- Google has reportedly trimmed about 1 million units, roughly 20 percent, from its 2027 orders for Broadcom's TPU 8i accelerators. Broadcom closed down 11.4 percent on Friday, and the pain spread: Micron down 11.4, AMD 11.2, Marvell 9.2, Nvidia 7.1.
- Taiwan's financial regulator banned short-selling of any stock that fell the previous session, effective Monday, after the TAIEX closed down 6.23 percent.
- Thinking Machines released Inkling 2 Mini with open weights, a 26.4-billion-parameter model with a roughly 20GB footprint that fits on a laptop.
- METR's updated time-horizon study found coding gains holding while judgment work lags, and the UN's 128-state weapons conference broke up without agreeing to negotiate a binding ban on autonomous weapons.
- If you want the shift underneath all this, start with where generative AI turns into functional AI.
For two years the AI buildout has run on one unexamined assumption: demand for compute is effectively infinite, so any amount of it gets bought. Friday was the day that assumption got priced as a question. The day's research pointed the same direction: these systems keep getting better at work you can grade automatically, and not much better at work needing judgment.
Two bills, same week. One from the market, one from the labs.
The Front Page: Google trims its 2027 TPU order and the chip trade buckles
Google has cut roughly 1 million units, about 20 percent, from its 2027 shipment plans for Broadcom's TPU 8i accelerators, most of it the pricier 8i Max, according to an industry survey from TF International Securities analyst Ming-Chi Kuo posted on November 20. Volume is said to be shifting toward the MediaTek-designed TPU 8t series. Broadcom closed down 11.4 percent and MediaTek fell 10 percent in Taipei, as reported here.
Worth slowing down on what this is. One analyst's supply-chain survey, not a filing, and Kuo said the details aren't final. Neither company has commented. Alphabet closed flat, so the market read it as a supplier problem, not a demand problem. The selling spread anyway: Micron down 11.4 percent, AMD 11.2, Marvell 9.2, Nvidia 7.1, with the Philadelphia semiconductor index off 8 percent intraday.
The aftershock landed in Taipei. Taiwan's regulator said on Saturday that from Monday it will bar short-selling of any stock that fell the previous session, its first blanket restriction since 2015, and lift the short-sale margin from 130 percent to 150 percent. The TAIEX closed Friday down 6.23 percent at 36,033.29, its third-largest single-session point loss on record. Focus Taiwan has the details.
What it means: Nothing here says AI demand collapsed. It says accelerator orders are now a variable, and one analyst note about one customer's 2027 plans can take a tenth off a supplier's value. That's a crowded trade losing its automatic story. If you build on this stack, watch inference prices: cheaper compute is the one outcome of a repricing that lands in your favor.
Releases & Features
Inkling 2 Mini, open weights, laptop-sized. Thinking Machines released a mixture-of-experts model on November 20: 26.4 billion total parameters, about 5.4 billion weight activations per token, in a roughly 20GB footprint the company says runs on consumer hardware. Mixture-of-experts means most parameters stay idle on any given token and only the needed slice wakes up, which is how a 26-billion-parameter model behaves like a much smaller one. The company's own table claims 77.9 percent on SWE-bench Pro v2 Hard, demoed offline on a 48GB MacBook Pro. The release post is here.
Those numbers are the company's, on the company's setups, and its own footnote concedes setups vary by model. The one figure from outside the house: an efficiency measurement researchers ran on an NVIDIA DGX Spark, where it scored 0.703 against 0.430 for Qwen3.6-35B-A3B.
Claude's mobile app got rebuilt. Anthropic began rolling out a redesigned phone interface on November 20, reorganized around code, projects and artifacts rather than one chat thread, covered here. Furniture moved, not capability.
What it means: The interesting release today is the small one. A capable coding model that runs on a laptop with no API key and no per-token meter changes the arithmetic for work that's repetitive, private, or both. Treat the benchmark table as marketing until somebody independent reruns it, then download the weights anyway.
In the Lab
METR published its Q4 2026 capability update, extending the time-horizon graph that has become the most-cited chart in the field. The original measured how long a task takes a human before a frontier model can finish it, and found that length doubling roughly every four months. The update adds longer tasks and, more importantly, open-ended ones: designing AI evaluations, and scoring philosophical arguments. The results there are blunt. Not one AI-designed evaluation was good enough for METR's own staff to call useful, and models fell short of humans on argument scoring, with Opus 5.5 described as borderline above the bar. MIT Technology Review has the write-up, and METR's post is here.
