THE CRUNCH
Tom's Hardware's latest Premium roundup frames the week around one theme: the AI buildout is reshaping the semiconductor industry from every angle. The headline item is a reported deal in which Tencent will lease high-performance AI chips from Oracle for $7 billion over five years, covering some 100,000 chips offshore. No specific chips have been named, but the pricing points toward a Hopper-class GPU, and the estimated per-hour cost works out around 43% below a standard one-year Nvidia H100 lease at $2.80 per hour. The report also explores how such rental arrangements could give Chinese firms a route to otherwise export-controlled hardware.
Elsewhere, OpenAI and Synopsys have partnered on a specialised model called GPT-Synopsys, which combines Synopsys' chip design IP and EDA (electronic design automation) tools with OpenAI's large language model expertise. The model is, in all likelihood, a chipmaking-specialised AI that could play a part in developing successors to OpenAI's Jalapeño AI ASIC, and teaming up could make the usual IP tangles around AI-designed chips a non-factor.
Hardware supply chains feature too. Japanese toolmaker Gigaphoton has debuted a neon recycling system with a claimed 50% recovery rate. Neon, a noble gas used in DUV (deep ultraviolet) lithography, has seen over 70% of global supply consumed by those older machines, with production historically concentrated in Ukraine and disrupted by the ongoing conflict there. Rival Kanten Techno has developed competing recycling machines, though newer EUV systems use far less of the gas.
In licensing news, Qualcomm has reportedly agreed to license Huawei's patents for the first time, including, according to one report, Huawei's LogicFolding chip-stacking methods, which target faster data transmission as a workaround for lacking access to advanced EUV machines. The Kirin 9050 Pro already uses the technique, with a teardown showing its lower die is mostly cache and I/O. Historically the arrangement ran the other way, with Huawei licensing Qualcomm's patents.
Finally, two former Groq employees and shareholders have filed a proposed class action against Groq's former board and CEO, alleging breaches of duty cost stockholders billions after Groq's $20 billion deal with Nvidia in December 2025. Nvidia is not a defendant, though most Groq staff have reportedly joined the company, and the Groq3 LPU is being showcased within Nvidia's Vera Rubin ecosystem.