These are the most-read DIGITIMES Asia stories in the week of December 22 to December 29, 2025.
AMD wins major Alibaba MI308 chip order to challenge Nvidia H200
AMD is poised to score a rare win against Nvidia in China after reports said Alibaba plans to buy about 40,000 to 50,000 units of AMD's China-specific MI308 AI accelerators, a deal that could generate hundreds of millions of dollars in revenue and mark AMD's largest single China transaction since US export controls took effect. The potential order comes as Washington allows both Nvidia and AMD to resume limited AI chip shipments to China in the second half of 2025, reopening competition that had been largely frozen.
While Nvidia has secured approval to reintroduce its H20 and later the more powerful H200, Chinese customers have reportedly cooled on the H20 amid performance concerns and regulatory scrutiny. By contrast, the MI308, which offers higher memory capacity, lower pricing, and no known security objections, is gaining traction as a practical alternative for Chinese cloud and AI workloads. Neither Alibaba nor AMD has confirmed the reported transaction.
SMEE wins China lithography order, yet ASML still controls advanced-node tools
Shanghai Micro Electronics Equipment (SMEE) has secured a CNY109 million, or about US$15.5 million, contract to supply a step-and-scan lithography system to a China government-linked buyer, highlighting Beijing's continued push to localize semiconductor equipment under tightening US-led export controls, while underscoring that advanced-node lithography remains firmly beyond domestic reach.
The tool targets mature process nodes between 0.18 microns and 90 nanometers, aligning with demand from automotive, power, industrial, and specialty chip production rather than cutting-edge logic manufacturing. The single-source, policy-driven procurement reflects China's priority on securing a stable supply for mature-node capacity that underpins much of its domestic semiconductor output.
US probes top SEA Nvidia buyer over chip smuggling
US authorities are investigating Singapore-based cloud provider Megaspeed International over concerns that Nvidia chips may have been diverted to China, a case that has drawn political attention as Republican Senator Tom Cotton urges tighter enforcement of US export controls. People familiar with the matter said the probe has not established any violations, and Megaspeed and Nvidia have denied wrongdoing, while regulators in the US, Singapore, and Malaysia say their reviews are ongoing.
Bloomberg's analysis of customs data shows Megaspeed has imported at least US$4.6 billion worth of Nvidia hardware since 2023, including large volumes of current-generation Blackwell chips, and rents computing capacity from data centers outside China to customers that include Chinese firms, a model generally allowed under current rules. US officials are examining Megaspeed's ownership and whether any chips were physically transferred into China without licenses, while Nvidia says repeated site checks found no evidence of diversion.
Samsung, SK Hynix reportedly hike HBM3E prices nearly 20% ahead of 2026 shift
South Korean memory makers Samsung Electronics and SK Hynix have raised prices for fifth-generation high-bandwidth memory by nearly 20% for 2026 deliveries, as demand for AI accelerators continues to outpace supply even as producers shift capacity toward next-generation HBM4, according to industry sources.
The increase for HBM3E chips defies the usual price softening ahead of a generational transition and reflects sustained orders from AI chip designers, including Nvidia, Google, and Amazon, which are preparing higher-memory AI processors for 2026. Continued demand is also being supported by Nvidia's planned resumption of H200 exports to China, which rely on HBM3E rather than HBM4, extending the lifecycle of the current standard.
TSMC's N-2 restrictions give Samsung US advanced chip an opening
Samsung Electronics is positioning itself to capture rising US demand for advanced chips as Taiwan's N-2 policy limits overseas deployment of TSMC's leading-edge processes and much of TSMC's near-term capacity is tied up by Apple and Nvidia, according to industry sources. Samsung has begun sample testing of its second-generation 2nm SF2P process with AMD and is working with Google on potential 2026 production, while its Taylor, Texas, fab is being lined up for 2nm mass manufacturing as early as next year.
By contrast, TSMC plans to reach 3nm volume production at its Arizona Fab 2 in 2027, leaving its US operations at least two generations behind its Taiwan base due to regulatory constraints. With Apple reportedly taking nearly half of TSMC's 2nm output and Nvidia planning to move 2nm production to TSMC starting in 2027, customers such as Qualcomm, AMD, and Google are increasingly evaluating Samsung as an alternative.
Google reportedly fires procurement execs amid HBM supply crunch
Google and Microsoft are intensifying efforts to secure high-bandwidth memory as production capacity at South Korean suppliers tightens, underscoring how AI-driven demand is straining the global memory supply chain, according to industry sources cited by Seoul Economic Daily and G-enews.
Major cloud companies, including Google, Microsoft, and Meta, have stationed procurement teams in South Korea to negotiate supplies of HBM, DRAM, and enterprise solid-state drives from Samsung and SK Hynix, whose advanced memory lines are nearing full utilization. Tensions have emerged as buyers seek long-term commitments that suppliers say they cannot meet, leading to stalled talks and, at Google, the dismissal of procurement executives after the company failed to lock in capacity for its in-house tensor processing units.
Samsung, SK Hynix reportedly accelerate HBM4 production to early 2026
Samsung Electronics and SK Hynix are accelerating sixth-generation HBM production to February 2026, moving months ahead of earlier schedules as they race to meet surging demand for AI infrastructure, according to industry sources and South Korean media, including Seoul Economic Daily and TheElec.
The revised timeline positions Samsung Electronics and SK Hynix to become the first suppliers to deliver HBM4 at scale, tightening their grip on the AI memory market before rivals can catch up. SK Hynix plans to start volume wafer input using 1b DRAM at its M16 and M15X facilities, while Samsung is expected to launch full HBM4 production at its Pyeongtaek campus after receiving positive evaluations from Nvidia, according to sources.
Article edited by Jerry Chen
