DAILY INTELLIGENCE BRIEF

The Vault — AI Edition

Monday, September 21, 2026
19 stories
18,746 words
86 min read

Business

5
01

Harvey’s Gross Margins Hit -50% on Frontier Model Bills. It Rebuilt on Kimi K3 and Went Positive.

Harvey, the legal AI startup valued at $15.6 billion, spent the first half of 2026 losing money on every additional query its customers ran.

5 minRead →
02

SoftBank Shelved a $50 Billion Data-Center IPO and Went to the Junk-Bond Market for $11 Billion the Same Day

SoftBank spent Monday doing two things that do not usually happen on the same day. In the morning, its US data-center developer SB Energy quietly pushed

5 minRead →
03

Shopify Turned On Agentic Checkout for Every Store by Default. The Agent Is Meta’s.

Every Shopify store woke up on Monday able to sell to a robot. Shopify announced Sept. 21 that Meta’s two-week-old Muse agent can now search

4 minRead →
04

Tencent Just Valued a 100-Person Startup at $1.42 Billion. It Has Never Pretrained a Model.

Naive AI has fewer than 100 employees, no public product, and an unresolved intellectual-property fight with its founder’s last employer.

5 minRead →
05

A Free iPhone App Added Roughly $200 Billion to Arm, Intel and AMD in One Session. Intel Says It Can Fill Only Half Its Orders.

Nobody at Arm, Intel or AMD announced anything on Monday. No earnings, no guidance, no new contract. What moved the three biggest names in CPUs was a consumer app

4 minRead →

Llms Genai

5
06

SpaceXAI Shipped Grok 4.7 at $2 In, $6 Out. On Harvey’s Legal Benchmark It Scored 19.6%.

SpaceXAI released Grok 4.7 on Monday and, for the third generation running, left the price alone: $2 per million input tokens and $6 per million output tokens...

4 minRead →
07

Xiaomi Put an MIT-Licensed Trillion-Parameter Model on Hugging Face and Says It Nearly Matches Opus 5. Most of the Scores Come From Its Own Harness.

Xiaomi, a company most people still associate with phones and electric cars, pushed two frontier-class models onto Hugging Face on Monday under the MIT license...

4 minRead →
08

Qwen-Image-2.1 Renders Native 2048x2048. It Is Also the First Qwen-Image You Cannot Use Commercially Without Asking.

Alibaba’s Qwen team has shipped the most capable image model it has ever released, and for the first time it has put a gate in front of anyone who wants to make money with it.

4 minRead →
09

Qwen Cut Simultaneous-Interpretation Lag From 2.8 to 2.3 Seconds Across 60 Languages. It Clones Your Voice While It Does It.

Half a second does not sound like much until you are the person waiting for it. Alibaba’s Qwen team says its new Qwen3.8-LiveTranslate model has trimmed the average lag...

4 minRead →
10

Z.ai’s Coding Tool Tried 564 Times to Upload a 313MB Archive of One Developer’s Repo. Now the Client Is Open Source.

A developer freeing up disk space on a 256GB MacBook Air found the file that would force one of China’s most prominent AI labs to publish the source code of its flagship coding tool.

5 minRead →

Research

4
11

Tim Dettmers Says a 125B Qwen Now Runs on One 24GB GPU, and a 550B DeepSeek on a MacBook

Tim Dettmers, the researcher behind bitsandbytes and QLoRA, says his lab has built an inference framework that runs Alibaba’s 125B Qwen on a single 24GB GPU.

4 minRead →
12

Any App on a Mac Can Steal Your Meta Muse Token. Patrick Wardle Used One to Locate a Linked iPhone in Barcelona.

Thirteen days after Meta launched Muse, and four days after it shipped the Mac client, Patrick Wardle showed that any program running on a Mac can quietly hijack the personal agent...

4 minRead →
13

The Jobs Most Exposed to AI Got the Biggest Raises Since 2021. They Are Also the Ones Being Cut.

The jobs generative AI is supposed to hollow out are the same jobs employers are paying the most to fill. New data from Indeed Hiring Lab, published Sept

5 minRead →
14

Google Pulled Agent Task State Out of Kubernetes Because etcd Could Not Take the Churn

Google has decided that Kubernetes, the system it invented to run the world’s containers, is the wrong place to keep an AI agent’s to-do list.

5 minRead →

Policy

5
15

The UN’s 40-Expert AI Panel Issued Its First Brief. It Says the Traditional Model of Safeguarding Is Unravelling.

The world’s first UN-backed scientific body on artificial intelligence used its first thematic brief on Monday to tell governments something blunt...

5 minRead →
16

British Columbia Says OpenAI’s Safety Team Flagged the Tumbler Ridge Shooter Eight Months Early. Now the Province Is Suing.

British Columbia became the first government outside the United States to sue OpenAI on Monday, filing a complaint in San Francisco federal court...

5 minRead →
17

Newsom Vetoed Data-Center Oversight Last Year. On Monday He Signed Seven Bills of It.

Eleven months after Gavin Newsom vetoed a bill that would have made data centers disclose how much water they use, the California governor signed seven of them on Monday...

5 minRead →
18

Two Days Before Altman Briefs the Security Council, OpenAI Put in Writing That It Will Not Pursue Fully Autonomous Self-Improvement

Two days before Sam Altman walks into the United Nations Security Council chamber, his company put a sentence on the record that the AI safety community...

4 minRead →
19

Politico Reconstructed the 19 Days the White House Forced Anthropic’s Fable Offline Over an Amazon-Flagged Jailbreak

For 19 days this summer, the most capable AI model ever sold to the public did not exist. On Sunday, Politico’s magazine desk published a

5 minRead →