Sign in Subscribe
AI Intelligence

Trump Waves Off AI Slowdown Calls as DeepSeek Ships Cheaper Model Anyway

Vijay Bhagwati

• 6 min read
Share

The president tells the FT "negative forces" are exaggerating AI risk days after Amodei's pacing essay, DeepSeek undercuts everyone with a smaller flash model, SoftBank locks in a $11.87B loan to keep funding OpenAI, and Anthropic discloses Yemeni rebels tried using Claude to help build guided missiles.

September 14, 2026 · 7 min read · Issue #265

Lead

Donald Trump gave his answer to Dario Amodei's pacing essay before the ink was dry. Asked about calls to slow AI development, the president told the Financial Times "you have a lot of negative forces that are bringing it up that shouldn't be bringing it up," dismissing the safety argument as a distraction from the race with China. It's the same posture Bloomberg reported him taking days earlier when he rejected warnings that AI could lead to human extinction, saying the US leads China "by a year" and would be "in a very bad position" if it stopped pushing. Where Amodei asked the industry to prove it can afford to slow down, the White House's answer is that slowing down is the actual risk.

That divergence framed a week in which the infrastructure kept moving regardless of who's right. DeepSeek launched V4.1-Flash, its smallest model yet, built on a new Causal Encoder-Decoder architecture with a 552-billion-parameter backbone and a 1-million-token context window — a smaller, cheaper model shipped days after Washington accused six Chinese firms, DeepSeek among them, of systematically distilling American models to build exactly this kind of product. OpenAI, meanwhile, confirmed it's expanding its Samsung partnership into production and research work on next-generation chips, calling Samsung "one of the largest-scale deployments of ChatGPT globally" — a reminder that the US-South Korea chip relationship is deepening at the same pace as the US-China one is fraying.

Money kept flowing too. SoftBank sealed an upsized $11.87 billion two-year loan facility — above its original $10 billion target, with roughly 20 banks participating — specifically to support its OpenAI investment, a sign that even at the top of the capital stack, funding an AI lab increasingly means underwriting debt, not just equity. And Anthropic disclosed that Iran-backed Houthi rebels in Yemen attempted to use Claude to assist in designing guided missiles, an incident the company says its safety systems caught and blocked, but one that lands the same week the administration is arguing safety concerns are overblown.

Briefs

Anthropic says Yemeni rebels tried to weaponize Claude. A new report finds Iran-backed Houthi rebels attempted to use Claude for technical assistance in developing guided missile systems; Anthropic says its monitoring detected and stopped the misuse before it produced usable output. The disclosure is notable less for the novelty — bad actors probing frontier models for weapons help is now a recurring finding — than for its timing, landing in the same week the White House is publicly minimizing AI safety concerns (New York Post).

SoftBank locks in $11.87B loan to keep funding OpenAI. Sources tell Bloomberg SoftBank sealed a two-year loan facility above its $10 billion target, pulling commitments from roughly 20 banks. It's a structural detail worth sitting with: the vehicle bankrolling one of OpenAI's largest investors is now debt-financed, adding a layer of leverage to a funding chain that was already unusually circular.

Nvidia's Huang defends "put in $1, get $100 back" investment logic as the stock keeps falling. At Goldman Sachs' Communacopia + Technology Conference, Jensen Huang rejected suggestions that Nvidia's roughly $50 billion in AI-lab investments and up to $105 billion in guarantees tied to OpenAI's Ohio data centre amount to circular financing. Nvidia shares fell a third straight session anyway, down 2.37% to $218.36, as investors including Allspring Global's Gary Tan flagged that "capital is increasingly being used to fund future AI customers and infrastructure deployments" (Invezz).

Ant International, Visa and Mastercard team up on an AI-agent payment standard. The companies are building interoperable "know-your-agent" standards — verifying agent identity, monitoring behavior, and letting an agent registered with one network transact across all three — betting on McKinsey's projection that AI agents will handle $3-5 trillion of consumer commerce by 2030. Ant's chief innovation officer Jiang-Ming Yang told CNBC the point is blunt: "we need to make sure people feel safe" given agents' hallucination risk (CNBC).

