Modernize Antitrust for AI Markets

Since 2024, Microsoft, Google, Meta, and Nvidia have spent billions of dollars hiring away AI startups' leaders without buying the companies. Federal enforcers say deals like these are going unreported, even when they eliminate a competitor. Congress should make these deals reportable, let cloud customers leave, stop bottleneck firms from squeezing rivals, and give enforcers fitting remedies.

The Problem

Three companies, Amazon, Microsoft, and Google, sell 63% of the world's cloud computing, the rented capacity on which AI is built.¹ They have invested billions in OpenAI and Anthropic, and the FTC found that the developers committed to spend a large portion of that money on the investor's own cloud.² Integration can cut costs. It can also let whoever controls the chips, cloud, and talent decide which rivals grow.

A newer pattern skips the purchase: hire the founders, license the technology or take a nonvoting stake, and leave the startup standing. Microsoft paid $650 million in the Inflection deal;³ Google, $2.7 billion in its Character.AI deal⁴ and $2.4 billion in Windsurf's;⁵ Meta, $14.3 billion for a nonvoting 49% of Scale AI;⁶ and Nvidia, about $20 billion in its Groq deal.⁷ That is about $40 billion in all.

Britain's competition authority reviewed Microsoft's Inflection deal as a merger, then cleared it.⁸ America's merger enforcers are still asking whether their own rules reach such deals at all.⁹ The law has three blind spots:

1. Deals escape review. The FTC and Justice Department report "an uptick" in unreported deals, including license-and-hire arrangements that leave a startup "intact but not competitively viable."⁹
2. Leaving costs extra. The big three clouds waive data-exit fees for customers who switch, but not for those who also use a second cloud, Britain's regulator found.¹⁰
3. Rivals pay more. Britain also found that Amazon and Google can pay Microsoft more for some software than Microsoft charges some of its own cloud customers.¹¹

Why legislation: Enforcers have the authority; they lack the view. The Clayton Act already bars acquiring "any part of the assets" of another firm where the effect may be substantially to lessen competition,¹² but premerger notice covers only acquisitions of voting securities or assets.¹³ In February 2026, a court vacated the FTC's updated premerger form for failing to show its benefits outweighed its costs;¹⁴ filers are back on what the FTC calls a "nearly 50-year-old" form.¹⁵ Commissioners of both parties voted 5–0 to publish the FTC's AI partnership study.² Companies that control the chips, clouds, and talent of a new industry should win customers on the merits, not buy their way out of competing.

The Solution

A four-step staircase: each step stands alone, and each step up adds obligations for the firms that control AI's bottlenecks. Steps 1 and 2 cover AI chip, cloud, and model providers above existing premerger size thresholds; Steps 3 and 4 reach only firms designated as bottlenecks after notice and judicial review; competition in social media and consumer chatbots is addressed separately.

Step 1 — Report deals that work like mergers. Require premerger notice, at the same dollar threshold as acquisitions ($133.9 million in 2026),¹⁶ for license-and-hire deals, nonvoting stakes, and exclusive compute, investment, and distribution agreements, counting connected agreements together so no one can split a deal to dodge review. The rules already disregard devices used to avoid filing,¹⁷ but enforcers can apply that only to deals they see.

Step 2 — Let customers leave. Require large cloud and model providers to give customers usable exports of their own data, clear termination terms, and migration help, and to charge no more than cost to move data to a rival, even while a customer uses both. The big three clouds already waive exit fees for customers who leave entirely, and Europe bans switching charges from January 2027.¹⁰

Step 3 — Stop the squeeze. For designated firms, prohibit coercive tying, retaliation against customers who use rival chips, clouds, or models, and use of business customers' confidential data against them, allowing security and efficiency defenses where no less restrictive option exists. Where a bottleneck blocks rivals, set narrow interoperability standards (never mandating access to model weights) and require independent technical audits.

Step 4 — Fit the remedy to the harm. Let the Justice Department and the FTC seek access duties, contract changes, or structural relief where evidence supports it, weighing harm to innovation and to workers' choice of employer alongside prices. Fund the technical and economic staff these cases require.

Where to start: Step 1 is the floor; it asks only for notice. Step 2 is the heart: when customers can leave, every provider has to keep earning them.

Administration and enforcement: The FTC and the Justice Department issue notice rules and designation criteria within 18 months, with 12 more months for demanding technical duties. Courts review designations and remedies; state antitrust law and private remedies remain.

Risks and Mitigations

• Punishing useful integration: Owning chips, cloud, and models together can cut costs. That is why conduct rules reach only designated bottlenecks, require proof of harm, and allow efficiency defenses; size alone never triggers divestiture. Enforcers can still misjudge a fast market.
• Chilling startups and talent: Founders need exits and engineers need mobility. Step 1 adds only notice and a short wait; Britain formally opened its Inflection inquiry in July 2024 and cleared the deal in September 2024.⁸ Extra filings still cost time and money.
• Security and rigidity: Mandated interfaces can open security holes, and rules can outlast the bottleneck they target. Limit interfaces to documented needs, require recipient security, and revisit designations periodically; some interoperability will prove too costly.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal

State

What this adds: Existing bills police online platforms (AICOA) or communications data (ACCESS). None makes license-and-hire deals reportable or gives cloud customers a statutory right to leave. This proposal reaches the AI supply chain itself: chips, cloud, models, and the talent deals that consolidate them.