Trump Demands the Impossible from Anthropic and a Drama at SAP

Trump Administration Demands Unfulfillable Security Guarantees from Anthropic

Trump administration officials have told Anthropic: If the company wants to re-release Claude Fable 5, it must prove that the model's guardrails cannot be bypassed. Independent security experts counter that jailbreaks in a large language model probably cannot be completely prevented at all. Anthropic had already made headlines at the end of March when the model was circulating internally under the name 'Mythos'. In parallel, OpenAI is beefing up its staff ahead of its planned IPO, bringing on board Noam Shazeer, one of the co-authors of the Transformer paper 'Attention Is All You Need', as well as former Trump AI policy man Dean Ball. Waymo is recalling nearly 4,000 robotaxis after the vehicles drove into restricted construction zones in at least 13 cases. And Baseten is finalizing a $1.5 billion round at a valuation of $11 to $13 billion. → StrictlyVC

Synthszr Take: The government is demanding proof from Anthropic that computer science cannot provide. No one can guarantee that a language model with billions of parameters won't break out of its guardrails under any conceivable prompt; that's about as realistic as assuring that a piece of software is bug-free. What's interesting is that the same administration was admiring China's pace back in May and is now putting up a hurdle for one of its own leading labs that effectively keeps the model on the shelf. Anyone who makes absolute security a condition for release is only shifting the risk: to China, where Manus is being bought back for 2 billion, or to the open-source world, where no guardrails apply at all. A measurable threshold would make more sense than an impossibility guarantee—defined attack classes, documented success rates, regular external audits. This could be operationalized in a work week, not after the next strategy paper from Washington. Security comes from constant testing and rapid improvement, not from a promise that is false the moment it is uttered.

Founder Drama at SAP, Stock Price Tumbles

André Christ (44) and Gero Decker (44) will soon be leaving SAP. With 110,000 employees, this would be a side note, but the two are not just administrators, they are the in-house founder stars: Christ came with the Bonn-based startup LeanIX for 1.2 billion euros, Decker with Signavio for 950 million euros. Both drove the cloud transformation of the Walldorf-based group. CEO Christian Klein (46) is thus losing exactly the founder DNA that SAP could urgently use in its top management. The departure comes at a delicate phase: The stock has plummeted because many investors no longer believe in 'Software as a Service', thinking that artificial intelligence will solve many customer tasks more cheaply in the future. The fitting buzzword is already circulating in meetings: 'SaaSpocalypse'. According to manager magazin, companies like Nemetschek from the construction sector, which have built a virtual moat against AI attackers, are comparatively protected. → manager magazin – Der Tag

Synthszr Take: Paying 2 billion euros for founder spirit only works as long as the founders stay, and they rarely stay for long. Christ and Decker sold, integrated, waited for the earn-outs to clear, and are now leaving. This is the recurring pattern of every corporate acquisition of hidden champions: The technology is digested, the people are made replaceable. Walldorf gets the LeanIX and Signavio platforms, the creators' DNA takes the elevator down with them. Klein bought the right companies, that's no coincidence. The honest question for the next acquisition is not the price, but whether the company's own immune system will let the foreign cells work at all. I know this from my own experience: ‘They bought us because we were different. Later, they blamed us for being different.’ Anyone who wants to buy velocity must protect it from their own apparatus, otherwise they'll pay two billion for software they could have just licensed.

Peter Thiel's Secret Club: Jens Spahn Attended 5 Times

A misconfigured website and a leaked registration list have unmasked Dialog, a private network that PayPal co-founder Peter Thiel set up in 2006 with data-broker manager Auren Hoffman. For two decades, the thing ran without a public member list and without a reachable website, until Swiss hacktivist Maia Arson Crimew (known for leaking the US No-Fly List) found an open directory. WIRED verified the leak and also obtained the registration list for the planned retreat in August 2026 near Dublin: 222 members, sorted by status. On the list are Jens Spahn from Germany, NATO Supreme Commander Alexus Grynkewich, Treasury Secretary Scott Bessent, Senators Ted Cruz and Cory Booker, and Palantir co-founder Joe Lonsdale, plus actors like Josh Brolin and Joseph Gordon-Levitt. Internal documents show that each member receives a letter score (A, B, or C) beforehand, rated on wealth, fame, and influence, including assets under management and Instagram follower count. Lower-ranked guests pay over $10,000 per retreat, higher-rated ones get a discount, and after each event, there's a 'Post-Retreat Code Review'. On the agenda in Dublin: nuclear energy, battlefield technology, preparation for World War III, and Taiwan's role in the AI race. → Techpresso

