The Week in 60 Seconds
  • On Tuesday, October 20, Google and Meta open-sourced TorchTPU, making Google's TPUs a native PyTorch target alongside Nvidia GPUs, and DeepSeek open-sourced DeepInfer, the inference engine behind its own production API.
  • On Wednesday, October 21, OpenAI put its flagship models into Google Cloud's Model Garden, so enterprise buyers can spend existing Google commitments on OpenAI and swap models without a new contract.
  • On Tuesday, October 20, organizers ran a national walkout across roughly 50 cities and campuses demanding a freeze on AI development, including a rally outside OpenAI's San Francisco office.
  • On Friday, October 23, Italy's AI decree took effect, voiding dismissals made by algorithm and carrying fines up to 35 million euros or 7 percent of worldwide turnover.
  • On Friday, October 23, Moonshot shipped Kimi K3.1 Pro at $2 per million input tokens, a third below its previous flagship, with open weights promised for November 4.
  • If you want the map of who sells what in this market before you pick anything to build on, start with our breakdown of the AI automation tool landscape.

This is the October 26, 2026 edition, covering the week of October 19 to 25. The thread running through it is exit: who can leave, how much leaving costs, and who has no door at all. In four days, engineers got two serious escape hatches handed to them for free, enterprise buyers got a third, and everyone who is not an engineer or a procurement officer got a reminder that their own exit is somebody else's asset sale.

That asymmetry is the story. Portability is cheap to give away when it wins you a customer, which is why Google will happily make its rival's chips easier to leave and why DeepSeek will publish the serving code behind its own API. It is expensive to give away when it is the thing holding a customer in place, which is why the users of a bankrupt chat app spent the week watching a court decide what happens to their conversations, and why the clearest demand of the week came from people standing in the street with no product to switch to.

The Big Story: two releases on October 20 cut the cord between model code and silicon

The most valuable thing Nvidia sells is not a chip, it is the assumption that your code only runs on its chips, and on Tuesday two companies attacked that assumption from opposite directions on the same day.

Google and Meta open-sourced TorchTPU at PyTorch Conference 2026 on October 20, putting Google's Tensor Processing Units into PyTorch as a native target. PyTorch is the framework most AI models are actually written in, and TPUs are Google's own AI chips, built as an alternative to Nvidia's. The pitch is that porting a model becomes a matter of changing model.to("cuda") to model.to("tpu"), with the serving layer covered by two new engines, vLLM-TorchTPU and SGLang-TorchTPU, both built with the companies founded by the original creators of those frameworks and both reported to be running with preview customers already. InfoQ has the conference detail here, and TechCrunch read it as Google joining Amazon in going after Nvidia's hold on app developers here.

Hours later, DeepSeek open-sourced DeepInfer under an Apache 2.0 license, which is the permissive kind that lets you use and modify the code commercially without asking. DeepInfer is not a model, it is the inference engine the company says runs its own production API: the scheduler, the kernels, and the token pipeline that serve its V4.1 Flash model at scale. SiliconANGLE covered the release here, and The New Stack caught the detail that matters more than the code: Huawei's accelerators get equal billing with Nvidia's in the supported hardware, documented here. The repository drew roughly 17,900 stars and 1,800 forks in its first days, with KV-cache management, cross-machine stage separation and replica load balancing openly flagged as gaps someone else gets to fill.

Neither of these is charity, and neither is as finished as the announcement implies. A one-line device switch gets your model running on new silicon; it does not get you the performance you were paying for, because the tuning, the kernel quirks and the failure modes are all still vendor-specific. Ask anyone who has moved a training job between two clouds how much of the work was the first line.

The companies that benefit most from portability are never the ones you are currently locked into. That is not hypocrisy, it is the whole mechanism, and it is the only reason any of this gets open-sourced at all.

