Wednesday, June 24, 2026

Prior editions tracked rated project bonds and local backlash; today the financing architecture itself shifted, with Meta executing a $26 billion sale-leaseback while Amazon, Google, Microsoft, and Meta diverge across owned, leased, and offtake-backed structures—against more than 300 US data-center development bans passed since 2023. An Electron Economics analysis compared widening AI-infrastructure CDS spreads to 2007 monoline stress, making those spreads, further sale-leaseback issuance, and whether flexibility measures like Google's 1 GW of flexible capacity and a Voltus deal of up to 100 MW soften siting opposition the variables to watch.

SK Hynix's ADR plan and its rise past Samsung were covered in prior editions; today the company filed a US prospectus targeting a $29 billion listing, OpenAI unveiled its Jalapeño inference chip with Broadcom to cut Nvidia dependence, and Nvidia Blackwell servers reportedly fetch over $1.1 million on China's black market—triple US prices. The grey-market premium quantifies how porous export controls have become in practice, and the open questions are Jalapeño's route from announcement to deployed silicon, the ADR's pricing outcome, and whether SambaNova completes its reported $800 million–$1 billion raise at roughly a $10 billion valuation.

Anthropic's server-side Claude Managed Agents runtime is now deployed at enterprises including Notion, Rakuten, Asana, Sentry, and Atlassian, and its Slack-native Claude Tag merges 65% of Anthropic's own product pull requests—yet security researchers at Air gained full control of more than 26,000 agents through a single malicious skill promoted with one Instagram ad. No security scanner flagged the skill and open-source community members actively vouched for it, so the gap between enterprise-scale agent deployment and skill-vetting infrastructure is the uncertainty that bears watching, including whether vendors introduce review or signing standards.

New in today's intake: the forty-year stock-bond hedge underpinning 60/40 portfolio construction is reported to be fracturing as inflation uncertainty shifts correlations, with simultaneous S&P 500, Nasdaq, and Nikkei sell-offs, while yen devaluation pressures the global carry trade and forces large institutional rebalancing that amplifies cross-asset volatility. Whether equities and bonds move together in the next drawdown—and how far the yen's slide extends—are the observable tests distinguishing a durable regime change from a passing episode.

The Claude Fable 5 suspension has run through every edition since June 15; today's deltas are the NSA losing access to an Anthropic model in a contractual dispute and prediction markets pricing 60% odds that Fable 5 access is restored by July 1. The geopolitical backdrop sharpened simultaneously—China's LineShine reclaimed the top supercomputer ranking for the first time since 2017 and EU allies joined a US pact to reduce Chinese AI supply-chain dependence—so the July 1 restoration marker and early evidence of pact implementation are the near-term observables.

• Nvidia Blackwell servers fetch over $1.1 million on China's black market, triple US prices; Nvidia is also extending its platform into biomedical research.?
• SK Hynix filed a US ADR prospectus targeting a $29 billion listing, having overtaken Samsung as South Korea's most valuable company.?
• OpenAI unveiled its Jalapeño inference chip with Broadcom to cut Nvidia dependence; SambaNova is raising $800 million–$1 billion at roughly $10 billion valuation.?
• Unitree's G1 humanoid marks China's robotics push; analysts at Crux Capital see edge compute and perception hardware as the most investible physical-AI layers.?

• Model-layer revenue in H1 2026 concentrates in an Anthropic–OpenAI duopoly, while Chinese open-source models like z.AI's GLM-5.2 gain US enterprise customers.?
• OpenAI disclosed at Cannes that roughly 20% of ChatGPT queries carry direct commercial intent, signalling a new advertising revenue path.?
• Cursor's coding harness lifted model benchmark scores roughly 11 percentage points on average; Claude Opus rose from 77% to 93% on one test.?
• Apple's on-device model fits roughly 3 billion parameters using 2-bit quantisation; ParallelKernelBench's best zero-shot multi-GPU kernel score was only 28/87.?

• Anthropic launched Claude Tag, a Slack-native multiplayer agent; the internal version merges 65% of Anthropic's own product pull requests.?
• Claude Managed Agents, a server-side runtime, is deployed by enterprises including Notion, Rakuten, Asana, Sentry, and Atlassian; Interactive Brokers integrated Claude, Grok, and ChatGPT.?
• Microsoft open-sourced the ASSERT evaluation harness; Matei Zaharia, Databricks CTO, open-sourced Omnigent, composing Claude Code, Codex, and other agent SDKs.?
• Bay Area Founders Club is hosting a Claude Code masterclass led by Chinat Yu, a Stanford Learning, Design & Technology educator.?

