All Things Agentic: November 9, 2026: Data Center Ballot Measures, Revolut GLM, Tencent Hy4

This is the November 9, 2026 edition, covering the week of November 2 to 8. Here is the thread: almost nothing that mattered this week was decided by a lab. A county board in Ohio decided. About 2,000 engineers at a London fintech decided. A newspaper's lawyers decided. A few hundred unpaid kernel maintainers had it decided for them. The people on the receiving end of AI set the terms, and they set them in units no model card has a field for.

Watch the units, because they are the tell. Acres. Megawatts. Gallons per day. The invoice total before and after. Unpaid review hours. Every one of those is a number the buyer or the neighbor controls and the vendor does not, which is why the week's announcements read oddly defensive next to the week's decisions. Benchmarks are a language the industry grades itself in. A 500,000-gallon-a-day threshold on a ballot in Grove City, Ohio is not.

The Big Story: Data Centers Lost 40 of 41 Votes

The AI buildout has spent two years as a financing story, and on November 3, 2026 it became a zoning story.

Voters in ten states decided 41 local ballot measures touching data center development, and favored limiting it in 40 of them, per the tally Ballotpedia published on November 6. Twenty-seven limiting measures passed: five full bans (one in California, one in Florida, three in Ohio), twelve partial bans, three six-month moratoriums in Maine, and two zoning requirements in Adams County, Ohio. Thirteen pro-development measures were defeated, including four rezonings in Wilmington, Ohio and a Boulder City, Nevada measure to permit data centers in a specific area. Stateline and Scientific American both read the night the same way. The single exception was Harwood, North Dakota, where voters turned down a proposed development review requirement.

Two of those results are worth reading closely, because they are the ones with teeth. Grove City, Ohio's Issue 28 now requires a popular vote on any development using more than 50 acres, 20 megawatts of electricity, or 500,000 gallons of water a day. Perry, Ohio's Issue 16 requires hyperscale developers seeking certain tax incentives to post financial guarantees to the community first. Janesville, Wisconsin added voter approval for a General Motors and JATCO site if the project exceeds $450 million.

Europe did the same thing with its bodies instead of its ballots. On November 6, roughly 100 activists from the Stop Dirty Data Centers coalition walked onto a Microsoft AI data center under construction in Amsterdam, some carrying placards reading "Screw your data center," as ABC News and Tom's Hardware reported. The coalition's map lists 85 locations across Europe, with actions in 16 countries including Dublin, Warsaw, and a Microsoft site in Denmark. Microsoft said it respects peaceful expression but does not condone unlawful action. Pure Data Centers, which owns the Amsterdam site, asked the protesters to leave and start a dialogue. In Nova Scotia, campaigners held a run of rallies pressing for a provincial moratorium.

Now the skeptical part, because the triumphalist read is wrong. Most of these are small measures in small jurisdictions, several are explicitly non-binding, and a township ban in Ohio does nothing to a gigawatt campus three states over. Hyperscalers site projects where the power and the tax abatements are, and a map with 27 new no-go dots on it is an inconvenience, not a wall. A six-month moratorium is, by its own terms, six months.

Nobody in this industry had "water allocation" on their 2026 risk register. They have it now, because a ballot in Grove City put it there in units an earnings call cannot reframe.

What it means: The cost of a new data center just acquired a line item nobody models well, which is local consent, and consent is slow in a way capital is not. The companies positioned for this are the ones with land, substations, and signed power purchase agreements already locked, because scarcity in siting is a moat for whoever got in before the vote. Everyone else is about to discover that the binding constraint on the AI buildout is not chip supply or capital, it is a planning commission. Expect the next round of announcements to lead with water recycling figures and community benefit agreements, not FLOPs.

What's coming: Two more measures land in Clarksville, Arkansas and one in Mandan, North Dakota in December 2026, so the tally keeps moving. Watch for state legislatures to preempt local bans in the spring sessions, which is the standard industry counter and the thing to actually track. And watch whether any operator starts publishing its own acre, megawatt, and gallon numbers voluntarily. The first one to do it will be trying to set the disclosure standard before someone sets it for them.

