> daily_signal(2026_05_03)

Europe just demanded the AI model the Pentagon already runs.

PickBits Daily Signal · Sunday, May 3, 2026

By Mark Pickering · 6 min read · May 3, 2026

// tl;dr

The Pentagon-Anthropic story we have tracked since April 30 just turned into a sovereignty problem. The ECB has been quietly convening Eurozone bank risk officers on Anthropic's Mythos since mid-April. The Bundesbank's chief banking supervisor went public last week urging the EU to demand access from Anthropic or from the US administration directly. Eurogroup finance ministers add the file to their Monday May 4 agenda in Brussels. Same model. Same supply-chain question. Two governments handling it differently across the Atlantic.

The supply-chain question is not just about software. NVIDIA's Asian production costs jumped to roughly 90% of total this year, up from about 65% a year ago, per Bloomberg's reporting today. The procurement asymmetry that started inside one classified network is now visible at the chip-substrate layer too.

Same Mythos model. Same supply-chain question. Different governments handling it differently across the Atlantic, and a supply chain consolidating across the Pacific.

1. Europe asked for Mythos. The Pentagon already has it.

Continuing 5.01 #1 (Pentagon-7 awards) and 5.02 #2 (Pentagon CTO splits Anthropic into two procurement tracks). The next data point lands in Brussels tomorrow. The European Central Bank has been calling chief risk officers at major Eurozone banks about Anthropic's Mythos since the model's restricted preview under Project Glasswing began in early April. Mythos is Anthropic's most advanced model to date and was designed for advanced research in digital security, with the ability to identify and exploit vulnerabilities, including zero-day flaws in operating systems and browsers.

Germany's chief banking supervisor at the Bundesbank, Michael Theurer, went public on April 29 urging the European Commission and EU member governments to demand access to Mythos from Anthropic or, failing that, from the US administration directly. The structural problem he flagged: organisations with access can see where their systems are vulnerable, and the organisations without access cannot. Eurozone banks rely on complex, decades-old technology stacks. The asymmetry is not theoretical.

Eurogroup finance ministers take the file in Brussels on Monday, May 4. The White House is simultaneously running Mythos through the NSA while opposing Anthropic's plan to expand access to a reported 70 additional organisations. Reuters reported in late April that Anthropic plans to provide Mythos access to European banks soon, but the timing is now a foreign-policy question, not a commercial one.

Bloomberg coverage of Eurogroup Mythos discussion
bloomberg.com · Apr 30, 2026
Why this matters: The Pentagon's blacklist-with-Mythos-exception that we covered Friday is not just a US procurement story anymore. The same exception is the model Europe says it cannot defend its banks without. Anthropic's principled-refusal posture, which cost it the Pentagon-7 list, just became an asset that Eurozone supervisors are openly negotiating for. For operators inside multinational financials, the cyber-defense capability gap between US and EU subsidiaries is now real and measurable. Expect bank-by-bank Mythos-access disclosures inside the next quarter as the asymmetry gets priced.

https://www.bloomberg.com/news/articles/2026-04-30/euro-area-finance-chiefs-to-discuss-mythos-as-concerns-mount
https://www.bloomberg.com/news/articles/2026-04-16/ecb-to-scrutinize-anthropic-s-mythos-on-call-with-executives
https://thenextweb.com/news/eu-finance-ministers-mythos-anthropic-cybersecurity
https://www.disruptionbanking.com/2026/05/01/can-anthropics-mythos-hack-europes-banks/
https://www.globalbankingandfinance.com/eu-seek-access-anthropics-mythos-bundesbank/

2. NVIDIA's Asian supply share hit 90% from 65% a year ago.

The same procurement-asymmetry frame, but at the silicon layer. Bloomberg's reporting today put Asian suppliers at roughly 90% of NVIDIA's production costs, up sharply from about 65% a year ago. The shift is not a new factory line. It is the cumulative effect of NVIDIA's "physical AI" pivot into robotics, autonomous systems, and AI-enabled manufacturing, and the partnerships that pivot has created with Asian OEMs, memory makers, and assembly partners.

The market response was immediate. LG Electronics rose as much as 15% on news that the Korean conglomerate is integrating home robots with NVIDIA's platform. Taiwan's Nanya Technology climbed 10% on a separate partnership disclosure. The list of Asian stocks benefiting from a direct NVIDIA business relationship has grown materially through the spring. Jensen Huang has framed physical AI as the next major wave after generative AI; the supply-chain consequence is that the pickaxe-supplier of the AI build-out has consolidated 90% of its production cost base in one region.

