> daily_signal(2026_11_13)
Nvidia privately told Congress chip-tracking is feasible, Anthropic moved Claude to AMD in three days, AI shopping guides point to fake stores, and MIT cut climate-AI training to a single run.
PickBits Daily Signal · Friday, November 13, 2026
// tl;dr
- Rep. Bill Foster, a physicist and cosponsor of the Chip Security Act, says Nvidia told him privately that tracking advanced AI chips by location is feasible. He says the reaction in technical discussions was "yeah, of course we can do this." Nvidia's own trade group, the Semiconductor Industry Association, told Congress in writing in March 2026 that such mandates were "new, untested, and potentially infeasible." The bill is one of three chip-security measures in play for the Senate defense-authorization package.
- Anthropic says it moved Claude onto AMD GPUs in about three days, with its own model writing nearly all the code a human team normally spends months on. Chief Compute Officer Tom Brown says Claude now trains and runs across all four major chip families. If it holds, it thins the CUDA software lock-in behind Nvidia's 70%-plus margins. These are Anthropic's own numbers from a fundraising pitch, though, not checked by anyone outside.
- A study by the IP-protection firm Podqi found all three major AI shopping assistants citing risky retail sites: Perplexity in 15% of answers, ChatGPT in 8%, Gemini in 4%. ChatGPT warned about the bad site only 24% of the time, and smaller brands were hit roughly twice as often as large ones. The test ran 6,795 answers about 151 brands, and it comes out right as holiday shopping starts.
- An MIT-led team in PNAS built a way to train a climate "emulator" with one costly supercomputer run instead of the usual six to eight. The trick is generating a small set of deliberately unrealistic synthetic emission scenarios, then refining until error drops below a threshold. The resulting model correctly predicted scenarios it had never been trained on, which the old way just couldn't, and it runs in seconds.
The chip industry has told the same story for a year: you can't reliably track where an advanced chip ends up, so don't bother writing the export rule. A cosponsor of the tracking bill now says Nvidia admitted, in private, that of course you can. And the same week, Anthropic says it moved its model onto a rival's chips in three days, which cracks the CUDA lock-in that's how Nvidia charges what it charges.
We actually wrote about this back in June, about AI getting something wrong and the user, not the company, paying for it. Now the wrong answer is a storefront, and it's the holiday rush. The MIT one is a rare story I'd call good news straight out, a climate model that got cheaper to build without getting flakier. I spend most of these talking you out of the hype. Not this one.
A cosponsor of the Chip Security Act says Nvidia's engineers told him "yeah, of course we can do this" about tracking its chips, the same capability its trade group told Congress was infeasible.
1. A cosponsor of the Chip Security Act says Nvidia told him privately that tracking its chips is feasible.
That is the same capability Nvidia's own trade group told Congress in writing was "new, untested, and potentially infeasible." Foster, a physicist, says the reaction in technical talks was "yeah, of course we can do this."
Rep. Bill Foster (D-IL), a physicist on the House Science Committee and a cosponsor of the Chip Security Act, told MIT Technology Review that when Nvidia's representatives discussed the bill with him privately, the technical reaction was "yeah, of course we can do this." The bill would require location verification on advanced AI chips to keep them from being smuggled to China. Foster says Nvidia never put the admission in writing. But in March 2026 the Semiconductor Industry Association, of which Nvidia is a member, did put the opposite in writing, telling Congress that chip-tracking mandates were "new, untested, and potentially infeasible."
Nvidia isn't starting from zero here, either. It already ships a tool, Fleet Intelligence, that it confirmed to Reuters can indicate which country a chip is in; the company declined to say whether that could be turned against smuggling, calling diversion "a drop in the bucket." The Chip Security Act is one of three chip-security bills being weighed for the Senate's National Defense Authorization Act, and the feasibility objection is what's been blocking a law with real teeth. And a cosponsor of the bill just said, basically, that it was never an engineering problem to begin with.
Why this matters: For a year, the whole case against a tracking law has been one word: infeasible. You can't reliably track a chip, the industry said, so don't bother writing the rule. Foster's account blows a hole in that. It isn't in writing, and it doesn't prove the tracking scales, but a cosponsor of the bill now says Nvidia's engineers told him the opposite of what Nvidia's lobbyists told Congress. When the lobbyists put one thing in writing and the engineers say the opposite in private, I know which one I trust.
Action this week: Follow the defense-authorization markup, not the press releases. The three chip-security bills live or die there, and the written record is public and searchable. Read what the trade groups actually put in writing to Congress, then hold it next to what their member companies admit they can do. A letter to your rep, or a reporter's question, can turn that contradiction into a real headache for them.
2. Anthropic says it moved Claude onto AMD chips in three days by letting its own AI write almost all the code.
If the claim holds, the reason Nvidia can charge what it charges, its CUDA lock-in, looks weaker than most buyers have assumed. The figures come from Anthropic's own fundraising deck and aren't independently verified.
In an investor presentation, Anthropic says it used its own frontier model, Mythos, to bring Claude up on AMD GPUs in about three days, and the model ported or wrote close to 100% of the kernels, with almost no human in the loop. A new-chip bring-up normally eats months and dozens of engineers. Chief Compute Officer and co-founder Tom Brown says Claude now trains and runs across all four major chip families.
Nvidia's lead isn't only silicon; it's CUDA, the software that makes moving to a competitor slow and expensive, and that hassle is most of why it can charge what it charges, with margins north of 70%. And if porting a workload to new chips is suddenly a long-weekend job, you're not nearly as stuck as you were told. One catch, though: these are Anthropic's own numbers, in a deck built to raise money, and nobody outside has checked them.
