Compute & Infrastructure

Federal prosecutors charged Super Micro Computer's co-founder and two others with conspiring to smuggle $2.5 billion worth of NVIDIA-powered servers to China, according to Bloomberg and Tom's Hardware. The alleged scheme involved using a hairdryer to transfer serial numbers from restricted servers onto thousands of dummy units, allowing the perpetrators to ship advanced AI hardware while bypassing export controls. Super Micro's stock lost a third of its value following the revelations, and the company is now overhauling its compliance operations.

The case reveals both the sophistication of circumvention tactics and the enforcement challenges at a critical infrastructure node. Super Micro is a major supplier of AI server systems integrating NVIDIA GPUs, making it a strategic point in the supply chain where restricted components can be diverted. The scale of the alleged operation—$2.5 billion suggests systematic rather than opportunistic smuggling—and the involvement of company leadership raises questions about corporate governance at firms handling controlled AI hardware.

SoftBank is planning a $60-70 billion AI data centre in Ohio requiring 10 gigawatts of power, necessitating a $33 billion natural gas plant equivalent to nine nuclear reactors, reports Tom's Hardware. The facility would be among the largest data centres globally if built, but the project remains in planning stages and faces significant infrastructure challenges. The power plant cost alone approaches half the total data centre investment, illustrating how energy infrastructure is now the dominant capital requirement for hyperscale AI deployment.

The project's reliance on natural gas rather than renewable or nuclear baseload power reflects the urgency of AI buildout timelines. Gas plants can be constructed faster than nuclear facilities, but this choice creates tension with sustainability commitments and exposes the facility to fuel price volatility. The 10-gigawatt figure is roughly equivalent to the total power consumption of a small city, and securing grid interconnection, water for cooling, and environmental permits will likely determine whether this project proceeds beyond planning stages.

Senators Elizabeth Warren and Richard Blumenthal are investigating NVIDIA's $20 billion licensing deal with AI inference startup Groq, questioning whether the structure improperly avoids merger review while consolidating NVIDIA's compute market power, according to Bloomberg. The deal involves NVIDIA licensing its technology to Groq rather than acquiring the company outright, potentially skirting Committee on Foreign Investment in the United States review while achieving similar strategic control. The senators' letter suggests concern that NVIDIA is using licensing arrangements to extend its GPU dominance into the inference layer where alternative architectures like Groq's language processing units could pose competitive threats.

NVIDIA's market position in AI training hardware is estimated above 80 percent, and the company's CUDA software ecosystem creates switching costs that entrench its advantage. If NVIDIA can use licensing deals to prevent alternative inference architectures from gaining market share, it would further consolidate control over the entire AI compute stack. The regulatory scrutiny comes as the Trump administration's new AI framework emphasises avoiding heavy federal regulation, creating tension between competition concerns and the White House's stated preference for industry self-governance.

Alibaba shipped 470,000 AI chips in recent quarters and projects $100 billion in future cloud and AI revenue, though investor response remained tepid, reports Data Center Dynamics. The shipment figure indicates significant production scale for Alibaba's Yitian ARM-based processors and potential in-house AI accelerators, suggesting China's largest cloud provider is reducing dependence on U.S. semiconductor imports despite export restrictions. The $100 billion revenue projection spans an undefined timeframe and lacks detail on infrastructure investment requirements, making it more aspirational than actionable guidance.

The 470,000 chip shipment volume is meaningful primarily as evidence of manufacturing capacity rather than technical competitiveness—these are likely lower-performance chips suitable for inference workloads rather than cutting-edge training hardware. However, the scale demonstrates that Chinese firms are building domestic alternatives at volume, even if they lag U.S. offerings by one or two generations. Investor scepticism likely reflects concerns about margin pressure as Alibaba invests heavily in infrastructure while competing with ByteDance, Baidu, and state-backed AI initiatives in an increasingly crowded domestic market.