The global antitrust wave over Big Tech’s artificial intelligence deals has reached Brazil. The question is whether Brazil’s competition authority is paddling into a real breaker—or mistaking regulatory chop for a swell.
Last month, Brazil’s antitrust authority, the Administrative Council for Economic Defense (CADE), entered that debate. It issued part of its long-awaited decisions on a batch of AI partnership agreements and “acqui-hire” deals between Big Tech firms and AI startups. The practical result was modest. The doctrinal signal was not.
CADE dismissed three cases: NVIDIA/Run, Microsoft/Mistral, and Google/Character.AI. It ordered ex post notification in one case, Microsoft/Inflection, and opened two new administrative proceedings involving Google/Windsurf and Google/Hume AI. One final case, Amazon/Anthropic, had been scheduled for decision at the same May Tribunal session, but Commissioner-Rapporteur José Levi Mello do Amaral Júnior withdrew it from the agenda, citing “notorious facts” that warranted further investigation.
The headline result was prudent. Three cases escaped formal notification, and the one that did not—Microsoft/Inflection—had already been cleared on the merits by the United Kingdom’s Competition and Markets Authority (CMA). But beneath that restrained bottom line, the Tribunal opened two doctrinal doors that deserve careful scrutiny.
The first is substantive. CADE held that “reverse acqui-hires”—bundles of nonexclusive licensing, key-team hiring, and substantial payments—can qualify as “concentration acts” under Article 90, II of Brazil’s Competition Law, even when no shares or direct assets change hands. In plain English, CADE signaled that a deal need not look like a conventional merger to be treated like one.
The second is procedural. CADE invoked its “call-in” power under Article 88, §7º to require ex post notification of a below-threshold deal, even though the agency still lacks clear, predictable criteria for when it will exercise that exceptional authority.
Both moves echo the July 2024 joint statement by the U.S. Justice Department (DOJ), Federal Trade Commission (FTC), European Commission, and the CMA on competition in generative-AI foundation models and AI products. CADE has, in effect, decided to surf the same wave.
This post explains how it got there. It starts with the basics of Brazil’s merger-notification regime, then examines how the Tribunal characterized “acqui-hire” and “reverse acqui-hire” arrangements as “concentration acts.” It then asks what distinguished Microsoft/Inflection from Google/Character.AI on the “call-in” question, summarizes the outcome of each case, and closes with a substantive critique of the killer-acquisition theory CADE has imported.
The question, in the end, is whether Brazil’s Competition Law has the doctrinal board to ride that wave safely.
The Merger Review Before the Merger Review
CADE opened the five APAC proceedings—short for Apuração de Ato de Concentração, or investigation of a concentration act—in the second half of 2024, shortly after the joint statement from the four other leading enforcers.
A brief procedural note helps. APACs are governed by CADE Resolution 24/2019. They are used mainly in gun-jumping cases—that is, cases involving transactions that should have been notified before closing but were not—and to determine, at a threshold level, whether a transaction qualifies as a “merger” under Brazil’s Competition Law.
An APAC does not involve a substantive competition analysis of the transaction. At the end of an APAC, CADE’s Tribunal may order formal notification of the merger, which triggers the Ato de Concentração (AC) procedure. That is the proper vehicle for substantive legal and economic review. Under the AC procedure, CADE’s General Superintendence conducts the competition analysis and, depending on the case, refers it to the Tribunal for a final decision.
That means none of these AI-related proceedings has yet received a meaningful substantive competition review. The reason is simple: an APAC is not built for that job.
Instead, CADE’s Tribunal faced three preliminary questions. First, does each transaction qualify as a “merger” under Brazil’s Competition Law? Second, if so, does it meet Brazil’s mandatory notification thresholds? Third, if it does not meet those thresholds, should CADE nonetheless require formal notification under its “call-in” power?
Even without a full merger analysis, the APAC proceedings surfaced a long list of questions and hypothetical concerns worth examining. They also share one central legal fact: CADE opened them ex officio—on its own initiative—because none of the parties self-notified. And none self-notified because none of the deals met Brazil’s mandatory filing thresholds.
