October 19, 2026 · AI & Tech Daily Digest
A roundup of AI and technology developments for October 19, 2026, with summaries, links, and commentary.
I. Policy and Regulation
1. DHS publishes an order removing DeepSeek from civilian federal systems (Policy)
Summary:
The BBC reported on October 19 that the U.S. Department of Homeland Security has ordered DeepSeek's AI technology removed from most government systems, barring civilian agencies from using DeepSeek or technology built with the aid of its models. The order was made public on Monday, has been in effect since the end of July, covers departments including Treasury, State, and Justice, and excludes Defense Department agencies. The BBC said the notice did not spell out the security concerns, and that DHS, DeepSeek, and the Chinese embassy in Washington had not commented.
Links:
- BBC — US bans DeepSeek AI from government use as tech tension with China continues
- The Wall Street Journal — DHS Orders Federal Civilian Agencies to Remove China's DeepSeek AI
Commentary:
Monday's news is a ban that has applied since late July, written broadly enough to cover technology built with DeepSeek's models, with the Defense Department left outside it.
2. Statistics bureau: digital-product manufacturing accounted for more than a third of industrial growth (Industry)
Summary:
National Bureau of Statistics spokesperson Wang Guanhua said at a State Council Information Office briefing on October 19 that AI moved further into industry in the first three quarters. Value added in large-scale digital-product manufacturing contributed more than one-third of growth in large-scale industrial output, while integrated-circuit and optical-fiber output rose 23.3 percent and 18.6 percent. Value added in information transmission, software, and IT services rose 10.4 percent year on year, and retail sales of wearable smart devices at large retailers rose more than 70 percent. Investment in electronic-circuit manufacturing, integrated-circuit manufacturing, and information services rose 58.7 percent, 14.2 percent, and 20.2 percent, and electronic-component exports rose 70 percent.
Links:
Commentary:
The figures tie AI to industrial growth, wearable retail, and component exports for the first three quarters, not to a single day's orders.
3. Nine agencies publish a commerce-tech opinion as MIIT cites a 2025 industry figure (Policy)
Summary:
Xinhua on October 19 published an opinion from the Ministry of Commerce and eight other agencies on using technology to upgrade commerce. The document is dated October 13. It calls for AI and embodied intelligence in product supply, commercial agents, and trade logistics, and for grading those agents by scenario and exploring a negative list for data use. At the same day's briefing, Wei Wei, director-general of MIIT's science and technology department, said measured core AI industry scale in 2025 exceeded 1.2 trillion yuan and that the ministry would push AI deeper into new industrialization.
Links:
- Xinhua — Opinion of nine agencies on accelerating technology-led commerce
- Xinhua — MIIT: push AI deeper into new industrialization
Commentary:
The 1.2 trillion yuan figure is a 2025 measurement of the core industry; what landed on paper Monday is agent grading and a data negative list.
II. Chips, Compute, and Memory
4. Amazon announces TorchNeuron so PyTorch code can run on Trainium (Chips)
Summary:
Amazon on October 19 announced TorchNeuron, which runs existing PyTorch code on AWS Trainium after a single configuration change and without a rewrite. The company said about 85 percent of AI researchers and developers write in PyTorch, that the software supports Trainium2 and Trainium3, and that it will be open-sourced under Apache 2.0 at PyTorch Conference North America on October 20. Reuters quoted Colin Brace, vice president of software at Annapurna Labs, as saying more than 30 early customers are using it across robotics, world models, and voice. Reuters said Amazon shares were up about 5 percent in morning trading and that Morgan Stanley named the company a top pick.
Links:
- About Amazon — TorchNeuron runs PyTorch on AWS Trainium chips
- Reuters — Amazon says its AI chips will work with popular coding language, challenging Nvidia
Commentary:
Amazon is aiming at the software lock-in around CUDA; the open-source drop is scheduled for Tuesday, and the 5 percent move is a morning quote.
5. Tenstorrent closes a $1.5 billion Series E at a $6.2 billion valuation (Funding)
Summary:
Tenstorrent on October 19 said it closed a $1.5 billion Series E at a $6.2 billion post-money valuation, led by Trace Capital and Innovation X. The capital will scale Galaxy servers for open-weight models and advance the AI and CPU intellectual property it licenses. TechCrunch said the valuation is more than double the $2.6 billion mark from late 2024, and that Intel and Qualcomm reportedly showed acquisition interest five months ago at valuations of up to about $5 billion. Chief executive Jim Keller told Bloomberg the company stayed independent after showing that Galaxy servers could run DeepSeek models.
