It was a volatile week for AI chips, with announcements spanning processors, memory, manufacturing and security—and renewed scrutiny of Nvidia-related export controls.

AMD unveiled new EPYC server and embedded processors, an AI GPU with 288GB of HBM4, and tamper-resistant FPGAs, while Google expanded its eighth-generation TPU family with systems for agentic AI and specialized security workloads.

One of Google’s new TPU systems pairs its accelerators with AMD EPYC CPU hosts, while Arm introduced a processor for machine-learning and signal-processing workloads in compact devices.

AMD also presented MEXT, a predictive memory system combining a transformer model with flash storage; in the described configuration, it used half the DRAM while retaining 95% of server performance.

AMD reported that MEXT had been deployed on more than 100 servers, including 145 using the technology, and said it plans to expand to 2,000 servers in 2027.

Benchmarks comparing Apple’s Mac Studio with its M5 Ultra chip and Nvidia’s DGX Spark showed that local AI results depend not just on hardware, but also on inference software, optimization, memory behavior and model size.

Access to computing and memory is increasingly uneven: the Lowy Institute estimated that U.S. AI data centers have nearly 40 times China’s computing power, while memory has risen to 54% of a low-end smartphone’s bill of materials, with prices expected to keep climbing through the first quarter of 2027.

As Chinese AI labs sought alternatives, Nvidia tightened overseas cloud rentals of its GPUs for Chinese customers, requiring partners to reserve 51% of deployed capacity for non-Chinese users; black-market prices for a B300 rack reportedly reached $2.8 million.

Some Chinese companies also sought Nvidia compute through rentals in Malaysia and Singapore, while Beijing’s requirement to pair foreign hardware with domestic silicon added cost and complexity for AI labs.

Nvidia faced renewed congressional and policy scrutiny over chip exports and alleged diversion to Megaspeed, with senators seeking records, House committee leaders requesting a briefing from Jensen Huang, and reports describing disputes over tighter restrictions.

Separate reports alleged that Jensen Huang stepped away from his chief executive post and resigned, and described scrutiny involving Alice Huang and Shanghai fund manager Kang Wei, whom they linked to an alleged $7.5 billion operation smuggling Nvidia AI GPUs.

Nvidia’s broader financial ties to the AI industry also drew attention, with the Financial Times reporting more than 120 investments during the year and about $124 billion in equity investments held or committed by July.