Revolut Moved 2,000 Engineers to a Self-Hosted Chinese Model and Says Its Bill Fell 90%

The most consequential model decision of the week was made by a customer, not a lab.

On November 3, 2026, David Abitbol, who runs AI operations and infrastructure at Revolut, told the AI Engineer conference in New York that the fintech had made GLM, an open-weight model from the Chinese lab Z.ai, the default coding assistant for roughly 2,000 software engineers, as reported here. It displaced Anthropic's Claude Code and OpenAI's Codex as the defaults, both of which stay available for the hardest tasks. Open weights means the model file is published, so Revolut runs it on its own US infrastructure, about 56 Nvidia B200 GPUs, rather than calling Z.ai's hosted service. The reported saving is roughly 90%, or a bill about one tenth of the previous one, with claimed reductions of 60% in coding time and 70% in lead time. Startup Fortune has a second account.

Read the asterisks. The 90% compares API invoices to API invoices and excludes cloud, staff, and monitoring costs, which for 56 B200s and the team babysitting them is not a rounding error. The speed numbers do not isolate the model swap from everything else that changed, including the removal of token limits, and a team told it can finally stop rationing tokens will look faster regardless of which model it is rationing. Abitbol also said out-of-the-box inference was "terrible" before tuning, which is the part a smaller team should sit with, and added that "the bottleneck for engineering is not necessarily coding."

What it means: This is the per-task routing decision arriving at enterprise scale: use the cheapest model that clears the bar for each step, and reserve the expensive one for the steps that actually need it. That is a workflow design problem rather than a model problem, and it is the same calculation behind every automation worth building, which is why our agent workflow directory is organized by task rather than by vendor. For the frontier labs, the exposed revenue is not the hard problems, it is the enormous middle of easy, high-volume calls that an open-weight model does adequately for a tenth of the price. Watch gross margin commentary, not benchmark scores.

An AI Wrote 1,169 Linux Root Exploits, and Volunteers Got the Bill

The cost of finding a security bug collapsed this week. The cost of fixing one did not move at all.

On November 5, 2026, OpenAI published results from an unreleased cyber model that, in a mid-October run lasting about 12 hours over roughly 7% of the Linux kernel, flagged 2,308 candidate bugs, reproduced 1,724 of them in isolated virtual machines, and produced 1,169 working root exploits grouped into 948 distinct vulnerabilities, plus 291 information-leak proofs of concept. The run consumed $1,213,430 in API credits, about $1,279 per exploit. Help Net Security has the full accounting and SecurityWeek covers the disclosure process. These are local privilege escalation flaws, meaning an attacker needs some access already and uses the bug to become root. About ten days after the run, 61 were patched upstream, 43 had staging patches, and 299 were in public discussion.

To OpenAI's credit, it did not just publish and walk away: security researchers reviewed every exploit and severity rating before anything reached maintainers, and the company says it started with the most serious flaws at the kernel community's own suggestion. It is also worth noting what went wrong first. An earlier run claimed 200 root exploits, and human review found 20 of them had used a different zero-day than the one assigned. The technical details are withheld for safety reasons, which means none of it is independently verifiable, and the headline valuation of $18.9 million to $66.3 million comes from applying contest bounty payouts to exploits that were never submitted to a contest.

Here is the part the press release does not price. Those reports land on people. Willy Tarreau, the creator of HAProxy and a longtime stable kernel maintainer, said in March that the kernel security list had gone from two or three reports a week two years earlier to five or ten a day. In May, Linus Torvalds called the list "almost entirely unmanageable" and "a waste of time for everybody involved," and told researchers to "read the documentation, create a patch too, and add some real value on top of what the AI did," as covered at the time. The kernel's merged security documentation now treats AI-found bugs as public disclosures routed straight to maintainers, with a verified reproducer required. Greg Kroah-Hartman, one of the kernel's most prolific maintainers, runs his own bug-finding system that writes the fix and submits the patch publicly, which is the shape Torvalds endorsed.