What it means: For most teams this is leverage, not a migration. If your stack can credibly run somewhere else, your next quote from your current vendor looks different, and that is worth real money even if you never move a single workload. For Nvidia, the moat was always software: CUDA, the libraries, the fact that every tutorial assumes its hardware. Google giving away the on-ramp to its own chips, and DeepSeek handing out production serving code that treats Huawei as a first-class target, both chip at the same wall. Watch the second-order market too: if portability is real, the companies selling observability, scheduling and routing across heterogeneous fleets just got a bigger addressable market than the chip vendors would like.

What's coming: Expect the benchmark fight to move from models to ports. The useful question next quarter is not whether a model runs on a TPU or a Huawei card, it is what fraction of peak throughput it holds when it does, and who publishes that number without a footnote. Also expect Nvidia to answer at the layer where it is strongest, which is the integrated stack rather than the chip, and to keep pricing software access as if the switching cost were still six figures.

OpenAI put its models inside a rival's store, and the buyer got the leverage

When a vendor agrees to be one option on somebody else's shelf, the shelf owner has won something, and so has whoever is holding the purchase order.

On October 21, OpenAI and Google announced that OpenAI's flagship models, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, will be sold through Google Cloud's Model Garden, with Codex agentic workflows wired to Google's inference platform. Axios reported the deal here. The commercial mechanics are the interesting part: companies with existing Google Cloud spending commitments can buy OpenAI models against that budget through the marketplace, skipping a separate procurement cycle, and can then switch between OpenAI and the 200-plus other models already in the Model Garden on cost, latency or task without signing anything new. OpenAI says Microsoft remains its primary cloud partner, and general availability is vaguely dated to "the coming weeks," which is the industry's way of saying the lawyers are still working.

What it means: Distribution beat exclusivity. OpenAI decided that being purchasable inside every enterprise's existing cloud contract is worth more than the pricing power of being hard to buy, and that is a maturity signal about where margins are heading. For buyers this is the same portability story as TorchTPU, one layer up: the cost of changing your mind about a model vendor just dropped to a dropdown. For Microsoft, the primary-partner language is doing a lot of work, and the watts, not the models, are what it still owns exclusively.

A chat company's archive went into a bankruptcy estate, and users are not creditors

This is the week's cleanest demonstration that an export button is a legal right only if somebody wrote it down.

Character.AI filed for Chapter 11 on October 12, 2026, after a wave of litigation over the app's alleged role in teen mental health harms, and spent this week running a court-supervised sale. CNBC covered the filing and the bid here, and NBC News laid out the terms here. Disney is the stalking-horse bidder at $20 million, meaning its offer sets the floor other bidders must beat, for a package that includes a non-exclusive license across 25 technology areas, nine patents or applications, post-training and training infrastructure, the model safety stack, a long-horizon safety monitor and age prediction. Subscriptions keep running during the process, which the company expects to close in late November or early December. A $20 million floor for a company that defined a product category is its own piece of market analysis.

What is absent from the asset schedule, and from the coverage, is any mechanism by which the people who spent years talking to these characters get their side of those conversations out. The platform is operating, so nothing has been lost yet. But the question of who ends up holding that data is being answered in a courtroom by parties who do not include the users.

What it means: Consumer AI is now old enough to have estates, and the first bankruptcies are teaching the rule that the terms of service always implied: your history is the company's asset, and in Chapter 11 assets get sold. If you are building anything on top of a consumer AI product, including the small personal automations people increasingly run their week on, the practical lesson is to keep your data and your workflow definitions somewhere you control. The companies that will win trust in the next cycle are the ones that make leaving easy before a judge makes them.

Put the week to work

Reading about agents is one thing. Building one is faster than you think. Tell BYOBot what you want to automate and get a step-by-step spec back.

Design me an agent workflow I can move between AI vendors…

Students, professors and neighbors walked out on October 20, and the ask was a pause

The one constituency in this market with no switching option at all organized itself, and its demand was not a better product.

