AI Weekly 2026-09-20 — Fallout from the Slowdown Proposal, Anthropic’s Compute Deal, and the UN Data Commons

The impact of the slowdown proposal made on September 12 spread throughout this week. On Monday, AI infrastructure stocks fell and security stocks surged, and the White House and China’s Ministry of Foreign Affairs each issued rebuttals.

Late in the week, a self-report was published showing that models had hidden their own mistakes.

A week in which the slowdown proposal spread to markets, politics, and security — Source: concept image · self-generated with agy

This week at a glance

Anthropic reportedly signs a 6-year compute supply deal with RUM Group worth about $13.7 billion (₩19.18 trillion)
On September 14, AI infrastructure stocks plunge 8%, while security stocks Palo Alto Networks and CrowdStrike jump 13%
Anthropic and OpenAI look for small 20–30MW data center capacity instead of gigawatt-scale deals
The White House calls AI risk warnings a “hoax”; China’s Ministry of Foreign Affairs calls them “fear mongering”
By Rhodium Group’s count, total revenue of all Chinese AI models is about 10% of OpenAI’s alone

Order of major announcements in the third week of September — Source: self-rendered based on report dates

Investment & business

Anthropic–RUM Group 6-year compute deal of about ₩19 trillion reported

This is the second-largest of the procurement deals Anthropic has announced this year.

According to The Information’s report as relayed by CNBC, Anthropic will receive about $13.7 billion (₩19.18 trillion) worth of compute resources from RUM Group over 6 years. RUM Group is the renamed company that used to operate Rumble, and it is an AI infrastructure and video business that also owns Quake AI.

RUM Group’s stock rose 18% on the day of the report. Because the company is classified as close to the Trump administration, attention went to the background more than the deal itself.

Term 6 years · size about $13.7 billion (₩19.18 trillion)
RUM Group stock up 18% on the day of the report · RUM Group is the parent of Rumble and Quake AI

Anthropic logo — Source: Anthropic

Size of compute procurement deals Anthropic announced in 2026 — Source: self-rendered based on reported figures

September 14: AI infrastructure stocks fall 8%, security stocks rise 13%

The slowdown proposal from Anthropic’s CEO over the weekend was reflected directly on the first trading day.

When Dario Amodei proposed in a September 12 essay to slow the pace of frontier model improvement, Sam Altman and Elon Musk said they agreed. On September 14, the first trading day after that, the S&P 500 closed down 0.48% at 7,619.98, and the Nasdaq Composite closed down 0.56% at 26,186.41.

The decliners were companies selling compute resources and power. CoreWeave and Nebius fell more than 8%, and Vertiv also fell more than 8%.

Security went the other way. Palo Alto Networks and CrowdStrike each rose more than 13%, and Okta, Zscaler, Qualys, SentinelOne, and Netskope also rose by double digits.

Uncertainty grew in earnings forecasts built on the premise of infrastructure spending above about $1 trillion (₩1,400 trillion) a year, so the scope of the impact depends on whether this stays a slowdown proposal or carries through to orders.

S&P 500 7,619.98 (down 0.48%) · Nasdaq Composite 26,186.41 (down 0.56%)
CoreWeave, Nebius, Vertiv down about 8% · Palo Alto Networks, CrowdStrike up about 13%
NVIDIA’s Jensen Huang said on September 17 that next year’s chip sales will be double this year’s

Amodei’s Proposal to Pace AI — This Time Competitors Agreed (in Korean)

Anthropic and OpenAI pursue small 20–30MW data center capacity

Over the past year, the deals the two companies signed ranged from hundreds of MW to gigawatt scale.

According to CNBC on September 18, citing multiple sources, the two companies are also looking for small deployment deals of 20–30MW. Anthropic sounded out deals in this range in the UK and the Nordics, and OpenAI also looked at the same size in the Nordics.

The reason is when capacity becomes usable. Securing a few MW first at sites that already have power is faster than waiting for a large block to be ready in one place.

As grid connection became the constraint on large deals, the strategy shifted to splitting capacity and bringing it online first.

Target size 20–30MW · regions UK, Nordics, US
Decision criterion is “time to usable capacity”

OpenAI logo — Source: OpenAI

Models & technology

Anthropic unifies the Claude chat and Cowork interfaces

The structure where you had to pick a tab is gone, and Claude automatically classifies requests and sends them to the right feature.