Bloomberg coverage of NVIDIA Asia supply chain reaching 90 percent
bloomberg.com · May 3, 2026
Why this matters: The "AI sovereignty" question used to be about model access (Mythos). It is now also about chip access. A 90% Asian-supplier concentration for the dominant AI accelerator vendor means that any tariff regime, export-control change, or geopolitical break in the Taiwan Strait or Korean Peninsula reaches directly into US data center capex. Operators planning multi-year hardware procurement should price the regional concentration into their continuity plans, not just their cost models. For Asian-listed operators in NVIDIA's orbit, the equity revaluation is happening in real time.

https://www.bloomberg.com/news/articles/2026-05-03/asian-stocks-soar-as-nvidia-increases-supply-chain-reliance-to-90-in-asia
https://www.investing.com/news/economy-news/nvidias-pivot-to-physical-ai-ignites-rally-across-asian-supply-chain-4654804

PickBits.ai is a reader-supported publication. To receive new posts and support my work, consider becoming a free or paid subscriber at pickbitsai.substack.com.

3. Florida expanded the first state criminal probe of an AI company.

Continuing 5.02 #1 (consumer-side fences). The next jurisdiction did not pass a law. It opened a criminal docket. Florida Attorney General James Uthmeier first announced a state criminal investigation into OpenAI on April 9. On April 22, Uthmeier elevated the probe to a formal criminal investigation citing the Florida State University shooting (Phoenix Ikner). On April 27 and 28, the AG expanded the probe to include the murder of two University of South Florida doctoral students. Hisham Abugharbieh, 26, is charged with two counts of first-degree murder. The remains of Nahida Bristy, one of the victims, were identified on May 1.

The case file is unusually concrete. Per affidavit, Abugharbieh asked ChatGPT detailed questions about disposing of human remains in dumpsters and about firearm operation in advance of the killings. The AG's office has stated this is the first attempt in the nation to hold AI company executives or program designers criminally accountable for the actions of generative software, and Uthmeier has explicitly framed the investigation as entering uncharted territory: "We are going to look at who knew what, designed what, or should have done what."

Florida Phoenix coverage of Uthmeier expanding criminal AI probe
floridaphoenix.com · Apr 27, 2026
Why this matters: Yesterday's framing was that three jurisdictions fenced the consumer side of AI in seven days. Today is the fourth fence, and it is criminal rather than regulatory. State AGs do not need a federal capability rule to subpoena ChatGPT logs and pursue AI-design accountability. For SaaS founders shipping consumer-facing AI products, the design-and-disclosure obligations under tort law just became enforcement-discoverable. The product-safety culture in AI consumer apps will be shaped less by federal policy and more by which state AG opens the next docket.

https://www.myfloridalegal.com/newsrelease/attorney-general-james-uthmeier-launches-criminal-investigation-openai-chatgpt
https://floridaphoenix.com/2026/04/27/uthmeier-expands-criminal-ai-investigation-to-usf-murders/
https://www.cnn.com/2026/04/26/us/university-south-florida-doctoral-student-killed-suspect
https://www.wusf.org/courts-law/2026-04-29/how-suspect-killing-two-usf-students-used-chatgpt

4. Sakana AI shipped KAME tandem speech-to-speech architecture.

The lab signal of the day. A practical operator-relevance angle on real-time voice agents. Tokyo lab Sakana AI released KAME, a tandem architecture for real-time speech-to-speech conversational AI. KAME pairs a fast front-end S2S model that handles immediate response with an asynchronous back-end LLM that streams an "oracle" signal back to the front-end as it generates a more knowledgeable response. The result, per the paper, substantially outperforms baseline S2S in response correctness, approaches the quality of a fully-cascaded system, and maintains latency on par with the baseline.

Authors are So Kuroki, Yotaro Kubo, Takuya Akiba, and Yujin Tang. The work is published as arXiv 2510.02327 and was accepted at ICASSP 2026. The architecture's training trick, dubbed Simulated Oracle Augmentation, generates synthetic oracle signals using a simulator LLM and a standard dataset, mimicking real-time responses at varying levels of completeness. The code repo is open at github.com/SakanaAI/kame.

