Nobody setting AI's speed limit this week was a regulator: Amodei said on Sept 12 the industry must slow down, three days later the three largest labs were confirmed to be drafting their own standards body, and California signed a kill-switch order on Sept 18 — while Anthropic's own disclosure put Claude in the lead on 26% of the work that builds the next Claude
Today is the general-AI slot, but this week's through-line lands squarely on the floor that clinical AI stands on. On Sept 12, Axios reported that Anthropic CEO Dario Amodei published "We Must Pace the Frontier," arguing openly that "we must slow the pace at which we improve the capabilities of AI models." Three days later TechCrunch confirmed that OpenAI, Anthropic and Google DeepMind had been in safety talks for weeks, including a proposal for an industry-run standards body. On Sept 18 Governor Newsom signed an executive order convening national experts within two months to consider embedding independent verification organizations onsite in labs and requiring an emergency shutoff for frontier models. In the same week Anthropic published internal metrics for the first time: Claude now leads more than a quarter of its own R&D. Put those four together and the medical question is not when the FDA will regulate AI, but who sets the model's safety floor long before any device reaches the FDA.
01 — Top Stories
Amodei argues for pacing the frontier: give outside evaluators employee-level access first, then ask rivals to match it
On September 12, Amodei published "We Must Pace the Frontier," laying out three steps: Anthropic will unilaterally grant external evaluators employee-level access to monitor its safety procedures and report incidents; it calls on the rest of the industry to do the same; and it asks democratic governments to coordinate on standards. Axios, 2026-09-12 quotes his warning that rogue AI swarms could "take over the entire internet," causing hundreds of billions of dollars in damage. Sam Altman agreed the industry needs slower frontier development and committed to the same evaluator access; Elon Musk said simply that "Dario is right."
This is the first time the largest labs have themselves proposed decelerating, and the core of the proposal is not a ban on anything but letting third parties inside the company. For medicine this is the upstream version of post-market surveillance: a large share of the risk in a clinical AI product comes from behavioral drift in the underlying foundation model, and no medical regulator currently has visibility into that layer.
This is a manifesto written by the company with the strongest incentive to define the rules. Investor Chamath Palihapitiya argued the essay aimed to consolidate power with Anthropic rather than purely address safety, and the essay coverage makes no mention of biological or medical risk at all.
The three largest labs spent weeks drafting a FINRA-style self-regulator; the fourth lab's CEO had already called it a cartel
On September 15, OpenAI global policy chief Chris Lehane confirmed the company had been working with Anthropic and Google DeepMind on AI safety for weeks, including discussion of an industry standards body, and said OpenAI supports the FRONTIER Act provision giving "independent verification organizations" access to frontier labs (TechCrunch, 2026-09-15). Lehane said the firms "don't need" an antitrust waiver for such coordination. Two days earlier, on September 13, Cohere co-founder and CEO Aidan Gomez had published an essay calling the proposal "a cartel by any other name," on the grounds that "a safety regime designed by a few labs will only be rigorous about the risks they have already built their safety systems to assess" (AI Weekly, 2026-09-13).
Gomez's objection cuts hardest in medicine. If three general-purpose model companies write the standard, the risks that get tested are their risks — jailbreaks, cyber-offense, autonomous replication — not clinical ones: performance gaps across subpopulations, hallucination inside an EHR workflow, diagnostic drift after a model version bump. If healthcare is not at that table, it will be handed a safety certificate about somebody else's problem.
California's executive order studies an emergency shutoff for frontier models and adds loss-of-control to the definition of a critical safety incident
On September 18, Newsom signed an executive order directing the Government Operations Agency to accelerate implementation of SB 813 and AB 1405 and to convene national experts within two months on strengthening state AI law. The measures under consideration: independent third parties writing frontier companies' safety plans; a required emergency shutoff mechanism for frontier models; designated independent verification organizations embedded onsite in labs for regular audits; and an updated definition of critical safety incidents that includes loss-of-control events (Office of the Governor of California, 2026-09-18).
This is the first time a government has written the labs' own "embed external evaluators" idea into an executive document — meaning the Sept 12 manifesto and the Sept 18 order describe the same mechanism, one voluntary, one on its way to becoming state law. What healthcare should notice: the order does not mention healthcare anywhere. Medical AI governance is still arguing at the device layer while a state government moves first at the model layer.
This is an executive order, not a statute: it orders study and the convening of experts, provides no technical specification, and leaves "kill switch" an undefined term. Whether any of it lands depends on legislation that has not happened yet.