The Thresholds Are the Point
Before turning to the specific cases, it is worth pausing over Brazil’s merger-notification system under Brazilian Competition Law (BCL) Law 12.529/2011.
Under Article 88 of the BCL, a transaction must be notified to CADE before closing only if it first qualifies as a “concentration act” under Article 90. That category includes mergers, acquisitions of control or parts of companies—including tangible or intangible assets—and the contested category of “associative contracts,” which mainly covers joint ventures and similar arrangements.
Article 90’s concept of a “concentration act” is broad by design. Congress did not want merger control to be trapped by the narrower labels of business or contract law. Still, the category does not cover ordinary day-to-day commercial contracts between firms. A supply agreement is not a merger merely because the parties shook hands with gusto.
Nor is classification as a “concentration act” enough to trigger CADE review. The transaction also must meet the revenue thresholds in Article 88. Mandatory notification arises only when both conditions are met: one party’s Brazilian group—typically the acquirer—recorded revenue of at least 750 million Brazilian reais (about $144 million) in the year before the transaction; and the other party’s group—typically the target—recorded revenue of at least 75 million Brazilian reais (about $14.4 million). If those thresholds are not met, the transaction is not subject to mandatory notification.
The thresholds are jurisdictional. Without them—and absent a “call-in” decision—CADE has no authority to review the transaction at all. In that respect, Brazil is closer to the European Union Merger Regulation and its jurisdictional rules than to the U.S. Hart-Scott-Rodino Act regime, which is reporting-based rather than jurisdiction-conferring.
Article 88, §7º creates the exception: CADE’s so-called “call-in” power. The statute provides that, “within one year from the date of consummation,” CADE may require notification of a transaction that otherwise falls below the thresholds. CADE has used that power sparingly. One study covering 2011—when the current BCL took effect—through 2023 found only seven uses. For perspective, CADE’s public dataset records at least 4,358 merger filings from 2015 through 2023 alone. Even that understates the full picture, because it excludes 2011 through 2014.
Until now, the call-in power has been understood as a narrow corrective, reserved for deals whose economic significance slips through the revenue test. The five AI cases test just how far that correction can stretch before it becomes something else.
In each of the four AI deals decided at the Tribunal’s May session, CADE’s General Superintendence issued its opinion in March of this year. It took nearly two years for the Superintendence to conclude that none of the transactions was subject to mandatory notification under Brazil’s merger-control regime. The acquirers—NVIDIA, Microsoft, and Google—comfortably cleared the 750 million Brazilian reais threshold. The targets—Run, Mistral, Character.AI, and Inflection—generated no meaningful revenue in Brazil.
That resolved the second question identified above: The transactions did not meet the mandatory notification thresholds. The Tribunal then turned to the remaining two questions. First, do these transactions qualify as “mergers” under Article 90 of the BCL? Second, as the Article 88, §7º inquiry was reframed, should the targets’ lack of Brazilian turnover be overridden because AI startups may have strategic value?
Those are interesting questions. But especially on CADE’s call-in power, they raise a more basic one: Are these the right cases in which to answer them? And is this the best use of CADE’s resources when no agency has found evidence of actual harm to competition in these deals?
The Amazon/Anthropic case adds another wrinkle. As noted above, CADE postponed that case for further investigation because of “notorious facts” that were not disclosed publicly during the Tribunal session. Why? Which facts or conditions justify different treatment?
Other agencies have not found concrete competitive harm from the Amazon/Anthropic partnership. The FTC and the CMA both examined the deal. The CMA cleared it outright. The FTC’s report—conducted under the Biden administration and approved 5-0—identified potential risks, including lock-in, input foreclosure, and information asymmetry, but stopped short of enforcement action.