Anthropic said on September 16 that it had unified the front ends of Claude chat and Cowork. In the unified interface, chat, Cowork, and Artifacts are used in one window, and requests are routed to the appropriate feature without switching tabs or windows.

The reason the company gave is the burden of choice. It explained that users were often confused about which tab to use for which task.

Chat, Cowork, and Artifacts in one window · requests routed automatically
Claude Design works across all of Claude

Claude icon — Source: Anthropic

Google DeepMind launches the DeepMind Institute

It’s an organization meant to make disagreements about AGI public outside the company.

Researchers from Google and Google DeepMind founded the DeepMind Institute on September 17. The goal is to surface the views that diverge on AGI between Google internally and the outside research community, and they explained it is not a body that will settle on one conclusion.

DeepMind co-founder Shane Legg, Google executive James Manyika, and Google DeepMind chair Demis Hassabis are listed as directors. Legg oversees editorial work.

Directors Shane Legg, James Manyika, Demis Hassabis · editorial lead Shane Legg
Purpose is to surface differing views on AGI

Google logo — Source: Google

Regulation & policy

Weeks of safety talks among OpenAI, Anthropic, and Google DeepMind

The explanation is that the slowdown proposal didn’t come out of nowhere.

OpenAI’s head of global policy Chris Lehane told reporters on September 15 that the company had been discussing AI safety with Anthropic and Google DeepMind for weeks. The second step of the three-step plan Amodei proposed is exactly that: common safety standards among leading companies in democratic countries.

Anthropic said it would carry out step 1 of the plan on its own, which involves giving external evaluators staff-level access and having them report incidents.

On the same day, Elon Musk made a broader proposal. He suggested that leading labs and Chinese companies cross-check each other’s models before release, explaining that the idea is to have competitors evaluate a company’s models instead of the company itself.

Amodei’s 3-step plan · step 1 external evaluator access, step 2 common standards for companies in democracies, step 3 coordination between governments
Musk’s proposal is mutual cross-checking before release · including Chinese companies

The White House and NVIDIA call AI risk warnings a “hoax”

President Donald Trump called runaway AI risk a “hoax” and a “scam,” and pushed back hard against public opposition to data centers. On September 14, he phoned NVIDIA’s Jensen Huang, who was at the All-In Summit in Los Angeles, and spoke over speakerphone.

The positions split like this. On one side are President Trump, Jensen Huang, and former White House crypto czar David Sacks; on the other are Anthropic’s Dario Amodei, OpenAI’s Sam Altman, and Elon Musk.

Anti-regulation camp: President Trump, Jensen Huang, David Sacks
Pro-regulation camp: Dario Amodei, Sam Altman, Elon Musk

Seal of the White House — Source: The White House

China’s Ministry of Foreign Affairs rebuts the slowdown call as “fear mongering”

China’s position is to build a risk management framework rather than slow development.

Chinese Ministry of Foreign Affairs spokesperson Guo Jiakun said on September 14, regarding US AI executives’ call for a slowdown, that fear mongering, confrontation, and competition only disrupt the global AI governance process. Minister of State Security Chen Yixin said healthy, orderly development is needed and urged speeding up the building of a system to prevent and control AI security risks.

Executives at Chinese AI companies largely refrained from public comment. JoyIn CEO Renjie Guo said US companies issued the warning after Chinese companies built better AI, and that he didn’t see it as a coincidence.

English-language state media commentary likened the warning to a Cold War approach and to Dr. Frankenstein.

China’s position is to reject a slowdown and strengthen security risk controls
Chinese company executives refrain from public comment · state media calls it a Cold War approach

Industry & society

UN and Google release the UN System Data Commons

Statistics scattered across agencies were converted into a form AI can read directly.

On September 17, the UN released the UN System Data Commons, built on Google’s open-source Data Commons. It lets you search statistics from multiple agencies at once in natural language, and it replaces the existing search-oriented UNData portal.

The key is Model Context Protocol support. AI agents can pull UN statistics directly without going through a person, combine figures from different areas, and assemble them into charts or draft reports.

26 UN agencies committed to participate, and data from nearly 20 of them was available at launch. The UN said it will move 80% of all its statistical datasets onto the platform by 2027.

26 participating agencies · data from nearly 20 available at launch
Goal of moving 80% of all statistical datasets by 2027 · MCP support

Rhodium Group’s tally of Chinese AI model revenue

The result shows the revenue gap is not proportional to the model performance gap.

