Open But Empty: Kimi K3's Record Weights Drop as Adoption Stalls at 11%

Capability & Integration

• Moonshot AI published the full weights for Kimi K3 today, making the 2.8-trillion-parameter, MXFP4-quantized model (roughly 1.4TB) the largest open-weight release to date; Artificial Analysis ranks it third globally, behind Claude Fable 5 and GPT-5.6 Sol, per Bloomberg and Tech Times. The model leans on memory bandwidth more than raw compute — a design choice Bloomberg says helped trigger a chip-stock selloff on release, per Bloomberg.

  Unverified

  Tech Times reports Moonshot has not disclosed hallucination-rate benchmarks for K3; no independent measurement of that metric was found.

  Review

  “published the full weights for Kimi K3 today… per Bloomberg and Tech Times” — neither cited source actually confirms a same-day release: the Bloomberg piece is dated July 17 and covers the initial model/API launch, and the Tech Times piece is dated July 25 and is forward-looking (“arrive Sunday”). The weights release itself is independently corroborated as roughly accurate for July 27, but the citations don’t support “today” as written — suggested action: cite a same-day source or soften to “this week.”

• A study led by MIT FutureTech and Carnegie Mellon researchers, scoring AI-related language across 4,400 firm-year S&P 500 filings, found just 11% of companies had “deeply integrated” AI into core operations by 2025 — up from 5% in 2022, but concentrated heavily in tech firms, per the arXiv preprint.

Robotics

• Figure AI’s BotQ factory in California now assembles one Figure 03 humanoid per hour, up from one per day roughly 120 days earlier, per Figure AI’s own announcement. That’s a manufacturing throughput figure, not a deployment count — the robots doing factory work are a separate, smaller tally, with roughly 40 units placed at BMW’s Spartanburg plant, per Collision Repair Mag.

• Update: Unitree says the H1 Pro rollout that began with commercial sales in Europe on July 22 (covered here Friday) continues to Asian markets August 5 and North America August 12; the Bluetooth exploit flagged in that coverage remains unresolved.

  Unverified

  The Asia/North America launch dates trace to a single industry-tracker aggregator, not a primary Unitree announcement or established outlet.

Hardware & Supply Chain

• Nvidia is reportedly reworking parts of its 2028 “Feynman” chip after failing to secure enough capacity on TSMC’s new A16 process, shifting non-critical components to the older 3nm N3P node; TSMC’s most advanced nodes are booked out to 2028, per Yahoo Finance.

  Unverified

  The redesign traces to unnamed industry sources circulated through trade press; neither Nvidia nor TSMC has confirmed it directly.

• Shanghai startup Dongfang Suanxin unveiled the DF1000, a 14nm AI chip it says delivers higher memory bandwidth than Nvidia’s H200 without relying on export-controlled HBM, with mass-production shipments planned by year-end, per TrendForce.

  Unverified

  The H200 bandwidth comparison is the company’s own claim; no independent benchmark was found.

Environmental & Cultural Impact

• Google consumed 10.9 billion gallons of water in 2025, up 34% from 2024 and more than double its 2021 level, driven largely by AI infrastructure growth; the company says it replenished 78% (7.7 billion gallons) through watershed-restoration projects, per its 2026 Environmental Report as covered by Axios.

• US data center electricity demand is projected to nearly double, from roughly 80 gigawatts to 150 gigawatts, between 2025 and 2028, per a Bloom Energy-based analysis reported by Utility Dive.

  Review

  “between 2025 and 2028” — the linked Utility Dive article’s own headline ties the electricity-doubling figure to 2030, not 2028 (2028 is the timeframe given for the separate water-use figure in the same headline); this looks like a mismatched timeframe, and the claim is single-sourced without an unverified flag unlike the other forward-looking projections in this brief — suggested action: fix the year or add [!unverified].

AI in the Wild

An AI-generated video of a deepfaked Michael Jordan reacting to LeBron James’s move to the Philadelphia 76ers spread widely this week, with fabricated “Jordan” dialogue calling the move shameless; the clip is openly synthetic, framed as commentary rather than disguised as real footage, per TotalProSports. Its spread shows how quickly synthetic reaction content piggybacks on real sports news, often outrunning any actual response from the person it depicts.

Takeaway