Key Takeaways
- Chinese language Academy hit 3.42% EUV effectivity in 2025, narrowing ASML’s chipmaking lead.
- ASML’s 2019 China ban nonetheless limits sub-7nm chips, pushing home EUV efforts towards 2028.
- Alibaba unveiled 2.4T-parameter Qwen3.8 as T-Head expands post-CUDA AI instruments in 2026.
Essentially the most precious bottleneck in superior chipmaking is the EUV gentle supply, and for years it has sat behind a single Dutch gatekeeper: ASML. Now a Chinese language Academy of Sciences staff on the Shanghai Institute of Optics and Fantastic Mechanics, led by ex-ASML scientist Lin Nan, has revealed outcomes for a laser-produced plasma EUV platform that hits internationally aggressive parameters, together with a reported 3.42% conversion effectivity in a March 2025 paper. The Diplomat notes the ability output continues to be solely about 100 to 150 watts, properly in need of ASML’s business benchmarks, however the course issues for a rustic blocked from shopping for ASML’s most superior EUV instruments since 2019. In Shanghai this week, Alibaba can also be leaning into the identical thesis from the software program facet, previewing its 2.4 trillion-parameter Qwen3.8 and having T-Head pledge to open-source SAIL as a CUDA various.
Cracking the EUV bottleneck
For US readers monitoring semiconductors, the actual drama isn’t just new chips, it’s the instruments that make them. A 12 months after a key report surfaced, China continues to be making an attempt to loosen a choke level that has formed all the pieces from iPhone provide chains to Nvidia-powered AI clusters. The newest progress facilities on the sunshine supply wanted for excessive ultraviolet lithography.
A staff on the Chinese language Academy of Sciences’ Shanghai Institute of Optics and Fantastic Mechanics, led by Lin Nan, constructed an EUV light-source platform {that a} analysis paper describes as working at internationally aggressive parameters, a step the Could 12, 2025 report framed as a barrier giving method. Lin beforehand labored at ASML within the Netherlands, together with as a competence proprietor for gentle sources utilized in metrology, a element that helps clarify why US officers watch these technical papers so intently.
A robust lab end result, in need of manufacturing unit actuality
In March 2025, Lin’s group reported an LPP (laser-produced plasma) EUV setup pushed by a solid-state 1-micron laser, somewhat than the CO2 lasers utilized in business techniques. The paper put its peak conversion effectivity at 3.42%, beating earlier tutorial outcomes from ETH Zurich (1.8% in 2021) and a Dutch analysis heart (3.2% in 2019). Industrial EUV gentle sources are nonetheless round 5.5% environment friendly.
Energy output stays the half that separates prototypes from manufacturing strains. A latest evaluation stated the EUV gentle generated by China’s strategies has been round 100 to 150 watts, versus 600 watts for ASML’s present business benchmark. That hole issues as a result of larger energy permits larger wafer throughput, which is how fabs generate profits.
Export controls, reverse-engineering, and the 2028 goal
US export strain has saved China from shopping for essentially the most superior ASML EUV techniques since 2019, and EUV is a key requirement for chip nodes under 7 nm. One report put China’s lithography-machine market as 99% managed by ASML and Japan’s Nikon and Canon, underscoring how a lot of the availability chain sits exterior Beijing’s attain.
Reuters reported in December 2025 that researchers in Shenzhen had assembled an operational prototype EUV lithography machine in early 2025 by reverse-engineering ASML know-how, utilizing elements salvaged from older instruments and different export-restricted parts. The machine was not producing useful chips, however the ambition was blunt: “The purpose is for China to ultimately be capable to make superior chips on machines which are fully China-made.”
Alibaba’s parallel push: larger fashions, tighter stacks
On the World AI Convention in Shanghai, Alibaba previewed Qwen3.8-Max-Preview on July 19, 2026, describing it as a multimodal mannequin with 2.4 trillion whole parameters utilizing a sparse Combination-of-Specialists design, per the July 20, 2026 write-up. Alibaba has not but disclosed benchmark outcomes, license phrases, or what number of parameters are lively at inference time.
A day earlier, Alibaba’s chip unit T-Head stated it will open-source SAIL, the software program stack behind its Zhenwu chips, positioning it as an on-ramp away from Nvidia’s CUDA ecosystem. That announcement follows T-Head’s Could 20, 2026 unveiling of the Zhenwu M890, which it stated delivers 0.6 PFLOPs FP16 with 144 GB of HBM3, and had already shipped 560,000 models to greater than 400 company prospects by April 2026.