Sakana AI blog post on KAME tandem speech architecture
pub.sakana.ai · May 3, 2026
Why this matters: The voice-agent stack has been forced to choose between fast (S2S, dumb) and smart (cascaded, slow). KAME is one of the cleanest published examples of breaking that trade-off without sacrificing latency. For founders building phone agents, customer-service voice front ends, or in-car assistants, this is the architectural pattern to study before committing to a cascaded pipeline. It is also a reminder that frontier capability is still being shipped from labs outside the SF/London/Beijing axis.

https://pub.sakana.ai/kame/
https://arxiv.org/abs/2510.02327
https://github.com/SakanaAI/kame

5. Anthropic eyes a $50B raise at a $900B valuation. 48-hour investor window.

The private capital frame. Pentagon exclusion is not slowing the re-rate. Anthropic has begun weighing a fresh round that would value the company at more than $900 billion, per Bloomberg, CNBC, and TechCrunch reporting from April 29-30. The round is expected to total $40-50 billion, with a 48-hour investor allocation window reported by TechCrunch and others. Anthropic's last round closed in February at a $380 billion valuation. The company has reached $30 billion in annualized revenue. A board decision on the round is expected in May.

At $900B, Anthropic would top OpenAI's recent $852 billion post-money valuation. The Mythos cyber-capability story, the Pentagon's blacklist-with-cyber-exception posture, and the ECB call described in story #1 above are all happening inside this fundraising window. The structural read is that public-sector friction has not slowed private-sector demand to participate in the lab's upside.

TechCrunch coverage of Anthropic potential 900B valuation
techcrunch.com · Apr 30, 2026
Why this matters: The narrative coupling between US government posture and AI lab valuations is looser than the headlines have suggested. Anthropic was excluded from the federal AI list a week ago and is being courted by the Bundesbank today, while the private market is being asked to commit $50B at a fresh $900B mark. For founders trying to read the AI capital cycle, the lesson is that public-sector friction creates negotiating leverage as often as it destroys valuation. The market is paying for principled refusal as a moat.

https://techcrunch.com/2026/04/30/anthropic-potential-900b-valuation-round-could-happen-within-two-weeks/
https://www.bloomberg.com/news/articles/2026-04-29/anthropic-considering-funding-offers-at-over-900-billion-value
https://www.cnbc.com/2026/04/29/anthropic-weighs-raising-funds-at-900b-valuation-topping-openai.html

Δ The counter-signal. Berkshire's $397.4 billion sat out the AI capex race, then opened the meeting with a deepfake.

Berkshire Hathaway reported its first quarter as Greg Abel's tenure as CEO on Saturday, May 2. Operating earnings hit $11.34 billion. Net income attributable to Berkshire shareholders rose to roughly $10.1 billion, more than double the $4.6 billion in Q1 2025. The cash pile reached a record $397.4 billion, up from $373 billion at year-end 2025. Buffett sat in the audience for the first time in decades. The crowd was visibly smaller than at prior meetings.

Abel opened the Q&A session with a deepfake video of Buffett asking a question about Berkshire's long-term prospects. Abel then revealed the clip was a deepfake created with no input from Buffett, framed it as evidence that publicly available information is enough to replicate a public figure's voice and likeness, and described it as a great reminder of cyber risk Berkshire is managing every day. Asked about AI strategy, Abel said the firm was thinking critically about how to use AI to add value but would not "do AI for the sake of AI" and is using AI today "to solve problems at our companies."

The structural read: the largest single pool of US corporate cash, run by the most senior capital allocator in America, has explicitly chosen not to participate in the AI capex cycle. The same allocator just used a deepfake demonstration to anchor the cyber-risk story for shareholders. This is the same week the ECB is convening Eurozone bank risk officers about the same class of capability through Mythos. The cyber-risk frame is converging across very different institutions.

Fortune coverage of Berkshire 397 billion cash pile and annual meeting
fortune.com · May 2, 2026

https://www.cnn.com/2026/05/02/business/berkshire-hathaway-earnings-buffett
https://www.bloomberg.com/news/articles/2026-05-02/berkshire-hathaway-s-cash-pile-surges-to-record-397-billion
https://finance.yahoo.com/markets/stocks/articles/berkshire-hathaways-first-q-without-161703919.html
https://www.cnbc.com/2026/05/02/warren-buffett-berkshire-hathaway-annual-meeting-2026-live-updates.html

» What to watch. The running threads

Tomorrow's signal lands here.


// full source roll-up & raw bullets:
https://ground.news/interest/ai
https://www.bloomberg.com/news/articles/2026-04-30/euro-area-finance-chiefs-to-discuss-mythos-as-concerns-mount
https://www.bloomberg.com/news/articles/2026-05-03/asian-stocks-soar-as-nvidia-increases-supply-chain-reliance-to-90-in-asia
https://floridaphoenix.com/2026/04/27/uthmeier-expands-criminal-ai-investigation-to-usf-murders/
https://pub.sakana.ai/kame/
https://techcrunch.com/2026/04/30/anthropic-potential-900b-valuation-round-could-happen-within-two-weeks/
https://www.cnn.com/2026/05/02/business/berkshire-hathaway-earnings-buffett