Break Up Nvidia

A strategy of flouting the law, putting China first, and flooding the swamp deserves a harsh response.

As a business success story, Nvidia’s is hard to top. Already worth $1 trillion in 2023, its share price has quadrupled in the past two years, adding roughly $3 trillion of market capitalization and making it the world’s most valuable corporation. These gains have been driven by an increase in net income from $4 billion to more than $70 billion, as profit margins expanded from 16% to 56% and profit per employee rose from less than $200,000 to more than $2,000,000. The Wall Street Journal refers to the company’s eye-watering price markups as the “Nvidia tax.”

Of course, such returns are consistent with a monopoly rather than free-market competition, and no one questions that Nvidia has monopolized the market for advanced semiconductors known as GPUs that are critical to training and running AI models: Estimates place its market share above 90% for GPUs in data centers. But that focus on tangible hardware undersells the broader dominance that the company has asserted over the entire AI ecosystem, including its CUDA development platform and NVLink technology for connecting clusters of chips.

Indeed, the chips themselves are not the core of Nvidia’s advantage. Manufacturing its product is perhaps the most sophisticated technological feat in human history, but not one that Nvidia performs. As a “fabless” semiconductor firm, it only designs the chips, then contracts Taiwan Semiconductor Manufacturing Co. (TSMC) to do the extraordinarily difficult and capital-intensive task of etching more than 200 billion transistors onto each shard of silicon.

On technical benchmarks, those chips tend to perform at the top of the field, but not to an extent where they are the only viable option for customers or will never be caught. For instance, GPUs from fabless competitor Advanced Micro Devices (AMD) “are convincing Nvidia alternatives,” according to a feature last year in Spectrum, the magazine of the Institute of Electrical and Electronics Engineers. The problem for AMD and other would-be competitors is that the “software ecosystem can’t rival Nvidia’s CUDA.” Google and Amazon have developed chips of their own, while Microsoft and OpenAI, among others, are racing to do the same.

Monopoly, monopolists are quick to remind us, is not itself an antitrust violation. You can just be minding your own business, running the world’s best lemonade stand, and everyone wants to buy their lemonade from you. No harm, no foul. But monopoly and its attendant market power frustrate the theoretical assumptions for how free markets should operate, instead creating the means, the motive, and the opportunity for anti-competitive practices that maximize one firm’s profits at the expense of investment, innovation, and consumer welfare. For instance, a dominant firm can insist that customers who want access to its products must use its products exclusively, or use “tying agreements” to force customers who want one product to also buy another. Prospects for such abuse are heightened when supply of a product is constrained, which is the case for AI chips, leaving the seller with the power to decide who receives the limited inventory.

Both competitors and customers have accused Nvidia of employing these practices in the United States and the federal government opened an investigation last year. An in-depth report by The Information cited “allegations that Nvidia has threatened to punish customers who also buy products from its competitors” by raising prices or restricting supply, and also that it had “pressured some of its chip customers, including cloud providers that rent Nvidia-powered servers to app developers, to buy additional Nvidia products.” Other concerns center on strategies that the company has allegedly employed to preclude customers from switching to other suppliers, through acquisition of key software tools and insistence on Nvidia-specific server racks. And more than a few eyebrows went up when Nvidia entered the cloud services market. “In what is perhaps a testament to Nvidia’s market power,” the Wall Street Journal observed dryly, existing cloud providers (who happen to be most desperate for access to AI chips) helped it develop its competing service.

Nvidia’s efforts at corrupting American politics are at least as concerning. A corporation with its level of economic dominance still remains subject to the nation’s laws, guaranteeing that it will pursue profit in a manner that also advances the common good and checking any abuses. But when market power becomes political power used to release the constraints on the market power, a vicious cycle begins and must be interrupted.

