The Investor

Begin with the arithmetic, since that is the part everyone reads wrong. The Shanghai program is reported at roughly five immersion DUV tools in 2026 and a target of about 20 in 2027, set against ASML's guided ~130 DUV immersion machines for the guidance year. That is not share loss on any horizon a 2026 earnings model has an opinion about. The tools reportedly lag on performance and reliability. The equity fell anyway.

What got priced was the disappearance of monopoly pricing power on the mere existence of a credible substitute — or rather, on the existence of something a customer could plausibly threaten to buy. That is a terminal-value event rather than an EPS event, which is exactly why it bites private semicap marks harder than the public tape.

The pincer nobody can diligence away

Techpresso supplies the second blade: the US Congress has separately proposed blocking DUV sales to China outright. So ASML loses the China channel by substitution or by regulation, and China was roughly 30% of its 2025 revenue per The Information's reporting. There is no branch of that tree where the revenue line arrives intact. Any IC memo underwritten on "no credible Chinese substitute" is describing a world that has closed.

The memory leg is the sharper signal

Morning Brew reports CXMT listed in Shanghai up 466% to more than $487B, which makes the world's number-four DRAM maker China's most valuable listed company on its first trading day, while it operationally trails Micron, Samsung and SK hynix and holds no demonstrated high-bandwidth memory position. Using that as a comp would be malpractice. Using it as a balance sheet would be prudent, because a capital-rich national champion with a mandate spends differently from a company answering to a cost of capital.

The detail that turns politics into a socket is Apple reportedly evaluating CXMT DRAM for China-market devices. Bifurcation has reached the design-win level. Every hardware holding with a bill of materials sold into China now needs a Chinese-second-source scenario inside the model rather than a paragraph in the risk section.

Where the sources diverge, and what it tells you

Both can be true, and the reconciliation is where the money sits. Chinese fabs running lower-reliability tools at volume generate structural demand for metrology, inspection, process control and service intensity. The reliability gap becomes revenue for somebody. This is probably wrong at the margins, but concentrating in EUV-adjacent, inspection and advanced-packaging assets while haircutting DUV-tier China terminal value is one trade expressed twice.

The market repriced a lithography monopoly on the existence of an inferior competitor. If that is the standard for moat skepticism, every terminal-value assumption in the book needs a rerun.

The genuine unknown is EUV. DUV was the beachhead; if the same playbook reaches extreme ultraviolet, the tiering thesis collapses rather than adjusts. Nothing in the available reporting suggests that, and nothing in it rules it out.

The mechanics are more interesting than the headline, which is usually where the money is. A 7.5x ask against a 1.25x realization is not a pricing event. It is a pricing-power failure, disclosed in public, in a market where every procurement officer reads the same reporting. The asset destroyed here is precedent: AI tooling list prices are now known to be negotiable by 80% or more.

Then follow the released dollars, which is the part nobody puts in a deck. The same IT consultancy, per The Information, expects to pay Anthropic more than $10M in 2026 through Claude Code, up from negligible. Sanofi's Chief Digital Officer said on the record that a 5x renewal proposal may go unrenewed because "I get a better deal with Claude Code." Druva is reportedly complaining too. The aggregator holds flat while the model owner scales to eight figures inside the same account.

What this does to a mark

Any position where more than roughly a third of revenue is third-party model resale gets squeezed from three directions at once: inference COGS above, first-party agents below, open harnesses beneath both. "We offer model choice" is a feature description wearing a moat's clothes. The numbers that actually settle the question are unglamorous and specific: realized net dollar retention, average discount-off-ask on 2026 renewals, inference COGS as a percentage of revenue. A company booking list and realizing a quarter of it is a markdown, and the GP would very much rather find it than the LP.

The contradiction worth holding

Weave sells model routing, which means it sees cross-vendor spend, and it reports that its customers' Cursor costs rose less than their Claude Code costs over six months. The loudest complaints are not tracking the largest actual increases. Some meaningful share of this narrative is therefore sentiment rather than invoice. It also means the neutral routing provider holds the only ground truth on AI coding economics, an asymmetry that is both a moat to underwrite and a diligence resource to rent.

