Thursday, June 25, 2026

Meta is building an AI-run prediction-market app. On the policy front, the Fable 5 / Mythos 5 standoff has a genuinely new development — Tom Brown is now leading Commerce talks and a Jun 26 deadline looms. (No Jun 24 digest ran, so this cycle covers Jun 23–25.)

TLDR

OpenAI + Broadcom unveil "Jalapeño," OpenAI's first custom inferenceRunning a trained model on new input to get an output, as opposed to training it.NVIDIA ↗ chip (Jun 24) — LLM-optimized ASIC, taped out in 9 months, first deployments targeted end of 2026 Infra
Qualcomm to acquire Modular for ~$4B (Jun 24) — buys the cross-hardware AI software stack to challenge Nvidia's CUDA lock-in Infra Trending
Meta building "Arena," an AI-run prediction-market app (Jun 24) — Llama generates the questions and resolves the markets; play-money to start Customer AI Trending
Fable 5 / Mythos 5 update: Tom Brown leads Commerce talks, Trump softens, Jun 26 deadline (Jun 24) — models still offline; Polymarket ~41% for US restoration by Jun 26 Policy
ByteDance announces Seedance 2.5 video model (Jun 23) — native 30-second clips, up to 50 reference inputs; enterprise beta now, public launch early July Trending

Action Items

Things your engineering team should look at or do this week.

Audit user-facing AI for the EU AI Act Article 50 transparency rules (effective Aug 2). If you serve EU users, AI-generated content and chatbot interactions need disclosure/labeling. Parliament approved amendments Jun 16, but the Aug 2 transparency deadline stands — scope your obligations now. (Sidley)
Hold any Fable 5 / Mythos 5 fallbacks in place. Restoration odds improved but there's still no confirmed date (Polymarket ~41% by Jun 26). Don't unwind contingency routing yet. (Korea JoongAng Daily)
Re-check provider SLAs for production traffic. Claude logged its third disruption of the month; 90-day uptime sits below the 99.9% enterprise bar across claude.ai, Code, and API. Confirm cross-provider failover, not just model fallback. (Thoughtworks)

Models & Tools

OpenAI and Broadcom unveil "Jalapeño," OpenAI's first Intelligence Processor (Jun 24) Infra — A reticle-sized ASIC architected specifically for LLMLarge language modelA neural network trained on huge amounts of text that generates text, answers questions and summarizes.Anthropic ↗ inference (kernels, memory movement, networking, serving patterns), co-developed from design to tape-out in ~9 months. Engineering samples are already running ML workloads in-lab at production frequency/power, including GPT-5.3-Codex-Spark; initial deployment is targeted for end of 2026, scaling to gigawatt-scale data centers. It's a clear signal OpenAI wants to own inference economics rather than rent GPUsGraphics processing unitA highly parallel processor that does the large matrix math behind training and running models.NVIDIA ↗. (OpenAI | CNBC | TechCrunch)
ByteDance announces Seedance 2.5 video model (Jun 23) Trending — Unveiled at Volcano Engine's conference: single native 30-second clips (no stitching), up to 50 reference inputs per generation, and region-level editing that re-draws part of a frame. Not yet GA — framed as a global enterprise beta, with public launch targeted for early July. Aims squarely at Sora, Veo, and Runway. (The Information | Seedance)

Industry & Business

Qualcomm to acquire Modular for ~$4B (Jun 24) Infra — Qualcomm announced it will buy Modular, whose software lets AI models run efficiently across Nvidia, AMD, Intel, and other silicon — a direct play at loosening Nvidia's CUDA grip. The deal (up to 19.2M shares, expected to close 2H 2026) is a ~2.5x markup on Modular's $1.6B valuation from under a year ago, and central to CEO Cristiano Amon's pivot beyond smartphones into AI infrastructure. (CNBC | TechCrunch | Bloomberg)
Meta building "Arena," an AI-run prediction-market app (Jun 24) Customer AI Trending — Per internal docs (codenames "Antwerp"/"FBForecast"), Zuckerberg has tasked a team with a standalone app where Llama auto-generates markets from trending topics, makes personalized recommendations, and resolves outcomes — i.e., the model decides what did or didn't happen. Launches with daily play-money allotments, taking aim at Kalshi and Polymarket. (NPR)
Vertical AI funding stays hot (Jun 24) Business — Assort Health raised a $120M Series C (Menlo, Lightspeed; ~$222M total) and Taktile, a financial-services AI firm, raised a $110M Series C (Goldman Sachs Alternatives, Balderton, Index; ~$184M total) — both landing the same day, underscoring continued appetite for applied/vertical AI. (Tech Startups)

AI in Production

Claude reliability under scrutiny after a third June disruption (Jun 23) Production — Analysis, not a new incident. Following back-to-back outages, Anthropic's own status page shows 90-day uptime of ~99.12% (claude.ai), 99.28% (Code), and 99.41% (API) — roughly 19–23 hours of downtime per service per quarter, below the ~99.9% most enterprise contracts require. The takeaway for teams treating Claude as infrastructure: budget for cross-provider redundancy and degraded-mode UX, especially for agenticDescribes AI systems where a model uses tools and decides its own steps instead of following a fixed script.Anthropic ↗ pipelines where a single dependency stalls the whole chain. (Thoughtworks | Status)

Policy & Safety

Fable 5 / Mythos 5: negotiations heat up, Jun 26 deadline looms (Jun 24) Policy — Update to an ongoing story. The models remain globally suspended (day 14), but the picture shifted: co-founder/chief compute officer Tom Brown took over Washington engagement (arriving at Commerce Jun 15 with policy head Sarah Heck), and by Jun 19 Trump said Anthropic was "no longer a national-security threat." Commerce Secretary Lutnick must respond to four members of Congress by Jun 26, a hard political marker; Polymarket puts US restoration by Jun 26 at ~41% (down from a 60% spike). Next structural dates: Jul 8 (ID-verification policy) and Aug 1 (EO deadline). (Korea JoongAng Daily | Fortune)
EU AI Act: Article 50 transparency obligations live Aug 2 (Jun 24) Policy — Providers and deployers of generative and interactive AI must meet disclosure/labeling requirements from Aug 2, 2026. The European Parliament approved material amendments Jun 16 (some obligations delayed), but transparency duties are not among the items pushed back — EU-facing teams should scope compliance now. (Sidley | Morgan Lewis)

Custom-silicon day: OpenAI and Broadcom unveiled OpenAI's first inference chip, and Qualcomm bought Modular for ~$4B to take a run at Nvidia's CUDA moat.