Thirty days. That is the maximum prerelease review Demis Hassabis proposed for frontier-class models. In the same week, Moonshot AI launched a 2.8T-parameter model and said its weights would follow by July 27. One clock was meant to slow release; the other was already counting down.

Open weights turn scale into a distribution strategy

Moonshot AI launched Kimi K3 and plans to release its weights by July 27. The promised weight release matters more than Moonshot’s benchmark claims. If the company follows through, developers can adapt the model without depending only on a hosted interface.

Thinking Machines Lab made architecture as important as scale with Inkling, an open-weight mixture-of-experts model with 975B total parameters but 41B active parameters.

Kimi K3 total parameters
Inkling active versus total parameters

Inkling’s total-to-active ratio is 23.8 to 1, making total parameter count a poor proxy for the compute engaged. A buyer comparing these releases must ask which weights are available and how many parameters activate, not simply which headline number is larger.

Release queues and app stores can outweigh benchmarks

Hassabis proposed a US standards body modeled after FINRA, with frontier labs submitting models up to 30 days before release. A shared queue would turn launch timing into a compliance cost and favor labs that can absorb the uncertainty.

The EU issued two DMA decisions ordering Google to give rival search engines and AI assistants comparable access to Android and some Search data. An open model can be abundant and still struggle commercially without a route to users. For model providers, distribution terms may matter more than a temporary benchmark lead.

AI spending is concentrating in physical chokepoints

ASML reported €9.3B in quarterly sales, above the €8.8B estimate, and raised its 2026 sales forecast from €36B–€40B to €43B–€45B. Model demand can rotate while ASML collects at the lithography layer.

IBM’s preliminary second-quarter revenue rose 1% to $17.2B, below the $17.9B estimate, and its shares fell 25%. CEO Arvind Krishna said customers were shifting spending to chips. AI budget growth can reward fabrication while punishing vendors farther downstream.

New York imposed a moratorium on new environmental permits for data centers above 50MW for up to one year. A developer planning a large facility must now absorb a permitting delay regardless of chip availability. Compute capacity depends on local permission as well as capital.

Code generation shifts the bottleneck to review

Users of AI coding tools are flooding open-source projects with low-quality contributions, overwhelming maintainers and threatening community engagement. Code generators cut the cost of proposing a change, but maintainers still bear the cost of deciding whether to accept it.

For companies that depend on open-source software, maintainer attention is now a supply-chain risk. More generated code does not improve throughput when the review queue grows faster than human capacity.

Hassabis framed thirty days as time to inspect a frontier model before release. By week’s end, it read as something else: the gap between how quickly capability can be distributed and how slowly it can be made usable. The release clock starts the race, but it no longer sets the pace.