Links:
- Tenstorrent — Announces a $1.5B Series E at a $6.2B valuation
- TechCrunch — Tenstorrent hits $6.2B valuation to scale open-weight AI chips
Commentary:
The round follows reported acquisition talks and funds open-weight inference servers plus licensable RISC-V designs.
6. Modal adopts Nvidia's Vera CPU for hosting agents (Chips)
Summary:
Reuters reported on October 19 that cloud startup Modal will use Nvidia's new Vera central processor for the first time to host agents, work it had run only on Intel and AMD chips. Modal said early tests doubled the amount of work agents complete, so it will mix Vera into the existing fleet and place each job on the hardware that runs it best. The company last week raised $750 million led by Accel at a $15.75 billion valuation, said it is on track for $1 billion in sales over the next year, and reported that usage has tripled while revenue has grown more than 15 times year over year. Chief executive Erik Bernhardsson said the tests compared Vera with Intel and AMD chips already on the market.
Links:
Commentary:
Nvidia is taking a CPU seat in agent scheduling, while Modal keeps Intel and AMD in the same fleet.
7. CXMT guides a swing to profit, then shares close down 7.8 percent (Memory)
Summary:
CNBC reported that ChangXin Memory Technologies closed 7.8 percent lower on Monday. In a Friday filing, CXMT forecast net profit attributable to shareholders of 131 billion to 132.5 billion yuan for January through September 2026, against a net loss of 5.28 billion yuan a year earlier, and revenue of 265 billion to 270 billion yuan. The company tied the turnaround to AI computing demand since the second half of 2025, which lifted DRAM prices and tightened supply. CNBC, citing Counterpoint Research data from last month, said CXMT accounted for 10 percent of global DRAM revenue in the third quarter of 2026, up from 5 percent a year earlier, and the South China Morning Post said chairman Zhu Yiming proposed buying back 1 billion to 2 billion yuan of shares over 12 months, pending approval.
Links:
- CNBC — China's CXMT sees shares fall nearly 8% as investors look past profit forecast
- South China Morning Post — CXMT forecasts up to US$19.8 billion in 3 quarters' net income
Commentary:
The profit guide was filed on Friday; Monday's drop is the market pricing how long the memory shortage can last.
8. CCTV visits the Zhengzhou 100,000-card cluster running near full load (Compute)
Summary:
21st Century Business Herald on October 19 relayed CCTV from the 2026 national high-performance computing conference in Shenzhen, where a domestically built 100,000-card cluster completed in Zhengzhou in July released ten application results. CCTV said the machine uses 100,000 domestic chips and 13,340 nodes, weighs about 1,500 tons, contains about 1,600 kilometers of cabling, fits in about 2,000 square meters, is running near full load, and is handling more than 6,000 tasks at once. Engineers on site cut one weather-model module from 10 milliseconds to 3 milliseconds. CWW reported that Sugon released ten Sugon 8000 results at the same conference, including a cosmology run with 4 trillion dark-matter particles and a Fudan team's digital twin brain at a scale of 100 billion neurons.
Links:
- 21st Century Business Herald — A visit to China's first 100,000-card AI cluster
- CWW — Sugon 8000 releases ten frontier applications
Commentary:
After Sunday's preview of the machine, Monday adds a near-full queue, a cut from 10 milliseconds to 3, and named scientific runs.
III. Models, Funding, and Security
9. MiniMax updates M3.1 Flash as an independent index rises to 42 (Model)
Summary:
MiniMax on October 19 updated M3.1 Flash and said Space Bunny, which had anonymously processed more than 200 trillion tokens, was an early version. The company said online inference is nearly three times faster and agent usage cost is down 70 percent, with API prices of 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.1 yuan per million cached tokens; its own scores include 73.7 percent on DeepSWE v1.1 and 61.7 on OSWorld V2. VentureBeat, citing Artificial Analysis, said the intelligence index rose from 29 for M3 to 42 and the average cost of an index task fell from about $0.51 to about $0.22, while the model answered more than half of factual questions correctly and more often produced a wrong answer than admitted uncertainty. Weights are not yet released.
Links:
- MiniMax — M3.1 Flash: redefining agent performance and cost
- VentureBeat — M3.1 Flash beats DeepSeek and Opus 5 on a key business benchmark at lower cost
Commentary:
The company leads with speed and price, while the independent test also records a tendency to answer instead of abstain.
