Nvidia’s board’s premature confidence entrenches Jensen Huang before the conflict review ends.
No disclosed evidence currently shows board corruption, bribery, or Jensen Huang’s participation in espionage. The record supports a narrower but serious governance indictment: the board has publicly blessed its founder while the independent investigation into a possible intelligence and conflict-of-interest problem remains open, and BBC sources said Huang’s removal appeared unlikely because substantial financial interests were at stake and no clear successor was available. The inference is not that the board has taken a bribe. It is that succession and conflict management have been made less independent than the review will require.
On Sunday, 8 November 2026, after New York Times reporting that US officials had concluded Alice Huang was recruited by China’s top intelligence agency and that Jensen Huang had maintained a romantic relationship with her, Nvidia’s board opened an independent investigation and said it had “confidence in Jensen and his leadership team.”
Alice Huang was a senior executive at Singapore-based Megaspeed, a data-centre company suspected of helping divert Nvidia chips into China; Nvidia says that concern was investigated and debunked. Surveillance reportedly surfaced communications between the two, and one American participant was later identified as Jensen Huang through “unmasking,” an identification process that attaches a name to an otherwise concealed account. The reporting also describes Huang’s admission of infidelity.
It is fair to call that a personal failure rather than proof of espionage. It is also fair that a man’s marriage is not by itself a policy matter, and I have no standing to say a word about it. The trouble is that a chief executive’s personal conduct can create a material governance and national-security risk. Deciding whether an independent conflict review exists, what records it can inspect, and whether Huang should remain central while it proceeds is not a private marital question. It is a board responsibility.
Nvidia says it knew of none of the intelligence concerns and put its position this way: “We trust that if we were the target of a genuine threat, the US government would notify us.” That is not a security control; for this conflict, it is the absence of one. A company selling chips under national-security export controls does not need to duplicate the intelligence service. It does need a mechanism for flagging a senior executive’s relationship with someone whom US officials believe was recruited to spy for China, and for reviewing whether that relationship creates a conflict. That mechanism must sit outside the chief executive’s chain of command. Otherwise, the person being reviewed becomes the review process.
The disclosed institutional facts are narrower than the conclusion one might want to draw. The board has called the review independent, has not announced an adverse finding, and has paired the investigation with a confidence statement. It has not yet said who will decide conflicts, what records the reviewers may obtain, whether Huang’s access to sensitive policy work will be restricted, or when the succession question will be resolved. BBC reporting that no clear successor is available is not proof that Nvidia has no internal succession bench. It does show how much weight the board places on continuity, and both of Huang’s children work for Nvidia.
Keeping the chief executive because replacing him is inconvenient is risk management performed backwards. Large stakes do not make a director less entitled to ask hard questions. They make the questions more necessary. A probe is a process; governance is what happens after it finds something.
The technical structure shows why succession at Nvidia is a market-structure question, not merely an executive-suite appointment. Nvidia’s advantage is not merely faster silicon. CUDA, its proprietary programming platform and toolchain, makes the hardware easier to use once a customer has organised code, libraries, debugging tools, and cloud deployment around it. Switching to another accelerator can then require migration of the operating environment around the chip; it is not simply a procurement form with a new vendor’s name on it.
Nvidia’s CUDA programming guide describes a platform-specific compiler, memory model, and collection of libraries. The compiler turns source code into instructions for a particular processor. The memory model tells software how it may use that processor’s memory. The libraries provide ready-made routines. Different processors therefore do not merely require different boxes. They require different translation software, different rules for using memory, and different ready-made routines.
AMD ships a HIPIFY conversion tool for translating CUDA source code to its own platform. That helps, but it does not reproduce Nvidia’s accumulated libraries, tools, code, and tuning for free. A competing accelerator is not necessarily a drop-in production input. Teams may have to alter source code, rebuild dependencies, retest workloads, and reproduce their tuning before they can demonstrate comparable performance.
The hardware side is contestable. AMD sells accelerators, and Google sells tensor processing unit cloud services built around its own custom chip. The harder gate is the cost of switching. Nvidia does not need every AI dollar to pass through its toll. It needs its software to remain the least troublesome path for organisations whose existing code and accumulated expertise make migration expensive. Market-share analysis alone therefore misses part of the rent. The rent includes the cost of switching away.
At the fabrication layer, Taiwan Semiconductor Manufacturing Company makes more than 90 per cent of the world’s most advanced chips, according to the BBC report. The AI economy therefore has a narrow fabrication gate on one side and a dominant software interface on the other. Nvidia’s leverage comes from controlling the latter, not from owning the former.