• Amazon, Google, Microsoft, and Meta use divergent owned, leased, and offtake-backed structures; Meta executed a $26 billion sale-leaseback reshaping AI infrastructure financing.?
• Google reached 1 GW of flexible data-centre capacity, signed a Voltus VPP deal up to 100 MW, alongside nuclear and geothermal partnerships.?
• Groq raised $650 million to scale its inference cloud toward 200 MW by 2027; RunPod raised $100 million after refusing acquisition offers.?
• More than 300 US data-center development bans passed since 2023; an Electron Economics analysis compared widening AI-infrastructure CDS spreads to 2007 monoline stress.?

• China reclaimed the world's top supercomputer ranking with LineShine, its first since 2017; EU allies joined a US pact reducing Chinese AI supply-chain dependence.?
• The NSA lost access to an Anthropic model amid contractual dispute; prediction markets priced 60% odds of Claude Fable 5 restoration by July 1.?
• Dario Amodei, Anthropic co-founder and CEO, proposed an international AI governance framework at the G7; the EU published its Technological Sovereignty Package June 3.?
• The Trump administration pressed Meta to submit AI models for voluntary review; a Lawfare analysis argued the executive order's voluntary evaluations become effectively mandatory.?

• The forty-year stock-bond hedge underpinning 60/40 portfolios is fracturing as inflation uncertainty shifts correlations, with simultaneous S&P 500, Nasdaq, and Nikkei sell-offs.?
• Yen devaluation is pressuring the global carry trade, forcing large institutional investors to rebalance in ways amplifying cross-asset volatility.?
• Kalshi, a prediction market exceeding $2 billion annualized revenue, sued Illinois arguing federal CFTC jurisdiction preempts the state's new licensing law.?
• CryptoQuant analysts argued Strategy should pause Bitcoin purchases and rebuild cash as its Bitcoin, MSTR, and STRC preferred structures weaken simultaneously.?

• Meta unveiled AI-powered smart glasses with EssilorLuxottica starting at $299, calling smart glasses 'the most exciting hardware category of the AI era.'?
• Meta is building a prediction-markets app to compete with Polymarket and Kalshi, though it cannot replicate their regulatory moat.?
• Meta paused an internal AI training programme that tracked employee keystrokes after an internal leak provoked staff backlash.?

• Salesforce agreed to acquire Fin, formerly Intercom, for $3.6 billion; $100 million of Fin's $400 million run rate is outcome-based AI support pricing.?
• Superhuman agreed to acquire GPTZero, the AI-detection company co-founded by Edward Tian and Alex Cui.?
• Databricks is reportedly eyeing a $175 billion valuation ahead of a potential 2027 IPO; its Genie data agent claims 3x generic-agent accuracy.?
• Companies are shifting from maximizing AI token use to rationing by ROI; an AI memory startup cutting token costs raised $98 million.?

• Security researchers at Air gained full control of over 26,000 agents via a malicious skill promoted with a single Instagram ad.?
• No security scanner flagged the malicious skill, and the open-source community actively vouched for it, exposing weak vetting in agent ecosystems.?
• Zvi Mowshowitz, independent AI researcher, asked whether a superhuman AI hacker could write code it could not break with 100× more effort.?
• Replies noted bug-free code is theoretically possible but social-engineering attacks have no perfect defense, and cryptographic systems already resisted 100× build effort.?

• Amazon Prime Day 2026 expanded to a four-day event; Airbnb's accelerating growth and potential super-app expansion into travel and ride-sharing were assessed.?
• Intercalation Station covered neutron and muon beam battery research at Oxfordshire's ISIS facility; Ankit Kalanki of RMI discussed air-conditioner testing-standard reform.?
• Several items were newsletter subscription promotions: PhotonCap's 20%-off annual deal ends July 6; AI CFO Office Insider pricing rises from $600 to $799.?
• An X post noted an unusual eight-day lull in major AI news; one Semi-Doped YouTube link returned a 403 error with no content.?