Shipped, Not Working Yet: Atlas Goes to Hyundai While the Line Waits Until 2028

Capability & Integration

• Update: GPT-6 Astra’s headline 99.9% ARC-AGI-3 score used OpenAI’s own “Provider Adapter” harness; ARC Prize’s independent standard-harness test put the same model at 62.7%, per Winbuzzer — a concrete number behind the harness-dependency critique flagged in the previous brief. Its context window now runs to 1.05 million tokens, per The New Stack.
• Anthropic shipped Fable 5.1 and Mythos 5.1 on September 1, cutting cache-read prices 75% and adding EU AI Act-mandated watermarking, per MacRumors. Anthropic reports 52.6% on a science-reasoning benchmark versus 24.7% for the prior model — a company claim, not an independent one, per VentureBeat.
• McKinsey’s State of AI survey (1,719 respondents, May–June) found 32% of organizations skipped buying a software product because agentic coding tools let them build it in-house — 41% in tech, per Yahoo Finance. Forrester research cited alongside it found most adopters still haven’t reached production deployment.

Robotics

• Figure signed a deal with Nscale on September 3 for up to 100,000 Nvidia Vera Rubin GPUs, an initial $3.5 billion compute commitment scaling toward $6 billion, per Nscale’s press release. It’s a training-compute bet, not a robot deployment — capacity targets the back half of 2027.

• Boston Dynamics began shipping production Atlas units to Hyundai (its majority owner) and Google DeepMind this year, per The Robot Report. Hyundai’s own timeline places factory-floor sequencing in 2028 and complex assembly in 2030 — shipped and working the line are, here, years apart.

  Review

  “Hyundai (its majority owner)” — Hyundai Motor Group completed a full buyout of Boston Dynamics around June 22, 2026 (the remaining ~9.65–9.9% stake bought from SoftBank), per Bloomberg, Hyundai’s own newsroom, and Boston Dynamics’ own news page. By this brief’s date, Hyundai is the sole owner, not the majority owner — suggested action: fix.

  The 2028/2030 timeline traces to secondary robotics trade sites rather than a Hyundai or Boston Dynamics primary statement.

Hardware & Supply Chain

• DRAM and HBM remain in a multi-quarter shortage as manufacturers reallocate wafer capacity from commodity DRAM toward AI-accelerator HBM, which consumes roughly three times the wafer capacity per gigabyte, per Network World.

  Reported price-increase percentages diverge sharply across analysts; the squeeze is well corroborated, its exact magnitude is not.

• Huawei’s Ascend line has gone from a negligible share of China’s AI compute market in 2023 to an estimated 50–60% in 2026, worth roughly $12.1 billion annually, after Beijing’s September 2025 directive closed off buying Nvidia through third parties, per Value Add VC. Export controls appear to be accelerating domestic substitution, not halting China’s buildout.

  Review

  50–60% share / $12.1B figure is single-sourced to one blog. Independent estimates diverge (50% Huawei Central, 55% other trackers, 60% WCCFTech/Capacity, and separate $12B vs $12.1B revenue figures) — comparable in spread to the memory-price claim already flagged above. Suggested action: add unverified flag.

Environmental & Cultural Impact

• Global data centers consumed 222 billion liters (59 billion gallons) of water for cooling in 2025 — Rystad Energy projects that could nearly triple to 644 billion liters by 2030 without further efficiency measures, or stay under double with them, per Bangkok Post. Microsoft and AWS report self-measured water-efficiency gains of 25% and 37% since 2022, even as both expanded their footprints.
• Those figures cover only on-site water. Water used to generate a data center’s electricity and manufacture its hardware can run roughly double the on-site cooling total, per the same reporting — the efficiency gains reported don’t touch that larger, upstream number.

AI in the Wild

ChatTJB looks like a chatbot interface; it’s actually artist and former Google PM Tucker Bryant typing every reply himself, with a billboard revealing “AI” stands for “average individual.” Launched in April, it drew fresh viral attention in August after more than 10,000 people volunteered to help him answer, per Fortune and PC Gamer. Its staying power says something plain: an interface that merely resembles intelligence is often enough.

Takeaway