What it means: Automated discovery is real and it is cheap, and the asymmetry it creates is not attacker versus defender, it is machine throughput versus human review capacity. Any company planning to point one of these at its own code should budget for the review queue, not the scan. For open source specifically, the generosity gap is now measurable: a lab can spend $1.2 million to find bugs in software maintained by volunteers who are paid nothing to read the results. The sustainable version of this is the Kroah-Hartman pattern, where whoever generates the finding also generates the verified fix. Anything less is outsourcing your homework to a mailing list.

Tencent Open-Sourced a Model That Can See Your Screen

The open-weight frontier moved again, and it moved toward models that operate software rather than just discuss it.

Tencent's Hunyuan team launched Hy4 on November 4, 2026, its first model with visual understanding and the ability to operate a computer, developed alongside the company's WorkBuddy agent, per Caixin. Tencent reports a hallucination rate of 7.2%, down from 12.95%, and guessing errors at 5.6%, down from 23.1%, with input pricing held at 6 yuan per million tokens. A second account puts the parameter count around 770 billion and frames the goal as taking the top open-source slot outright. Two caveats: Caixin's figures come partly from an AI-generated summary with the body paywalled, and every number here is Tencent's own. Treat the parameter count and the license terms as unconfirmed until Tencent's model card says so.

It did not happen in isolation. Caixin also notes Chinese models held the top four spots in OpenRouter's token rankings for the week of November 4. Moonshot's Kimi K3.1 Pro, a 3-trillion-parameter-class model with a 1-million-token context window, had its full weights slated for release on November 4. DeepSeek open-sourced a coding agent in the first days of the month.

What it means: Computer use, meaning a model that looks at a screen and clicks things, was the last capability the closed labs could reasonably call a moat. If a 770-billion-parameter open model does it adequately at 6 yuan per million input tokens, the moat is a toll booth with a bypass road next to it. The people who benefit are exactly the Revolut-shaped buyers in the story above, which is not a coincidence: the same week a Western bank picked a Chinese open model, Chinese labs shipped the capability that makes that choice easier next quarter. The constraint on adoption is no longer availability, it is whether your compliance team will sign off on the provenance.

Publishers Stopped Suing the Labs and Started Suing the Plumbing

Three years of copyright litigation aimed at model makers just changed targets, to the far less glamorous layer that feeds them.

On November 2, 2026, the New York Times sued Brave Software in federal court in Manhattan, with the Daily News, Chicago Tribune, and Orlando Sentinel filing a similar action the same day, reported by TechCrunch and heise. The complaint alleges Brave built an index of more than 30 billion web pages using an unidentified crawler that publishers cannot block, then distributed the contents to AI firms the suit names as Anthropic, Perplexity, Mistral, xAI, and Cohere, calling Brave a "supplier, dealer, and distributor of stolen copyrighted content." It also targets a browser feature that bypasses paywalls, and Brave's practice of blocking publishers' ads while selling its own. The Times seeks an injunction and damages. Brave did not immediately comment.

In the same stretch, Meta pulled down the terms of service that would have let it scrape news sites broadly, after already delaying the rollout in October under legal criticism.

What it means: Suing a lab means arguing about fair use, which is slow, expensive, and genuinely uncertain. Suing the index means arguing about a crawler that would not identify itself and a paywall that got walked around, which is a far more familiar kind of case. If it works, it is the more efficient attack: five AI companies lose a supplier from one judgment. Every company whose AI product quietly depends on a third-party web index should be asking, this week, whose crawler collected it and whether that crawler respected a robots file. Provenance just became a procurement question.

The Actual Product News Was Small and Cheap

While everyone watched the frontier, two releases moved the price curve, which is where adoption actually happens.

