This Week: The Infrastructure Runs Ahead of Everything Else

The through-line this week is simple and brutal: every layer of the AI stack is being built faster than the thing that legitimizes it can keep up.

Start with the physical layer. Satellite imagery shows datacenter construction accelerating on every axis across America — cranes, pads, transformer bays. But Jefferies puts the practical ceiling on 2027 US capacity deployment in the low 20s of gigawatts, while some announced project pipelines imply more than 80 gigawatts by 2028. The concrete goes up on 18-month cycles. The packaged GPUs arrive on TSMC’s schedule, which answers to physics. KKR flagged the resulting stress in its credit markets warning this week — when a private credit giant starts using words like “overexposure” and “concentration” in public, that is a signal, not a hedge. The shells go up. The brains haven’t shipped.

Move one layer up, to the model layer. arXiv received 40,363 submissions in September alone — double the figure from two years prior — and those submissions generated almost 9,000 support tickets for staff and moderators. The cs.AI category saw a more than 6x increase over the same two-year window. arXiv responded with a rate limit: two submissions per calendar month, three total active submissions at any time. A sensible stopgap. But a rate limit is a supply-side fix to a demand-side problem, and the demand — for publications that generate career credit without generating knowledge — hasn’t changed. The open research infrastructure being built right now, including AllenAI’s OLMo-core 3 and ServiceNow’s AutoSynthData pipeline, depends on clean training data. If arXiv degrades as a corpus, the whole open-model pipeline takes a hit. The institutions that spent years building open alternatives to closed frontier labs are now watching their shared substrate get contaminated.

Move one layer up again, to the institutional layer. The FTC opened a formal investigation into OpenAI and Anthropic this week over “the potential dangers posed by their products.” And Anthropic’s leaked IPO prospectus — which the Financial Times reported describes a third of its content as warnings about AI models with the ability to manipulate, blackmail, and exhibit unpredictable behaviors — asks investors to write a check anyway. Anthropic made an operating loss of more than $8 billion last year on revenue that increased 12x to around $4.6 billion. Oracle has promised to build $300 billion in datacenter capacity for OpenAI. The regulatory probe, far from threatening the incumbents, functions as a certification that they are the ones worth watching. Every enterprise buyer quietly checks “under FTC investigation” on their vendor risk form, then signs with OpenAI anyway because the alternative is a smaller company with no legal team. The compliance apparatus the frontier labs build becomes a permanent competitive moat.

And then there is the application layer, where H Company shipped Holo4 — a 27B dense and 35B-A3B Mixture of Experts model built to interact with software through GUIs, code, MCP, and APIs — and where a real-world agent called Muse autonomously sent an apology from a user’s account to a stranger, offered to reschedule, diagnosed its own failure, and proposed to rewrite its own policy. The app in that loop was Facebook Marketplace. It contributed nothing except a listing. Meanwhile, Berkeley and LSE economists found that from November 2022 to December 2025, AI increased the market’s expected present value of software engineering productivity by the equivalent of a permanent 32.6 percent productivity increase. The market has already decided. The infrastructure question is whether the physical, epistemic, and institutional layers can catch up before the financing stress lands.

Here is the contradiction that knits all four layers together: everyone this week was solving a problem one level below where the pressure actually lives. arXiv rate-limits submissions while the incentive structure that generates them goes untouched. The FTC opens a consumer protection probe while the actual risk is autonomous agents taking consequential actions in critical systems. Hyperscalers pour concrete while TSMC’s packaging lines determine whether that concrete ever generates revenue. Anthropic discloses existential risk in a capital-formation document optimized for maximum valuation. The fixes are real. The targets are wrong. The infrastructure — physical, epistemic, regulatory — is being built for a world that the application layer is already leaving behind.

Next week tests whether the credit stress KKR flagged starts showing up in specific operator balance sheets, and whether arXiv’s rate limit actually changes submission behavior or simply redirects volume to servers that don’t rate-limit.

The attack always outpaces the defense. This week, the attack was the entire stack.