Anthropic publishes an internal automation metric for the first time: Claude leads 26% of its own R&D, and the fully autonomous share is 0%
On September 17, Anthropic published an "R&D Automation Index" stating that Claude now leads more than a quarter — 26% — of the company's research and development work, up from under 1% in February. Quartz describes the bar as Claude being able to "complete most of a given task end-to-end from a high-level prompt while a human supervisor remains in the loop." Aggregated coverage adds that the index spans roughly 15,000 tasks with about 30,000 agents running concurrently inside the company, that over 90% of the work sits at the "collaborates" level or above, and that the fully autonomous share is 0% (buildfastwithai roundup, 2026-09-18).
The value of this index is not the 26% but the 0%. It hands healthcare a scale shape it can borrow directly: break "AI involvement" into assists / collaborates / leads / fully autonomous, then admit the top tier is currently empty. That graded vocabulary is exactly what hospitals lack when deploying ambient documentation or diagnostic support, where the debate still tends to collapse into a binary about whether AI was used at all.
This is a vendor-reported internal metric with no independent audit, a taxonomy defined by the party publishing it, and raters who are employees working with the model they know best. "Leads 26%" is not "replaces 26% of the headcount," and no conversion between the two is offered.
Novo Nordisk picked a general-purpose model over a bio-specialist: Claude Science enters drug discovery, with no figures disclosed
On September 16, Novo Nordisk announced a collaboration with Anthropic to use Claude for accelerating drug discovery and development, advancing scientific reasoning in R&D workflows, supporting biological reasoning and understanding of drug mechanics, and for AI-driven software development and agentic engineering; the two will also test Claude Science on specific R&D workflows. CEO Mike Doustdar called it "another testament to our ambition to become the world's most AI-driven healthcare company," saying AI can "compress the path from research to marketed product." Amodei restated his standing claim that AI could "compress a century's worth of biological and medical breakthroughs into a decade" (BioSpace, full press release, 2026-09-16). Novo employs more than 67,000 people globally.
Europe's largest pharmaceutical company chose a general-purpose frontier model over a bio-specialist, which is a structural signal: what pharma is buying is not structure prediction but reasoning and agentic work spanning literature, lab records, regulatory documents and internal code. Structure prediction is already crowded; the layer that stitches the whole R&D pipeline together has no winner yet.
The press release discloses no deal value, timeline, hours saved or number of programs covered, and describes data governance only as "robust data governance and human oversight." Until a verifiable milestone appears, this is news at the level of a letter of intent.
Nature: 20,000-plus proprietary structures from drug firms lift protein–ligand prediction from a third of cases to more than half
On September 14, Nature news reported that a consortium including AbbVie and Astex Pharmaceuticals trained protein-folding models on more than 20,000 proprietary structures: across 1,056 test structures the consortium model reached high-accuracy protein–ligand interaction prediction on more than half, against roughly a third for the public OpenFold3 and about 40% for Boltz-2. For context, the Protein Data Bank holds some 200,000 experimentally determined structures but only about 10,000 involving drug-like molecules. The UK government is backing the OpenBind project with up to £8 million (US$10.8 million). Columbia computational biologist Mohammed AlQuraishi called the result "a pretty big bump in performance."
This is the hardest result of the week, and it contradicts a popular story: that the bottleneck on model capability is architecture. Here the bottleneck is data. Public databases hold only ~10,000 drug-like-molecule structures; 20,000 more out of pharma's drawers moved the hit rate from a third to over half in one step. The same shape applies to hospitals: what is genuinely scarce is not models but labelled clinical data locked inside institutions.
Nature flags it explicitly: the work so far exists only in a blog post, is not peer reviewed, and the model is not publicly available. The test set is 1,056 structures, the "high accuracy" threshold is the consortium's own, and the comparison against public models has not been independently reproduced.
GLM-5.3-FlashX claims ~200 tokens/s across roughly 100,000 domestic accelerators — with the inference stack written by the model itself
On September 17–18, Zhipu/Z.ai detailed the serving stack behind GLM-5.3-FlashX: a 320-billion-parameter mixture-of-experts model activating 18 billion parameters per token, running at roughly 200 tokens per second across about 100,000 domestic accelerators, with serving optimization performed by what the company calls an "Infra Agent" (Pandaily). The predecessor GLM-5.3-Flash launched August 26 with a one-million-token context at $0.15 per million input tokens and $0.50 output, against $1.40 and $4.40 for the full GLM-5.3 (implicator.ai).
