By the Second: BMW Sets a Hard Production Bar for a Humanoid Robot
Figure's second-generation humanoid takes on real BMW logistics work with an 84-second cycle time and a zero-intervention target — unusually granular numbers for humanoid deployment claims — while a $6.3 billion compute lease and a pending New York moratorium show capital and backlash both accelerating around AI infrastructure.
Capability & Integration
- Anthropic set Claude Sonnet 5 as the default model for Free and Pro users starting July 1, replacing Sonnet 4.6. The company says agentic performance now approaches Opus 4.8, at introductory pricing of $2/$10 per million input/output tokens through August 31, rising to $3/$15 after.
Unverified
“Performance close to Opus 4.8” is Anthropic’s own characterization, relayed by press coverage; no independent benchmark reproduction is cited.
- Google shipped Gemini 3.1 Flash-Lite Image (“Nano Banana 2 Lite”) on June 30 — 4-second generation at $0.034 per 1,000 images, built for high-volume enterprise throughput rather than peak image quality.
Robotics
-
Update: Figure 03, successor to the Figure 02 unit that spent nearly a year in BMW’s Spartanburg body shop, has moved to a new task: sorting unsorted parts containers into assembly-ordered trolleys. BMW has attached explicit targets — 84-second cycle time, 37-second load time, >99% placement accuracy, zero human interventions, across 10-hour shifts five days a week. This confirms the deployment status flagged as open last time: it’s a bounded production task with hard numbers attached, not a broad rollout — but the specificity itself is new. Most humanoid claims to date have stayed vague enough to resist verification.
-
NVIDIA and Agility Robotics are using Halos for Robotics, a safety stack drawing on NVIDIA’s autonomous-vehicle work, to pursue IEC 61508 and ISO 13849 certification for Digit’s human-detection system. This is a certification-track announcement, not a new deployment: Digit already works at Amazon, GXO, Schaeffler, and Toyota sites — Halos formalizes safety claims those deployments already rest on.
Hardware & Supply Chain
-
SpaceX will lease Colossus 2 compute capacity to Reflection AI for up to $6.3 billion through 2029 ($150M/month, 90-day exit after the first quarter), giving the open-weight startup access to Nvidia GB300 chips. NVIDIA sits on both sides of the deal, having invested $800 million directly in Reflection AI.
-
TSMC’s CoWoS advanced-packaging supply-demand gap is projected to narrow from 20% to 10% by end of 2026 as monthly capacity scales from ~75,000–80,000 wafers toward a 120,000–130,000 target — the bottleneck behind AI accelerator shortages is easing, not resolved.
Environmental & Cultural Impact
-
New York’s Responsible Data Center Development Act — a one-year moratorium on new data centers with peak demand above 20MW, plus a ratepayer-protected electric and water rate class — passed the legislature June 4 and awaits Governor Hochul’s signature. It would be the first moratorium of its kind in the country if signed; as of this writing it is not yet law.
-
More than 300 data center bills have been filed across 30-plus states so far in 2026. At least 18 states now require large power users to guarantee payment for 85% of requested capacity over 10 years — shifting grid-expansion costs off residential ratepayers.
Unverified
The “300 bills / 30 states” and “18 states” figures are MultiState’s own aggregation; not independently cross-checked against state legislative databases.
Review
“At least 18 states now require…” — MultiState’s own reporting describes many of these as introduced bills rather than enacted law, and the 85%/10-year terms appear specific to Virginia’s rate class rather than uniform across all 18 states — verify enactment status; suggested action: reword to “introduced or enacted.”
AI in the Wild
Videos of young women in 2026 World Cup crowds — framed as candid stadium reaction shots — have drawn hundreds of millions of views on TikTok and Instagram this month. The clips are AI-generated, not filmed. What the volume reveals: a crowd-reaction shot is exactly the footage current AI video tools render most convincingly, and that viewers are least equipped to individually verify.
Unverified
Reporting on this trend traces to a single outlet (Brussels Times); view-count figures are platform-reported and not independently audited.
Takeaway
Takeaway
The most concrete claim in today’s brief is a robot’s cycle time in seconds. That’s a real signal — humanoid robotics is starting to produce falsifiable performance data that lets outsiders judge progress instead of taking a company’s word for it — even as the industry’s bigger moves this week (a $6.3B compute lease, a pending state moratorium, a new default model) still run almost entirely on trust in the announcing party.