エピソード

  • Chinese Models Hit 46% Token Share — The Cost War Reshapes AI

    Chinese Models Hit 46% Token Share — The Cost War Reshapes AI
    2026/07/08
    (00:00:00) Chinese Models Hit 46% Token Share — The Cost War Reshapes AI
    (00:00:42) Z.ai GLM 5.2 Record Adoption
    (00:01:24) Lindy Drops Claude for DeepSeek
    (00:02:13) OpenAI GPT-5.6 vs. Anthropic Timing
    (00:02:39) GSA Procurement Rules Confusion
    (00:03:20) UN AI Report and Global South

    Nearly half of all AI tokens processed on OpenRouter now come from Chinese models — and this week's episode breaks down exactly why that number signals a structural shift, not a trend.

    DeepSeek and Z.ai together account for up to 46% of OpenRouter token usage, up from 11% twelve months ago. The driver is price: 60–90% cheaper than OpenAI or Anthropic on comparable tasks. Z.ai's GLM 5.2 launch recorded 27x daily token growth and 80x customer growth in its first week on Vercel — the fastest adoption on the platform in 2026. AI startup Lindy went further, moving 100% of its traffic from Claude to DeepSeek in June, projecting millions in monthly savings with no meaningful performance drop.

    On the US frontier side, OpenAI is targeting a July 7–9 window for GPT-5.6, with prediction markets putting the probability at 68–74.5%. The timing aligns directly with Anthropic retiring its Fable 5 subscription tier — commercial strategy made visible.

    In governance, the US General Services Administration released updated AI procurement regulations on June 17th, but undefined terms around ideological neutrality and unclear compliance mechanics have left the industry confused. Only six formal comments were submitted in three weeks. A July 14 listening session precedes an August 3 deadline.

    Finally, the UN released its first global AI governance report, flagging that AI development is concentrated in a handful of private firms, English-language bias limits Global South access, and over 100 countries remain outside active governance conversations.

    Six stories. Clear stakes. Everything you need to stay ahead.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • 1.7M Pentagon AI Users, 12-Year Delays & Mistral's Sovereignty Bet

    1.7M Pentagon AI Users, 12-Year Delays & Mistral's Sovereignty Bet
    2026/07/07
    (00:00:00) 1.7M Pentagon AI Users, 12-Year Delays & Mistral's Sovereignty Bet
    (00:00:44) GAO Weapons Delay Report 2026
    (00:01:37) ATO Automation Gamble
    (00:02:17) Pentagon's AI Talent Crisis
    (00:02:58) Mistral's Sovereignty Bet
    (00:03:58) What to Watch Next

    The US Department of Defense has scaled its GenAI.mil platform to 1.7 million users and over 100,000 custom AI agents — yet a new GAO report reveals 104 major defense programs average more than twelve years behind schedule. Today's briefing examines that core tension: rapid AI adoption inside an acquisition system fundamentally incompatible with AI's pace of change.

    We break down the GAO's 2026 weapons delay findings, why the Pentagon's fast-track 'middle-tier acquisition' pathway is fielding immature technology, and what it means strategically when frontier AI capabilities refresh every few months but weapons systems take a decade to arrive.

    On the workforce front, the Defense Department lost over 24,000 technical employees in FY2025. Its War Force recruiting campaign targets GS-14 software engineers at salaries that compete with mid-tier tech roles — not frontier AI labs. Whether patriotic framing bridges that compensation gap is an open question.

    Also in focus: the Pentagon's pilot to automate Authority to Operate compliance using AI agents, compressing a two-year approval cycle — and the risk that faster approvals mean less genuine security scrutiny.

    Finally, Mistral AI closes a €11.7 billion Series C led by ASML and commits €4 billion to French and Swedish data centre infrastructure, positioning itself as Europe's sovereign AI alternative ahead of a major open-weight model release this summer.

    Two watchpoints to track: whether Pentagon ATO automation delivers real cycle-time cuts, and whether Mistral's summer release is technically competitive enough to make the sovereignty argument economically durable.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • AI Capex Crisis, Tesla's Grok Mandate & Mistral's $23B Sprint

    AI Capex Crisis, Tesla's Grok Mandate & Mistral's $23B Sprint
    2026/07/05
    (00:00:00) AI Capex Crisis, Tesla's Grok Mandate & Mistral's $23B Sprint
    (00:00:49) Enterprise Spending Caps Go Mainstream
    (00:01:30) Fable 5 Export Ban Resolved
    (00:02:18) Jailbreak Framework and Safety Standards
    (00:02:53) Mistral's $23B Valuation Sprint
    (00:03:34) Claude Science Beta Launch

    The structural tension at the heart of the AI industry snapped into focus this week. Capex growth is outpacing revenue growth by 46 percent — a gap wider than anything recorded during the 2001 telecom bust. The Silicon Data LLM Token Expenditure Index dropped 20 percent from its May peak, signalling the first sustained demand weakness after months of expansion. Whether it's enterprise budget exhaustion, migration to cheaper models, or a ceiling on frontier AI pricing, the revenue model is under real strain.