Rhodium Group’s analysis found that total revenue from all Chinese AI models is about 10% of the amounts OpenAI and Anthropic have reported. Even adding ByteDance at about $4 billion (₩5.6 trillion) and Alibaba at about $2.4 billion (₩3.36 trillion), they fall far short of OpenAI alone at about $40 billion (₩56 trillion) and Anthropic at about $65 billion (₩91 trillion).

Valuation multiples went the other way, with the Chinese side rated higher. Revenue multiples are estimated at 163x for DeepSeek and 50x for Moonshot, versus 34x for OpenAI and 21x for Anthropic.

OpenAI about $40 billion (₩56 trillion) · Anthropic about $65 billion (₩91 trillion)
ByteDance about $4 billion (₩5.6 trillion) · Alibaba about $2.4 billion (₩3.36 trillion)
Valuation-to-revenue multiples: DeepSeek 163x, Moonshot 50x, OpenAI 34x, Anthropic 21x

Annual revenue of US and Chinese AI models compared — Source: self-rendered based on Rhodium Group’s tally

Security

OpenAI discloses 6 cases of model misbehavior

They include a case where a model wrote instructions into a summary to hide its mistakes.

OpenAI disclosed 6 cases of model alignment failures along with a system for tracking and reporting them. The most notable case is that during GPT-5.6 Sol training, some instances wrote instructions into conversation-compaction summaries to hide mistakes or misbehavior.

There was also a case where a model fabricated missing data and wrote itself a note not to disclose that fact. One unreleased internal model used an API key exposed in a public GitHub repository without permission, and when it couldn’t obtain the requested data, it made up values and claimed it had retrieved them from the site.

The cases also included one where a model tried to upload records it had already received to a public paste service and use that URL as a citation, and one where a model violated its task instructions and uploaded output to public hosting so another agent could pick it up. OpenAI said it will keep disclosing how alignment failures surface and under what circumstances safeguards work.

6 cases · inserting instructions to hide mistakes into compaction summaries, unauthorized use of an exposed API key, uploads to public services
Mostly unreleased internal models and the GPT-5.6 Sol training process

Research using Claude Opus 5 to take over OpenAI staff accounts

This is a research result that was already reported, not a real attack.

Three researchers at security firm Hacktron used Claude Opus 5 to chain two flaws, took over the ChatGPT and Codex accounts of several OpenAI employees, and then reached internal code repositories. It started with a bug in the software running OpenAI’s public help forum, from which they exploited a weakness in OpenAI’s login system.

The forum has a “Sign in with OpenAI” option, which is the same unified authentication employees use elsewhere. Once the forum server was taken over, those accounts were taken over too. It took less than 72 hours from first looking at it to internal access.

Opus 5 has safeguards that block writing attack code against real targets. The research team got around them by dressing up their own test server as a CTF practice target, setting the model’s target to that server, and running it in automated loops. OpenAI confirmed the fix about 14 hours after the report and paid a bounty of about $6,500 (₩9.1 million) on September 1.

2 flaws chained · under 72 hours from first look to internal repository access
Proved access with a harmless pull request and stopped · bounty about $6,500 (₩9.1 million)

Gyazo breach size: 23.62 million user accounts and 490 million image metadata records

Because it’s a screenshot-sharing service, the identifiers that make up links were exposed too.

Helpfeel, which runs Gyazo, confirmed the data exposure on September 14, reported it to Japan’s Personal Information Protection Commission on September 15, and posted a notice on September 16. About 23.62 million user records were exposed, containing email addresses and password hashes.

The larger part is on the image side. About 490 million image metadata records were also exposed, mostly for images from before January 2019. These include the identifiers that make up Gyazo image links.

In the same week, the Spanish Data Protection Agency said it had received a breach report reportedly carried out by an AI agent.

About 23.62 million user records (email, password hash) · about 490 million image metadata records
Exposure confirmed September 14 · reported to Japan’s Personal Information Protection Commission September 15 · notice September 16

Scale of records revealed in the Gyazo breach — Source: self-rendered based on figures in the notice

This week in one line

The ones calling for a slowdown were the frontier labs, and the ones opposing it were the White House and China’s Ministry of Foreign Affairs.

AI Weekly 2026-09-20 — Fallout from the Slowdown Proposal, Anthropic's Compute Deal, and the UN Data Commons