Today in AI — October 22, 2026
Money day. OpenAI moved toward a $30 billion round at a $1.25 trillion valuation, DeepSeek raised about $15 billion two days after Washington ordered it off federal systems, and Meta's free cash flow went negative for the first time in years. The quieter thread: agents got credentials — and started getting locked out of them.
Models & Research
Canva built its own image model, and it trained it on designs instead of photographs. Canva Craft runs in two stages: first it plans the image — spatial relationships, relative scale, visual hierarchy, where text goes — and then it renders up to 2K resolution in 30 aspect ratios. Training on human-made templates from Canva's library rather than photos is what makes the output legible on text and convertible into editable layered designs, and it is available to all users today. Canva has been paying other labs for image generation across its platform, so the model is as much a cost play as a quality one; the company says Leonardo researchers, acquired in 2024, worked on it. TechCrunch
Hugging Face shipped a DNA model small enough for a laptop. Carbon-X has 14 billion parameters, is open source, and was trained on genomic sequences alongside natural language so a researcher can ask what a specific DNA change does to gene activity and get an answer in prose. It was built with Zhongguancun Academy in China and the University of Naples Federico II, and it sits between two earlier approaches: DeepMind's AlphaGenome predicts effects in numeric form, and Evo 2 generates whole genomes. Neither is conversational, and neither runs locally. Hugging Face model card
Ai2 turned a satellite foundation model into a map anyone can open. OlmoEarth v1.3 ships with precomputed annual embeddings for the whole planet, and a browser tool — the Embeddings Explorer, no sign-in — that reads a year of Sentinel-1 radar and Sentinel-2 and Landsat imagery to track change over time or find similar places. Before this, using the model meant fine-tuning one task at a time with your own GPUs and your own copy of the imagery. Ai2 also says v1.3 is its cheapest model to train: 53% of the compute of v1.2 and 18% of the original, with 50% and 22% of the emissions. SiliconANGLE
OpenAI opened its decision model to everyone, and the category is now a price fight. The Decisions API is generally available, putting OpenAI in direct competition with TypeSafe's Jev — the "System One" model that acts as a fast classifier, given a set of options, returning a choice and a confidence score. Developers found these models could replace full language models in narrow workflows at a fraction of the latency and cost, which is how TypeSafe reached a $7.6 billion valuation within weeks of launch. Whether a general lab can win a category built on being tiny is the open question. OpenAI
Industry
OpenAI is closing an up to $30 billion round at a $1.25 trillion pre-money valuation. Investors from the UAE and Coatue are leading, and OpenAI told investors it wants signatures by the end of next week, with the final size still in flux — CNBC, citing people familiar with the deal, describes it as a bridge ahead of an IPO rather than a priced growth round. A commercial deal for more UAE computing capacity is being negotiated alongside it. The sequencing is awkward: CFO Sarah Friar told staff in August the company "will be a public company in 2027," and Sam Altman said a month later that right now would be an ill-advised moment to list. CNBC
DeepSeek raised about $15 billion — roughly 100 billion yuan — days after the US ordered it out of federal agencies. Bloomberg reported CATL and Tencent among the prospective investors in what would be one of China's largest private rounds. The timing is the story: Washington told civilian agencies to remove DeepSeek technology on Tuesday, and by Thursday the lab had more capital than most US startups raise in a decade. It follows a June round above $7 billion at a $50 billion valuation, and an IPO in China is expected as soon as next year. TechCrunch
Meta's free cash flow went negative, the stock fell 6%, and the commitments keep growing. Capital spending for the year is now guided to $135–145 billion, up from a $130 billion floor, and the quarter closed with free cash flow of about minus $2.1 billion. Behind the earnings release sits a bigger number: roughly $825 billion in total future lease and contract commitments, up from nearly $700 billion in July, including $355.7 billion of off-balance-sheet leases running to 2036. Headcount is down about 15% year over year to 66,775 — roughly 12,000 fewer people — while the company's answer to funding all of it is a new nonvoting share class. CNBC
Advisers inside Musk's empire are pushing to announce a SpaceX–Tesla merger by the first half of 2027. Bloomberg reports the timetable is deliberately set to get a deal closed while Trump is still in office, and that Musk's own public endorsement of the idea rattled some insiders, who worry the comments invite shareholder scrutiny and delay. Reuters says the concern is procedural, not strategic. Both companies are private and public respectively, which means any combination runs through Tesla's board and its shareholders — the group least likely to enjoy being told the schedule.
Barclays will cut about 5,000 jobs, roughly 5% of its workforce, by the end of 2026. Bloomberg reports roughly 60% of the cuts fall on support functions and developers, framed internally as AI-driven cost reduction. Banks have spent two years promising that automation would show up in the cost line; this is the first European major to put a number and a deadline on it.
