Elena Voss€43 billion is now the low end of ASML’s 2026 sales forecast. The floor had been €36 billion. That €7 billion reset arrived as giant models moved toward wider access and New York made large data centers harder to permit. What connects a higher equipment floor to wider model access and a new permit barrier?
Chip budgets ran into permits
ASML reported €9.3 billion in quarterly sales, above the €8.8 billion estimate, and raised its full-year forecast for the second time this year.
IBM showed what that allocation displaced. Revenue rose 1% to $17.2 billion, below the $17.9 billion estimate, and shares fell 25%. CEO Arvind Krishna said customers were shifting spending toward chips. Customers kept spending, but moved budget toward compute.
New York then imposed a moratorium on new environmental permits for data centers above 50 megawatts for as long as one year. Data-center developers must now price that delay into where and when they build. Chip availability no longer sets the schedule alone.
Open weights shifted the buyer’s problem
Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that the company says rivals Opus 4.8 and GPT-5.5. It plans to release the weights by July 27, making the model inspectable and deployable beyond one vendor’s interface.
Thinking Machines Lab debuted Inkling, an open-weight mixture-of-experts model with 975 billion total parameters and 41 billion active. Those figures do not establish benchmark parity, but they do change procurement. Buyers can inspect and deploy open weights outside a single interface, while assuming more responsibility for integration, inference cost, and safety controls.
Platform owners made defaults the product
Apple put its new Siri AI into public beta across iOS, iPadOS, macOS, watchOS, and tvOS. OpenAI is reportedly taking the opposite route: building a moveable, screen-free smart speaker with a camera and sensors. Apple owns the endpoint and is adding intelligence; OpenAI owns the assistant relationship and is seeking an endpoint.
The EU targeted that control point directly. Two DMA decisions ordered Google to give rival AI assistants and search engines comparable access to Android and some Search data. A model company can win a benchmark and still lose the user moment when the platform controls the default.
Maintainers inherited automation’s review bill
Demis Hassabis proposed a US-based standards body for frontier AI, modeled after FINRA. Participating labs would provide models for review as much as 30 days before release. The proposal gives voluntary commitments an institution, a timetable, and access to the systems being evaluated.
AI coding tools are already flooding open-source projects with low-quality contributions, leaving maintainers to triage, review, and reject generated patches. Frontier labs can schedule review before release; maintainers receive output patch by patch after deployment. Teams adopting coding agents must budget for the reviewers and rejection queues their output creates, not only for generated code.
The €7 billion added to ASML’s forecast floor was the week’s clearest price tag for what model abundance cannot supply: machinery, permission, distribution, and human judgment.