NVIDIA defined every layer of the AI stack at GTC 2026 while an active DOJ antitrust investigation examines whether the loyalty programs, bundling, and acquisitions that enforce that dominance constitute anticompetitive behavior. Six days before Jensen's keynote, Michael Burry called the company's practices 'mafia-like.' The game-maker meets the referee.
On March 16, Jensen Huang stood on stage at GTC 2026 in San Jose and spent two hours defining every layer of the AI stack. NVFP4 numerical format. Vera Rubin at three hundred thirty-six billion transistors. Groq inference integration. A full-stack vertical integration so comprehensive that The Definition described it as one company attempting to own the conversation between silicon and software at every level.
What Jensen did not mention: the Department of Justice has an active antitrust investigation into whether NVIDIA's dominance is maintained through practices that violate federal competition law.
The Investigation
The DOJ opened its antitrust probe in August 2024, after complaints from AMD, Intel, and other rivals. By September 2024, the department had escalated from informal questionnaires to legally binding subpoenas — civil investigative demands issued to NVIDIA and third-party companies. NVIDIA's stock lost two hundred eighty billion dollars in market value in a single trading session on the news.
The investigation centers on three categories of alleged conduct.
First, loyalty programs. Regulators are examining whether NVIDIA gives preferential supply and pricing to customers who use its technology exclusively — and penalizes those who don't. Microsoft reportedly expressed concern that purchasing networking equipment from competitors would cause NVIDIA to delay GPU shipments. In a market where a single quarter's worth of chip access can determine a company's competitive position, delayed shipping is not an inconvenience. It is an existential threat.
Second, bundling. Customers who buy NVIDIA's GPUs alongside its networking cables and software stacks receive lower prices on each product. Bundling is not per se illegal. But when a company with ninety percent market share in AI training chips uses that position to foreclose competition in adjacent markets — networking, software, orchestration — antitrust law has a word for it.
Third, the CUDA ecosystem. NVIDIA's proprietary programming framework has become the default language of AI development. Rewriting CUDA code for AMD's ROCm costs months of engineering time and hundreds of thousands of dollars. The switching cost is not a feature of NVIDIA's product. It is the moat. Google and Meta are collaborating on TorchTPU as an open-source alternative, but the installed base of CUDA-dependent code grows faster than any alternative can erode it.
The Acquisition
In April 2024, NVIDIA announced the acquisition of Run:ai, an Israeli startup that makes GPU orchestration software — the layer that manages how AI workloads are distributed across GPU clusters. The purchase price was approximately seven hundred million dollars.
The concern was structural. Run:ai sits at the scheduling layer — the software that decides which jobs run on which chips, when, and how efficiently. If NVIDIA owns the chip, the networking, the programming framework, and the scheduler, the entire compute stack reports to one company. The acquisition would complete the vertical integration.
The European Commission approved the deal without conditions in December 2024, finding no competition concerns in the EEA. The DOJ also cleared it. Run:ai closed on December 30, 2024. As a concession, NVIDIA committed to open-sourcing Run:ai's software, allowing AMD and Intel to adapt it for their hardware. Whether an open-source release backed by a company with every incentive to prioritize its own chips constitutes genuine competitive access is a question the market will answer faster than the courts.
The Allegations
Six days before GTC, Michael Burry — the investor who shorted the housing market before the 2008 financial crisis — publicly accused NVIDIA of anticompetitive behavior. On March 10, Burry alleged that NVIDIA paid a one hundred fifty million dollar deposit to a site developer to block AMD from securing a hardware contract tied to the OpenAI-Oracle Stargate data center project in Abilene, Texas. He called NVIDIA's practices "mafia-like" and called for a formal antitrust case.
The allegation is unverified. It comes from Burry's public statements, not from DOJ filings or confirmed reporting. But the timing is significant. GTC is NVIDIA's annual demonstration of market power — the event where the company defines the technological future for thirty-nine thousand attendees. Burry chose the week before that event to make his claim. The message was addressed not to regulators but to the market: the company defining the AI stack is doing so through practices that may not survive legal scrutiny.
Four days before Burry's statement, Jensen Huang announced that NVIDIA was pulling back from investments in OpenAI and Anthropic. He said a thirty-billion-dollar OpenAI investment "might be the last" and described a hundred billion dollars as "probably not in the cards." TechCrunch noted the explanation raised more questions than it answers. Antitrust scholars had already flagged the conflict: NVIDIA's financial interest in OpenAI could create incentives to sell chips on less favorable terms to OpenAI's competitors. The pullback may be commercial. It may also be a company clearing its antitrust exposure before the investigation reaches its next phase.
