view is probably wrong if those Copilot bills keep climbing, but here it is: the revenue level is intact and its quality is weaker. Marks tend to follow reported totals, and the totals have not moved. The useful work now is finding which holdings depend on a buyer that just got a cheaper option, before the cohort data shows it.

Ed Skoudis made the point in one line in SANS NewsBites: as the cost of producing a plausible vulnerability report approaches zero, the cost of proving one real does not fall. Google's answer was to stop paying for reports, or rather to stop funding the triage that came stapled to them. Adrian Sanabria notes HackerOne absorbed the same flood first, and John Pescatore says managed bounty providers hit it too. We have watched this shape before, in markets where the input got cheap and the check stayed expensive, and it usually ends badly for whoever bundled the check into the price. Wherever a vendor's price includes triage (sorting real reports from bad ones), triage cost grows with every submission while revenue does not.

The same wall in five places

The code figures come from The Pragmatic Engineer, the math and benchmark figures from AINews, and the arXiv figures from MIT Technology Review.

Why this is a valuation problem for vertical AI

ByteByteGo's review of research on sycophancy (models telling users what they want to hear) shows approval metrics can move opposite to correctness. In one cited study, agreement with the user's view predicted preference ratings. Humans and preference models alike sometimes chose a convincing falsehood over an accurate correction.

An AI app citing CSAT or thumbs-up rates as proof of quality in medical, legal or financial work is citing a metric that can be gamed, and the best-resourced lab in the business did not catch the problem either. OpenAI rolled back a GPT-4o update in April 2025 after it turned sycophantic, and admitted it had no evals built to catch sycophancy before deployment.

Where the value lands, and where it doesn't

What gets paid when generating is cheap and checking is not:

• Reproducibility gating: proving an exploit works in a sandbox before a human reads the report.
• Independent eval auditing: AINews notes Artificial Analysis, Vals and Surge gaining influence as vendor claims and independent results drift apart.
• Verification-first developer tooling: every company Gergely Orosz spoke to says it is prioritizing output validation.

Sanabria argues something narrower about Anthropic's Mythos. In AI vulnerability discovery, he says, the model is the least important and most interchangeable part, which is an awkward sentence for anyone paying a model-company multiple. The asset is the expert researcher plus the harness around the model.

Not every domain gets the cost collapse, and the thesis can break in two directions here. François Chollet asks whether gains in fields that can be checked automatically, like math and code, carry over to work that can't. If they carry over, the moats erode faster than anyone is pricing; if they stay put, the moats hold. Our bet is that domains without a cheap automated verifier keep their human-data moats, and AINews supplies the exhibit: the decision model Jev matched GPT-6 Astra's 97.5% on claim verification at about 1/500th the cost, then ranked last on LegalBench.

Caveats: GitHub, Linear and Factory AI supplied the agent-PR numbers, and all three benefit from the agent narrative. The 75M figure is an estimate. If OpenAI's headline math results fail scrutiny, that strengthens this thesis rather than weakening it.

Read the Nvidia sequence the way a credit committee would. The program attracted the borrowers who needed it most. The strongest operator, Nebius, opted out because it didn't want Nvidia taking a cut of its margins and could raise debt elsewhere. That is textbook adverse selection. Nvidia's response, per The Information, is to tighten term sheets and contracts. The implied promise that Nvidia would stand behind mid-tier neocloud credit is getting weaker. The program is being restructured, not cancelled, and final terms are not public. The same reporting cycle includes data center debt concessions stalling build-outs and several IPOs postponed in choppy markets.

Sale-leaseback investors aren't buying Amazon credit

Amazon's reported sale-leaseback of about $8B in Nvidia chips comes from unnamed sources and is unconfirmed. If it happens, the special-purpose vehicle owns the hardware and leases it back to Amazon. The vehicle's main risk is not that Amazon defaults. It is that Nvidia's yearly chip releases cut resale and re-lease values faster than the model assumes.

Two second-order effects follow. First, GPU-backed paper becomes an asset class anchored by hyperscaler issuers. That should tighten spreads for top lessees and widen them for neoclouds competing for the same capital. Second, if even Amazon wants depreciation off its balance sheet, valuations that assume hyperscaler capex keeps compounding deserve a harder look.

Demand is not the problem

Nothing here says AI compute demand is fading. Bloomberg reports that Nvidia partner Hon Hai posted a 47% jump in quarterly sales. The Pragmatic Engineer reports that scarcity has spread to CPUs. Spot capacity is nearly impossible to get without long-standing relationships, reservations are being declined, and customers are prepaying for capacity that comes online from December. Whoever holds supply has pricing power. The stress is in how unbuilt or leveraged capacity gets financed, not in whether the capacity gets used.

DayOne is the counterexample: capacity that is already built

DayOne Data Centers' Nasdaq filing shows what the market will fund. Per The Information, H1 2026 revenue was $512M. That is more than 3x the prior year and already above full-year 2025 revenue of $484M.

Losses are shrinking as capacity fills:

• 2025 net loss: $369M, about a -76% margin.
• H1 2026 net loss: $82M, about a -16% margin.

The growth is in Southeast Asia, outside power-constrained US markets. Fortune's Term Sheet reports $845M in sales for the year ended June 30. By our arithmetic, that matches the two half-year figures.

Valuation is still open. Bloomberg reported a raise of about $5B in August, which the filing has not confirmed, and the excerpts include no EBITDA. Coatue, HillHouse and GDS each own about 19%, so roughly 57% of the cap table could come to market after lockup. GDS's Shanghai heritage is also a diligence question for US buyers.

The smart move is to draw the same line through your infrastructure book. Value capacity that is energized and contracted. Treat promised megawatts and vendor-supported credit as options, not assets.