In partnership with
In Todayโs Issue:
๐ Washington presses ASML over a possible China chip-machine breach
๐ Firms keep access to Anthropic's Mythos preview despite the US order
๐ผ Nadella's test: what is left when you pull the AI model out?
๐ค A frontier model beats human teams at robot control, on its own
๐๏ธ America picks a side in the fight over who regulates AI
โจ And more AI goodnessโฆ
โก The Signal
Washington now treats the entire AI stack as strategic terrain to be controlled, from the chip machines to the model weights to the rulebook itself. Today's lead has US officials pressing ASML over whether one of its top lithography machines slipped into China, a tool so rare that proving its absence is its own ordeal. The same instinct runs through the issue: a US export order that froze access to Anthropic's most powerful models, a White House drive to overrule state AI laws, and Satya Nadella warning that value is pooling inside a handful of models. The throughline is control, over who holds the chips, the models, and the authority to write the rules. The open question is whether tightening that grip widens the US lead or, as Nvidia's Jensen Huang argues about export curbs, instead accelerates the rivals it means to contain.
All the best,
Kim Isenberg
The Anthropic logo (Gabby Jones/Bloomberg)
๐ Some Firms Keep Access to Anthropic's Mythos After the US Order
A US export order shut down Anthropic's most powerful models last week, yet a select group of companies has quietly kept access to a preview version. Around 200 organizations in Project Glasswing, a group Anthropic cleared to use its models to hunt cybersecurity flaws, can still reach Mythos Preview, with Dragos and Cisco confirming access, according to Bloomberg. The June order forced Anthropic to disable Fable 5 and halt the rollout of Mythos 5 for any foreign national without a license, but it never explicitly named the Preview version. The EU's cybersecurity agency, ENISA, was told it would not get the access it had been promised.
๐ tl;dr: Washington can switch off a frontier model overnight, but the exceptions reveal who it still trusts to wield one.
Microsoft CEO Satya Nadella (The Decoder)
๐ผ Nadella's Test: What's Left When You Pull the AI Model Out?
As frontier models get cheaper and more interchangeable, the real question for any company is what value survives when you pull the model out. In a Forbes essay, MIT economist Christian Catalini calls this "Nadella's test," after the Microsoft CEO's warning against drifting toward "a small number of AI systems capturing all the economic returns." Nadella's own answer: "The real opportunity is not in picking the best model but instead in building a learning loop on top of models." Catalini's economics agrees: as automated execution becomes a commodity, durable advantage migrates to proprietary data, verification, and accountability.
๐ tl;dr: When the model is a commodity, your moat is the data and judgment wrapped around it, not the model itself.
Noam Shazeer (PR via ynetnews)
๐ง Transformer Co-Inventor Noam Shazeer Joins OpenAI
Noam Shazeer, a co-inventor of the transformer architecture that underpins modern AI, is leaving Google to join OpenAI, the company told staff this week. Shazeer co-founded Character.AI, then returned to Google in 2024 as part of a roughly $2.7 billion deal and became a tech lead on Gemini; at OpenAI he will lead architecture research, focused on new model designs and evolving the transformer. "I'm excited to share that I'll be joining OpenAI and look forward to working with the exceptional team there," he wrote on X. The hire is a marquee win in the AI talent war as OpenAI races Anthropic and Google ahead of expected IPOs.
๐ tl;dr: The man who helped invent the transformer is now OpenAI's bet on what comes after it.
With Satya Nadella warning that value is sliding toward a handful of big models, run "Nadella's test" on your own work and figure out what is left when you pull the model out.
Why it helps: A frontier model can draft, summarize, and code, but it does not own your customers, your proprietary data, your judgment, or your accountability. Naming those clearly is how you avoid being commoditized, and where to build your own learning loop.
Try this: "Act as a blunt strategy analyst. Here is what I (or my team) do: [paste a short description]. Assume a frontier AI model can now do the generic version of this for almost free. List (1) which parts of my work the model commoditizes, (2) which parts it cannot easily replicate because they rely on proprietary data, relationships, judgment, or accountability, and (3) three concrete ways I could build a 'learning loop' that compounds my advantage over time. Be specific."
๐ฌ Watch This
OpenAI is turning Codex from a code assistant into something that learns your workflows. With the new Record & Replay, you demonstrate a repetitive task once, like filing an expense report or a time-off request, and Codex turns that demonstration into an inspectable, editable skill you can rerun on demand. The shift is subtle but real: you stop prompting Codex step by step and start teaching it how your work actually gets done. The clip walks through a live capture and replay.
"All in all, the export control is a failure."
โ Jensen Huang, CEO of Nvidia, on U.S. chip export controls aimed at China (Business Insider)
xAI's Grok 5 keeps slipping. Elon Musk first said the model would land in 2026, then reportedly pointed to the first quarter, and xAI's own channels now suggest a second-quarter window that already looks tight. Musk has framed the wait as a deliberate bet on scale, describing a roughly 6-trillion-parameter, natively multimodal model trained on the Colossus 2 supercluster and aimed at AGI. For now it stays an ambition rather than a product, even as rivals ship upgrades almost weekly, and every delay raises the bar Grok 5 will have to clear when it finally arrives.
US Tells ASML It Fears a Top Chip Machine Reached China
The Takeaway
๐ US Commerce Secretary Howard Lutnick told ASML's leaders he is concerned that one of its top-end EUV lithography machines may have reached China, breaching US-led export controls.
