Nvidia's black-market chips, a premier's warning, and the price of a US-China tech split

On 24 June 2026, two data points arrived within hours of each other and, taken together, describe the same contest. The Financial Times reported, via Reuters at 09:40 UTC, that Nvidia accelerators barred from export to China under successive US controls are now selling on Chinese grey markets for roughly twice their US list price. At 08:31 UTC the same day, Hong Kong Free Press carried Premier Li Qiang's call for stronger artificial-intelligence governance so the technology does not "lose control." One fact describes supply; the other describes the political ceiling a Beijing-led AI sector is now trying to build for itself. Neither can be read without the other.

The story is not that chips are expensive in China. The story is that a five-year-old US export-control regime, designed to slow Chinese frontier-AI capability, is now producing the opposite of its stated effect on price. Banned silicon is scarce; scarcity is rent. The rent flows to intermediaries in Shenzhen, to brokers who can move hardware across borders, and to the legitimate Chinese firms willing to pay a premium for guaranteed delivery. Meanwhile, in the same news cycle, China's number-two official is publicly telling domestic AI labs and platforms that the technology has moved faster than the rulebook. The two facts sit inside one policy environment, and that environment is the story.

The price signal

The FT reporting, surfaced by Reuters, is the kind of detail that sounds small and ends up being diagnostic. A piece of Nvidia silicon with no legitimate route into mainland China is now worth more, not less, than the same part sitting in a US cloud warehouse. That is a textbook inversion of what export controls are designed to produce. The intended effect is to widen the capability gap between the sanctioned buyer and the rest of the field. The observed effect, at the consumer end, is the opposite: a black-market premium that funds further procurement and rewards precisely the brokers and resellers the controls were meant to starve of oxygen.

This is not the first time a US technology control has produced a price inversion. The pattern is familiar from advanced lithography, from high-end machine tools, and from earlier rounds of restrictions on Chinese telecoms equipment. What is new is the scale. Frontier AI training runs are now large enough, and hungry enough for memory bandwidth and interconnect, that even a small legal supply of compliant chips cannot keep up with internal Chinese demand. A premium of roughly 100 percent on a banned part is, in that context, an equilibrium signal. It tells the market — and tells Beijing — that legal supply is structurally short.

The structural read is straightforward and worth stating plainly: when a sanction regime raises the price of a controlled good inside the target jurisdiction, it has either set the price floor too low or has not closed enough of the supply routes to matter. Which of those is the operative explanation here is contested. The FT report and the Reuters wire both note the rise but do not, in the surfaced excerpts, attempt to disaggregate. That work is left to the reader.

Beijing's parallel constraint

If the price line describes the supply side, Premier Li's remarks describe the demand side, and the constraint that even a sanctioned Chinese AI sector is now running into. Hong Kong Free Press reported at 08:31 UTC on 24 June that Li, speaking at a high-level meeting in Beijing, urged AI governance frameworks that prevent the technology from "losing control." The phrasing is deliberate. It signals two things at once: that the Chinese state intends to remain the senior partner in any domestic AI deployment, and that senior Chinese policymakers are themselves no longer fully confident that the technology's trajectory is predictable.

That second signal is the more important one. Chinese officials do not, as a rule, voice concern about their own industrial sectors in public unless the concern is real and politically usable. A premier raising governance as a binding constraint is, among other things, an admission that the sector has outrun the rulebook. The same week that black-market chips trade at twice the US sticker, Beijing is telling its own AI firms that the runway is not unlimited and that compliance, data localisation, model-evaluation regimes, and content controls will tighten before they loosen. The two announcements are not contradictory. They are complementary. Supply-side pressure from Washington and demand-side discipline from Beijing describe a single Chinese policy: catch up fast enough to be inside the global frontier, but never so fast that the technology outruns the party's ability to set its terms.

A counter-reading deserves airtime. It is possible to argue that Premier Li's governance push is, in effect, an industrial-policy signal to domestic capital: the state will regulate, but it will regulate in ways that favour Chinese incumbents and lock out foreign competition under the cover of safety. That reading is structurally plausible. It is also the same reading one could apply to EU AI Act enforcement or to US export licensing, where safety language and competitive language often travel in the same sentence. The honest position is that the Chinese governance push and the Western export regime both mix safety claims with competitive intent. Neither side has a monopoly on mixed motives.

