The DOJ's case against Google's search monopoly took roughly five years from filing to a remedies order. The European Commission's cases against the same company have run more than a decade across multiple cycles. In the same span, frontier AI training compute has grown 2 to 3x per year for eight years running — meaning the AI stack triples in scale roughly every twelve to eighteen months. The two clocks are not on the same axis.
The structural mismatch
Antitrust is ex post and case-specific by design. A complaint is filed. Discovery runs for years. Expert reports are exchanged. A trial happens. A judgment is rendered. A remedy is appealed. By the time anything binding lands, the market that was being adjudicated has been replaced by a market that wasn't.
This was tolerable when the underlying technology cycled in decades. Standard Oil was a multi-decade fight; the railroads were similar; even Microsoft's antitrust ordeal in the late 1990s addressed a stack that remained recognizable a decade later. AI does not cycle in decades. The big four cloud companies are projected to spend up to $660 billion on AI capex in 2026 alone — more than the annual GDP of most OECD economies. That capital is purchasing positions in a stack that will be entrenched long before any case filed today reaches a remedies order.
The Yale Law & Policy Review's antimonopoly analysis of AI is blunt: the AI stack is severely concentrated at multiple layers simultaneously — Nvidia for chips, three hyperscalers for cloud, a handful of labs for frontier models — and each layer reinforces the others. By the time a court rules on any one layer, the others have moved.
What is being capitalized today
The next decade's monopoly is not being built. It is being capitalized. Microsoft is entangled with OpenAI through a multi-tier investment and revenue-share structure. Amazon committed billions to Anthropic and routes its training onto AWS. Google trains on its own silicon and rents it back to itself. Every hyperscaler is locked into Nvidia, and Nvidia rations supply.
These are not allegations. They are the ordinary terms of business of the most valuable companies on earth. The vertical integration is the strategy, and it is being financed at a rate that an antitrust filing five years from now cannot retroactively undo. By the time a court could order Microsoft to divest its OpenAI position, OpenAI's frontier models will already have been deployed across Microsoft's enterprise base and integrated into a million Excel files. The remedy that would have mattered in 2024 does not fit 2029.
The political path looks worse than the legal one
The political path looks worse than the legal one, for three reasons. AI labs have the capital to lobby and the rhetorical advantage of "we are building the future." They have a national-security framing — "we have to win the AI race against China" — that is genuinely true and aligned with their commercial interests. And the political coalition that historically counters monopoly, organized labor, is being structurally weakened by the same technology, with entry-level work down roughly 35% since January 2023.
That is the precise template of the cyberpunk corporation in fiction: too useful to break up, too critical to regulate, too entangled with state power to confront. The fiction treated entanglement as the bug. In the political economy of 2026, it's the deal.
What a serious response would look like
A response calibrated to the actual clock would not be a single five-year case. It would be:
- Ex ante structural rules on vertical integration in the AI stack, treating GPU supply, cloud, and frontier model training as separable layers with required separation between them.
- Mandatory data-portability and model-context export rules, so that user lock-in cannot harden into the next generation of switching cost.
- Public investment in open-weight model research and public compute infrastructure, on the public-utility model that built the early internet and the interstate highways.
- Compute and concentration audits that publish quarterly, not multi-year discovery cycles that publish at the end of an appeal.
None of this is exotic. All of it is the digital-era version of fights already fought, and largely won, for electricity, telephony, and the early internet — before those technologies got captured.
The question is not whether antitrust will eventually act. It will. The question is what shape the market will be when the ruling lands, and whether anyone outside the firms named in the caption will still have leverage by then.