The Second Front
The DOJ investigation is not the only regulatory pressure NVIDIA faces.
In September 2025, China's State Administration for Market Regulation announced a preliminary finding that NVIDIA violated China's anti-monopoly law in connection with its 2020 acquisition of Mellanox Technologies — the seven-billion-dollar networking deal that gave NVIDIA control of InfiniBand, the high-speed interconnect technology that links GPUs in data center clusters. SAMR found that NVIDIA failed to comply with conditions imposed at the time of the merger, specifically around InfiniBand competition. The investigation continues.
The Chinese case and the American case share a common structural concern: NVIDIA controls the chip, the interconnect, the software framework, and the orchestration layer. Each acquisition extends the stack. Each extension makes the whole more difficult to compete against. The question is whether vertical integration at this scale and in this market constitutes illegal monopoly maintenance or simply reflects a company that builds better products.
The Precedent
The closest historical parallel is United States v. Microsoft Corporation, filed in May 1998.
The similarities are structural, not superficial. Microsoft held over ninety percent of the desktop operating system market. NVIDIA holds eighty to ninety percent of the AI accelerator market, and over ninety percent of the AI training market specifically. Microsoft's moat was the Windows API — the programming interface that made applications Windows-dependent by default. NVIDIA's moat is CUDA — the programming framework that makes AI code NVIDIA-dependent by default. Microsoft bundled Internet Explorer with Windows; NVIDIA bundles chips with networking, software, and orchestration tools.
The 2001 ruling forced Microsoft to open its APIs to competitors. The consent decree permanently altered the competitive landscape — not by breaking Microsoft apart, but by creating interoperability requirements that opened doors for Google, AWS, and the mobile ecosystem. The open Windows APIs made it possible for browsers other than Internet Explorer to compete. The question for NVIDIA is whether a similar requirement — forced interoperability with CUDA, mandatory chip allocation transparency, restrictions on bundling — would produce a similar structural opening in the AI compute market.
One difference is critical. The Microsoft case was filed as a formal lawsuit in May 1998, tried, and decided. The NVIDIA investigation has not reached that stage. As of March 2026, no formal complaint has been filed. The investigation remains in the evidence-gathering phase. Under the current administration, which issued Executive Order 14179 in January 2025 focused on removing barriers to American AI leadership, the political appetite for an antitrust case against America's most valuable semiconductor company is unclear.
Burry himself expressed skepticism that the current administration would pursue a case. The investigation may resolve through settlement, behavioral commitments, or quiet closure. But the investigation's existence — and the specific practices under examination — has already changed the competitive calculus. Jensen's investment pullback, Run:ai's open-sourcing, and the defensive framing of NVIDIA's GTC announcements all suggest a company that knows it is being watched.
The Question
The infrastructure bet series has tracked a six-hundred-fifty-billion-dollar capital cycle built primarily on one company's silicon. The Vertical documented NVIDIA investing four billion dollars in two neocloud companies to extend its reach through the stack. The Definition showed Jensen defining every layer from numerical format to agent software. The Mint showed Alibaba reorganizing its entire corporate structure around the token — a primitive that NVIDIA's hardware manufactures.
The antitrust question cuts through all of it. If NVIDIA's dominance is maintained through loyalty programs that penalize competitors' customers, bundling that forecloses adjacent markets, and an ecosystem moat that costs hundreds of thousands of dollars to cross — then the infrastructure bet is not just a bet on AI demand. It is a bet on whether one company will be permitted to define the terms of that demand.
The Microsoft precedent suggests the answer matters less than the timing. Microsoft remained dominant after the antitrust ruling — it is a three-trillion-dollar company today. But the consent decree opened cracks that became canyons. Google Search, Amazon Web Services, Apple's mobile ecosystem, Chrome — none of these required Microsoft's destruction. They required Microsoft's interoperability. The APIs that the DOJ forced open became the surface area on which the next generation of technology companies was built.
If the DOJ eventually requires CUDA interoperability, chip allocation transparency, or restrictions on bundling, NVIDIA will not be destroyed. It will remain the dominant AI hardware company for years. But the cracks will open. AMD's ROCm would compete on a level playing field for the first time. Custom silicon from Google, Amazon, and Microsoft would have access to the software ecosystem their engineers currently cannot reach without rewriting millions of lines of code. The value chain that today flows through NVIDIA at every layer would develop alternative paths — not because NVIDIA's products are inferior, but because the on-ramps would no longer be controlled by the same company that owns the highway.
Jensen spent two hours today defining the road. The subpoena asks who owns it.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)