When a Hiring Spree Starts Looking Like a Merger
In short, CADE’s Tribunal concluded that acqui-hires and reverse acqui-hires can qualify as mergers under Article 90 of the Brazilian Competition Law and may therefore require mandatory notification when Article 88’s revenue thresholds are met. That conclusion is significant. But the reasoning behind it—and the factual elements needed to turn a bundle of contracts into a “concentration act”—remains somewhat murky.
As Selcukhan Ünekbas has explained, “acquihires are transactions aimed primarily at acquiring a firm’s workforce, rather than its products or other assets. They differ from ‘license-and-hire’ agreements, in which the acquirer also licenses the target’s technology.”
The controversy is not new. It traces at least to a paper by John Coyle and Gregg Polsky published more than a decade ago. So the debate over how Big Tech firms acquire, absorb, or otherwise siphon talent from startups—and what that means for competition law—has been around for a while, at least in the literature.
The newer label, or perhaps the newer lens, is the “reverse acqui-hire.” A traditional acqui-hire involves the direct acquisition of a small startup, with the workforce as the main prize. A reverse acqui-hire tries to reach much the same practical result without formally acquiring the company.
Reverse acqui-hires typically combine three sets of obligations, often spread across several contracts:
(1) some compensation paid to the team—conventionally structured to create incentives for the team to stay with the acquirer for a longer period of time; (2) a non-exclusive license to the startup’s intellectual property; and (3) some kind of waiver from the target company that it will not sue the acquirer for poaching the talent.
Unlike a traditional acqui-hire, the startup generally survives the deal: “In contrast to traditional acquisitions or acqui-hiring, RAH transactions neither involve equity transfer nor full buyout of technology nor full takeover of the people.” The startup usually reshapes its business afterward, but it remains a distinct entity, in some form, from its would-be acquirer.
In Google/Character.AI, Commissioner Camila Alves examined the hallmark features of a reverse acqui-hire and how they bear on Article 90’s definition of a “concentration act.” In her view, the transaction should not be assessed as a pile of disconnected pieces, but as an integrated whole. Licensing intellectual property, hiring part of a technical team, terminating a prior investment, and making a substantial payment to the company might each fall outside Article 90 if viewed in isolation. But when bundled together as “part of a single economic package” (§ 59), CADE may treat the operation as a concentration act under Article 90, II of the BCL—either as the acquisition of “intangible assets” or as the acquisition of a “part of a company” (§ 72).
Alves stressed that this is not a broad rule making every nonexclusive technology license or hiring decision notifiable. Rather, in her view, what matters is the coordinated structuring of those elements into a single negotiated package. Put less delicately: one contract may be a hire, another may be a license, and a third may be a payment. Together, CADE may see a merger wearing a fake mustache.
Commissioner José Levi reached a similar conclusion in Microsoft/Inflection. Microsoft argued that none of the deal’s individual elements—the hiring of Inflection co-founders Mustafa Suleyman and Karen Simonyan, the hiring of a substantial part of Inflection’s team, the licensing arrangements, and the related corporate contracts—amounted on its own to the acquisition of Inflection’s business or assets. José Levi agreed that the pieces should not be viewed in isolation. Taken together, he concluded, they fit Article 90, II.
For José Levi, Microsoft/Inflection matched the reverse acqui-hire pattern described in the literature (§§ 46-51): leading technology firms effectively absorb a startup’s business activity by hiring its key personnel and compensating investors through substantial licensing payments, thereby producing the economic effect of a conventional acquisition outside the formal channels of corporate law. Because these arrangements replicate the economic logic of a conventional acquisition, CADE cannot decline jurisdiction merely because the transaction used an unconventional form.
CADE therefore characterized Google/Character.AI and Microsoft/Inflection as reverse acqui-hire transactions based on the combination of elements in each deal, even though neither involved a direct acquisition of shares or assets.