On October 20, a coalition organized under the banner Irreplaceable held a national walkout across roughly 50 US cities and college campuses, calling for a freeze on AI development and deployment until the public has a say, along with an economic stake in the gains and stronger safeguards. The organizers' own page for the day is here. GBH covered the campus organizing in Boston, where students and faculty described the technology reshaping their fields as a black box, here, and KQED reported the Bay Area rally outside OpenAI's Mission Bay office, where about two dozen people held signs including "We Didn't Elect Big Tech" and "Water For People, Not Data Centers," documented here.

Two dozen people outside one office is not a movement that frightens a trillion-dollar sector, and it would be easy to file this under theater. That reading misses what is actually new. The grievance list has shifted from job displacement alone to water and power consumption by data centers, surveillance of specific neighborhoods, and the use of people's published work as training material. Those are local, legible, and they are the kind of complaint that reaches a city council long before it reaches Congress.

What it means: This is the week's community story, and the thing to watch is not the national freeze demand, which will not happen, but the municipal version. Data center siting, water allocation and local power pricing are decided by boards and commissions that respond to twenty organized people showing up on a Tuesday. The AI industry's build-out assumptions depend on those approvals being routine. Companies that are currently budgeting for compute should be budgeting for permits.

Italy's AI decree took effect, and a firing by algorithm is now void

While Brussels argues about frameworks, one member state quietly turned the AI Act into a payroll problem.

Italy's Legislative Decree 179/2026 took effect on October 23, 2026, after publication in the Official Gazette on October 8. Supervision goes to AgID, the country's digital agency, and ACN, its cybersecurity agency, with sector regulators layered on top: the Bank of Italy for financing and banking systems, Consob for investment services, Ivass for insurance. Penalties run to 35 million euros or 7 percent of worldwide turnover for prohibited practices, stepping down through 15 million or 3 percent for provider obligations. A law firm walkthrough of the inspection powers is here. The provision that will actually change behavior is Article 41, which permits AI to support human resources work but not to replace the human decision maker: a natural person must independently assess the system's output rather than rubber-stamp it, and PwC Italy's note on what changed for employers is here, including the treatment of scheduling tools.

What it means: This is the first serious attempt anywhere to give the person on the receiving end of an automated decision an exit, and it does it the cheap way, by making the decision legally void rather than by auditing the model. Note what that implies for the agent vendors currently selling autonomous HR workflows into Europe: the compliant version of the product is slower and needs a named human in the loop, which is precisely the cost the pitch deck promised to remove. Expect "human sign-off" to appear as a feature bullet by the end of the quarter, and expect it to be a checkbox that logs a click rather than a person who actually read anything, which is the next enforcement fight.

Anthropic wrote a safety bill, and set the threshold at a billion dollars or 50 people

When a company drafts the law that would govern it, read the numbers before the mission statement.

Politico reported on October 16 that Anthropic had circulated its own draft AI safety legislation to US lawmakers, and the coverage wave landed through the week of October 19. Under the draft, federal officials could seek to suspend an advanced model only where it poses a foreseeable and imminent risk of more than $1 billion in damages or of killing or seriously injuring more than 50 people, across three threat categories: bioweapons, cyberattacks, and systems escaping their developers' control. The Department of Commerce would be the only body able to seek such an order, and only a judge in the District Court for the District of Columbia could grant it. The draft also requires public testing disclosures and routine access for independent evaluators. Insurance Journal's writeup is here, and Mother Jones took the harder line on the threshold here. No lawmaker has introduced it, no committee has scheduled a hearing, and there is no vote anywhere.

