The FTC Probed the AI Labs. The Real Target Sits One Layer Up.

Contrary to the headline, the most consequential fact about the Federal Trade Commission's industry-wide inquiry into [OpenAI], Anthropic, and their peers is not that it exists. It is that the coverage of it exists almost entirely without evidence. No docket number. No cited legal instrument. No defined scope. No timeline. Five information points, and three of them are cushioned by the word "may." During the 2017 ICO boom, I audited fifteen ERC-20 whitepapers for a Frankfurt fintech desk and found mathematical inconsistencies in eight of them; the discipline that produced that result was simple — judge a document by what it refuses to say. Applied here, that discipline returns an uncomfortable verdict: a probe into the capital plumbing of artificial intelligence has been announced, and the announcement is already being misread as a verdict. It is not one. The interesting question is not whether the FTC is looking. It is where it is looking, and why the visible target and the real target are not the same entity. The FTC does not open "probes" in the abstract. Its toolkit matters, because the legal instrument chosen determines everything downstream — cost, duration, and consequence. The most plausible reading of an "industry-wide" inquiry into capital-linked laboratories is a Section 6(b) study: a research authority that compels firms to produce information for policy analysis, carries no allegation of wrongdoing, and produces no penalty on its own. This is a fundamentally different animal from an antitrust enforcement action, which arrives with named defendants and remedies. The distinction is not academic. A 6(b) study can run for months, terminate quietly, and reshape industry behavior purely through the anticipation it generates. An enforcement action either lands or it does not. Confusing the two is the error baked into the framing here. The structural backdrop is what makes the omission so costly. Over the past three years, the dominant organizational form in frontier AI has become a triangle: a hyperscaler supplies capital, the same hyperscaler supplies compute, and the same hyperscaler receives preferential access to the resulting models. Microsoft and [OpenAI]. Amazon, Google, and Anthropic. The arrangement is elegant because it solves two problems at once — it gives the laboratory the two things it cannot buy on a startup's budget, and it gives the cloud provider a lock on the most valuable inference workload of the decade. Any serious inquiry into this industry is, by construction, an inquiry into that triangle. Reporting that names only the laboratories and omits the investors has therefore inverted the map. Here is where I want to be precise, because the technical substance of the story is not in the AI labs at all — it is in the incentive geometry of capital-for-compute swaps. When a hyperscaler invests in a laboratory, the capital is frequently recycled as cloud credits. The laboratory trains on the investor's silicon, deploys inference on the investor's regions, and in many arrangements grants the investor rights of first refusal on model access. Trace the cash and you find it barely leaves the building. This is not fraud; it is a rational structure that produces a specific competitive artifact: the appearance of an independent laboratory sitting atop a de facto vertically integrated stack. The ICO parallel is exact in shape if not in scale. In 2017, tokens were sold to create the appearance of decentralization while the treasury, the multisig, and the roadmap remained under one hand. The math did not lie then, and it does not lie now — the difference is that the instrument of concealment has shifted from tokenomics to corporate structure. Charting the entropy of digital scarcity was the task of the last cycle; charting the entropy of compute scarcity is the task of this one. From an audit perspective, three variables matter. First, exclusivity: does the arrangement contractually bind the laboratory's compute to a single provider? Second, control: does the investment carry board seats, information rights, or veto power that convert a minority stake into effective governance? Third, precedence: does the investor receive model access ahead of the market, converting a financing relationship into an information asymmetry? Each is independently checkable, and each is precisely the kind of thing a 6(b) study is designed to surface. Following the code where the humans fear to tread, the trail runs straight into the investor's data center, not the laboratory's. The reporting that triggered this essay supplies none of the three. The inclusion of nonprofit and benefit-corporation structures in the scope is the detail most worth pausing on, because it is the one the coverage treats as a footnote. [OpenAI] operates under a capped-profit subsidiary beneath a nonprofit parent; Anthropic styles itself a public benefit corporation. Both structures exist to let the firms pursue a mission while raising market capital — and both, from a regulator's chair, look like a mechanism that converts mission language into competitive insulation. If the inquiry reaches governance, the question stops being "did these labs compete fairly" and becomes "is the nonprofit wrapper itself a device for avoiding the scrutiny that a conventional for-profit would face." That is a far more durable line of attack than any single investment clause, and it is the one the source material never raises. This is also where the crypto-native reader has a stake, even if the article never mentions them. The capital-for-compute triangle has a mirror image in decentralized compute networks — Render, Akash, and their peers — which exist precisely because the price of centralized compute is set inside relationships that outside buyers cannot see. My own modeling of AI training demand against node profitability, which I began in 2025, keeps returning to one asymmetry: when compute is allocated by contract rather than by market, the visible price stops carrying information. Regulators are now circling the same observation from the competition side that cryptographers reached from the incentive side. If the triangle is constrained, the marginal dollar of AI compute migrates toward open, price-discoverable markets — not because decentralization is virtuous, but because regulated relationships create demand for arms-length alternatives. There is a second-order effect the reporting ignores entirely, and it is the one I would flag to any portfolio manager. A constraint on deep binding is asymmetric. It hurts the laboratory that rents its compute and helps the firm that owns it. Google, with its in-house TPUs, and Meta, with its open-weight strategy and owned clusters, sit outside the triangle's dependency. The inquiry, if it hardens into remedy, redistributes competitive position toward the self-funded — a structural trade invisible in a story that names only the two laboratories most exposed. Now the contrarian cut. The instinctive read is that regulation is coming for AI, and that the safe assumption is "stricter." I would invert it. A research inquiry is a mechanism for gathering facts, not for punishing them, and the fact that a probe exists tells you almost nothing about whether it will convert into action. The variable that actually governs the outcome is political, and the source material is silent on it — deliberately, I suspect, because naming it would deflate the drama. Leadership at the FTC turns over with administrations, and priorities turn over with leadership. The same file that one commission pursues aggressively can be shelved by the next without a single hearing. The reporting's repeated use of "may" in the body, paired with a title that reads as certainty, is a textbook case of certainty inflation — the hedges quietly dropped between the evidence and the headline. The most likely near-term outcome is not a breakup of Microsoft–[OpenAI] or Amazon–Anthropic. It is a chill: firms pre-emptively slowing new strategic investments, tightening the language of exclusivity clauses, and routing deals through structures that survive a subpoena. That chill is real, it is already priced into how lawyers advise, and it will never appear in a headline because nothing dramatic has to happen for it to bite. The deeper blind spot is scope. The reporting frames this as an American AI story. It is not. The European Union's AI Act, the United Kingdom's approach, and parallel frameworks elsewhere mean that any capital structure reviewed in Washington is being reviewed in Brussels under a different theory. Regulators are converging on the same target from different directions — competition law in the United States, product liability in Europe — and a single-jurisdiction reading misses the compound risk. The same compute triangle that a Section 6(b) study probes for competitive harm is the one the AI Act implicitly regulates for safety and accountability. Two instruments, one structure, no coordination between them. That is the exposure a reader should be tracking, and it is absent from the source. The signal worth holding is not the probe. It is the question of who, exactly, is being examined — the laboratory on the label, or the investor behind the glass. Watch for the FTC's first formal document; that single filing will tell you more than a hundred aggregation pieces, because it will name the legal instrument and the object, and everything downstream — duration, remedies, and the competitive reshuffle — is determined by those two words. Until it appears, treat the story as an open ledger with the most important line redacted. The architecture of value in a system this opaque is not revealed by the headline. It is revealed by whoever is finally compelled to disclose.