The other two cases differed. NVIDIA/Run was a straightforward full acquisition of Run by NVIDIA. Microsoft/Mistral involved a €15 million investment in instruments convertible into shares that, upon conversion, would represent 0.31% of Mistral’s equity. Under Articles 9 and 10 of CADE Resolution 33/2022, acquiring more than 5% of a company’s shares—combined with the Article 88 revenue thresholds discussed above—triggers mandatory notification. The acquisition of call options is treated as equivalent to a direct acquisition of shares for merger-control purposes. Those two cases therefore did not raise the acqui-hire questions that drove the analysis in Google/Character.AI and Microsoft/Inflection.
To be sure, Article 90, II of the BCL is exceptionally broad, as noted above, and deliberately so. It provides:
Art. 90. For the purposes of Article 88 of this Law, a concentration act shall be carried out when: […] II—one (1) or more companies acquire, directly or indirectly, by purchase or exchange of stocks, shares, bonds, or securities convertible into stocks or assets, whether tangible or intangible, by contract or by any other means or way, the control or parts of one or more companies.
Given that language, it is defensible to read reverse acqui-hire arrangements—combining personnel hiring, nonexclusive licensing, and monetary payments—as a single concentration act when those elements occur together, even though none would qualify on its own.
The harder question is what, exactly, the contracts must contain before CADE can call the package an acqui-hire and, by extension, a concentration act. Is hiring one person enough, if that person is the chief executive officer? Must the acquirer hire part of the team, a substantial share of the team, or nearly everyone? What obligations must the nonexclusive licensing agreement include? Is an upfront payment necessary? If no compensation changes hands, can the transaction still be an acqui-hire? Must the contracts include a no-sue waiver by the target for talent poaching? And if one element is missing while the others are present, does the transaction still count?
Those questions remain unresolved. That uncertainty is the predictable consequence of defining the concept by combining elements that, standing alone, would not trigger merger review. CADE’s Tribunal will need to refine these factors in future cases if companies are to have the legal certainty they need when deciding whether a transaction qualifies as a merger under Brazilian competition law.
CADE Calls One In, Lets the Others Surf By
With the first question resolved (reverse acqui-hires can qualify as “concentration acts” under Article 90 of the Brazilian Competition Law) and the second settled as well (none of these AI partnership transactions triggered mandatory notification because they failed to meet Article 88’s revenue thresholds) the analysis turns to the third: CADE’s “call-in” power under Article 88, §7º.
In the end, CADE required formal notification only in Microsoft/Inflection. The remaining cases were not called in. Why?
One caveat: The analysis that follows focuses mainly on Microsoft/Inflection and Google/Character.AI, because the opinions in the other cases were not yet public at the time of writing.
Article 88, §7º does not give CADE a roving license to second-guess every transaction—or every “concentration act”—that falls below the mandatory revenue thresholds. As both rapporteurs acknowledge, invoking the provision requires a juízo de conveniência e oportunidade: a discretionary assessment of whether a call-in is convenient and opportune in the specific case.
Commissioner Alves frames the requirement clearly in Google/Character.AI:
The provision (call-in power) functions as an exceptional mechanism for correcting the mandatory notification system. The revenue criteria confer objectivity, predictability, and administrative rationality on merger control, but they do not exhaust all hypotheses of possible competitive relevance. […] This possibility does not authorize the conversion of Article 88, §7º into an ordinary instrument for broad review of every non-notifiable operation. Its application requires caution, specific reasoning, a demonstration of competitive plausibility, and observance of the legal limits. It is an exceptional jurisdiction to be exercised when the circumstances of the case indicate that the ordinary revenue criteria may not have adequately captured the economic relevance of the operation. (§§ 21-22)
Alves reiterates the point in her conclusion: Article 88, §7º must be preserved “as an exceptional instrument for correcting the revenue-based notification system, and not as a mechanism of ordinary use in any atypical operation” (§ 133). On the standard itself, Alves (§§ 21-22 and § 133) and Commissioner José Levi (§ 60) appear to agree: The call-in power is exceptional; the transaction must present a plausible competitive concern; and the Tribunal must show that the ordinary thresholds failed to capture the deal’s economic significance.
