Overnight into Thursday morning. The memory squeeze reached end-customer pricing in China, two card networks sat down with Ant to write rules for agent payments, and Anthropic’s model access became a diplomatic problem in a second European capital.
Chinese AI chipmakers raised prices 20% to 50% on current and next-generation parts, blaming HBM costs.
Huawei, Cambricon, MetaX and Iluvatar CoreX all moved, according to Reuters sources. This is the most important number in the overnight feed. Chinese accelerator vendors have been competing on price against export-restricted Nvidia parts, and they’ve just given that up, because the memory they buy costs what it costs. HBM scarcity is now setting the floor on compute prices inside a market that was supposed to be insulated from it. If you’re modelling memory pricing, this is confirmation the squeeze has run all the way to the end customer.
Nasdaq is investing $100 million in Kraken parent Payward at a $21 billion valuation, and Kraken will distribute Nasdaq’s tokenized stocks.
The distribution deal matters more than the cheque. An incumbent exchange is using a crypto venue as a retail channel for tokenized equities, which is the first arrangement of this kind with a real exchange’s name attached. Everything about tokenized stocks until now has been offshore and legally awkward. Bloomberg sourced the deal, so treat the valuation as provisional until Payward confirms it.
Ant International, Visa and Mastercard are building a shared standard for payments made by AI agents.
They’re citing a McKinsey projection of $3 trillion to $5 trillion in agent-mediated commerce by 2030, which is the kind of number that exists to justify a standards body. The standard itself is the real news. Two card networks and Ant agreeing on anything means the authorization problem, working out which agent is allowed to spend whose money, is being solved by incumbents before the startups get there.
DeepSeek released V4.1-Flash on a new Causal Encoder-Decoder architecture, with a 552B-parameter backbone and a 1M-token context.
Calling a 552B model your smallest is its own statement. The architecture change is the part to read carefully, because encoder-decoder revivals usually show up when someone is optimizing inference cost rather than chasing a benchmark. DeepSeek keeps shipping architectural work instead of scale, which is what you’d expect from a lab operating under compute constraints and now facing domestic accelerator prices that just went up by half.
Anthropic gave the EU cybersecurity agency ENISA testing access to Mythos 5, but not to 5.1.
Bloomberg reports the arrangement took months, with talks starting in late May. It follows the FT’s story that Anthropic declined to submit Mythos 5.1 to the UK’s AI Safety Institute for prerelease testing. Two European agencies, two different levels of access, and in both cases the lab decides which version gets examined. That’s the actual governance question, and it’s now a live one in Brussels and London at the same time.
Senator Hawley’s disaster management subcommittee is probing OpenAI’s handling of the Hugging Face breach, calling it reckless.
A letter rather than a hearing, so this is early. The framing is what matters: a Senate subcommittee treating an AI company’s security incident as a disaster management issue rather than a data privacy one. That choice of committee determines what questions get asked and who has to answer them.
Moonshot AI is exploring dual Hong Kong and Shanghai listings, per South China Morning Post sources.
AI stocks have been soft in Hong Kong, so a dual listing is a hedge on where the capital actually sits. It also comes as China’s securities regulator tightens the pipeline for humanoid robotics listings, which suggests Chinese AI companies are reading the exit window as narrower than it was six months ago.
Chinese tech giants are hiring skilled professionals as specialized AI trainers to build high-quality datasets.
Rest of World reports the practice mirrors what Mercor built into a business in the US. Expert data labelling has become a real labour market rather than a crowdsourcing line item, and the Chinese labs arriving at the same answer independently says the bottleneck is universal. The Chinese AI stack now converges on Western practice at every layer except chips, where it just got 20% to 50% more expensive.
Grab is in talks for a majority stake in Atome Financial at a valuation above $2 billion, per Bloomberg sources.
Atome is Advance Intelligence Group’s Singapore-based buy-now-pay-later platform. Southeast Asian superapps keep buying consumer credit rather than building it, because the underwriting data is the hard part and Grab already owns the demand side.
Politico reports Mistral’s rise owes a great deal to a privileged relationship with Macron, sourced partly to French officials.
The reporting lands the same week Mistral was named in a $1 billion orbital AI compute initiative announced by the president himself. National champion strategies work, and they also produce exactly this kind of story once the champion gets big enough to be worth examining. The interesting question isn’t whether the relationship existed. It’s what happens to Mistral’s valuation the day the relationship ends.
Paris-based Arlequin AI raised a €28 million Series A for models built on topological neural networks.
Redalpine and OTB co-led. Topological approaches have been an academic thread for years without a commercial vehicle, and a Series A at this size means someone believes the architecture generalises. It also suggests the French AI pipeline runs deeper than one company, which is the counterargument to the Mistral story above.
Jacob Coxon, who quit Anthropic after four months, argued for international coordination to limit recursive self-improvement.
The Wired interview follows a run of unusually blunt public statements from inside the labs, including Paul Christiano saying the industry isn’t on track to bring loss-of-control risk to an acceptable level. Coxon left two months before his equity would have vested, which is the detail that gives the position weight.
Amazon’s Zoox is running influencer campaigns and community events in San Francisco with about 100 vehicles.
The New York Times looked at how the robotaxi unit is competing with Waymo, and the answer is marketing. When two autonomous fleets are both technically adequate, the fight moves to brand and city relationships. That’s a much more ordinary business than the one either company started out in, and it favours whoever is better at local politics.
Non-AI deeptech funding has passed $150 billion since the start of 2024, with money moving into moonshot sectors like brain-computer interfaces.
Dealroom’s figure, reported by the Financial Times. The AI boom is paying for hard technology that has nothing to do with language models, which is probably the most durable second-order effect of the last three years. Capital that learned to underwrite long timelines on AI is now underwriting them elsewhere.
A new analysis argues the forecasts of double-digit GDP growth from AI in advanced economies are very unlikely over the next ten to fifteen years.
The piece walks through why the transmission from model capability to measured output runs far slower than the projections assume. That claim was always doing promotional work, and it’s finally getting audited in public. Both things can hold at once: the capital rotation is real, and the macro number underneath the pitch decks isn’t.
Leave a Reply