The policy of the United States across successive presidential administrations, Democrat and Republican, has been to deny China access to advanced AI chips that it cannot produce itself and that would not only strengthen its economic competitiveness but also accelerate the modernization of its military. Thanks to this policy, the U.S. share of global AI computing power has reached 74% while China’s has fallen to 14%. Nvidia does not like this policy, which prevents it from attempting to establish leadership in the Chinese market and earn large profits there. China also dislikes the policy, which prevents it from developing and running leading-edge AI models and preparing its military to defeat America’s.

Nvidia of course has the right to petition the government to redress its grievances, in the language of the First Amendment, or to sic its lobbyists on the Capitol and the White House, in modern parlance. But along with that right, as the beneficiary of and participant in a liberal democracy, Nvidia has the obligation to use facts that are true and arguments that have logical coherence. Its failure to do so, preferring instead to rely on contradictions, outright lies, nonsensical name-calling, and efforts to ruin careers, has raised more than eyebrows—it raises the question of how a political system should treat a company that thinks it’s a mob boss.

Nvidia and its CEO, Jensen Huang, call export controls a failure, but also claim they are airtight, saying “there's no evidence of AI chip diversion” despite overwhelming evidence to the contrary. On its quarterly conference call last month, Mr. Huang proudly celebrated that “everything’s sold out,” but when legislators proposed the following week that U.S. firms get first dibs on chip sales before anything is routed to China, Nvidia tried to take it back and declared that widely reported supply constraints were “fake news.” It issues statements like, “commercial chip sales pose no more national security threat than uncontrolled products like commercial jetliners and consumer pickup trucks.”

Through it all, and while working to open a research center in Shanghai, Nvidia has called anyone who would limit AI chip exports a “doomer,” attempting to conflate the somewhat fringe and philosophical fear that AI progress will destroy humanity with the rather more pragmatic concern that the United States should not give any economic advantage and military assistance to its primary geopolitical rival, especially at the expense of its own innovators eager to use the same limited supply of chips. David Sacks, the White House’s “AI & Crypto Czar,” has echoed the “doomer” line and even encouraged a pressure campaign to get opponents fired.

The program here is not to advocate or persuade but to leverage and threaten. Misbehaving with impunity is itself an assertion of power, intended to demonstrate that the rules do not apply. Note also the silence from Nvidia customers as Congress considered legislation that would put them ahead of China in line to buy chips—a policy in which they would have an enormous stake. Either these firms no longer care about buying as many AI chips as they can get their hands on, they’ve decided to disavow political advocacy on policy issues of particular concern, or they are terrified to say anything that might antagonize their supplier. One of these explanations seems more likely than the others.

Should Nvidia be broken up? Probably. Its excessive concentration of power is exactly what antitrust law intends to combat, and its valuation suggests that investors don’t expect its position to change any time soon. In many cases, a large and dominant firm at least has countervailing benefits to consider: for instance, it may rely upon economies of scale, network effects from having as many users as possible on the same platform, or large profits to reinvest in capital-intensive pursuits. Nvidia has none of these—if anything, its emergent status as a money-printing machine is more likely to foster complacency and rent-seeking than a desperate desire to win through innovation.

Three separate Nvidias starting with the same intellectual property and competing to gain market share would be more innovative and more likely to pursue open standards that would allow customers to integrate chips and software in whatever combinations they might choose. All three would have more than sufficient cash flow to fund any investment needs. None could dictate how the market operates.

Short of a breakup, which takes time, greater scrutiny through investigations, hearings, and reporting can also be an end unto itself. Thus far, bad behavior has been rewarded, prompting more. If bad behavior instead bore consequences, the incentives would quickly look quite different. With genuinely better conduct, a breakup might be avoided.

Thus far in his second term, President Trump has not hesitated to use federal power to confront institutions that undermine the national interest. An agenda like Nvidia’s—flouting the law, putting China first, and flooding the swamp—is an appropriate candidate for a harsh response. Invitations to meet should be postmarked from the Department of Justice, not the Oval Office.