JadidJust-in AI Developments, Incidents & Deadlines
July 2026

The circular-financing era went supersize overnight: Nvidia is in talks to guarantee ~$250B of OpenAI's Ohio data-center lease — with up to $350B more discussed for chips — a deal even Michael Burry mocked as Nvidia underwriting demand for its own silicon.

6 stories

Meanwhile Kimi K3's open weightsA model whose trained weights are published for anyone to download and run.NTIA ↗ actually landed at 00:00 UTC (yesterday's "imminent" is now live), and GitHub's bug-bounty payout cuts take effect today, a direct reaction to the flood of AI-generated vuln reports.

TLDR

  • Nvidia in talks to backstop ~$250B of OpenAI's $500B Ohio data-center buildout — plus up to $350B more to finance chip purchases; negotiations are early-stage and could still collapse (Jul 26) Infra Trending
  • Kimi K3 open weights went live at 00:00 UTC — the 2.8T-param model is now downloadable on Hugging Face, making yesterday's imminent release real; ~1.4TB self-hosting footprint and ~51% hallucination rate still apply (Jul 27) Open Source Trending
  • GitHub's bug-bounty payout cuts take effect today — public rewards drop ≥50% at every severity (critical $20–30K+ → fixed $10K); top rewards move to an invite-only VIP tier, in response to AI-generated report spam (effective Jul 27) Security
  • CuspAI raises $450M at $2.6B to scale AI materials discovery — Kleiner Perkins/NEA-led, with an "AI Materials Foundry" of 45+ orgs incl. Nvidia, Meta, Samsung; ~80% of research aimed at semiconductor materials (Jul 21) Data

Action Items

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

  • Evaluate Kimi K3 now that the weights are actually live. Confirm you're pulling from the verified Moonshot moonshotai Hugging Face org, check the license carefully (reports conflict between Apache 2.0 and a Modified MIT license), and price out the MXFP4 hosting footprint (~1.4TB of fast memory before context) before any self-host pilot. Gate any pilot on your own factuality evals — independent testing puts its hallucination rate near 51%. (VentureBeat | Kili Technology)
  • Finish MCPModel Context ProtocolAn open standard that lets AI apps connect to data sources, tools and workflows through one common interface.MCP ↗ SDK-beta validation today — the final 2026-07-28 spec lands tomorrow. The stateless spec (no session header, breaking changes, servers repositioned as OAuth 2.1 resource servers) moves from release candidate to final Jul 28. If you run or consume MCP servers, run your beta tests now rather than after the cutover. (MCP Blog)

Models & Tools

  • Kimi K3 open weights are live (update) (Jul 27) Open Source Trending — Yesterday's "publishing tonight" is now real: Moonshot's 2.8T-parameter MoEMixture of expertsA model made of many sub-networks, called experts, where a learned router sends each input to only a few of them.Shazeer et al. 2017 ↗ (1M-token context, ~16 of 896 experts active per tokenA small unit of data, such as part of a word, that AI models read and generate. Many AI services price by tokens in and out.NVIDIA ↗) is downloadable on Hugging Face in MXFP4. It's the largest open-weight release to date and instantly the strongest open model by Intelligence Index, but self-hosting still needs ~1.4TB of fast memory, and Moonshot's charts omit the independently-measured ~51% hallucination rate (up from 39% on K2.6 even as factual accuracy improved). Licensing is being reported inconsistently — verify before you build. (VentureBeat | Interconnects)

Security

  • GitHub slashes public bug-bounty payouts, effective today (effective Jul 27) Security — GitHub is moving to a two-tier program: public submissions now pay ≥50% less at every severity (critical drops from $20–30K+ to a fixed $10K; low-severity ~59% lower), while a permanent invite-only VIP tier pays $30K+. Reports filed before today keep the old terms. GitHub says the change is a direct response to the flood of low-effort, AI-generated vulnerability reports burying its security team — a signal of how much AI submission-spam is reshaping bounty economics. (The Register | The Hacker News)

Industry & Business

  • Nvidia in talks to backstop ~$250B of OpenAI's Ohio data center (Jul 26) Infra Trending — Per WSJ (via Reuters/Bloomberg), Nvidia would guarantee ~$250B of lease and construction debt for a 10-gigawatt campus SoftBank's SB Energy is building on a former uranium-enrichment site in Piketon, Ohio — part of a >$500B total buildout — and is separately discussing up to $350B to finance OpenAI's chip purchases. It would be OpenAI's first big step toward owning infrastructure rather than renting from Microsoft/Amazon/Oracle. OpenAI lacks an investment-grade rating and is projected to lose ~$14B on ~$25B revenue in 2026; Michael Burry blasted the structure as Nvidia guaranteeing spend on its own chips. Talks are early and could collapse. (Bloomberg | Reuters via Yahoo Finance)
  • CuspAI raises $450M Series B at $2.6B valuation (Jul 21) Data — The Cambridge, UK materials-discovery startup (up from a $520M valuation last September) raised a Kleiner Perkins/NEA-led round with Bezos Expeditions, Lux Capital, and AMD Ventures, and launched an "AI Materials Foundry" pooling compute and lab access across 45+ orgs including Nvidia, Meta, Samsung, and Hyundai. ~80% of its research this year targets semiconductor materials, including replacing supply-constrained metals like ruthenium and iridium in chipmaking. (Reuters via Yahoo Finance | TechNode Global)
  • AI-cited tech layoffs pass a grim milestone; July cuts up 48% MoM (Jul 2026) Hiring — Trackers now count ~168K tech jobs cut across ~467 events in 2026, with a majority of July events explicitly citing AI/automation; AI chips/hardware/semis drove ~39% of the month's total. Notable: Oracle has shed ~21K roles over 12 months even as it plans thousands of AI-datacenter hires, illustrating the "cut non-AI, hire for AI infra" reshuffle. (TechCrunch)

Policy & Safety

  • Hawaii enacts its first AI guardrails (signed mid-Jul) Policy — Gov. Josh Green signed two bills: SB3001 requires AI companion platforms to disclose they're machines, add self-harm/suicide-response protocols, and file annual reports to the state Health Department; HB2137 (enacted as Act 247 on Jul 14) bans non-consensual realistic AI imitations of people for commercial/harmful use, mandates labeling, and creates a private right to sue. If you ship user-facing companion or generative-media AI, add Hawaii to your state-compliance matrix. (Hawaii Governor's Office | Honolulu Star-Advertiser)

Type to search · ↑↓ to move · enter to open