TSMC and Samsung commit to ASML's $400 million High NA EUV machines
The chip supply chain just locked in its next decade. Two of the three biggest buyers of advanced lithography committed to ASML's High NA EUV tools on the same day, and AI is the reason both gave.
The world's two largest chipmakers have both committed to ASML's High NA EUV lithography machines, ending years of speculation about whether the industry's most expensive tool would ever reach volume production. Samsung said it will introduce High NA into DRAM high-volume manufacturing in 2028 — an industry first — while TSMC said it expects to adopt the technology in 2030 for advanced logic, attributing the demand to "the increasingly complex transistor architectures required for AI applications." Each machine costs around $400 million, and both companies are also joining ASML in an initiative to move the industry from 6-inch photomasks to 12-inch, a format shift that ASML says will raise fab productivity and lower per-chip costs. Samsung and TSMC join Intel, which ASML confirmed in July is already using High NA for advanced manufacturing.
The significance is less the machines than the dates. For years the open question was whether High NA was a science project that Intel alone would pay for — a way to look ambitious without a roadmap. Now the two foundries that actually run the world's leading-edge volume have put years on the record, which converts High NA from an optional upgrade into the industry's default path past 2028. Barclays called the announcements "a positive" that gives ASML "more visibility on adoption which has been a key debate," and flagged the follow-on decision: whether ASML expands EUV capacity beyond the roughly 30% increase it has already guided for 2027. The honest caveat is that a 2030 logic date is four budget cycles away and commitments at this range are cheap to make and easy to slide. But the alternative — that leading-edge scaling stalls because the printing tools can't resolve the next node — is now visibly off the table, and that is what the AI buildout is betting on.
Hyperscalers are quietly scouting Argentina's Patagonia for mega data centers, drawn by cold air, shale gas, and the absence of local opposition. Reuters reports that Pampa Energia is pitching a data center next to its Loma de la Lata power plant in Neuquén, drawing up to 500 megawatts from the Vaca Muerta shale formation, with grid infrastructure for that load costing almost $900 million. Poland's Green Capital is planning a Chubut facility that would start at 300 MW for $3 billion and eventually reach 3,000 MW, and U.S. firm FlexDomes is seeking investors for a $1.4 billion, 120 MW first stage. Investors want a small pilot first — 20 to 40 MW — before committing to anything at scale.
What makes Patagonia interesting is not the climate, it's the politics. The U.S. buildout is running into organized local resistance over grid strain and utility bills; Argentina has no such backlash yet, because nobody has felt a large data center there. That is a temporary condition, not a durable advantage, and the people quoted by Reuters know it — one adviser notes that "saying you have power is not enough," and several investors are waiting on Argentina's 2027 presidential election before committing. OpenAI's up-to-$25 billion Sur Energy project, announced last October, still hasn't signed a definitive contract. Read this as land-banking, not capacity.
A startup founded by Huawei's former Noah's Ark Lab head has released a model trained on its own agents' mistakes. TokenRhythm (基元律动), founded by Wang Yunhe, shipped NeoHorse-1 in 4B and 9B sizes, both post-trained on Alibaba's open Qwen3.5 base models. The training data is the interesting part: rather than scraped text, the core corpus is execution traces generated by the company's routing harness — which model was picked for a task, what tools it called, what feedback it got, whether it recovered from errors. The team reports the highest unweighted average among 4B-class models across 11 benchmarks spanning agent execution, tool interaction, code, and instruction-following, with the 4B beating the Qwen3.5-9B base on five of them.
The gains are real but narrow, and the report is honest about where they land: tasks with clear process, observable feedback, and verifiable outcomes. Complex state maintenance, long debugging sessions, and failure recovery still favor bigger models. TokenRhythm frames this as one turn of recursive self-improvement — agent runs, evals expose gaps, training adjusts, the updated model goes back into the harness. One documented turn is not a flywheel, and self-reported benchmark tables from the lab that built the model deserve a wait-and-see. But using agent trajectories as training data is a genuinely different input from the internet, and it's a path that any lab with a production agent can now copy.
What to watch: whether ASML commits to EUV capacity beyond its 2027 target — that number will tell you if these dates are real.
If a 4B model trained on agent failures beats a 9B base on agent tasks, does model size still matter for the work you actually ship? Tell us in the comments.
Sources: CNBC · Samsung Newsroom · Techzine · Reuters · Techmeme · Leiphone (雷峰网) · NeoHorse-1 (GitHub)
