Sign in Subscribe
AI Intelligence

Anthropic Weighs a New Model Days After Calling for a Slowdown

Vijay Bhagwati

• 8 min read
Share

Sources say Anthropic is considering a fresh model release to answer OpenAI's Astra momentum ahead of its own IPO, Huawei says it can't build Ascend chips fast enough to meet Chinese demand, Cohere and Aleph Alpha close a $20 billion merger, and Newsom orders California to study an AI "kill switch."

September 19, 2026 · 8 min read · Issue #267

Lead

Anthropic is weighing whether to ship a new frontier model to counter the momentum OpenAI has built since launching Astra, according to people familiar with the matter, with the decision landing in the same window as the company's push toward a roughly $2 trillion IPO. The timing is awkward by design or by accident, depending on which Anthropic you're reading about this week. Days earlier, CEO Dario Amodei published an essay arguing the industry should slow down and submit to third-party safety verification. Days after that, the company disclosed a bioweapons-uplift report on its own models that drew public skepticism from outside scientists. Now come reports that Anthropic is considering doing the one thing its own CEO's essay was structured to argue against: shipping fast to keep pace.

The contradiction is easier to parse once you separate Anthropic's public safety posture from its commercial one. Reuters also reported this week that Anthropic has quietly built a physical biology lab in the Bay Area, beyond the "in silico" work it has talked about publicly — wet-lab space where the company wants Claude to eventually direct robotic units through experiments with limited human oversight. Eric Kauderer-Abrams, Anthropic's head of life sciences, told Reuters the company is "in the very early innings of using AI to automate the execution of lab work." The lab sits alongside a $400 million acquisition of Coefficient Bio, a Model Hardware Standard for AI-operated lab equipment, and a board seat for Novartis CEO Vas Narasimhan. None of this reads as a company pacing itself; it reads as a company building out every commercial surface it can reach — drug discovery, agentic coding, enterprise deployment — while its CEO argues in public that the industry as a whole needs to slow down.

That's not necessarily hypocrisy so much as a bet that safety credibility and commercial aggression aren't mutually exclusive, as long as the former is procedural (external evaluators, disclosure frameworks) rather than a literal brake on shipping. Whether that bet survives contact with an IPO roadshow is the open question. Investors underwriting a $2 trillion valuation are not going to reward "we chose not to ship a competitive model." The forward-looking test is simple: does Anthropic's next model launch come bundled with the third-party verification Amodei has been publicly requesting from everyone else, or does it ship on the same unverified basis as every prior release, essay notwithstanding.

Briefs

Huawei says it can't produce enough Ascend chips to meet Chinese demand. Rotating chairman Eric Xu told reporters at Huawei Connect in Shanghai the company is "limiting overseas sales" because domestic orders already outstrip capacity, and argued Ascend's China market share "should be bigger than Nvidia's" — though Huawei provided no data to support the claim. The company detailed an accelerated roadmap: the Ascend 960DT chip now due Q1 2027, three quarters earlier than planned, plus a new "Peerium" architecture designed to eventually link up to 1 million AI processors into a single system. Nvidia's CUDA software moat remains the one advantage Huawei hasn't matched (Reuters).

Cohere and Aleph Alpha close their long-flagged $20 billion merger. The combined company keeps the Cohere name and runs dual headquarters in Toronto and Berlin; Aleph Alpha co-chief Ilhan Scheer becomes Cohere's COO. Schwarz Group, the German retail conglomerate that owns Lidl, is putting in €500 million directly and up to €13 billion into a 100,000-chip German data-centre campus through its cloud arm StackIT. Cohere reported $240 million in annual recurring revenue last year; Aleph Alpha reported under €1 million in 2023. The deal arrives a week after Mistral closed €3 billion at a valuation above €21 billion — Europe's AI sector is consolidating capital into fewer, better-funded vehicles rather than spreading it thin (Reuters).

Newsom orders a two-month study into an AI "kill switch" for California. The governor's executive order creates a working group tasked with recommending guardrails, including whether frontier-model developers should be required to build a kill switch and whether independent third parties should draft companies' safety plans. It's California testing, at state level, the same procedural-verification idea Amodei has been pitching at the federal level — with a two-month clock rather than a essay's worth of advocacy (Bloomberg).

Manus, the Chinese agentic-AI startup Beijing forced to split from Meta, is raising toward a $4 billion valuation. The planned $500 million round would double Manus's valuation and make it China's most valuable agentic-AI startup, according to people familiar with the talks. Existing backers Tencent, HSG, and ZhenFund are involved; new investors remain undisclosed. The raise is the company's first since the forced Meta breakup — a geopolitical setback that, on this valuation trajectory, doesn't appear to have cost it much (Bloomberg).

US and Chinese security researchers propose nuclear-style red lines for military AI. Brookings' Melanie Sisson and Fudan University's Tianjiao Jiang, part of a standing US-China AI dialogue, are urging both governments to bar AI from autonomously attacking nuclear command systems and to set up a dedicated hotline for AI-related military incidents. The proposals land ahead of a planned September 24 Trump-Xi meeting, though neither government has endorsed them, and Johns Hopkins' Carla Freeman notes the existing US-China military hotline has already failed once — Chinese officials didn't answer calls during the 2023 spy-balloon incident (Reuters).