For medical AI the significance is the price band. At $0.15 per million input tokens, feeding an entire chart, a whole batch of imaging reports, or a department's year of notes into a model shifts from a luxury to a line item. Whichever ecosystem makes long-context inference cheap first bends its hospital adoption curve first — and this time the cheap path is the one that does not depend on Nvidia.
implicator.ai is explicit: Z.ai named neither the chip model nor the vendor, and published no power consumption, exact throughput, utilization rate or normalized Nvidia comparison, and "none of the serving results has been independently audited." More to the point, it demonstrated inference only — not training on domestic hardware.
Same week: the US House votes 417–3 to make data centers pay for their own grid upgrades, and Crusoe raises $3.9B at a $30.9B valuation to build more of them
On September 16, the US House passed the Ratepayer Protection Act 417–3, requiring states to "consider" standards for loads above 100 MW: full cost recovery for generation, transmission and distribution upgrades, financial assurances before upgrades begin, and guaranteed recovery if customers exit contracts early. ClearView Energy Partners judged the bill would "largely reinforce" a transition already underway, since only 13 states currently lack large-load tariffs (Utility Dive, 2026-09-16). The next day, Crusoe announced a $3.9 billion Series F at a $30.9 billion post-money valuation led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Nvidia, Founders Fund, GIC, QIA and TPG participating; its truck-transportable modular data centers, Spark, are the headline product, customers include OpenAI (Abilene, Texas), Meta, Microsoft and Oracle, and Jane Street recently signed a five-year, $13 billion contract for GPU and AI infrastructure (TechCrunch, 2026-09-17).
417–3 is the cleanest political signal of the week: in the US, "AI is fine but should not be subsidized by my power bill" is now bipartisan consensus. Hospitals are large, price-sensitive loads that cannot go dark, competing on the same grid with data centers for capacity and for price; how the cost-allocation rules get written lands directly in hospital operating expense.
The bill requires states only to "consider" the standards, not to adopt them, and Utility Dive notes Senate passage before the midterms is unlikely. On Crusoe, TechCrunch provides no megawatt capacity or revenue figures at all.
Pew: American worry about AI flips partisan for the first time — 56% of Democrats against 49% of Republicans, with overall concern up from 37% in 2021 to about half
On September 16, the Pew Research Center published a survey of 3,488 US adults on the American Trends Panel, fielded June 22–28, 2026: 56% of Democrats and 49% of Republicans are more concerned than excited about AI's growing role in daily life, and 75% of Democrats against 68% of Republicans expect AI to reduce the number of jobs. Democratic concern rose from 46% in 2023 to 56%, a 25-point climb since 2021, while Republican concern fell 10 points since 2023. Among liberal Democrats it jumped from 45% in 2023 to 63%. Across all Americans, concern is up from 37% in 2021 to roughly half.
Adoption of medical AI is ultimately settled not by accuracy but by whether patients accept being handled by it. With "more concerned than excited" now at about half of all Americans and rising on both sides of the aisle, disclosure of AI use, informed consent and documented human review stop being compliance overhead and become preconditions for deployment.
The field dates are late June, nearly three months before publication, and the question asks about AI in daily life in general, not in a medical setting. Extrapolating it straight onto patient attitudes toward clinical AI over-reaches.
02 — Product Analysis
Claude Science
Scientific reasoning and agentic R&D workflows · Anthropic (US)
Function and position. Not a structure-prediction model but a reasoning layer stitching together literature, lab records, regulatory documents and internal code. This week's proof-point buyer is Novo Nordisk: the two will test Claude Science on specific R&D workflows covering biological reasoning, drug mechanics and agentic engineering (press release, 2026-09-16).
- Strength : one substrate does both scientific reasoning and software engineering, so pharma need not maintain separate vendors for research and IT — and Anthropic's own R&D Automation Index, with Claude leading 26% of internal R&D, is the in-house stress test of exactly that pattern.
- Concern : no public performance benchmark exists. Novo's release gives no hours saved, hit rate or program count, and Anthropic has published no comparison of Claude Science against specialist models like OpenFold3 or Boltz-2 on any biology or chemistry benchmark. What is verifiable today is who bought it, not how good it is.
GLM-5.3-FlashX
Cheap long-context inference · Zhipu / Z.ai (China)
Function and position. A 320B mixture-of-experts activating 18B per token, with a one-million-token context on the preceding release; the pitch is price plus a supply chain that does not depend on Nvidia. The prior Flash priced at $0.15 per million input tokens and $0.50 output against $1.40 and $4.40 for the full GLM-5.3 (implicator.ai); FlashX claims ~200 tokens/s across about 100,000 domestic accelerators (Pandaily).
- Strength : this price band turns "put the whole chart in" into a budgetable line rather than a pilot, and the supply chain is not hostage to the cadence of US export controls.