Stephen Witt’s biography Thinking Machine supplies the history and the image. In 2012, when machine-learning researchers began using Nvidia’s chips, Huang recognised that the field would become a massive industry and directed enormous resources toward building the hardware market around it. Nvidia had been founded in 1993 to make graphics chips for games. In the early years there were no customers for the new machine-learning hardware; the first was OpenAI, then a company of thirty people, to which Huang gave the equipment. AlexNet—the first real neural network of the modern era, in Witt’s account—was trained on two Nvidia cards in a bedroom. The modern equivalent uses between two hundred thousand and three hundred thousand cards in a building full of computers. Nvidia became the default toll collector on much of the buildout, although custom chips and competing accelerators remain outside that collector.
Witt’s summary is technically apt: “Jensen runs a tollbooth. All of that goes through him.” The toll is not merely the price of a chip. It includes the cost of leaving its environment.
In “Amazon’s Antitrust Paradox”, Lina Khan showed how control over a critical point in a transaction can shift costs onto sellers and reduce their competitive options in ways a simple price comparison misses. Here the critical point is the accelerator and the software environment surrounding it. The parties that cannot route around it without undertaking a migration project include university laboratories, hospitals, national laboratories, smaller software firms, and provincial governments that need compute but cannot treat the proprietary stack as interchangeable with every competing chip.
Who receives the benefit? Nvidia was worth about $4.5 trillion when this story broke, and Huang’s estimated fortune was $155 billion. A stock price is how a toll is made of: the valuation is a claim on future revenue from an industry whose companies cannot always substitute one accelerator platform for another.
The pattern is old enough that I do not need a theory of technology to explain it. My father was a journeyman millwright at Manitoba Rolling Mills, thirty years on the bar mill; when Gerdau bought the mill in 1995 he kept his job while several of my uncles did not, and he retired in 2011 on a much-reduced pension. MRM had maintenance crews and transferable knowledge when ownership changed. It did not need a founder’s biography to continue.
It is a warning, not a finding about Nvidia: the inheritance of extraction has changed its interface—from steel bars to CUDA and cloud contracts—but not its governance trick. Make one person the centre, then call dependence devotion.
Harold Innis’s staples economy is useful here, but the staple is not grain. A modern AI system still requires scarce industrial capacity, and the decisive control lies less in the object than in the communications and administrative system deciding where it travels, who may buy it, and whose software can direct it. Nvidia’s position resembles the railway, the port, or the refinery: its power grows as other firms must move their work through controlled interfaces. Export controls are one of those interfaces. They do not merely restrict trade; they determine which buyers can reach the tollbooth and which are denied the market.
Since 2022, the United States has used various export controls to keep advanced technology out of China, and the BBC reported that those controls have largely locked Nvidia out of what it described as the world’s most lucrative market. Huang once said Nvidia’s Chinese share fell from around 95 per cent to effectively zero. At the same time, he shuttled between Washington and Beijing seeking an approved route back into that market and, the BBC reported, shaped US policy towards China.
Whether every official decision served Nvidia is the wrong question. The disclosed structural fact is enough: Nvidia’s chief executive retained privileged access to the rulemaking that excluded his company from the Chinese market. That permits influence without bribery, and no bribe is alleged. A chief executive whose private relationships now raise an intelligence inquiry while he is helping shape the rules governing access to the Chinese market makes the independence of that review a governance question rather than a procedural ornament.
At a recent White House AI summit attended by forty-two technology executives, Donald Trump was flanked by Elon Musk and Huang. Asked about the investigation, Trump said: “I just know he’s a fine person, but you never know what’s going on,” and added that Huang had “done an amazing job for our country.” That was a presidential endorsement, not a determination about the intelligence allegations.
Nobody has to allege a favour to see the structure, and I would rather not allege one. Set the record side by side without any theory attached: Nvidia says it knew of no threat; its board has opened an independent investigation and expressed confidence in Huang; Trump has praised him; the intelligence inquiry has not closed; and congressional committees are asking Huang to testify. These are distinct positions, not one institutional exoneration. Pretending otherwise would erase the unresolved matter rather than analyse it.
It is also true that China gaining an AI lead could strengthen Beijing’s military and surveillance capabilities, and that Nvidia’s own conduct cannot make a legitimate security control disappear. National security is not an exemption from competition analysis; it is the reason competition analysis matters.
Export controls are a sovereign instrument and a legitimate one. Advanced compute is a strategic asset, the American advantage was built on companies such as Nvidia working with TSMC, and the fear that China takes the lead in AI and converts it into military and surveillance capability is not hysteria. Nvidia’s own record in this passage is a good one: it built the market that made modern AI possible at all, out of nothing, on a bet nobody else made, and gave hardware to a company of thirty people when there were no customers. It has also borne the cost of its political closeness, locked out of China quarter after quarter despite Huang’s access to the administration and his role in shaping its China policy. A locked-out firm is entitled to complain, and to complain loudly.