On November 4, 2026, Cognition launched SWE-2 Flash, a coding model built on its Inkling-Small architecture that runs at roughly 200 tokens per second, about three times faster than SWE-2 at one fifth the price, aimed at its Devin agent, documented here. On November 5, Google shipped Gemini Omni 1.2 Flash, extending generated video clips to 15 seconds at $0.07 per second, roughly 30% below the previous Flash pricing, per Google's announcement and MarkTechPost.

What it means: A "Flash" tier is a vendor admitting that most requests do not need the flagship, and pricing accordingly before a customer works it out and leaves. Three times the speed at a fifth of the cost changes which automations are worth building at all, because a workflow that was uneconomic at frontier prices is often obviously worth it at a fifth. The pattern to notice across this entire edition: every vendor who shipped a cheap tier this week was responding to the same pressure that pushed Revolut onto open weights.

Washington Built an AI Prosecutor, and Lawyers Found the Hole in a Day

The US government created an enforcement body for AI misuse without creating anything new to charge people with.

On November 5, 2026, the Justice Department announced the Super Intelligence Task Force, led by Associate Attorney General Stanley Woodward and Associate Deputy Attorney General Stephen Weiss, in its own release. Its stated remit covers foreign state actors, transnational criminal organizations, and corporate fraud involving advanced AI, plus a line about ensuring "companies exercise vigorous internal controls so that SI technology behaves as intended." Attorney General Todd Blanche framed it as protecting American innovators. Government Executive covered the launch.

Within a day, the structural problem was public: no new criminal offense was created, so the task force prosecutes under existing computer-crime statutes, and those statutes require human intent. Autonomous agents are precisely the case where intent gets hard to locate.

What it means: Read the corporate internal controls language twice, because that is the part aimed at legitimate companies rather than at criminals, and it has no rulemaking behind it. In practice that means discretion, and discretion means the first test case sets the standard. If you deploy agents in a regulated context, the defensible position is the boring one: logs showing what the agent touched, when, and who approved it. That was good advice before this task force existed and it is now the only concrete thing anyone can act on.

A Humanoid Robot Company Went Public on $174,000 of Revenue

Public markets got their first clean look at embodied AI economics, and the numbers are a reality check.

Agility Robotics began trading on the Nasdaq on November 5, 2026, the first pure humanoid robot company to list in the US, via a SPAC reverse merger that raised about $465 million. For the first six months of 2026 it reported $174,000 in revenue, mostly from related parties, against a $362 million net loss, per TechCrunch's read of the filing; Quartz and TechSpot cite different revenue and loss figures, so compare them against the filing itself before quoting any of them. CEO Peggy Johnson and co-founder Jonathan Hurst point to customers including Schaeffler, GXO, Toyota Motor Manufacturing Canada, and Mercado Libre, and about $300 million in planned contracts for Digit 5, a machine that has not been publicly demonstrated and is slated for CES 2027. The current Digit places totes with 98% accuracy and cannot do complex manipulation. The pitch of roughly $100,000 in annual savings per robot rests on uptime nobody has measured yet.

What it means: This is the first humanoid valuation anyone can short, which makes it the sector's price discovery mechanism whether Agility wanted the job or not. The useful signal will be the quarterly filings, because they will show whether proof-of-concept deployments convert into repeat orders, the single number the whole category rests on. Warehouse operators now have a public comparable to negotiate against, which is a better deal for buyers than for the industry.

Lightning Round

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

• DeepMind agents poked at RSA. Google DeepMind said on November 6, 2026 that AI agents surfaced two weaknesses in legacy 1024-bit RSA encryption; 2048-bit keys, which is what almost everything current uses, are unaffected. Source.
• The same cyber model found 81 Firefox flaws. Reported alongside the Linux work on November 5, 2026, which makes browsers the next obvious target. Source.
• DeepSeek open-sourced a coding agent that watches team chats. Released November 2, 2026 as a research prototype that acts on explicit requests from allow-listed users, with auto-merge left off by default. Source.
• Kimi K3.1 Pro's weights landed. Moonshot's 3-trillion-parameter-class model, with a 1-million-token context window, was slated for full open-weight release on November 4, 2026. Source.
• AI deals led a $454.5 million fintech funding week. Week ending November 6, 2026, which is a useful read on where agent money is actually landing. Source.
• JPMorgan and Mastercard backed a shopping-agent startup. Seattle's Firmly raised on the premise that agents will transact on your behalf, which requires the payment networks to play along, which is presumably the point of the investor list. Source.
• Hugging Face launched ProtoStar for scientific data. Announced November 4, 2026, aimed at the fragmentation that makes research datasets hard to use. Source.

