Ting-Wei "Willy" Sun pleaded guilty in a Manhattan federal court on Thursday to four counts, including conspiring to violate US export controls, smuggle goods out of the country, defraud the United States, and obstruct justice. The plea came seven months after prosecutors charged him alongside two other people tied to Super Micro Computer in a scheme to move roughly $2.5 billion worth of AI hardware into China without licenses. A court filing made the plea public on Friday.
The details matter because this is no longer a paperwork case. Prosecutors say Sun worked as a broker and "fixer" inside a scheme that started around October 2023, staging "dummy" servers in December 2025 and then lying to a Commerce Department officer who came to inspect supposedly stored equipment that had already left the country. Nvidia's B200, H100 and H200 GPUs and the servers that carry them were all covered by the restrictions Washington imposed on China over national security concerns. A contractor just told a judge he helped defeat those controls.
What the plea settles, and what it doesn't
Sun's guilty plea resolves his own exposure but not the wider case. Co-defendants Yih-Shyan "Wally" Liaw, who co-founded Super Micro in 1993 and joined its board in 2023, and Ruei-Tsang "Steven" Chang, a former sales manager at the company's Taiwan office, still face their own proceedings. Liaw's lawyer did not immediately respond to a request for comment. Chang's lawyer could not be reached.
Super Micro has kept its distance from all three. The California company said it was not named as a defendant in the indictment and that the case had no impact on its business operations, and it terminated Sun and cut ties with the other defendants earlier this year. That is a clean line on paper, though the company also launched an independent probe into the matter in April, which suggests the board was not treating it as a closed book either.
The controls have teeth now
For years the debate over US chip export policy ran on theory. Economists argued about whether restrictions would slow Chinese AI labs or just push them toward domestic alternatives. The Sun plea gives that debate a number and a name. About $2.5 billion in hardware, most of it Nvidia silicon that Chinese buyers cannot legally purchase, was allegedly routed out of the United States by people who understood the paperwork better than the inspectors did.
The Commerce Department officer who showed up to look at the stored servers found an empty room and, according to the indictment, a person willing to say the boxes were still there. That single scene captures the enforcement problem in this sector. The rules are on the books and the parts are identifiable by serial number, but the burden of proving where a server physically sits falls on agencies that have to catch up after the fact.
It is not clear how much this verdict changes that arithmetic. One guilty plea does not add inspectors or shorten the lag between diversion and detection. What it does do is raise the price of participating, because Sun now faces sentencing on four counts that include obstruction.
Demand is why the pressure exists
The reason anyone bothers to smuggle a rack of H100s is that AI builders cannot get enough of them. Reuters reported this week that investors put about $10.7 billion into semiconductor startups in the first five months of 2026, already past the $12.2 billion raised globally in all of 2025, according to Crunchbase data. Intel said earlier this year that it could not manufacture enough chips to meet demand and ended up selling parts it had previously judged too faulty to ship. AMD described swelling CPU sales on a recent analyst call.
Those two facts sit uncomfortably together. A $2.5 billion smuggling operation and a shortage at the world's largest chipmakers are the same shortage, seen from opposite ends. When legitimate allocation runs dry, a secondary channel opens, and the people who fill it tend to be the sort who understand customs forms.
The equity market has noticed. Shares in Intel and AMD have each risen more than 180% this year.
Where the money is going next
Capital is chasing the part of the stack that has been left behind. Nvidia's GPUs absorb most of the attention, but they do not run a data center alone. They need CPUs to route traffic and to execute autonomous agent software, and that bottleneck is now a funding thesis.
Nuvacore, a six-month-old San Jose startup with no shipping product, is raising hundreds of millions of dollars at a roughly $2.5 billion valuation, according to two people familiar with the round who spoke to Reuters. The company was founded by Gerard Williams, John Bruno and Ram Srinivasan. Williams, a former Apple executive, sold his prior startup Nuvia to Qualcomm for $1.4 billion in 2021. A seed round earlier this year was led by Sequoia Capital.
Nuvacore's pitch is architectural rather than incremental. Last month the company said it would build the core functionality of its chip before committing to an instruction set, avoiding the constraints of x86, the architecture Intel and AMD use, or Arm, which powers Amazon and Microsoft custom data center CPUs. The goal is a part tuned for the workloads that dominate AI infrastructure rather than one that has to serve everything. Nuvacore declined to comment on the fundraising.
Nvidia launched a CPU of its own this year, called Vera, and CEO Jensen Huang said the company expects to sell $20 billion worth of it this fiscal year, which ends in January. That figure alone is larger than most estimates of the entire data center CPU market before the AI buildout began.
The read
Two things are happening at once. Enforcers are finally landing convictions in the chip diversion cases, and venture money is pricing the resulting scarcity into every layer of the supply chain above and below the GPU. Sun's plea is the first half. The $2.5 billion valuation on a company that has not yet taped out a chip is the second.
Whether the scarcity that drove both stories holds past 2027 is a different question, and one nobody in this story is answering. Fab construction takes years. Sun's sentencing has not yet been scheduled. More on the case in our semiconductors coverage, and the original reporting on the plea is at Reuters. See also our archive of recent industry stories.