Synthszr Take: The real punchline is the letter score. A group that presents itself as an intellectual salon ideal rates its guests by Instagram followers and assets under management and conducts a 'code review' after each meeting, like in a poorly managed product development department. This is Palantir logic applied to social life: everything becomes a data point, every person gets a confidence score, and anyone who falls below the threshold no longer gets an invitation. It gets spicy with the 222 names, as none of the government officials used an official email address, elegantly slipping the participation out of disclosure requirements. This is precisely the point that bothers me about this casualness: a NATO supreme commander and a treasury secretary discussing war preparations and Taiwan in a format that, by design, leaves no paper trail. In May, we wrote that the UAE wants to hand over decisions to AI; here we see the other side of the same coin, namely power operating outside of public accountability. Transparency cannot be circumvented with a second, private email address without someone eventually finding the open directory.

Washington Claims ASML's Most Important Machine is in China

According to Bloomberg, U.S. Commerce Secretary Howard Lutnick told senior ASML managers in several conversations that he was concerned one of the Dutch company's EUV lithography machines could have ended up in China. EUV systems are the only tools in the world that can print the finest circuit patterns for state-of-the-art chips, and they have been banned for China since the first Trump administration. The administration claims to have evidence of delivered EUV components and transport equipment but has shown it to neither Bloomberg nor ASML. ASML clearly contradicts this: such a machine does not exist in China and has never existed there. The company says it tracks every machine it has ever delivered, has built an internal firewall around its EUV technology, and expects about 20% of its revenue in 2026 from already approved sales of older technology to China. With a market cap of around $700 billion, ASML is Europe's most valuable publicly traded company and a de facto monopolist; every top chip from TSMC for Nvidia and Apple depends on these machines. The spicy detail: Lutnick's department invested up to $150 million in taxpayer money late last year in the startup xLight, which is working on the core technology of ASML's monopoly. → Techpresso

Synthszr Take: A machine that took two decades and billions to develop can't just disappear in a container without someone noticing. ASML's argument is sober and convincing: you can't reverse-engineer something you've never had your hands on, and 80% of the machine is based on decades of accumulated knowledge; the truly new problem, generating EUV light, took 20 years alone. Economically, the accusation makes little sense, as ASML would be risking 20% of its annual revenue and its status as Europe's most profitable monopolist for a single illegal delivery. What makes the situation interesting is the minister's other hand: while Lutnick publicly pressures the monopoly, his department has money invested in xLight, a challenger for its very core technology. Nothing public connects the two, and without evidence presented, the accusation remains just a claim. Still, something tangible can be noted: anyone who puts regulatory pressure on a monopoly while betting on its competitor should put their evidence on the table before a $700 billion market cap trembles over a rumor.

Google Copies Nvidia's Financial Trick to Crack Nvidia's Fortress

In the competition for AI chips, Google is resorting to the very financial mechanics Nvidia has used to fuel its own demand for years. According to a Wall Street Journal investigation, Google is providing a $3.2 billion financial guarantee at its Lake Mariner site in western New York so that developers TeraWulf and FluidStack can lease computing power from thousands of Google TPUs to Anthropic. This is complemented by classic circular financing: money flows into projects like the $7 billion River Bend data center near Baton Rouge and $1.4 billion in guarantees in Colorado City, Texas, and returns as chip purchases. On top of that is a roughly $35 billion private credit deal from Apollo and Blackstone, which buys Google TPUs and leases them to Anthropic. Google has been selling its TPUs directly since May, has struck a $5 billion deal with Blackstone against CoreWeave and Nebius, and aims to raise $85 billion in equity for AI infrastructure. Citadel Securities, a former user, now runs workloads 30 percent cheaper and up to four times faster on TPUs. Jensen Huang remains relaxed, pointing to CUDA and over 90 percent market share. → Techpresso

Synthszr Take: Nvidia's moat was never just CUDA; it was the ability to manufacture demand through balance sheets. Google is now replicating this very flywheel, and Google has the bigger balance sheet: $85 billion in fresh equity, $35 billion in private credit, a $3.2 billion guarantee for a single cluster on Lake Ontario. The catch is in the word 'circular': if the money Google invests comes back as TPU purchases and Anthropic remains the one big external customer (Huang has a point there), then the demand is largely self-financing. This works brilliantly as long as inference workloads grow exponentially and the 30 percent cost advantages remain real. It will tip brutally as soon as the supercycle takes a breather and the debt-financed data centers stand empty. Anyone planning between TPUs and GPUs today should take Citadel's performance numbers seriously—and the financing chain behind them just as seriously, because both hang by the same thread.

