News · · 10 min
Period: September 13 — morning of September 14.
Today's theme: competition is increasingly shifting from the "best LLM" to the entire production pipeline — lithography, cheap inference, harness, action verification, engineering skills, and open model distribution channels.
What happened. Reuters reports today that current ASML EUV machines costing about $200 million are virtually sold out through the end of 2027, and major manufacturers have begun giving multiple commitments for High-NA EUV systems costing approximately $400 million each. High-NA allows printing features roughly 40% smaller than current EUV.
Intel has already processed over 1 million wafers using High-NA, including individual layers for mass-produced Panther Lake. Samsung plans to use High-NA for mass production of DRAM starting in 2028, TSMC — for advanced-node production starting in 2030.
Why it matters. The AI-compute bottleneck resides far higher in the supply chain than Nvidia GPUs:
lithography → logic/HBM → accelerators → clusters → models → inference.
The spread of High-NA to DRAM is especially important: further scaling of AI requires not only compute dies but also increasingly dense and fast memory.
What the data showed. This is production/manufacturing evidence, not an AI benchmark. Intel says throughput, availability, and overlay already meet its expectations; individual High-NA Intel 18A layers yield results no worse than equivalent layers on standard EUV.
What to know already: EUV, High-NA, photomask, HBM/DRAM, process node.
What this could change. For a regular team — observe for now. For businesses with large compute demand, it's important to understand: reducing AI cost depends not only on models and GPUs but also on multi-year expansion of the entire semiconductor supply chain.
Date: significant industry update — September 14. Primary sources: ASML + Intel, ASML + TSMC. Additional source: Reuters.
deepseek-v4-pro endpoint todayThis continues a previously covered story; the actual switch date is today.
According to the official DeepSeek schedule, as of 04:00 UTC on September 14, all calls to deepseek-v4-pro are routed to V4.1 Flash and billed at Flash pricing until V4.1 Pro is released.
Why it matters. The endpoint remains the same, the underlying model does not. This is a good example of hidden model dependency: an application might not change a single line of code but receive different reasoning behavior, latency, and failure profile.
DeepSeek justifies the switch by stating that V4.1 Flash outperforms V4 Pro on agent-oriented workloads and requires four times less HBM and eight times less SSD for KV-cache compared to the previous generation. This is vendor evidence, already analyzed in a previous issue.
What to know already: model routing, regression testing, KV-cache, endpoint compatibility.
What this could change. If your production uses deepseek-v4-pro, it's worth running your own evals today. API compatibility does not mean behavioral compatibility.
Date: September 14, 04:00 UTC. Primary source: DeepSeek — V4.1 Flash
What happened. Shanghai AI Laboratory announced on the evening of September 13 the full opening of the full-stack service of its scientific agent platform "Shusheng Duan Yan", describing the architecture as a combination of "compute — prove — verify".
The platform itself was presented earlier. Its architecture connects scientific foundation models, specialized agents, experimental data, and automated lab equipment, closing the loop from hypothesis to physical verification.
Why it matters. This clearly illustrates a more general agent pattern:
LLM builds a hypothesis
→ evidence is sought
→ the result is verified by an external system
→ a real experiment is conducted if possible.
For AI for Science, this is a literal lab cycle. For enterprise agents, analogs could be a compiler, database, API, simulator, policy engine, or another source of ground truth.
What the data showed. The new announcement lacks a new independent benchmark to quantitatively measure the quality improvement from the full-stack architecture itself. Therefore, this is a significant architectural deployment, but not proof of system superiority.
What to know already: AI for Science, agent harness, external verification, provenance, closed-loop experimentation.
What this could change. A useful architectural template for long-horizon agents: avoid forcing a single LLM to simultaneously devise a solution, evaluate its own solution, and declare it correct.
Date: September 13, 23:19 China Standard Time.
What happened. OpenAI published a Perplexity case study today. The company uses Astra not only for code generation: the model writes communications, modifies software, and monitors production systems. For testing, Astra builds its own small simulation harness around the application, generating responses from external APIs and connectors, then verifies the workflow end-to-end.