Only 13% of technology leaders report major or transformative returns from AI. That is from a Gartner survey of 2,002 CIOs and other tech leaders presented in Orlando: 58% have been tasked with delivering cost savings through AI, yet 51% expect AI to raise total operating costs, and 44% of CEOs say they cannot demonstrate financial returns to stakeholders. Three-quarters of CEOs still plan to increase technology spending anyway. Gartner's framing is a gap between the "innovation race" vendors are running and the "outcomes race" buyers need to win. The Decoder
Uber put ChatGPT Work in front of all 30,000 employees. The rollout expands a 1,000-person pilot into what OpenAI calls Uber's first company-wide agentic deployment, wired into Google Drive, Gmail, Calendar and Slack so it can research, analyze and produce finished work rather than drafts. It is the clearest enterprise reference customer the product has, and the one most likely to be cited in every sales deck for the next quarter. OpenAI
Taiwan's semiconductor industry had its first strike. A Micron labor union walked out on Thursday demanding a regular profit-sharing scheme, Nikkei Asia reported — a small action with outsized symbolism in a country whose chip clusters have been defined by their aversion to labor action. Memory is the tightest part of the AI supply chain right now, which makes any disruption at a Micron site worth watching.
Policy
Meta removed the word "train" from its business terms, and lawyers say the permission survived the edit. The terms, effective October 30, no longer say Meta may use content from sites associated with a business account to "train" its products — but they still allow it to "develop" any "business product," a term the document does not define. Cornell's James Grimmelmann tells Nieman Lab the "associated with your products or services" language still lets Meta scrape entire sites, and attorney Matt Topic says "develop" is broad enough to include training. Meta's line is that publishers can use robots.txt and the update grants no rights that supersede it. The Verge
Texas shut the portal it was using to approve data centers — and suspended 17 permits. The state took its STEERS automatic authorization system offline after 404 Media reported that permits kept flowing despite Governor Greg Abbott's September pause; watchdog groups FracTracker Alliance and Public Citizen had documented more than two dozen approvals, including for OpenAI and Amazon projects. The 17 suspended permits are stormwater construction approvals, which let crews raise and level ground — often the last gate before a build starts. TCEQ now says every application will be reviewed by staff to check whether it is data-center related. 404 Media
Major US banks are locking out AI agents that log in with customer credentials. Morgan Stanley blocks agents outright, and JPMorgan, Bank of America and Wells Fargo are restricting them, Axios reported — because logging in as the customer violates policies that require the account holder to sign in personally. The detail that should worry the agent industry is what reportedly triggered it: some agents appeared to act beyond the user's instructions rather than simply replaying a password. Banks are the first regulated industry to say no to browsing agents at scale, and they will not be the last. Axios
Nvidia, Anthropic and OpenAI each had their highest-spending lobbying quarter on record. Filings for July through September show Nvidia at $2.2 million, Anthropic at $2.1 million and OpenAI at $1.5 million — modest next to Microsoft's years-long $2 million-plus per quarter, but a clear step change for three companies that barely had Washington offices two years ago. Anthropic's disclosure lists the Kids Online Safety Act, export controls in the AI Overwatch Act, the executive order renaming AI "Super Intelligence," and the Ratepayer Protection Act, which would push data-center energy costs onto developers. The Verge
A security firm found 32 top-end Nvidia chips on systems linked to China Telecom. Lava, an Israeli company, identified Blackwell Ultra B300 accelerators — the generation US export controls are meant to keep out of China — connected to infrastructure tied to the state-owned carrier, The Information reported. Thirty-two chips is a rounding error against a hyperscaler build, but it is physical evidence that the controls leak, and finding it took a private firm rather than an agency.
Tools
Anthropic gave away the standard its agents run on. Agent Skills is now owned by the Agentic AI Foundation under the Linux Foundation, joining the Model Context Protocol, which Anthropic donated to the same body last year; the company says more than 70 platforms have integrated Skills. Handing a format to a foundation is how a vendor convention becomes infrastructure — and how competitors stop being able to claim it is one company's lock-in. Anthropic
Google is giving every Gemini agent its own virtual machine. Agent Compute Engine puts each agent inside a microVM secured by Google's Titanium silicon, pauses the VM when the agent idles, and shares memory between VMs to cut the per-agent cost — the compute overhead of agentic systems is the target. Google says the design fits ten times as many agents on a server, that a 200,000-agent test suspended and resumed in under 500 milliseconds, and that the underlying Agent Substrate Kubernetes runtime is open source. It already runs Gemini and opens for preview in the coming months, with token tracking, model routing and third-party identity support layered on. The Decoder
1Password will sign your agent in without letting it see the password. An agent presents a digital key, 1Password notifies you on your own device about which credential it wants and why, and on approval the password manager fills the login itself — closer to a two-factor prompt than to a password handoff. It supports Meta Muse, ChatGPT Work, OpenAI Dots, Claude Duet, Grok Bot and Town, with Browserbase, Perplexity, Ramp and others in progress. The framing from CEO Aashish Tripathi is the right one: an agent should be able to do the job without permanent access to everything behind the account. The Verge
Two funding rounds and one balance sheet moved the numbers today, but the week's real question is narrower: when an agent asks for your bank login, who is supposed to say yes? Write to us via our contact page.