Amazon Wants to Directly Challenge Nvidia by Selling Its Own AI Chips

According to AI chief Peter DeSantis, AWS is in talks with Bloomberg about selling its self-developed Trainium chip to other data center operators. The idea comes from Andy Jassy's shareholder letter in early April: if the chip business were a standalone entity, its annual revenue would be around $50 billion. That would be about the size of Intel, but would hardly shake Nvidia with its revenue run rate of $326 billion. So far, AWS has refused to sell because the real profit lies in the waterfall behind it: storage, security, networking, and monitoring around the processed AI tokens. In addition, Trainium capacity was sold out almost immediately, as was that of Trainium4, which won't be available for over a year. Anyone wanting to sell to third parties would have to put their own customers on a waiting list or fight for more volume at TSMC, where Nvidia has just overtaken Apple as the largest customer. In parallel, Jensen Huang has declared a new $200 billion market for CPUs, pushing into Intel and AMD territory himself. → Techpresso

Synthszr Take: The real leverage isn't in the chip, but in what AWS sells around it. This is precisely why Amazon was hesitant for years: giving Trainium away as a hardware box means giving up the lucrative service layer that monetizes every processed token again. The fact that Jassy is now thinking out loud about direct sales says more about the scarcity at TSMC than about Amazon's ambition. Foundry capacity is the real moat, and Nvidia now sits on the biggest chair there. $50 billion sounds like Intel's scale, but compared to Nvidia's $326 billion, it's more of a signal than a threat. What's more interesting is the direction: Amazon wants to move away from pure cloud dependency and make Trainium a standalone business, while keeping its own customers on waiting lists. Anyone who wants to sell chips they can't get in sufficient quantities themselves should first secure production and then open up the market.

CATL Wants to Become the Battery of Data Centers

For energy storage, CATL is copying the pattern with which the group already dominates the lithium battery market: securing critical links in the value chain, channeling products through its own channels, and keeping a larger share of the margin in-house. Founder and Chairman Robin Zeng openly stated this intention in a Reuters interview in November 2024. CATL plans standalone energy systems for data centers and entire cities, and Zeng estimated the market to be 'ten times' the size of the EV battery business. In June 2026, CATL confirmed this course to industry representatives. The leverage is clear: the power hunger of AI infrastructure requires storage on a scale that makes the automotive business look small. Whoever holds cells, systems, and sales in one hand controls more than just a component. → Hello China Tech

Synthszr Take: CATL learned in batteries that you can control the chain without owning every node. Exactly this pattern is now running in fast-forward through the energy supply of AI data centers. Power is the scarce resource of inference, and whoever finances the DC systems, the megawatt capacity, and the model itself sits at the lock-in points where cost and availability are decided. In May, DeepSeek was still negotiating a $45 billion valuation; now it's over $50 billion, and CATL is in the thick of it, not just a participant. What's fascinating is the leanness of the structure: minority stakes, tied voting rights, no operational baggage. This is how you build a moat that doesn't look like control but is. For anyone in Europe talking about sovereignty in AI, this is the uncomfortable lesson: the data layer belongs to you, but the oxygen—power and capacity—is currently being bought up by someone else.