Why it matters. Coding agents are gradually crossing the boundary:
write a diff
→ verify the system
→ work with a production system.
This dramatically increases the value of the model's ability to build its own test doubles and verification environment, but simultaneously increases the blast radius of an error.
What the data showed. Perplexity says it now checks the agent's work "much less frequently" than with previous model generations. But OpenAI does not publish intervention rate, defect rate, rollback frequency, or a controlled comparison. Therefore, this is a vendor-sponsored production anecdote, not a proven productivity benchmark.
What to know already: end-to-end testing, test double, production access, human supervision, rollback.
What this could change. If coding agents receive production authority, you need to measure at least human interventions, escaped defects, rollback rate, and mean time to recovery — not just code writing speed.
Date: September 14. Primary source: OpenAI — Perplexity trusts GPT‑6 Astra with end-to-end systems
What happened. Chinese Z.AI/Zhipu completed a funding round of approximately $5 billion: about $2 billion via share placement and $3 billion via convertible bonds.
The use of proceeds is especially important: the company stated that about 60% of net proceeds will be directed towards R&D for the next generation of models and its own fully self-training system, with another 15% for business expansion.
Why it matters. This is not a technological breakthrough, but a good indicator of where competition among Chinese frontier labs is heading.
Money is going not just into another training run, but into the system:
model → training signal/data generation → evaluation → post-training → next model.
If such a cycle truly becomes more automated, the competitive advantage shifts to the speed of the model-improvement loop.
What the data showed. $5 billion is a confirmed capital raise. No data yet demonstrates the effectiveness of the "fully self-training system".
What to know already: self-training, post-training, training compute, convertible bond, model improvement loop.
What this could change. For model users — observe for now. But after GLM‑5.3 Flash, Z.AI shows that the Chinese model ecosystem has capital to continue aggressive competition on capability and price.
Date: September 13. Source: Reuters citing company exchange filings.
A fresh, small but indicative project — rizqinrr/viserys-agent, created on September 12 and already garnering about 630 stars by the morning of September 14. The star count itself proves nothing; the project's construction is more interesting technically.
Viserys contains 28 separate engineering skills, four reviewer personas, fixtures for evals, validation scripts, and its own eval runner. The workflow is formalized as:
DEFINE → PLAN → BUILD → VERIFY → REVIEW → SHIP.
Why it matters. This continues yesterday's signal from Claude Code's plugin eval: the developer community is beginning to treat skills not as "good prompts" but as versionable and testable software artifacts.
First, the competitive layer was the model. Then the harness. Now another layer is emerging:
engineering process encoded as skills.
What to know already: Agent Skills, evals, regression testing, coding-agent harness.
What this could change. Team rules for coding agents should be stored in a repository, versioned, and validated with evals. For now, Viserys itself is an early project without production evidence; choosing it just for stars is not advisable.
Signal date: September 13–14. GitHub — Viserys
On September 13, at the BRICS summit, China proposed establishing a joint AI open-source community, supporting joint LLM development and application, conducting AI training, and building a common open ecosystem. A BRICS digital ecosystem cloud platform was also separately proposed.
Key technical details are currently missing: which models, licenses, compute, repositories, standards, and governance would be part of the system. Therefore, this is currently a policy/distribution signal, not a new open-source platform.
But the potential chain is interesting:
open models → developer training → regional cloud infrastructure → local deployments → reduced dependency on US APIs.
For companies operating in Global South/BRICS countries, this is worth monitoring as a possible future distribution and procurement layer.
Date: September 13. Primary source: PRC Ministry of Foreign Affairs — Statement at BRICS Summit
The AI stack continues to fragment into independent competitive layers: lithography → inference-efficient models → harness → verification → skills → production authority → distribution ecosystem.
One trend is especially noticeable: the more real work an agent gets, the more important external verification, a testable process, and clearly defined authority become over additional reasoning tokens. Simultaneously, upstream infrastructure shows that even rapid model improvement remains dependent on multi-year chip production cycles.