METR flags the caveat itself: completion times for the new tasks are mostly estimates, and the argument scoring was rated largely by one person. Thin measuring stick, strong claim.
What it means: This is the day's most useful finding for anyone deciding what to hand a model. Capability isn't advancing as one smooth frontier. It's advancing fastest where a machine can check the answer. Code compiles or it doesn't. Tests pass or they don't. A good evaluation design has no unit test, and that's where the curve flattens. Build so the model does the gradeable part and a person keeps the judgment call.
The Oversight Desk
The 128 states party to the UN Convention on Certain Conventional Weapons wrapped up in Geneva early Saturday with no mandate to negotiate a binding ban on autonomous weapons that can kill without a human making the final call. Almost 100 countries had asked for exactly that. The United States opposed a specific mandate, as did Israel, Belarus and Britain. Instead, government experts will make recommendations by 2028 on whether a treaty should follow. Campaign groups called it the lowest common denominator, per this dpa report.
The timing is uncomfortable, because the same week produced a working demonstration of machines operating without a human in the loop. Amazon Threat Intelligence disclosed that a suspected China-linked actor used Cairn, an open-source framework for coordinating AI agents built for security research, to find and exploit an unknown path-traversal flaw in MONAI Label, an open-source medical imaging app. The agents went from mapping the API to exploiting the bug in 11 minutes, later planted a reverse shell to reach root, then got pointed at a second university network running the same software. A fix shipped November 18. BleepingComputer has the breakdown.
What it means: Hold the two next to each other and the shape is clear. Governments gave themselves until 2028 to decide whether to start writing rules, while the capability is already cheap, open-source and in use. The honest caveat: these agents weren't superhuman. The same reporting describes a 16-hour, 148-action failure against a Redis server. Tireless rather than brilliant, and tireless is the part that should worry you. Patch windows were written for attackers who sleep.
Today's research says models are strongest on work you can check automatically. That's a design hint, not a limitation. Describe the repetitive part of your week and get a spec back that keeps the judgment calls with you.
On the Radar
Smaller moves worth a glance, with their sources.
- OpenAI put out a drug-chemistry benchmark. MedChemBench runs 99 problems drawn from 117 papers and patents. The company reports 84.5 percent for its top configuration, but says its models earned only about half the credit on picking which compound to test next and predicting numerical properties, the two parts a chemist gets paid for. Source.
- China wants to meter AI tokens like electricity. National Data Administration director Liu Liehong said the agency will set standards for measuring, billing and quality-checking tokens, after national daily use passed 500 trillion by June, up from 140 trillion in March. Source.
- Broadcom reports on December 9. That call is where the TPU order story gets confirmed, denied, or carefully not addressed, and the next real data point on whether Friday was a repricing or a panic. Source.
The Bottom Line
Strip out the tickers and the day tells one story twice. The market is asking which compute is genuinely needed, and METR's data is showing which tasks genuinely benefit. Both land on the same answer: the value sits in work that can be specified and checked, not in the general-purpose promise that got the chips ordered. You can watch the repricing play out, or spend the same hour building the narrow, checkable thing that stays worth having whatever the chip orders do next.
Frequently Asked Questions
-
Not on this evidence. It's one analyst's supply-chain survey covering 2027 plans for a single chip family, and it describes volume moving to a different supplier rather than disappearing. Neither company has confirmed it. What changed is market psychology: accelerator demand is now something investors check rather than assume. For people building workflows rather than trading chip stocks, the useful consequence would be cheaper inference if the bidding cools. For somewhere concrete to start regardless of the chip cycle, the agent workflow directory is organized by the job you want done.
-
Open weights means the trained model file is published for you to download and run yourself, rather than rented through an API you send your data to. Size matters because a model's memory footprint decides what hardware it fits on, and Inkling 2 Mini's roughly 20GB footprint puts it inside a well-specified laptop. That removes the per-token bill and keeps your inputs on your own machine, which is the difference between experimenting and deploying for private or high-volume work. For how self-hosting sits against the hosted platforms, see our map of the AI automation tool landscape.
-
AI Daily Newsstand is BYOBot's daily AI news brief, published every night. It covers the day's model releases, new features and capabilities, research, and oversight news, then tells you what each move means for people building with AI, in plain English and without the hype.
