The DMA Has No Box for Europe's AI Halls

Europe is pouring AI halls on a legal calendar that still has blank rows. The Commission’s cloud files on Azure and AWS are due in November — the first serious test of whether hyperscale infrastructure is a Digital Markets Act core platform service. Microsoft, in the same season, is still raising AI-related capex after a quarter in which Azure grew more than 40 percent. The EU’s own EuroHPC call for up to seven “AI gigafactories” is out to tender. Wire copy treats those as one story: a durable, increasingly risky global investment cycle with European weather on top.

They are not one story. Capex is a purchase order. The DMA is a sorting machine. The weather report — U.S. chip controls, Chinese efficiency, Dutch and Irish grid queues — only bites once you know which legal box the product landed in.

The residual is narrower than “will Brussels regulate AI?” It is this: if the DMA is the relevant statute, what mechanism decides whether an agent, a cloud region, or Copilot-as-a-feature falls under existing gatekeeper duties — or needs a different enforcement approach? Until that mapping is written, headline spend in Europe is a claim on future watts, not a classified service.

Mapping, Not a New Statute

The DMA never asked whether a product was intelligent. It asked whether it was an important gateway: operating system, search, app store, advertising, intermediation. Designation attaches to a core platform service, not to a company, and not to a technology generation. Microsoft learned that twice. Windows and LinkedIn wear the badge. Edge met the quantitative thresholds and still walked, because Brussels accepted it was not an important gateway — a rebuttal the General Court has now left standing against Opera. The lesson from Edge’s escape is not deregulation. It is discretion at the category boundary.

AI products are being forced through those same apertures. When the Commission needed a handle on ChatGPT, it did not invent an “agent” core platform service. It tagged the product as a search engine under the Digital Services Act and pointed DMA search-data sharing at a January 2027 start — a calendar, as we wrote, not a working feed of Google’s logs. Retrieval looks like search; so the box is search. Copilot, sitting on Windows and Azure, can be argued as a feature of already-designated or under-review services. A standalone agent that books, buys, or routes around app stores looks like an intermediary — if anyone opens that file. Cloud is the live gap: Azure and AWS are in a market investigation, not yet wearing the gatekeeper label that would drag interoperability, self-preferencing, and data-use duties onto the training and inference stack.

The DMA does not ask whether a product is AI. It asks which existing core platform service it most resembles — or whether it is merely a feature of one already designated.

That is why Article 53 matters. The Regulation requires the Commission to evaluate the statute and report — the only clean statutory moment to add a box, refuse one, or admit that mapping is the policy. No published evaluation has produced a consistent classification rule across agents, cloud, and integrated AI. The kill-conditions for this piece are unmet: Brussels has not said the old categories apply without material ambiguity, and it has not issued a uniform enforcement key. Uncertainty is not a vibe. It is the absence of that key.

Capex Clock Versus Category Clock

The physical clock does not wait for Luxembourg. Hyperscalers announce capex ahead of demand, grid connection, and commissioning. Depreciation and financing start when the hall is capitalized; cloud revenue starts when the interconnect is live and the SKU is sellable. Europe’s permitting and transmission queues stretch that gap. The utility test already asks who pays for racks that never fill. Classification risk is the legal twin: a region built as “cloud” may later be a designated CPS, an integrated Copilot feature, or an undesignated warehouse of GPUs. Same concrete. Different duties. Different counterparties.

Geopolitics writes the bill of materials inside that delay. Beijing’s 15th Five-Year Plan treats AI as an economic form, not a sector. Chinese efficiency work showed that frontier results need not follow Silicon Valley’s watt-first path — which is another way of saying U.S. export controls and custom-silicon lock-in, the stack we tracked in Broadcom’s quiet XPU role, can strand a European hall even when the capex line stays fat. Crates sit in bonded warehouses for a reason. Five Eyes’ warning that frontier-model cyberattacks are months away adds a security overlay Brussels will not ignore when it decides whether cloud is a gateway. None of that is priced if you only read “AI capex up.”

What still mis-prices is the fusion. Credit and equity treat European hyperscale as a delayed version of Virginia: same cycle, slower permits. Underwriters treat DMA risk as a fine overlay on Search and Play. The residual is the missing map. A November cloud designation would pull Azure and AWS into ex-ante duties while the gigafactory tenders are still ink. A non-designation would leave the largest AI landlords outside the statute that was supposed to civilize gatekeepers, and push agents back into search-or-feature mapping. Either outcome reprices the halls already poured.

The testable claim is not a metaphor. Watch the November Azure/AWS files for whether cloud becomes a core platform service. Watch the Article 53 evaluation for whether “AI agent” or “integrated AI” appears as a new category or is expressly refused. If both stay silent, the mechanism is confirmed: Europe builds first and classifies by resemblance, and the capex clock runs in the blank rows.