    That strain is showing up in company budgets directly. Uber burned through its entire 2026 AI budget by April. Meta, Amazon, and Walmart have implemented spending caps or shifted staff to lower-cost model tiers. This week Tesla joined them — capping Claude, OpenAI, and Google AI spend at $200 per engineer per week, while exempting Elon Musk's Grok entirely. Engineers reportedly prefer Claude. The policy favours Grok. That's a conflict of interest embedded in corporate procurement.

    On the regulatory front, Anthropic's Fable Five model was offline for 19 days under a Commerce Department export ban before a new safety classifier secured its reinstatement on July 1st — setting a documented precedent for regulators taking frontier models offline. Anthropic, Amazon, Microsoft, and Google also jointly released a four-criteria jailbreak scoring framework, a move that positions Anthropic to help write the industry's safety standards.

    Mistral closed a $3.5 billion raise at a $23 billion valuation — up from $11.7 billion a year ago — as enterprise buyers seek alternatives to U.S. regulatory exposure. And Anthropic launched Claude Science, a beta product targeting genomics, proteomics, and cheminformatics with auditable research pipelines and 3D protein visualisation.

    All roads lead back to the same question: can AI infrastructure spending find a revenue base to justify it?

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • Aramco's Compute Bet, OpenAI's Equity Play & Meta's Agent Reality Check

    Aramco's Compute Bet, OpenAI's Equity Play & Meta's Agent Reality Check
    2026/07/04
    (00:00:00) Aramco's Compute Bet, OpenAI's Equity Play & Meta's Agent Reality Check
    (00:00:56) OpenAI's Government Equity Proposal
    (00:01:30) Meta's Agent Reality Check
    (00:02:06) Anthropic Samsung Chip Talks
    (00:02:36) Google Power Surge and Microsoft Frontier
    (00:03:12) Kling AI, Nvidia Financing, and FTC Watch
    (00:03:52) Key Watchpoints Going Forward

    Today's episode cuts through seven major AI developments reshaping who controls compute, chips, capital, and implementation at scale.

    Saudi Aramco Ventures led an $800 million funding round for Together AI, pledging 500 megawatts of dedicated compute capacity. This is a sovereign infrastructure play, not a startup bet — treating compute the way nations once treated oil reserves. Together AI's ATLAS speculative decoding architecture claims cost advantages of 6x to 60x over alternatives, though those figures remain unvalidated.

    OpenAI has proposed offering the US government a 5% equity stake, valued at roughly $42.6 billion, framing it as a sovereign wealth model designed to align government incentives with AI development rather than pure regulation. Whether the Trump administration accepts — or whether Anthropic and Google follow — is unresolved.

    Mark Zuckerberg acknowledged internally that Meta's agentic AI progress is slower than expected, a rare admission that signals the gap between demo-ready agents and enterprise-ready systems is wider than industry narratives suggest.

    Anthropment has entered preliminary talks with Samsung on custom 2nm AI accelerators, part of a broader frontier-lab push to reduce Nvidia dependency. Google's environmental report revealed a 37% year-over-year electricity increase driven entirely by AI expansion. Microsoft launched Frontier Company, embedding 6,000 engineers with enterprise clients for $2.5 billion — betting the AI bottleneck is now implementation, not models.

    Finally, Kuaishou's Kling AI raised $2.8 billion at an $18 billion valuation, Nvidia shifted toward GPU financing models for startups, and the FTC opened a public comment period on AI accuracy — a reliable precursor to enforcement action.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分
  • Microsoft's $2.5B Deployment Bet & AWS's Counter-Move

    Microsoft's $2.5B Deployment Bet & AWS's Counter-Move
    2026/07/03
    (00:00:00) Microsoft's $2.5B Deployment Bet & AWS's Counter-Move
    (00:00:36) AWS Moves Forty-Eight Hours Earlier
    (00:01:12) OpenAI and Anthropic Partnership Model
    (00:01:44) Microsoft Multi-Model Pivot
    (00:02:24) Sovereign Cloud and Regulated Markets
    (00:03:01) Frontier Company Early Clients

    Two cloud giants. One week. One converging thesis: deployment is the new moat.

    Microsoft today launched the Frontier Company, a dedicated outcomes-driven AI implementation unit backed by two and a half billion dollars and six thousand engineers. Forty-eight hours earlier, AWS committed one billion dollars to its own AI deployment venture. When two hyperscalers move on the same thesis in the same week, that's not coincidence — it's a structural shift in how the AI industry generates revenue.

    This episode unpacks what that shift means for enterprise AI adoption, and how it compares to the approach taken by OpenAI and Anthropic, both of which are raising external capital for forward-deployed engineering rather than building internal operating units. Which model actually delivers measurable ROI at scale is the central question for the next eighteen to twenty-four months.