All of that is true, and none of it reaches the point, which is not about export controls and not about the marriage. Nvidia now dominates a layer of the AI hardware market and owns the software environment through which much of that hardware is made usable. The same man heads both.
Cory Doctorow’s account of how platforms die identifies four constraints that once kept a gatekeeper honest: competition, regulation, interoperability and self-help, and organised labour. Here competition exists at the hardware level, but switching costs blunt it. Regulation has been most active on export controls and national security, not on accountability for Nvidia’s proprietary software environment. Self-help is available in principle, but migration away from CUDA is expensive. Labour has less leverage after layoffs. The four constraints are not absent; they are poorly aligned with this particular gate. The result is not a market without alternatives. It is a market in which the alternatives do not remove the software toll.
Doctorow’s chokepoint capitalism describes firms positioned between creators and the people who want their work, able to extract concessions from both sides. Nvidia sits on a different part of the chain, but the economic geometry is familiar. Developers need its tools, cloud operators need its accelerators, the US government needs a domestic supplier capable of serving it, and each side has an incentive to preserve the gate rather than weaken it. A better chip without a usable software environment does not merely cost more to adopt; it may not be deployable at scale without rebuilding the surrounding stack.
Tim Wu’s choke-point analysis is useful for the same reason. His concern is not a company charging a high price for a product with many substitutes. It is control of an indispensable passage in a larger system. Nvidia’s relation to AI is now both product and passage: it sells the accelerator and supplies the environment in which the accelerator becomes useful. The antitrust question is therefore not whether customers would prefer a cheaper Nvidia chip. They might. It is whether they can build and run without Nvidia’s system, and whether exclusionary conduct has preserved the answer to that question. The 2020 House Antitrust Subcommittee’s Investigation of Competition in Digital Markets examined cloud concentration and digital-infrastructure power; the chips beneath those clouds make that concern less abstract.
The board remedy and the market remedy are distinct, and neither requires a verdict on what Alice Huang passed. Nvidia’s board should appoint an independent special committee outside Huang’s chain of command, give it records access and authority to review disclosures, restrict sensitive access, recommend action, require a succession plan, and publish what national security permits. Huang’s role in export-policy advocacy should not remain solely under his control while the review proceeds. A probe is a process; governance is what happens after it finds something. Nvidia does not need a morality play about its founder.
The market remedy is portability. Make it a condition of federal research money, public compute purchases, and government procurement. Require supported vendors to expose standard interchange formats for trained models, publish documented compiler interfaces—the software layer that translates code for particular chips—release portable source and build recipes, demonstrate equivalent performance under common benchmarks, and pass portability tests before an award. Open an antitrust file on the software tie, governed by an independent committee with the power to order a remedy rather than merely express confidence in the incumbent.
Canada already understands the underlying duty, though fibre is not a perfect analogy for accelerators. Canada’s local-loop incumbents—Bell, Rogers and Telus—were made subject to wholesale-access obligations because the regulator did not treat ownership of the copper as permission to deny rivals access. The recent Canadian fibre file shows what implementation requires: CRTC Telecom Decision 2023-358 addressed temporary wholesale fibre-to-the-premises access, Telecom Order 2024-261 set interim rates, and Telecom Regulatory Policy 2024-180, issued on 13 August 2024, set the direction for wholesale fibre access.
Fibre is not a perfect analogy for accelerators, but it isolates the duty worth writing. Ottawa should condition public compute purchases on portability because public money can buy access to Nvidia’s chips without buying the capacity to leave them.
The Jen-Hsun and Lori Huang Foundation, founded in 2007, is worth an estimated more than nine billion dollars. Its assets surged because Nvidia’s shares surged. That arithmetic needs no corruption to work.
Congressional committees have asked the chief executive to testify next week. That is legitimate oversight, but testimony cannot substitute for an independent board committee with authority to restrict access and recommend removal. A legislature cannot run Nvidia’s succession plan by schedule.
That is the part of the process with a deadline attached, and deadlines are the only part of regulatory procedure the regulated reliably respect. It is worth doing anyway, for a reason that has nothing to do with scandal: somebody ought to have to say the toll out loud, in a room, on the record.
The board must be able to answer three questions without recourse to the founder’s biography: whether the conflict review is genuinely independent, whether a succession plan exists, and whether customers can leave Nvidia’s software environment on workable terms. The record currently supports entrenchment, not corruption, bribery, or espionage. It also supports a toll on the rest of the AI economy.
A toll is not a price. It is a payment for permission, and nobody has audited the permission. A tollbooth with one key is not a monument; it is a bottleneck waiting for a breakdown.