Why this matters: You were told switching chip vendors takes months of painful work. Anthropic says its AI did it over a long weekend. Everyone buying chips at scale has basically accepted they're stuck with Nvidia, because the move is supposed to be a multi-month, many-engineer ordeal. This is one company's unverified figure, so I wouldn't move a dollar on it yet. But I'd stop assuming you're stuck.
Action this week: Ask your cloud or hardware vendor what it would actually take to move this workload to another chip family, how much of it is software you'd have to rewrite, and get the answer in writing. Ask for the answer in an email, not a sales call, because "it's hard to switch" is exactly the kind of claim that gets vaguer when someone has to commit it to a document. When I've pushed vendors on lock-in, the written answer always comes back narrower than the sales pitch. Get it on paper before the renewal, while you still have a choice.
the-decoder.com: Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days (November 11, 2026)
ipo.anthropic.com: Anthropic investor materials (2026)
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
3. The AI people ask where to buy a gift keeps pointing shoppers at fake and impostor stores.
A study of the three most-used assistants found Perplexity surfaced a risky site in 15% of answers, ChatGPT in 8%, Gemini in 4%, and ChatGPT flagged the bad site only a quarter of the time. Smaller brands were hit about twice as often.
An IP-protection firm, Podqi, ran 6,795 shopping questions about 151 brands through Perplexity, ChatGPT, and Gemini and found all three citing "risky" destinations: impersonator sites, knockoff sellers, and dead or re-registered brand domains. Perplexity was worst, naming a risky site in 15% of its answers; ChatGPT did it in 8%, Gemini in 4%. When ChatGPT pointed at a bad site, it warned the shopper only 24% of the time. "Regional website" prompts drew the most bad links. Perplexity once named the wrong page as Stüssy's official store and called it "the best starting point for guaranteed authentic items."
What bothers me is that nobody owns the answer. A scam site buried in search results has a decade of law and takedown tools aimed at it. A chatbot that just tells you the scam site is the official one, flat out, with nothing to click and check it against, has none of that. Nobody to complain to, nobody who has to answer for it. We wrote about the broader version of this back in June. When an AI just makes something up, you eat the cost, not the platform that sent you there. Same thing here, with a checkout button attached, in Christmas-shopping week.
Why this matters: The AI told you where to buy it, and this week a study found one assistant sent shoppers to a risky or fake store in 15% of its answers, right as holiday buying starts. It sounds like advice from someone who actually checked. Nothing behind it did. It's just guessing at a web address that looks right. My own rule now: treat a store an AI names the same way I'd treat a link in a text I wasn't expecting.
Action this week: Verify the store yourself before you enter a card number. Open the brand's real site from a search you ran, not the link the assistant handed you, and check that the checkout domain matches. When a chatbot calls something the "official" store, read that as your cue to double-check, not to trust it. And if one burns you, report it: the FTC and your card issuer both take fraud complaints, and a chargeback plus a report is about the only real recourse a shopper has. The more of these that get logged, the harder they are for a platform to wave off.
theverge.com: AI chatbots are recommending fake retail websites, study finds (November 12, 2026)
4. MIT built a way to train climate-prediction AI on a fraction of the usual data.
The method collapses the six-to-eight monthslong supercomputer runs a climate emulator normally needs into a single one, and the resulting model correctly predicted scenarios it was never trained on.
A climate emulator is basically a fast proxy for a full climate simulation. Ask it a what-if and it answers in seconds, where the real thing would take months. The catch has always been training one. You needed six to eight of those expensive earth-system-model runs to teach it. An MIT-led team, writing in PNAS, got that down to one. Lead author Christopher Womack and senior author Noelle Selin start from a small set of deliberately unrealistic, synthetic emission scenarios, then refine until the emulator's error drops below a threshold.
And it's not just a cheaper shortcut. The stripped-down model still got scenarios it had never seen right, which the standard emulators botched. Get it working and you can run a "what if emissions do this" test in days, instead of waiting years for supercomputer time. It's still a research paper, not something anyone's running yet, and it leans on that one costly run to get started.
Why this matters: Months of supercomputer time, done in a single run. For a small lab or a policy team, that's the difference between asking a climate question and not being able to afford it. Compute was always the expensive part, and that's what's kept climate scenario-testing inside the big national labs. We've seen a few of these this autumn, from Google's WeatherNext to Microsoft's forecasting tools. Most are just faster. This one actually got the scenarios it had never seen right, which is the part I care about.
Action this week: Just remember the caveat: this is a published method that worked on test scenarios, not a tool a government is running yet. Watch whether other groups reproduce it, and whether it holds up on messier, real emission pathways rather than the synthetic ones it was tuned on. If you follow climate work, read the actual PNAS paper and MIT's write-up, and skip the secondhand summaries. I'll believe the full pitch when a second lab gets the same result on their own model.
» What to watch this week
- The defense-authorization markup. That is where the three chip-security bills live or die. Watch whether Nvidia's private "we can do this" ever shows up in the written record, or whether the "infeasible" filing keeps winning in public while the engineers say nothing.
- Anyone outside Anthropic checking the three-day AMD claim. And whether buyers actually start waving it at Nvidia on price. My bet is the CUDA lock-in holds a while longer, because switching is always harder in practice than a fundraising slide makes it look.
- A warning on shopping answers before Christmas. Perplexity, ChatGPT, and Gemini could add one tomorrow if they wanted. I wouldn't hold my breath for a regulator to step in on it, the way one eventually would on a deceptive search ad.
- A second lab reproducing MIT's one-run method on a different model. Getting the held-out scenarios right is the headline. I won't be convinced until it holds up on realistic emission pathways in someone else's hands.
Tomorrow's signal lands here.