A Startup Claims to Have Cracked the LLM Compute Bottleneck

AI startup Subquadratic came out of stealth mode last month with a big announcement: it has solved a mathematical bottleneck that has been holding back large language models for nearly a decade. The alleged breakthrough consists of drastically reducing the number of calculations a Transformer needs to perform for an answer. The result is a faster, cheaper LLM that consumes far less energy than anything currently on the market. Many experts remained skeptical, but Subquadratic is now presenting initial evidence suggesting that a closer look might be worthwhile. Fittingly, the NYT reports that tech workers, who had first maximized their AI usage, are now switching back to saving: 'tokenmaxxing' has become 'tokenminning'. So the compute question is not academic; it directly hits the electricity bill. → The Download from MIT Technology Review

Synthszr Take: If the claim holds up, it's a pretty big deal, because the Transformer has been dragging its quadratic computational load with it since the 2017 Attention paper. Every longer input costs disproportionately, and that's exactly what's driving the absurd inference costs everyone is groaning about right now. Still, skepticism is warranted. We've seen several 'we've replaced the Transformer' promises that crumbled at real model sizes, and while Subquadratic is sharing evidence, it's not yet sharing a model you can torture yourself. The real point is the direction: cheaper inference is the lever that will bring AI from expensive cloud demos into the routine processes of administration, schools, and healthcare, where the multiplier effect counts. Jevons' paradox tells us anyway that cheaper doesn't mean less demand, but more use cases. Subquadratic should now put an open model on the table; then we'll talk.

Deutsche Bank Shows What AI Investments Really Yield

Denis Roux, CIO of Investment Banking at Deutsche Bank, said on June 18 that AI is cutting the processing time for some tech projects from two years to three months. The bank deliberately uses simpler models for routine tasks and is currently building tools that automatically extract financial data and link external events to the portfolio. The cost model is interesting: engineers are allocated a token budget and can request more if they can prove the added value. According to PYMNTS Intelligence, 85 percent of financial and insurance firms with over a billion dollars in revenue want to increase their AI budgets in the next twelve months. The reasons cited are productivity (65 percent), strategic positioning (65 percent), and risk reduction (55 percent). The most used applications are in the structured, auditable back office: closing entries, credit risk scoring, sales forecasting. Nvidia reports that nearly 90 percent of financial institutions are using or evaluating AI, and KPMG sees 70 percent of bank CEOs with a 10 to 20 percent AI budget. → MyClaw Newsletter

Synthszr Take: The real lesson isn't in the two years to three months, but in the token budget. Deutsche Bank is practicing compute discipline: if you want more computing power, you have to prove the return. This is exactly the mature approach we predicted at the end of December when we wrote about the shift from efficiency tool to business revolution. It's smart that the bank is starting with closing entries and credit risk scoring, because these functions are auditable and success can be measured in hours and euros (not a boardroom fairy tale, but a calculation). However, the 85 percent who are ramping up their budgets are all chasing the same back-office lever, thereby commoditizing each other. The real moat for a German bank lies in its industry knowledge of mid-sized business loans and risk models, which no platform from the Valley can replicate. Whoever now pours this domain knowledge into agent-capable products will win the race that is currently open; whoever only speeds up their month-end closing will have the same advantage tomorrow as everyone else.

Agentopia: When 100 Agents Simulate a Life for Ten Years

A team led by Xintao Wang has built Agentopia, a framework in which 100 LLM agents live autonomously over ten simulated years: pursuing personal growth, building social relationships, and fulfilling needs and goals. Previous society simulations mostly ran for days, which limited the depth of interactions and any long-term growth. The researchers define a 'Life Reward' that represents human well-being and use it to train the underlying LLM via rejection sampling. The long-term simulation results in rich emergent social behaviors, and the reward training measurably improves the model. The effect carries over to downstream role-playing benchmarks with a +15.6 percent improvement. The paper is 79 pages long with 19 figures, submitted on June 5, 2026. → Techpresso

Synthszr Take: Ten simulated years instead of a few days—that's the real lever here. The exciting thesis lies in the reward design: optimizing a model for 'well-being' over a lifetime instead of for the next token apparently gives you better social intelligence almost for free (+15.6 percent on role-playing benchmarks is not noise). In January, Moltbook was celebrating agents building their own social network; here it goes a level deeper, to the question of whether synthetic social experience is a viable training signal. The answer seems to be yes, and that has an uncomfortable consequence: high-quality human data is expensive and becoming scarce, while simulated life stories scale indefinitely. Anyone building GenAI products with character and relationship memory should take a close look at rejection sampling on lifetime rewards before burning the next expensive fine-tuning budget. The open flank remains validation: a simulation rewards what it defines itself, and whether 'Life Reward' truly captures human well-being or just a plausible shadow of it will only be decided by the real user.