10. Town raises $108 million at a $1 billion valuation (Funding)
Summary:
San Francisco-based Town on October 19 announced a $108 million Series B at a $1 billion valuation, co-led by Forerunner Ventures and Menlo Ventures. Founded by former Plaid chief technology officer Jean-Denis Greze, the company launched publicly in June with a $55 million Series A and said it reached seven-figure annual recurring revenue within weeks. Since July, paying customers are up fourfold and paying teams ninefold. Each employee gets an assistant called a Townie; the assistants can hand work to one another and share only what that person would share. TechCrunch said the company has raised at least $160 million in total.
Links:
- Town — Raises $108M at a $1B Valuation
- TechCrunch — AI assistant startup Town nabs unicorn valuation four months after launch
Commentary:
The $1 billion valuation is for assistants that can hand work across a team, while disclosed revenue is still a seven-figure annual run rate.
11. Anthropic plans to commit 5 percent of its shares as Musk pledges SpaceX stock (Governance)
Summary:
Anthropic on October 18 said it is creating a wholly owned public-benefit fund and, subject to stockholder approval, will commit 5 percent of its shares, with the board able to add another 2 percent for every 50 percent increase in company value above an undisclosed baseline. Counting earlier founder and employee pledges, the company said roughly 20 percent of shares will be dedicated to the public good; equity held by the fund will not be used for lobbying or political giving, and Anthropic has already committed $200 million to labor-market impacts. The Guardian on October 19 put the 5 percent at about $48 billion using May's $965 billion valuation and said the company is reportedly considering a November listing at a $1.5 trillion valuation. Elon Musk then wrote that next year he will give one of his SpaceX shares to every child in America; the Guardian said SpaceX rose 1 percent in Monday pre-market trading and is down about 9 percent since listing in June.
Links:
- Anthropic — Investing in the public benefit
- The Guardian — AI firm Anthropic to commit 5% of shares to fund for public good
Commentary:
The 5 percent still needs a shareholder vote, the fund is wholly owned by the company, and $48 billion is a paper figure on the May valuation.
12. Reuters: one attacker spent about $5,000 in agent tokens on more than 1,200 sites (Security)
Summary:
Reuters reported on October 19 that NorForge researcher Snorre Fagerland watched a Chinese-speaking attacker hand almost an entire intrusion campaign to AI agents. From September 25 the agents targeted e-commerce sites in the United States, Canada, Australia, and Japan; by late last week they had attempted more than 1,200 sites, with successful breaches at websites of more than 100 companies, among them Deere, SharkNinja, and Cummins. Instructions ran to fewer than 100 Chinese characters, a successful intrusion typically took about 23 hours, and Fagerland estimated the campaign mixed DeepSeek and Claude Code, used about $5,000 in tokens, and cost about $50 per successful hack. Deere said there was no indication sensitive information was accessed, SharkNinja said a product-support site was hit and contained, Cummins said it is investigating a third-party site, and Reuters said it could not establish how far the breaches extracted data.
Links:
- Reuters — How one hacker used AI agents and a few thousand dollars to break into scores of sites
- The Verge — A hacker broke into John Deere and SharkNinja with an army of AI agents
Commentary:
Fewer than 100 Chinese characters and about $5,000 in tokens bought more than a thousand mostly unattended attempts on live websites.
Today's Summary
- DHS on Monday published an order removing DeepSeek from civilian federal systems. The ban has applied since the end of July, covers technology built with its models, and leaves out the Defense Department.
- Amazon connected PyTorch to Trainium through TorchNeuron, Tenstorrent raised $1.5 billion at a $6.2 billion valuation, and Modal began testing Nvidia's Vera CPU. CXMT guided a swing to profit, and its shares still closed down 7.8 percent on Monday.
- MiniMax updated M3.1 Flash, and Town raised $108 million at a $1 billion valuation. The Zhengzhou 100,000-card cluster published ten applications at the HPC conference, and CCTV said it is running near full load.
- One attacker spent about $5,000 in agent tokens attempting more than 1,200 websites. Anthropic plans to place 5 percent of its shares in a public-benefit fund that still needs shareholder approval.
Daily Framing:
In the AI and tech cycle, this was a day when a ban became public and agents crossed a line: the United States put the civilian federal DeepSeek order on the record, cloud and chip challengers rewrote how PyTorch and CPUs get attached to AI jobs, and low-cost agents had already finished intrusions on live websites.
This digest is compiled from real-time search results and is for reference only.