- Concern : every number that matters is missing — chip model, vendor, power draw, utilization and a normalized Nvidia comparison are all unpublished, the serving results are unaudited, and only inference was demonstrated, not training. Medicine adds another layer: records are high-sensitivity personal data, and the compliance conditions for cross-border inference in Taiwan and the EU are far harder than the price.
03 — Companies & Competition
| Company | Recent state & numbers | Position & moat |
|---|---|---|
| Anthropic Frontier models plus author of the safety agenda |
Sept 12: Amodei publishes "We Must Pace the Frontier" (Axios). Sept 16: drug discovery deal with Novo Nordisk (BioSpace). Sept 17: discloses Claude leading 26% of internal R&D (Quartz). | The moat is authorship of the safety narrative: it defines the rule, demonstrates the self-restraint and books the pharma contract at once. That is also the weakness — outsiders cannot easily separate the safety claim from the competitive one. |
| OpenAI Largest scale, following on policy |
Sept 15: policy chief Lehane confirms weeks of safety talks with Anthropic and Google DeepMind, backs the FRONTIER Act's independent verification provision, and says the coordination "doesn't need" an antitrust waiver (TechCrunch). | Distribution and user scale are the moat; on this safety round it follows rather than drafts, so if the standards body forms, Anthropic's framing gets there first. |
| Cohere Enterprise models, the outsider |
Sept 13: CEO Aidan Gomez publishes an essay calling the three-lab standards plan "a cartel by any other name," offering four pillars instead: publicly funded evidence-based risk frameworks, mandatory transparency, independent testing confined to genuinely dangerous capabilities, and conflict-free assurance modeled on finance and aviation (AI Weekly). | Smaller than the top three, but outsider status gives its antitrust argument credibility — and the aviation/finance analogy it offers is the language medical regulators can pick up most easily. |
| Novo Nordisk Europe's largest pharma, the buyer |
Sept 16: announces the Anthropic collaboration, with CEO Doustdar aiming to become "the world's most AI-driven healthcare company"; over 67,000 employees globally; no deal value or timeline disclosed (BioSpace). | A buyer's moat is its own data and clinical network. Choosing a general model over a specialist one is a bet that the reasoning layer can be outsourced while the data layer must be held. |
| AbbVie / Astex / OpenBind Proprietary-structure consortium |
Trained on 20,000-plus proprietary structures; on 1,056 test structures it reached high accuracy on more than half, against roughly a third for OpenFold3 and about 40% for Boltz-2; the UK government backs OpenBind with up to £8 million (Nature, 2026-09-14). | The moat is purely data: public databases hold only ~10,000 drug-like-molecule structures. The weakness is that the work is unreviewed and the model unreleased, so nobody outside can reproduce it. |
| Z.ai（智譜） Cheap long context, the non-Nvidia path |
GLM-5.3-FlashX claims ~200 tokens/s across roughly 100,000 domestic accelerators (Pandaily); the prior Flash prices at $0.15 per million input tokens and $0.50 output, with serving results unaudited (implicator.ai). | The moat is cost structure and a domestic supply chain. The weakness is verifiability: without published chip and power figures it cannot clear an enterprise procurement review, let alone a medical one. |
| Crusoe Vertically integrated AI infrastructure |
Sept 17: closes a $3.9B Series F at a $30.9B post-money valuation with Nvidia among the investors; customers include OpenAI (Abilene), Meta, Microsoft and Oracle, plus a five-year $13B contract with Jane Street (TechCrunch). | Truck-transportable Spark modules let it route around siting fights and community opposition, a real moat now that data centers are political. The weakness is capital intensity and a head-on collision with ratepayer politics. |
Today's competitive structure is a game about who writes the rules, not a model leaderboard. The three largest labs are coordinating a standard, the fourth is calling it a cartel, a state government moved in the same week, and the only performance jump with hard evidence behind it came from proprietary data handed over by drug companies rather than from any model vendor. The frontier is migrating toward data at one end and governance at the other; the middle stretch of whose benchmark scores a few points higher was the least important thing this week.
04 — Taiwan Angle
(1) Taiwan is actually ahead of California, but stopped at the level of principles. The Legislative Yuan passed the Artificial Intelligence Basic Act on its third reading on December 23, 2025, establishing seven governance principles — sustainable development and wellbeing, human autonomy, privacy and data governance, information security and safety, transparency and explainability, fairness and non-discrimination, and accountability. The Ministry of Digital Affairs is drafting an AI risk classification framework alongside it, but the announcement publishes no specifics and mentions neither medical applications nor any designated high-risk domain (MODA press release). Set against California's Sept 18 move to write loss-of-control into the definition of a critical safety incident and to study an emergency shutoff for frontier models, Taiwan has principles and no thresholds. Which tier the risk framework assigns to healthcare is the single cell worth watching over the coming year.