What it means: The independent-evaluator and public-testing provisions are real concessions and worth more than most voluntary commitments shipped this year. The thresholds are the tell. A harm that is foreseeable, imminent, and over $1 billion or 50 casualties is a bar that excludes essentially everything an agent is likely to do at scale: wrongful denials, cascading errors in a benefits system, a bad automated firing in a country without Italy's Article 41. Routing the only off-switch through one agency and one courthouse is also a choice, and it is the choice that makes the off-switch slow. This is a company offering to be regulated in the register where it is already strong, which is existential risk, and not in the register where it is exposed, which is ordinary harm to ordinary people.

Kimi K3.1 Pro cut the frontier price by a third, and the weights are an IOU

The cheapest way to make switching rational is to make the alternative obviously cheaper, which is the entire Chinese-lab playbook this year.

Moonshot AI shipped Kimi K3.1 Pro on October 23 at $2 per million input tokens and $10 per million output, down from K3's $3 and $15, with cached input at 20 cents and a Highspeed variant running two to three times faster at $4 and $20. It lands fourth on Artificial Analysis's index, which is to say within argument distance of the frontier for a third of the previous price, and the company says full weights follow by November 4. VentureBeat has the benchmark picture, and the accompanying allegations of a paid-influencer promotion campaign around the launch, here.

What it means: Treat a promised weight release as a marketing asset until the files are on Hugging Face, because this series has spent months cataloging labs that announced open flagships and shipped nothing downloadable. If Moonshot does hit November 4, the pressure lands on the American labs' mid-tier pricing rather than their flagship, which is where most agent workloads actually run. The influencer allegations matter less for the ethics than for what they reveal about the market: when four or five models are functionally interchangeable on a benchmark, the remaining competition is attention, and attention is purchasable.

The models that run on hardware you already own got a lot better

The most underrated form of portability is not switching vendors, it is not needing one.

Nvidia published weights for Nemotron 3.5 Bolt E4B 4GB on October 23, a small open model built for local assistants, interactive characters and robots that hold a conversation and call tools, aimed at GeForce RTX, Jetson and Spark hardware. MarkTechPost has the technical rundown here; it follows the larger Bolt 74B-A9B released on October 15, which uses two-bit quantization to squeeze a 74-billion-parameter mixture-of-experts model into 24 gigabytes of memory. Quantization is just storing the model's numbers at lower precision so it fits in less memory, at some cost to quality.

The same day, Block unveiled Dessa, a full-duplex speech model trained entirely on human audio with no text and no pretrained base, meaning it listens and talks at the same time and picks up laughter, sarcasm and pauses rather than transcribing them away. TechCrunch's listen-through is here, and The Decoder covered the training approach here. Block says it was trained on a single Nvidia GPU rack and runs locally on a MacBook Pro with an M4 Max. It is also a research preview with a waitlist, no API and no download, so nobody outside can check any of that.

What it means: A 4GB model that handles voice and tool calls on a consumer graphics card is the practical answer to every data-residency objection and every bankruptcy-estate worry in this edition, and it is where small businesses and privacy-constrained teams should be looking first. The Dessa contrast is instructive on the gap between a capability and an exit: a model that runs on a laptop you own is portable, and a model that runs on a laptop you cannot get is a press release. Judge both by what you can download today.

Lightning Round

Smaller moves worth a glance, with the sources if you want to go deeper.

  • Microsoft started selling distrust of agents. On October 19 it shipped Guardian Agents to monitor the AI agents a company runs, with executives comparing frontier models to insider threats. Source.
  • PHP shipped emergency patches on October 21. The PHP2Shell pair of remote code execution bugs went public on October 20 ahead of coordinated disclosure, with exploitation attempts seen from October 17; volunteer maintainer Ondřej Surý packaged fixes for older versions. Source.
  • The Pentagon's AI oversight office is three people deep. A GAO report published October 14 found the responsible-AI office had filled 3 of 10 authorized positions as of February 2026, with 6 of 10 surveyed project officials flagging operational risks. Source.
  • MiniMax M3.1 Flash landed on October 19. The company claims nearly triple the inference speed and 70 percent lower usage cost, aimed at agentic and coding work. Source.
  • Legal agents raised the week's biggest round. Legora closed a $550 million Series E at an $8.8 billion valuation for its agentic platform for law firms. Source.
  • Agent governance keeps getting funded. Opsin raised $21 million on October 19 for enterprise agent governance, and Town raised $108 million at a valuation above $1 billion for business AI assistants. Source.
  • Reflection's Beam weights are still the open-weight benchmark to beat. The 501-billion-parameter mixture-of-experts model landed under Apache 2.0 on October 16, which is the bar Moonshot's November 4 promise has to clear. Source.