    Also covered: Microsoft's quiet pivot to model pluralism — customers can now choose between Microsoft, third-party, and open-source models — and a deeper look at the Kyndryl-Microsoft sovereign cloud offering entering government procurement cycles, with a two-billion-plus framework being restructured around data residency and sovereignty bundles.

    The Frontier Company's early client list spans the London Stock Exchange Group, Unilever, Land O'Lakes, and Accenture — a deliberate stress test across financial services, consumer goods, agriculture, and professional services. The proof point to watch: whether outcome-driven contracts hold up at enterprise scale when the first renewals land.

    This episode includes AI-generated content.
    続きを読む 一部表示
    4 分
  • Regulated Industries Making Irreversible AI Bets | Ep. 1

    Regulated Industries Making Irreversible AI Bets | Ep. 1
    2026/07/02
    (00:00:00) Regulated Industries Making Irreversible AI Bets | Ep. 1
    (00:01:12) Haleon's $5B Microsoft Commitment
    (00:02:15) Government AI Gatekeeping Hardens
    (00:03:18) Claude Sonnet 5 Mid-Tier Agentic Push
    (00:03:48) California Statewide Claude Deal
    (00:04:24) Engram's $98M Memory Layer Bet
    (00:04:56) Key Signals to Watch

    Regulated industries are locking in major AI commitments before the governance frameworks around them are fully formed — and today's episode maps exactly where those bets are being placed and what's still missing.

    The FDA is replacing its static device-approval model with a lifecycle framework built around Predetermined Change Control Plans, letting medical AI algorithms evolve continuously without fresh submissions. The architecture is ambitious. The enforcement path when an algorithm quietly drifts in a clinical setting is still murky.

    Haleon — maker of Sensodyne and Advil — just signed a five-year, multi-billion-dollar Microsoft deal deploying Copilot agents and Azure infrastructure across 25,000 employees. The story isn't the size; it's the single-vendor lock-in logic spreading across pharma and healthcare, and the unresolved liability question when Copilot agents start generating product claims.

    Government gatekeeping of frontier models is hardening into a pattern without statutory footing. GPT-5.6 shipped to roughly 20 authorized organizations under national security review. Anthropic's Fable 5 export restrictions were lifted, while Mythos 5 stays restricted. Labs are complying with government pulls on model access that have no clear legal framework — a real planning risk for any developer building on these platforms.

    Also covered: Claude Sonnet 5 hits 63.2% on agentic coding benchmarks at introductory pricing that narrows the mid-tier-to-premium gap faster than expected. California makes Claude the first AI tool available statewide. And Engram emerges from stealth with $98M from General Catalyst, Kleiner Perkins, and Sequoia to build an enterprise memory layer — with internal benchmarks only so far.

    Two signals to watch: the FDA's first enforcement action under its drift-detection framework, and whether government model-gatekeeping criteria ever get formalized.

    This episode includes AI-generated content.
    続きを読む 一部表示
    6 分
  • Copilot's $750 Bill, DeepMind Exodus & China's 1.6T Model

    Copilot's $750 Bill, DeepMind Exodus & China's 1.6T Model
    2026/07/01
    (00:00:00) Copilot's $750 Bill, DeepMind Exodus & China's 1.6T Model
    (00:01:05) Google DeepMind Talent Exodus
    (00:02:03) Meituan 1.6T Parameter Model Claim
    (00:02:43) Pentagon War Force AI Recruitment
    (00:03:15) Generative AI Revenue and Agent Economics
    (00:04:01) Cursor iOS and What Comes Next

    The first full month of GitHub Copilot metered billing is in, and some developers are staring at invoices between $750 and $3,000 — a stark jump from the $29–$50 flat-rate era. Today's briefing unpacks why agentic workflows consume roughly 1,000 times more tokens than standard chat, and what that means for every AI coding tool on the market.

    At Google DeepMind, the talent story is accelerating. At least six senior researchers have departed since February, including AlphaFold co-creator John Jumper, who has joined Anthropic. Pre-IPO equity windows and a widening gap in AI-authored code rates — Anthropic at roughly 100%, Google at 50% — are driving both the exits and a structural reorganisation of DeepMind's AI coding teams.

    China's Meituan has open-sourced a claimed 1.6 trillion-parameter model trained on domestic silicon under Apache 2.0. If independently verified, it represents a meaningful challenge to US export control strategy. Independent benchmarking has not yet confirmed the performance claims.

    Elsewhere: the Pentagon's War Force initiative is pulling hundreds of AI engineers into two-year government deployments, tightening an already constrained talent pool. Generative AI hit $110 billion in revenue this year, scaling three times faster than the early internet. And Cursor has entered public beta on iOS, making mobile-native agent management a serious infrastructure story.

    Google DeepMind is committing $10 million to multi-agent safety research covering coordination failures and prompt injection — because deployment is already outpacing the safety map.

    This episode includes AI-generated content.
    続きを読む 一部表示
    5 分