(2) The non-Nvidia inference path cuts both ways for Taiwan. Z.ai's claim that GLM-5.3-FlashX runs at ~200 tokens/s across roughly 100,000 domestic accelerators (Pandaily) is a signal Taiwan's advanced-process and AI-server supply chain has to take seriously as an alternative route — even though so far it demonstrates inference only, not training, and has not been independently audited (implicator.ai). The other side: if Taiwanese hospitals want to push whole charts into a long-context model, the price band really is falling — but cross-border inference touches highly sensitive personal data, so the first procurement question is not the per-million-token rate, it is which machine the data lands on.
(3) The politics of who pays for the grid is a question Taiwan will face too. The US House voted 417–3 to make large loads cover their own grid upgrade costs (Utility Dive), and a margin that lopsided means cost allocation is now bipartisan consensus. Taiwan's single grid has to carry advanced-process fabs, AI data centers and hospitals that cannot go dark — and hospitals are both price-sensitive and without bargaining power. As other jurisdictions start writing "who pays for the upgrade" into explicit rules, a Taiwanese large-load review that still runs on case-by-case negotiation leaves the health system as the last constituency anyone counts.
05 — Further Reading
-
Drug firms' secret data supercharge AI protein models — Nature (2026-09-14)
The only result this week with a control arm, and it swaps "the bottleneck is architecture" for "the bottleneck is data" — a reframing that matters more to a hospital's data strategy than any model version bump.
-
Anthropic, OpenAI CEOs call for slowdown in AI development — Axios (2026-09-12)
Read it for the mechanism — the access level granted to external evaluators — rather than the slogan about slowing down; the mechanism is what every subsequent bill will copy.
-
Cohere CEO Gomez calls rival labs' standards plan a 'cartel' — AI Weekly (2026-09-13)
The week's most useful dissent. His line about a regime being rigorous only about risks its authors already built for is the test question medicine should apply to any industry self-regulation proposal.
-
Democrats are now more worried than Republicans about AI and its impact on jobs — Pew Research Center (2026-09-16)
When the constituency to convince is patients rather than regulators, this trend line predicts adoption speed better than any accuracy report.
-
Z.ai served GLM-5.3-Flash entirely on Chinese AI chips — implicator.ai
A model of how to report a major claim while itemizing every figure the claimant withheld. Whether the interest is compute politics or procurement diligence, the methodology alone is worth copying.
06 — References
- Anthropic, OpenAI CEOs call for slowdown in AI development. Axios, 2026-09-12. axios.com
- Cohere CEO Gomez Calls Rival Labs' Standards Plan a 'Cartel'. AI Weekly, 2026-09-13. aiweekly.co
- Drug firms' secret data supercharge AI protein models. Nature, 2026-09-14. nature.com
- OpenAI, Anthropic, Google have been in talks on AI safety for weeks. TechCrunch, 2026-09-15. techcrunch.com
- Novo and Anthropic will collaborate to advance drug discovery with Claude. BioSpace（新聞稿全文 / full press release）, 2026-09-16. biospace.com
- Democrats are now more worried than Republicans about AI and its impact on jobs. Pew Research Center, 2026-09-16. pewresearch.org
- House passes ratepayer protection bill to limit data center cost shifts. Utility Dive, 2026-09-16. utilitydive.com
- Crusoe raises $3.9B to build massive data centers and small modular "AI factories". TechCrunch, 2026-09-17. techcrunch.com
- Zhipu Opens GLM-5.3-FlashX Near 200 Tokens/s on ~100k Domestic Accelerators. Pandaily, 2026-09. pandaily.com
- Z.ai Served GLM-5.3-Flash Entirely on Chinese AI Chips. implicator.ai, 2026-09. implicator.ai
- Anthropic says Claude leads 26% of its AI R&D work. Quartz, 2026-09-18. qz.com
- AI News Today September 18 2026: 14 Biggest Stories（彙整報導 / aggregated roundup）. buildfastwithai, 2026-09-18. blog.buildfastwithai.com
- Governor Newsom issues executive order to accelerate independent oversight and advance the creation of an AI kill switch. Office of the Governor of California, 2026-09-18. gov.ca.gov
- 立法院三讀通過《人工智慧基本法》 構築我國 AI 創新與安全治理基石. 數位發展部 moda, 民國 114 年 12 月. moda.gov.tw
- Google, Anthropic, and OpenAI Unveil Cyber AI Models, Safeguards, and Access Programs（背景脈絡 / background context）. The Hacker News, 2026-09. thehackernews.com