Why a licensing agreement feels like an acquisition

There's a common scenario in labs. A student writes half a set of training code, and a senior colleague says, "Let me borrow it for now, I'll give you credit, but no transfer of repository ownership." Two weeks later, the senior poaches the core authors. Although the repo nominally remains open source, maintenance, interfaces, and scheduling all belong to the new employer. You ask if this counts as taking it away, and people hesitate because there was no acquisition signed and no equity changed. But deep down, you know the thing has already changed hands.

These past two days, watching the deal between Nvidia and Groq, this exact scenario popped into my head. News reports say the US Department of Justice is investigating whether Nvidia circumvented antitrust review through a licensing agreement. On the surface, this deal doesn't look like a traditional acquisition. Groq calls it a "non-exclusive license agreement," allowing Nvidia to use its custom chips for AI tasks. But another detail is glaring. Groq CEO Jonathan Ross and COO Sunny Madra went to Nvidia. The contract didn't buy the company, but the people came over first.

Antitrust law used to focus on equity, boards of directors, and voting rights. How much stock you bought, whether you controlled decision-making, whether you absorbed competitors into your financial statements—these could all be calculated. But the most valuable assets of an AI chip company aren't necessarily factories and trademarks; they are architectural understanding, compilers, customer adaptation, inference system tuning, and the group of people who know where the pitfalls are. Technology can be licensed, teams can join, and customers can be gradually redirected. If a startup's key people and critical capabilities are taken by a big tech firm, what's left?

I've recently been doing reinforcement learning experiments, using PyTorch for about a month, and messing with GPUs throughout this month. Domestic cards like Biren, MetaX, and Moore Threads, I only started trying these past few days. Once running, I feel that chips aren't just a single card issue. You buy a card, but you also need drivers, operator libraries, framework adaptation, communication stacks, and debugging tools. Customers won't just ask about peak compute power; they'll ask if this model can run, if that inference service can migrate, and if anyone looks at error logs. Ecosystems are often built on engineers' patience and trust.

So these types of deals are hard to investigate because the contract structure is made too cleverly. One company hasn't bought another; it just took a "non-exclusive" license. Sounds gentle—everyone can continue using it, no one gets swallowed. But if the technology license comes with core talent flow, long-term engineering collaboration, and customer migration paths, then the three words "non-exclusive" might not reassure people.

Such arrangements typically allow avoiding automatic antitrust reviews implemented by governments for traditional M&A transactions.

This sentence is key. Reviewers need to see if market control has transferred, not just stare at the word "acquisition." If a company can absorb a competitor's most critical innovations into its own platform without triggering merger filings, it has found a faster expansion path. Later entrants wanting to make independent products will face huge pressure. Because big firms don't need to slowly wait for them to grow; they can rent technology first, acquire the team next, and finally fill the market position.

I previously wrote about domestic GPUs, thinking we shouldn't just look at valuations and benchmarks, but see who can stably sell cards into server rooms. Now looking at the Nvidia-Groq deal, my thoughts have advanced a bit further. Compute competition is simultaneously hardware competition and organizational absorption capability competition. Whoever can keep innovation teams, whoever can reduce developer pitfalls, can stick to the market. If antitrust only stares at equity ratios, it might miss actions that change the competitive landscape.

Of course, we can't apply a blanket ban. Technology licensing itself isn't a sin. Many companies license patents, collaborate on chips, and poach talent. Business society cannot demand big firms only build everything from scratch.

The problem lies in boundaries. If a license causes the licensor to lose key resources for independent competition, and gives the licensee near-acquisition-level actual control, it shouldn't continue hiding outside merger reviews. Senators demanding Nvidia provide more details is a reasonable direction. Everyone wants to know how technology, talent, customers, and capacity flow beyond the contract name.

What worries me more is something else. If AI chip startups realize it's too hard to become independent giants, they might choose to be licensed earlier. Founders might stop insisting on selling servers, building public clouds, and creating developer ecosystems, instead packaging architecture and teams for a licensing fee and a big-tech offer. This way, the company survives well short-term, and investors have an exit. But the market loses a batch of people willing to grind through the full stack. Eventually, the industry looks highly collaborative, but in reality, only a few platforms define the rules.

I've only just started looking at regulations like the EU AI Act these past few days, so I dare not misuse terminology. But there's a similar feeling. Regulation must look at results, but also structure and risk. An AI system not causing accidents doesn't mean it can be ignored. A chip deal not forming a traditional merger doesn't mean competitive impact doesn't exist. Regulators need to answer whether, in the next three years, there will still be another company able to independently produce next-generation inference chips and convince customers to leave Nvidia.

This matter may not reach a quick conclusion. Contract clauses, talent flow, license scope, customer migration—each item can be argued over for a long time. But one point seems to have emerged. The old ruler of antitrust is a bit short for measuring AI chip deals. Equity is hard control, licensing is soft control, and talent is covert control. If regulators only recognize the first type, they will miss the latter two.