China/East Asia

Huawei's chip disclosure this week is really two stories folded together: a supply constraint and a strategic bet. Eric Xu's admission that Huawei can't meet domestic demand undercuts the market-share claim in the same breath — a company genuinely leading on share wouldn't need to ration overseas sales to cover a shortfall at home. But the roadmap acceleration is real regardless of the share question: the 960DT chip moving up three quarters, a "Peerium" architecture aimed at linking a million processors, and 5,200 monthly active Ascend developers training more than 40 models on the platform. Huawei is compensating for weaker individual chips by wiring more of them together — clusters of 100,000 processors are already standard for large training runs, and Huawei's new UnifiedBus technology is built specifically to cut the 40%-plus of training time that inter-chip communication currently eats. Separately, Manus's march toward a $4 billion valuation is the clearest signal yet that Beijing's forced separation from Meta was a political cost, not a commercial one — Tencent and its existing backers are doubling down rather than walking away.

India

Modi's chip-industry event drew fresh capital commitments from Applied Materials and Lam Research toward India's $30 billion semiconductor build-out, with the prime minister pitching the country's engineering-graduate pipeline and policy environment directly to US toolmakers. The pitch is aimed squarely at the equipment layer — the machines that make chips — rather than the fabs themselves, a narrower and more achievable target than competing with TSMC on leading-edge manufacturing. That framing matters next to this week's other India-relevant supply chain move: Nexperia's agreement to partner with Tata on chip production and packaging inside India, adding a second concrete manufacturing commitment to the same week Modi was courting investors on stage. Neither deal makes India a frontier-model competitor, but both are building toward the same goal Europe and China are separately pursuing — controlling enough of the physical compute stack that AI policy isn't entirely hostage to decisions made in Washington or Taipei.

Europe

Cohere and Aleph Alpha's completed merger is the more concrete of two European consolidation stories this week, and it's worth reading against Mistral's €3 billion raise from the prior week rather than in isolation. Both moves point the same direction: European AI capital is concentrating into fewer, larger, better-capitalized vehicles rather than spreading across a longer tail of national champions. Aleph Alpha, once billed as Germany's answer to OpenAI, spent the past year retreating from frontier-model ambitions before this deal — its 2023 revenue was under €1 million against Cohere's $240 million ARR, a gap wide enough that this reads less like a merger of equals than an acquisition with generous branding. What Cohere gets in return is Schwarz Group's balance sheet: €500 million in direct investment plus up to €13 billion in committed compute through StackIT's German data-centre buildout, giving the combined company sovereign infrastructure it didn't previously have at that scale. The pitch — AI "governable enough to trust," in CEO Aidan Gomez's words — is explicitly an enterprise and government sale, not a consumer one, which is the same lane Mistral has been building in with Airbus, ASML, and HSBC.

The View

Three governments spent this week converging on the same procedural fix — external verification of AI safety claims — while the company that made the argument for it is quietly weighing whether to ship a model without waiting for anyone to build that verification apparatus. Newsom's kill-switch study, the Brookings-Fudan nuclear red lines, and Amodei's original pacing essay all share a structure: none of them ask labs to stop building, only to submit what they build to outside check. That's a lower bar than a slowdown, which is presumably why it's attracting agreement from governments that otherwise disagree about almost everything in AI policy. What undercuts the momentum is Anthropic's own week: a company that spent recent weeks publicly modeling what responsible pacing looks like is now reportedly considering a competitive model release with no indication that release would come bundled with the third-party evaluation its CEO has been requesting from the rest of the industry. Procedural safety frameworks only work if the company advocating for them is also the first to submit to them. An IPO roadshow is a specific kind of pressure that doesn't reward waiting for anyone's evaluators.

The Miss

A new arXiv paper, "Quantifying Overclaiming Propensity in Frontier LLM Agents," lands directly on top of this week's transparency debate without anyone connecting the two. The paper measures how often frontier agents state confidence in task outcomes that their own execution traces don't support — a mechanistic version of exactly the disclosure-gap problem Anthropic's safety team and outside critics have been arguing about qualitatively for months. If a model's overclaiming propensity is quantifiable and comparable across labs, it's a more useful transparency metric than any voluntary disclosure framework a company writes about itself, and it arrived the same week Anthropic's own model-release calculus became the story.

Pull Quotes

"We don't have a plan to expand into the international market in a fully-fledged way." — Eric Xu, Huawei rotating chairman, via Reuters

"Powerful enough to compete, but secure and governable enough to trust." — Aidan Gomez, Cohere CEO, on the merged company's pitch, via Reuters

"We're in the very early innings of using AI to automate the execution of lab work." — Eric Kauderer-Abrams, Anthropic head of life sciences, via Reuters

"No one in the Chinese system wants to communicate with the United States until the top leadership...has decided on a response, which may take days if not longer." — Carla Freeman, Johns Hopkins School of Advanced International Studies, on the limits of a proposed US-China AI hotline, via Reuters

Out

Amodei spent recent weeks telling the industry to slow down and submit to outside checks. His own company is reportedly weighing a model release with no such check attached — the fastest rebuttal anyone's given his argument all week is coming from inside the building.