This Week's Daily AI News Coverage

Every story above was tracked as it broke in the daily AI news editions for the week of November 2 to 8, 2026.

• November 3: GDPx Benchmark, DeepSeek Super-Agent, Apple Intelligence China
• November 5: Tencent Hy4, Revolut GLM Switch, NYT Brave Lawsuit
• November 6: Data Center Backlash, OpenAI Kernel Exploits, SWE-2 Flash
• November 7: Meta Scraping Terms, GPT-6 Cyber, Kimi K3.1 Pro
• November 8: DeepMind RSA Cryptanalysis, Gemini Omni 1.2 Flash, DOJ AI Task Force

Previous All Things Agentic Roundups

The weekly archive, newest first, if you want to see how these threads developed.

• All Things Agentic: November 2, 2026
• All Things Agentic: October 26, 2026
• All Things Agentic: October 19, 2026
• All Things Agentic: October 12, 2026
• All Things Agentic: October 5, 2026
• All Things Agentic: September 28, 2026
• All Things Agentic: September 21, 2026
• All Things Agentic: September 14, 2026

The Bottom Line

Put the week in one line and it reads like this: the decisions that stuck were made by people measuring in units the industry does not publish. Acres, megawatts, and gallons a day in Grove City. An invoice total cut to a tenth at Revolut. Unpaid review hours on a kernel mailing list. A six-month revenue line against a $362 million loss on a Nasdaq filing. None of those appear on a model card, in a benchmark table, or in a safety evaluation, and all of them determined what actually got deployed, built, or blocked this week. That is a different thing from the usual story about who holds power. It is a story about who holds the measuring tape, and this week the measuring tape was held by neighbors, buyers, maintainers, and auditors.

So the practical read is straightforward. If you are building with this stuff, pick your own units before someone picks them for you: cost per completed task, not cost per million tokens; review hours per thousand findings, not findings; consent secured per site, not capacity announced. Vendors optimize what you measure them on, and right now almost nobody is measuring them on anything that appears in this edition. The teams that quietly went and did the arithmetic this week, the ones who found out their expensive model was doing cheap work, are the ones who ended the week better off. That arithmetic is worth an afternoon, and it is the kind of thing worth doing before the next frontier release makes it feel urgent.

Frequently Asked Questions

• Is a self-hosted open-weight model actually cheaper than paying for an API? +
  Sometimes, and the honest answer depends on what you count. Open weights means the model file is published, so you can download it and run it on hardware you control instead of calling someone else's service. Revolut's reported 90% saving on November 3, 2026 compares API invoices against API invoices and excludes cloud, staff, and monitoring costs, which for roughly 56 high-end GPUs and the people who tune them is not a rounding error. Self-hosting tends to win when you have steady, high volume on a narrow set of tasks, and to lose when volume is spiky or the team is small. Our map of the AI automation tool landscape walks through where each shape of tool makes sense.
• What does it mean that an AI found 948 Linux kernel bugs? +
  It means automated bug hunting now has a price per finding, roughly $1,279 per working exploit by OpenAI's own accounting on November 5, 2026, rather than a pace set by a researcher's week. It does not mean Linux is suddenly unsafe: these are local privilege escalation flaws, which require an attacker who already has some access, and maintainers had patched dozens upstream within about ten days. The bottleneck moved from finding bugs to reviewing and fixing them, and that work is still mostly done by volunteers. We track developments like this as they break in the daily and weekly AI news coverage.
• What is All Things Agentic? +
  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.