Two adjacent numbers that change the underwriting

• The reported $60B SpaceX interest in Cursor does not belong in a comp table. That is strategic-acquirer logic priced against on-the-record churn risk. Both branches model cleanly. If it closes, pricing gets subsidized and the displacement window narrows. If it breaks, the app-layer comp set compresses and marks follow.
• Banyan surveyed 260 executives at sub-$50M-revenue software firms and roughly half report fewer than 25% of customers using the AI features they shipped this year. Coding tools get consumed voraciously. Generic embedded AI features languish. Paying a multiple for features nobody opens is a choice, and usage telemetry belongs on the board KPI page.

A vendor that asks $1.5M and books $250K has published its own leverage. The aggregator layer has no moat; the harness underneath it is where the pricing power appears to sit.

This is probably wrong somewhere, so the falsification case, stated plainly: the harness savings claims all come from interested parties, 25% from one operator and 97% from another, implying wildly different baselines. And the open-harness cost case rests on open weights, a live policy dependency sitting inside a gross margin line.

The preconditions are what make the Bun result underwritable rather than quotable. Jarred Sumner ported 535,496 lines of Zig to Rust in 11 days with 64 parallel agents and $165,000 of tokens, displacing something like three engineer-years, which is the number everyone repeats and roughly the least useful one. It worked because three things happened to be true at once: the original author was driving, the test suite is written in TypeScript and therefore indifferent to which language sits underneath it, and the harness was solid enough that a passing test counted as evidence rather than encouragement. Remove any one of them and the same attempt costs a quarter and six figures.

Which promotes a line item that has never appeared in a technical diligence template: test-suite architecture is now an underwritable asset, filed next to the org chart. It is the variable deciding whether an AI-native rewrite is 20x compression or a write-off.

Where this leaves seat-based pricing

Anthropic reportedly caps projects at two engineers, each running three to ten parallel agents continuously, while Bun's pipeline absorbs 100+ pull requests a day with two AI review vendors stacked on the same PR. Headcount has decoupled from output. A portfolio tool billing per seat gets to watch workload rise tenfold while revenue sits still, and it turns up in expansion revenue about two quarters before anyone in the room names the cause.

The same shape appears in physical AI

a16z handed Applied Intuition's CTO a platform to argue that model capability is not the binding constraint in robotics. Teams take delivery of a meaningfully better model and then spend roughly two quarters proving it is safe to ship. A company whose validation cycle runs in months is throttling frontier intelligence down to the speed of its own process, and loses to a rival running identical models on a faster loop. Two unrelated domains, one conclusion. The bottleneck moved downstream of generation, and the budget followed it.

The base rates that kill autonomy pricing

Devshot's ledger is blunt. Computer-use agents reach 26.9% task success when handed a program's actual files and data, up from 20.6% on screenshots alone, at roughly nine times lower cost per task. The 9x cost collapse is the interesting number, not the accuracy. Agents at these rates are not replacing workers. They are making one specific class of attempt nearly free. Businesses monetizing cheap attempts with a human left at the decision point compound. Businesses whose unit economics assume autonomous completion are underwriting a 27% success rate as though it were 90%.

The platform risk to score first

Anthropic shipped Claude Managed Agents in April with multi-cloud and customer-self-hosted sandbox support, after watching customers bolt harnesses together themselves. Anything in the book selling that bolt-on is selling a roadmap a frontier lab already executed. The scoring question fits on one line: what does this company do that Managed Agents does not?

Generation multiples are still being paid for a market whose value already migrated to proving the work was correct.

Counter-case, and it is not a weak one: if the labs fold independent review into the model itself, the verification layer never gets to charge, and this thesis ages badly. Anthropic paying two AI reviewers on the same pull request is the evidence against that, for now.