China/East Asia

DeepSeek's V4.1-Flash launch is the cheaper, smaller sibling to its flagship line, built on a new Causal Encoder-Decoder architecture and shipping with a 552B-parameter backbone and 1M-token context window (Reuters, via Techmeme) — arriving days after a joint FBI-NSA-CISA advisory named DeepSeek among six Chinese firms accused of systematically distilling US models since 2024. Separately, Moonshot AI, maker of the Kimi model family, is exploring dual listings in Hong Kong and Shanghai to raise capital, according to sources cited by the South China Morning Post, a move that would let the company tap both onshore and offshore investor pools as China's AI financing race accelerates (SCMP). And OpenAI confirmed its Samsung partnership is expanding beyond deployment into production and research collaboration on Samsung's next-generation chips, with OpenAI describing Samsung as "one of the largest-scale deployments of ChatGPT globally" (Reuters, via Techmeme) — a US-aligned chip relationship deepening in the same news cycle as the US-China distillation fight.

India

The Ant-Visa-Mastercard agentic payments standard is the most direct India-relevant story of the week: India's digital-payments ecosystem, dominated by UPI and e-wallet volume rather than card rails, is exactly the kind of market where interoperable "know-your-agent" frameworks will determine whether AI shopping agents can transact at all. Ant International's Alipay+ network already links to more than 50 e-wallets globally, and Worldpay data cited in the reporting shows digital wallets already account for 56% of global e-commerce value — a base India's own UPI-linked wallets are part of, making the emerging KYA standard a live question for how Indian consumers' AI agents will be authorized to spend on their behalf (CNBC).

Europe

Mistral AI's Arthur Mensch is deepening his relationship with Emmanuel Macron's government as the French champion works to secure political and financial backing amid an increasingly capital-intensive global race, according to Politico's reporting on the company's positioning inside French AI policy circles (Politico). It's the European counterpoint to the US funding-and-leverage story playing out with SoftBank and Nvidia: Europe's leading model lab is betting its capital strategy runs through Paris rather than Wall Street.

The View

Two governments are now making opposite bets on the same underlying fact pattern. Anthropic's CEO argues capability has outrun the industry's ability to vouch for its own alignment work, and wants coordinated, capability-linked checkpoints. The White House's answer, delivered days later through the FT and previously through Bloomberg, is that any such caution is itself the strategic risk, because it cedes ground to China. Both positions treat "the gap with China" as the load-bearing variable, and this week supplied a data point that complicates the argument for pacing more than it helps it: DeepSeek didn't slow down or get shut out by export controls — it shipped a smaller, cheaper, technically competitive model days after being named in a federal distillation advisory. If the US case for restraint depends on a lead that's wide enough to spend on caution, DeepSeek's Flash release is exactly the kind of evidence Washington's hawks will point to as proof the lead isn't there to spend.

The Miss

Almost no outlet connected Anthropic's Yemen/Claude disclosure to the pacing debate the same company's CEO started days earlier. Amodei's essay argued for third-party evaluators and capability checkpoints partly because rogue use cases are becoming more sophisticated and harder to catch pre-deployment; the Houthi missile-assistance attempt is precisely the kind of incident his framework is designed to prevent — and Anthropic's safety systems reportedly did catch it. That's arguably the single best piece of evidence Amodei has for his own argument this week, and it went almost entirely unremarked as a data point in that debate rather than as a standalone security story.

Pull Quotes

"You have a lot of negative forces that are bringing it up that shouldn't be bringing it up." — President Trump, on calls for an AI slowdown, to the Financial Times

"It's not circular because we put a little bit of money in, and a lot of money comes back... I put in one, and a hundred comes back." — Jensen Huang, at Goldman Sachs' Communacopia + Technology Conference, via Invezz

"Interoperability across Know-Your-Agent frameworks is essential to making agentic commerce work at scale." — Pablo Fourez, Chief Digital Officer, Mastercard, via CNBC

Out

Trump told the FT the slowdown crowd is exaggerating. DeepSeek answered with a smaller, cheaper model days after being named in a federal distillation alert — the fastest rebuttal anyone's given his argument all week.