So what made Microsoft/Inflection convenient and opportune in a way that Google/Character.AI was not? Reading the two opinions side by side, three distinguishing features emerge—though the line they draw is hardly neon-bright.
The first is the intensity of the team transfer. Alves flags the distinction: “the precedent (the Microsoft/Inflection case analyzed by the CMA in 2024) is not identical. In Inflection, the team transfer was more intense, involving almost the entire team of the company; in Character.AI, Google itself reports that [REDACTED] employees were released, in addition to a non-exclusive license over certain technologies” (§ 68).
José Levi makes a similar point in Microsoft/Inflection, relying on the CMA’s parallel findings: “from the data in the CMA’s decision on the same operation, it is possible to infer that the employees hired by Microsoft corresponded, in fact, to almost the entire team of Inflection” (§ 72). The redacted portion of Camila’s analysis suggests Google hired a meaningful subset—but not the whole team—of Character.AI. The scale of the team transfer therefore appears to have mattered to CADE’s decision to require formal notification in Microsoft/Inflection.
The second distinction is what happened to the target’s product strategy after the deal. This is the point on which José Levi rests most heavily. Before the transaction, he explains, Inflection focused on developing foundation models—general-purpose AI models that can support many downstream applications—and used them as the basis for Pi, its consumer-facing chatbot. After the team responsible for growth in that segment departed, Inflection allegedly shifted its focus to customized generative AI models for corporate clients (§ 75).
José Levi treats that pivot from business-to-consumer to business-to-business as “circumstantial evidence that the operation could have produced a reduction of rivalry in the segment of FM development and consumer-facing chatbots” (§ 76)—a segment where Microsoft operates through Copilot. Character.AI, by contrast, continued operating its consumer-facing chatbot platform after the Google deal. As Alves notes, it “preserved formal independence” and “continued its downstream application” (§ 140).
The third distinction is horizontal overlap. José Levi’s emphasis on Microsoft Copilot and the consumer-facing foundation-models segment frames Microsoft/Inflection as a deal that may have reduced current rivalry, not merely future rivalry, in a market where the acquirer already competes (§ 77). Alves’ analysis of Google/Character.AI leans more heavily on potential competition and innovation trajectory—the idea that Character.AI might have become a more significant rival in the future (§§ 81-82, 138-139). Whether the Tribunal believes this distinction matters under Brazilian Competition Law remains unclear.
Despite these differences, Alves is explicit that the dismissal of Google/Character.AI rests not only on the absence of competitive plausibility, but also on timing and proportionality. She writes:
I understand that it is neither convenient nor opportune to determine the notification of the operation. This conclusion does not derive exclusively from the time elapsed since consummation, although that element is relevant in the exceptional design of Article 88, §7º. It also weighs the convenience of an ex post requirement at this moment, the already advanced stage of implementation of the operation, the absence of consolidated administrative criteria for reverse acqui-hires, the proportionality of the measure, and the possibility of directing institutional action toward more recent analogous operations, in a specific and timely procedure. (§ 141).
José Levi acknowledges similar factors in Microsoft/Inflection but reaches the opposite conclusion. He leans instead on the killer-acquisition and talent-hoarding literature (§§ 65-69), the intensity of the team transfer, and the alleged reduction of rivalry in the consumer-facing foundation-models segment.
A skeptical reader is entitled to find the resulting line less than crystalline. Both deals were announced and consummated in 2024. Both involved nonexclusive licenses, substantial payments, and the hiring of part—or nearly all—of a target’s technical team. Both were treated as “concentration acts” under Article 90, II. And under Alves’ opinion, both were at least plausibly relevant from a competition standpoint.
The different outcomes therefore seem to rest less on a clear legal rule than on a bundle of factual gradations and ex post implications: how many team members were hired, what happened next at the target, and how closely the target’s business overlapped with the acquirer’s existing products.
That may be enough to decide these cases. It is less helpful for the next company trying to plan around CADE’s call-in power.