The litigation file

A third data point from 23 June sits at the edge of the same picture. Music platform Jamendo filed suit against Nvidia, alleging that its catalogue was used to train AI systems. The complaint itself is a single data point in a much larger litigation wave — suits against OpenAI, Meta, and several smaller model providers have piled up in US and EU courts over the past 18 months — but it has a specific bearing on the China story. The legal pressure on US frontier-AI firms is now thick enough that it changes how they negotiate with Chinese counterparts. A US firm selling into, or licensing into, the Chinese market is also negotiating with a regulator in Beijing that is now explicitly told by its premier to harden the rulebook. Two regulatory regimes, both tightening, both framing themselves as safety regimes, both producing competitive externalities. The Jamendo suit is a reminder that the bottleneck on US AI firms is not only Washington, and the bottleneck on Chinese AI firms is not only Washington either.

What the controls are actually buying

The contested question is what the US export-control regime is actually buying for American policy. The strict answer is: time, and a slower diffusion of frontier weights and architectures into Chinese hands than would otherwise occur. That is real, and the people who designed the controls deserve credit for it. The looser claim — that the controls will keep China permanently behind — is harder to defend on present evidence. Two facts cut against it. First, the price premium on banned parts is large and persistent, which means brokers and intermediaries are well compensated and the supply network is healthy. Second, the Chinese domestic fab and accelerator ecosystem, while not yet at parity with Nvidia's leading edge, has had five years of state direction and subsidy to catch up. SMIC, Huawei's HiSilicon, and a string of well-funded startups have moved the needle on memory bandwidth and interconnect even if they have not closed the leading-edge logic gap. Export controls buy time; they do not, on their own, buy permanent separation.

The counter-narrative from Beijing, heard through state-aligned outlets and echoed by Chinese officials in private, is that the controls are an attempt to lock in American technological primacy indefinitely and to deny China the catch-up it is entitled to under any honest reading of the WTO's technology-transfer principles. That framing is also too neat. The controls are a national-security instrument operating in a sector where national-security claims are unusually elastic. The fair position is somewhere between the two: the controls buy time and they are politically sustainable in Washington for that reason, but they do not, on their own, settle the long-run competitive question.

Stakes over the next twelve months

The trajectory is not hard to read. Over the next year, the US is likely to keep tightening chip-export rules while widening the licensing perimeter for friendly jurisdictions. Beijing is likely to keep tightening AI governance while pouring state capital into domestic accelerator and memory suppliers. Both sides will frame their tightening as safety. Both will be partly right and partly strategic. The black-market premium on banned Nvidia silicon will, in that world, stay positive. The Chinese premier's warning about losing control will, in that world, become more pointed rather than less.

The plausible wild cards sit on two axes. On the litigation axis, a successful plaintiff verdict against a US frontier-AI firm — Jamendo or one of the larger pending suits — would reset how US firms license into the Chinese market and would, ironically, push more training data and more model development into Chinese jurisdictions where the US litigation regime does not reach. On the industry axis, a credible Chinese accelerator at H100-class performance from a domestic fab would change the price signal overnight. Neither has happened yet on the public evidence. Both are plausible within the twelve-month horizon.

The nuance worth preserving is that the sources do not, in the surfaced excerpts, give us a clean read on which of the two readings of Premier Li's remarks is correct. The governance push could be a genuine safety warning, or it could be a competitive shield dressed in safety language, or — most likely — some weighted blend of both. The FT/Reuters reporting on the chip price gives us a number, not a mechanism. The 100 percent premium is consistent with several different stories about how the supply network actually functions. What is not in doubt is that the premium exists and that the premier's remarks were made. Those two facts, on the same day, are themselves the story.

This publication framed the day's three threads as one story rather than three. The Reuters wire, the Hong Kong Free Press report on the premier, and the litigation note are read here as a single policy environment in which supply-side pressure from Washington and demand-side discipline from Beijing are running in parallel — a read that gives equal weight to the safety framing on both sides and to the competitive framing on both sides.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• http://reut.rs/4vygUZa
• http://reut.rs/4w7St4I
• https://x.com/polymarket/status/2067805389526114304