Big AI companies learned to buy the useful part without always buying the company.
That is the point of the acquihire wave. The buyer licenses technology, hires founders and core researchers, leaves the corporate shell outside the deal, and avoids the clean shape of a conventional acquisition. The startup may still exist. The competitive asset has moved.
Brazil’s competition authority is now treating that structure as a merger-control question.
On May 13, 2026, Brazil’s Administrative Council for Economic Defense, CADE, said its tribunal reviewed a set of digital-market and AI cases focused on new forms of acquiring technology, talent and competitive capability. The main decision was not subtle: CADE ordered notification of the Microsoft/Inflection operation and opened two new APAC proceedings into reported operations involving Google/Windsurf and Google/Hume AI.
That is AI competition governance, not just antitrust housekeeping.
The regulator is asking whether control over an AI business can move through people, model rights, code, cloud access and roadmap influence even when the legal entity does not move in the ordinary way.
The asset is not always the company
CADE did not say every AI partnership is a merger. It said some arrangements can move enough competitive substance to deserve review even when they do not look like a standard acquisition.
That distinction matters.
In the Microsoft/Inflection APAC, CADE described contracts involving Microsoft’s license to use Inflection technology and the hiring of almost all of Inflection’s former employees. The reporting commissioner’s logic was economic, not cosmetic: in AI startups, the main assets are often intellectual property and talent. If those move together, the transaction can replicate the function of an acquisition without using the acquisition wrapper.
This is the part AI deal lawyers should underline.
Merger law was built around visible handles: shares, control, revenue, assets, ownership. AI talent deals blur those handles. A buyer may not need the legal entity if it can get the research team, the model know-how and the product direction. A startup can be hollowed out commercially while its cap table remains formally outside the buyer.
Brazil is saying form will not end the analysis.
That is a useful posture for AI markets. The scarce asset is often not current local revenue. It may be a team that knows how to train a model, a codebase that shortens a roadmap, a benchmark-leading capability, or a group of engineers that would have become a rival if left outside the incumbent.
Revenue thresholds miss that by design. A frontier-adjacent AI startup can be strategically important before it is large in Brazil.
CADE also showed restraint
The stronger part of the May 13 decision is that CADE did not grab every AI deal in sight.
At the same session, CADE closed three APACs involving NVIDIA/Run:ai, Google/Character.AI and Microsoft/Mistral AI. That matters because lazy AI enforcement is easy. “Big Tech plus startup” is not a theory of harm. It is a headline with legal stationery.
For NVIDIA/Run:ai, CADE said the parties’ Brazilian revenue was below mandatory notification thresholds and found no relevant Brazilian effect that justified a call-in. Run:ai, in CADE’s account, did not operate in Brazil, and the record did not show market-power effects there.
For Microsoft/Mistral, CADE described an arrangement involving investment, Azure supercomputing infrastructure and availability of Mistral models on Azure. It closed the APAC after finding no acquisition of control and no identified harm to the competitive environment.
Google/Character.AI is more instructive. CADE said the transaction involved technology and IP licensing plus the release of specialized professionals for later hiring by Google. It acknowledged that this type of structure can matter competitively, especially where technology, specialized teams and innovation capability are involved. Then it still closed the proceeding in that case, citing proportionality, elapsed time and alternative institutional routes.
That is the enforcement line taking shape: unusual AI form does not automatically trigger notification, but unusual form no longer ends the analysis.
Article 88 is the pressure valve
The legal mechanism is doing real work.
Brazil’s Law 12.529/2011 sets ordinary merger-notification thresholds in Article 88. The live CADE FAQ says those thresholds, updated by Interministerial Ordinance 994/2012, are R$750 million for one group and R$75 million for another group. But Article 88, paragraph 7, gives CADE a pressure valve: it may require submission of concentration acts that fall outside Article 88 within one year of consummation, according to the statutory text.
CADE used that valve on Microsoft/Inflection.
The authority said the operation did not meet the ordinary revenue requirement, but it still ordered notification under Article 88, paragraph 7, with submission due within 30 days of publication of the decision. CADE’s Resolutions page also points to Resolution 24/2019, which disciplines APAC procedures under Article 88, paragraphs 3 and 7, and to Resolution 33/2022, which disciplines notification of acts under Article 88.
This is the Brazilian version of a broader competition problem.
AI markets can concentrate through paths that conventional screens were not built to catch. A platform can take the team and leave the cap table. A cloud provider can finance a lab while shaping compute distribution. A search company can hire a coding-agent team while preserving formal separation from the original company. The structure may be hard to classify because ambiguity is part of the product.
CADE’s May 13 statement explicitly named technology, intangible assets, specialized teams, strategic personnel and potential competition as reasons digital and AI-market operations may demand review even where Brazilian revenue is not enough on its own.
That is a broader map than “who acquired whom?”
It asks what moved.
The Google probes are the forward edge
Microsoft/Inflection is the template case. Google/Windsurf and Google/Hume AI are the forward edge.
CADE did not resolve those two Google matters on May 13. It opened new APACs. That matters because the agency is not only retrofitting one famous deal. It is building a repeatable path for future AI talent and technology transfers.
Windsurf is a useful test because coding agents sit close to distribution power. If a dominant platform can recruit a critical engineering team, license core technology or redirect product development without triggering acquisition review, AI software markets can consolidate before revenue data tells the story. By the time thresholds wake up, the people who would have built the competitor may already be inside the incumbent.
Hume AI points to another flank: model capability in voice, affective interfaces and human-computer interaction. The same structure applies. A firm does not need to buy a startup outright if it can obtain the key people and rights that made the startup strategically relevant.
The point is not that Google did anything unlawful. CADE opened inquiries, not conclusions.
The point is that AI deal architecture itself is now an enforcement object.
The implication
Brazil is giving other regulators a clean template.
Do not pretend revenue thresholds are useless. They still screen out noise. Do not pretend every AI partnership is a disguised acquisition. That would turn competition law into a clipping service. But when a transaction transfers the team, IP, model rights or future competitive option, the authority should ask whether control has moved in economic substance even if it has not moved in corporate form.
This will make AI dealmaking slower. Good.
The current market rewards speed and ambiguity. Buyers want talent before competitors bid. Startups want rescue terms before the runway ends. Investors want capital returned without a long merger review. None of those interests is identical to competitive market structure.
The risk is overreach. If every licensing deal or hiring wave becomes a merger case, AI collaboration gets legally noisy and small firms lose exit routes. CADE’s simultaneous closures reduce that risk. The agency is saying it can distinguish a weak case from a structural concern. It now has to keep proving that discipline.
For AI companies, the operating lesson is simple: if the deal memo says “not an acquisition,” ask whether the business reality says otherwise.
CADE just did.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.