Based on the decisions so far, it remains difficult to articulate a clear legal test for when CADE will require formal notification. For companies assessing whether their own license-and-hire arrangements may attract scrutiny, the emerging criteria remain opaque. CADE’s Tribunal will need to clarify them in future cases—and it will soon have the chance to do so in Google/Windsurf and Google/Hume AI.
CADE Catches the AI Wave—Now Comes the Hard Part
After this brief tour of the cases, the outcomes were as follows.
In NVIDIA/Run, Commissioner-Rapporteur Carlos Jacques Vieira concluded that Run does not operate in Brazil, that the transaction does not produce effects in Brazil, that the parties lack market power in the country, and that there was no harm—or risk of harm—to competition sufficient to justify invoking Article 88, §7º. The European Commission reviewed the same case in December 2024 and approved it, finding that it raised no serious doubts and would not harm consumers. Like the other cases discussed below, NVIDIA/Run was decided unanimously by the four current members of CADE’s Tribunal.
In Microsoft/Mistral, Commissioner-Rapporteur José Levi found that Microsoft’s stake in Mistral fell below the 5% threshold, that the partnership agreement did not confer control, and that there was no evidence of competitive harm in the relevant market. He therefore dismissed the case and declined to require formal notification under Article 88, §7º. That outcome tracks the analysis and predictions advanced by Dirk Auer and Mario Zúñiga when the investigations were first opened. The CMA had reviewed the same transaction in May 2024 and similarly concluded that, given the deal’s specific features, “the CMA therefore does not have jurisdiction to review the Partnership in its current form.”
In Google/Character.AI, Commissioner-Rapporteur Alves voted to dismiss but wrote a substantial opinion on AI partnerships and acqui-hiring agreements. As discussed above, Alves treated the combined deal—a nonexclusive technology license, the “release” of part of Character.AI’s technical team for Google to hire, the cancellation of a prior investment, and a payment of roughly $2.7 billion—as a “concentration act” under Article 90, II of the BCL. She nonetheless declined to require formal notification after weighing the costs and benefits of an ex post filing requirement.
In the same decision, Alves ordered CADE’s General Superintendence to open two new APAC procedures involving the more recent Google/Windsurf and Google/Hume AI cases. CADE’s debate over whether acqui-hiring arrangements may qualify as concentration acts under Brazilian merger-control law is therefore very much alive.
Finally, in Microsoft/Inflection, José Levi required formal notification. As discussed above, he treated the transaction as following the same basic template as Google/Character.AI: the hiring of nearly the startup’s entire team, combined with a nonexclusive license and a substantial upfront payment. He found that package to be a “concentration act” under Article 90, II of the BCL. Microsoft must now formally notify the transaction under Article 88, §7º within 30 days.
CADE will therefore have to assess the transaction’s substantive effects, including the relevant market, the degree of market concentration, and the potential for unilateral and coordinated effects. It will also likely have to confront, at least to some extent, the “killer acquisition” theory of harm—the claim that an incumbent buys, absorbs, or neutralizes a nascent rival to prevent future competition.
That is notable because CADE required formal notification even though the CMA had already cleared Microsoft/Inflection on the merits in September 2024, finding no realistic prospect of a substantial lessening of competition. Microsoft/Inflection will thus become the first AI partnership agreement to receive full review under CADE’s regular Ato de Concentração merger-control procedure.
CADE’s first wave of AI partnership decisions produced a restrained practical outcome but an ambitious doctrinal framework. The Tribunal expanded the interpretation of concentration acts to encompass reverse acqui-hires and signaled a greater willingness to use its exceptional call-in power for transactions involving AI startups.
Those uncertainties are not merely procedural. They reflect a deeper substantive debate now emerging internationally and increasingly shaping Brazilian competition law. The next question is whether the theories driving this enforcement agenda—especially the theory of digital killer acquisitions—rest on solid empirical ground.
That question is the focus of the second article in this series. For now, CADE has paddled into the AI wave. The harder part is proving it can steer.