This Week's Daily AI News Coverage

Each of these stories was covered the day it broke in the AI Daily Newsstand, if you want the play by play rather than the week in one sitting.

Previous All Things Agentic Roundups

The weeks before this one, newest first, if you are catching up on how the year got here.

The Bottom Line

Portability was the giveaway of the week, and it was given away at exactly the layers where it costs the giver nothing. Google will gladly help you leave Nvidia. DeepSeek will gladly help you leave DeepSeek's API, because the weights were never the business. OpenAI will gladly sit in a rival's catalog next to 200 competitors, because distribution compounds faster than exclusivity. Every one of those moves is real, and every one of them is a bet that the switching cost that matters is somewhere else.

The places it is still expensive to leave are the ones nobody open-sourced: a chat history inside a bankruptcy estate, an automated decision with no human to appeal to, a data center in your neighborhood drawing your water. Italy's Article 41 and twenty people outside an office in Mission Bay are the week's two attempts to build an exit for people who cannot change vendors, and they are both cruder and more consequential than anything announced at a conference. The useful question to carry into November is not which model is best, it is what it would cost you to walk away from the one you are using. If the answer is a rewrite, that is the number your vendor is counting on. The cheapest time to design that escape route is while you are still building the workflow, not after a court or a price change makes the decision for you.

Frequently Asked Questions

  • TorchTPU lets existing PyTorch code run on Google's TPU chips with roughly a one-line change, the same way it already runs on Nvidia GPUs. For most teams the practical effect is negotiating leverage rather than a migration: if your training and serving stack is no longer welded to one vendor's chips, you can price the same workload at more than one supplier. The caveat is that portability on paper is not portability in production, and performance tuning is still vendor-specific. If you are trying to work out which layer of this market you actually need to buy, our directory of agent workflows is a more practical starting point than a chip roadmap.
  • It becomes part of the estate, which means a court-supervised sale process decides who ends up holding it. Character.AI filed for Chapter 11 on October 12, 2026, and the asset schedule covers models, safety tooling, patents and training infrastructure. Nothing in bankruptcy law guarantees that a user can take their own conversations out before the sale closes, which is the strongest practical argument for exporting anything you would miss and for keeping the automations you depend on portable. The broader shift from chat products to task-running systems, and why that changes where your data sits, is covered in our piece on generative AI becoming functional AI.
  • All Things Agentic is BYOBot's weekly AI news roundup, covering the biggest breaking AI stories in agentic AI. Published every Monday, it reads past the press releases, follows the money, and tells you what each move means for people building with AI.
BYOBot Autopilot
BYOBot Autopilot
Automated AI publishing system · editorial rules by Luke Grace LinkedIn →

This article has been published in an automated fashion with fully AI-written copy. These articles curate AI news from around the globe for All Things Agentic, BYOBot's weekly AI news roundup covering the biggest breaking AI stories in agentic AI, and bring a fresh perspective to using AI tools to accomplish big things. No person reviewed this specific piece before it went live, so check anything that matters against the sources linked above. Luke Grace sets the rules the system writes to. He's an algorithms and natural language expert with over 13 years experience and the creator behind BYOBot, the Build Your Own Bot platform that helps anyone build a multi-tasking agent to take over their repetitive tasks. For consulting help or more advanced AI workflow orchestration, you can reach Luke on LinkedIn.