Battlefield Nvidia
Strategy and skulduggery with trillions on the line
Nvidia is the most valuable company in the world. Its GPUs hold an estimated 80% share of the AI training chip market, and the data centers and supercomputers on which frontier models are trained and run rely on its chips to function.
Last week, Jensen Huang told a capacity crowd at GTC that the company expects its cumulative order backlog to reach one trillion dollars by 2027. Two days later, a co-founder of Supermicro, one of Nvidia’s closest server manufacturing partners, was arrested for allegedly diverting $2.5 billion of that hardware to China.
The United States has restricted exports of advanced Nvidia chips to China since 2022. These controls sit within a wider context of great-power tension: a trade war with China, military operations in the Middle East in which AI tools have been deployed, and an accelerating global race for compute capacity. The US restrictions reflect the hardware’s strategic value.
Infrastructural ambition unveiled
Nvidia’s GPU Technology Conference ran March 16–19 in San Jose. Huang’s two-hour keynote opened with a strategic retrospective on CUDA, the parallel computing platform that Nvidia introduced twenty years ago and that now underpins an installed base of hundreds of millions of GPUs. The keynote’s organizing concept built on that foundation: the token as the fundamental unit of AI production, and the data center as the factory in which tokens are produced.
The centerpiece announcement was Vera Rubin, a platform now in production that packages seven chip types into five rack-scale systems: Vera CPUs, Rubin GPUs, NVLink 6 switches, ConnectX-9 NICs, BlueField 4 DPUs, and Groq LPUs, all configured to operate as a single AI supercomputer. The NVL72 GPU rack alone delivers 20.7 TB of HBM4 memory, while the Vera CPU rack maintains 22,500 parallel sandboxes for AI agent workloads. The unit of sale is the integrated system. Nvidia has moved from supplying chips to supplying factories.
Huang also previewed Feynman, the next-generation architecture after Vera Rubin, scheduled for 2028. Feynman will include a new GPU, the LP40 LPU built with the acquired Groq team, a new CPU called Rosa, BlueField 5, and co-packaged optical networking for scale-up interconnects. Huang projected that cumulative orders between 2025 and 2027 could reach one trillion dollars, doubling his estimate from a year ago. The roadmap extends Nvidia’s planning horizon to mid-decade and assumes that demand for AI compute will continue to accelerate.
Nvidia accompanied the hardware announcements with moves into software and new deployment environments. NemoClaw, an open-source platform for building and governing enterprise AI agents, extends Nvidia’s reach into the agent orchestration layer. The Space-1 Vera Rubin Module brings data-center-class AI compute to orbital environments, with partners that include Axiom Space and Planet Labs. DLSS 5 uses neural networks to generate entire frames for gaming and real-time simulation, pushing AI inference onto consumer GPUs. Taken together, the GTC announcements span the full vertical of AI infrastructure, from consumer hardware to orbital computing to enterprise governance software.
Smugglers in the pipeline
On Thursday, a Manhattan federal court unsealed an indictment charging Yih-Shyan “Wally” Liaw, co-founder and senior vice president of Super Micro Computer, with conspiring to smuggle Nvidia-powered servers to China. Prosecutors charged two associates alongside him: Ruei-Tsang Chang, Supermicro’s Taiwan general manager, who remains a fugitive, and Ting-Wei Sun, a third-party contractor, who was arrested. The alleged scheme diverted approximately $2.5 billion in servers between 2024 and 2025, routing them through an unnamed Southeast Asian company before their final delivery to Chinese buyers.
The operation worked through layers of deception. Liaw and Chang directed the Southeast Asian company to place purchase orders with Supermicro as though the servers were for its own use. After assembly in the United States and shipment to Supermicro’s Taiwan facilities, the pass-through company received the servers, repackaged them in unmarked boxes, and forwarded them to China. To satisfy compliance audits, the defendants allegedly staged thousands of non-working “dummy” servers at a warehouse while the real hardware was already in transit. One compliance auditor, according to the indictment, was off-site at the pass-through company’s expense while Sun sent photographs and videos of the staged equipment in his place.
Liaw co-founded Supermicro in 1993 and controlled $464 million in company shares at the time of his arrest. Supermicro is one of Nvidia’s primary server manufacturing partners, assembling the GPU-powered systems on which much of the AI infrastructure buildout depends. The alleged smuggling operated within the company’s own sales and logistics channels, using its Taiwan facilities and internal approval processes to move the hardware.
Supermicro shares fell 33% on the day that the indictment was unsealed. Liaw resigned from the board the following day. Each defendant faces up to twenty years on conspiracy charges under the Export Controls Reform Act, with additional counts of smuggling and fraud. The indictment arrives amid an existing pattern of governance problems at the company: a 2018 accounting scandal that led to an SEC penalty, the resignation of auditor Ernst & Young in 2024, and a damaging short-seller report from Hindenburg that same year.
Independence at all costs
The Supermicro case is a symptom of a structural shortage. Export controls have constrained legal supply of advanced AI chips to China, while Chinese demand for that compute has grown in parallel, driven by state policy and competition among domestic firms. The gap creates the conditions in which a $2.5 billion smuggling operation becomes viable.
China’s 15th Five-Year Plan, released earlier this month to coincide with the National People’s Congress, designates AI as central to economic strategy for 2026–2030. The plan calls for hyper-scale computing clusters powered by abundant domestic electricity and for breakthroughs in high-end AI chip design. Embodied intelligence, meaning robotics integrated with AI, appeared as a distinct strategic priority for the first time in a document of this level. The plan aims to build domestic compute infrastructure sufficient to reduce reliance on imported hardware.