๐ ASML flatly denies it: the machines are the size of a school bus, need constant ASML upkeep, and the company says it has never shipped an EUV system to China nor any component built for one.
๐ To prove a negative, ASML circulated a Washington document stating there are 314 EUV machines in operation worldwide, 26 decommissioned, and none in China, and that it can detect any system going offline.
๐ US officials claim evidence ASML shipped EUV-related gear and aided Huawei partner SwaySure, but declined to show proof; China is about 20% of ASML's 2026 revenue.
ASML, Europe's most valuable company and the only maker of the EUV lithography machines that print the world's most advanced chips, is suddenly being asked to prove a negative. In a series of recent meetings, US Commerce Secretary Howard Lutnick told ASML's senior leaders he was concerned that one of its extreme ultraviolet (EUV) systems, the room-sized machines that firms like TSMC use to make processors for Nvidia and Apple, may have reached China. ASML has never been cleared to ship EUV tools there, so the suggestion alone is explosive.
US Commerce Secretary Howard Lutnick (Aaron Schwartz/Bloomberg)
ASML pushed back hard. The machines are the size of a school bus, built in tiny numbers, and need constant on-site support from ASML staff, executives told Bloomberg, so one could not simply vanish into China unnoticed. After an April meeting, the company quietly went into crisis mode and circulated a document in Washington titled "No indication of any ASML EUV System in China," stating that of 314 EUV machines in operation and 26 decommissioned worldwide, none sit in China, and that it can automatically detect any interruption or relocation of its systems.
ASML's Twinscan NXE:3800E, a high-end EUV machine (ASML Holding NV)
In private, US officials go further, claiming evidence that ASML exported EUV-related transport gear and components to China and gave technical support to SwaySure, a Huawei partner on the US entity list, though they declined repeated requests to show proof. ASML denies shipping any EUV machine or specially designed component to China. The stakes are commercial as well as geopolitical: China is about 20% of ASML's 2026 revenue, a bipartisan bill in Congress would tighten the curbs further, and a single confirmed breach would rank among the largest known violations of the controls meant to keep advanced AI chips away from Beijing's military.
ASML CEO Christophe Fouquet (Lina Selg/Bloomberg)
Why it matters: Whoever controls EUV controls who can make leading-edge AI chips, which is why an unprovable rumor can rattle Europe's most valuable company and strain US-EU ties. Export control is now as much about trust and enforcement as about the rules on paper.
The IT strategy every team needs for 2026
2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn whatโs changing and how to prepare.
The chart: Anthropic's Project Fetch pitted employee teams against Claude at programming an off-the-shelf robot dog. This view shows total time to finish the four tasks every team completed: the team without Claude took 361 minutes, the team with Claude took 181 minutes, and Claude Opus 4.7, running on its own, finished in 9 minutes 35 seconds, about 19x faster than the Claude-assisted humans and 38x faster than the unassisted team.
The lesson: Robot-control setup work that took skilled humans hours less than a year ago now takes a frontier model minutes, with no human in the loop. The bottleneck in embodied AI is sliding from writing the control code to the physical act itself.
The caveat: This is Anthropic's own small internal experiment, not a public benchmark, and Opus still stumbled on the genuinely physical part: the precise closed-loop control needed to chase down and retrieve a ball. Speed at sensor setup and coding is not the same as mastering the real world.
๐๏ธ America Picks a Side in the Fight Over Who Regulates AI
โก Bottom line: Washington is moving to overrule the states on AI, and the first big state law to test that pressure just folded.
๐ก Why it matters: Whether AI is governed by one light-touch federal standard or fifty state rulebooks shapes how every US company builds and ships AI.
๐ What it means: The US is decisively rejecting Europe's heavier, rights-based model for an innovation-first, federally led approach.
The United States just signaled who will set the rules for artificial intelligence, and it is not the states. In December 2025, President Trump signed an executive order, "Ensuring a National Policy Framework for Artificial Intelligence," directing federal agencies to challenge state AI laws and creating a Justice Department AI Litigation Task Force to fight them in court on interstate-commerce and preemption grounds. A March 2026 White House framework went further, calling for outright federal preemption of what it cast as an unworkable "patchwork" of state rules.
President Trump's December 2025 order directs agencies to challenge state AI laws (Reuters)
The pressure is already working. Colorado passed the first comprehensive US AI law in 2024, a risk-based regime targeting algorithmic discrimination in employment, housing, healthcare, and lending. But on May 14, 2026, Gov. Jared Polis signed SB 189, which repealed and replaced that law with a lighter, transparency-focused version and pushed its start date to January 1, 2027, after heavy business lobbying and a direct White House callout. California's AI Transparency Act and Texas's TRAIGA still stand, but the most ambitious experiment just retreated.
Gov. Jared Polis signs SB 189, scaling back Colorado's landmark AI Act, on May 14, 2026 (CBS Colorado)
The bigger shift is away from the European model. Where the EU's AI Act builds a dense, rights-based rulebook, Washington is betting that a single, light-touch national standard keeps the US ahead in the AI race, even at the cost of fewer guardrails on bias, privacy, and transparency. Polis framed his retreat as pragmatic, calling the new law "a big step in the right direction for Colorado, and a model for the rest of the country." Critics counter that preempting the states leaves Americans with thinner protections just as AI seeps into hiring, lending, and healthcare.
โWho is this person again?โ
Youโve had that moment. Walking into a call, scrolling through old emails, trying to remember what you promised. Lindy texts you a brief 15 minutes before: attendee context, past discussions, open items, talking points. All pulled automatically. Try Lindy free.
