The Spec Sheet Gap: Humanoid Hype Meets a Memory Budget
A thin news Monday leaves the robotics story in focus, where Tesla's shifting Optimus chip memory and a skeptical MIT Technology Review take suggest the bottleneck is supply and capability, not announcements. Estimates of data center power use also turn out to disagree by about 8%.
Capability & Integration
- Update: the Gemini 4 timing flag from Friday remains open. LLM Gateway’s release timeline lists a limited-access “Gemini 4 Argon” on September 30, while Hans India says Gemini 4 is still in post-training with no announced date. These accounts conflict, and I found no Google announcement.
Unverified
Both accounts are secondary aggregators; neither links a Google source.
- LLM Gateway also lists Mistral Large 4 on October 6, described as a roughly 1-trillion-parameter model, API-only for now, with no published benchmarks and weights promised about three weeks later. Without benchmarks there is nothing to compare yet.
Unverified
Single aggregator source; I did not reach Mistral’s own announcement.
Robotics
- MIT Technology Review published a piece on October 8 arguing that robotics AI breakthroughs won’t change daily life soon. Search summaries of it say Tesla’s January claim was that only “some” Optimus robots were doing simple factory tasks, far short of earlier talk of thousands. I could not open the article itself, so treat these details as secondhand.
Review
Tesla’s January claim that only “some” Optimus robots do simple factory tasks, and the article’s thesis — drawn from search summaries of an article the author could not open; the description and takeaway lean on this “skeptical MIT Technology Review take” — add unverified flag or read the article before publishing.
- On deployment status, the same coverage describes the field as moving from demos to pilots. 1X’s NEO is still a preorder product that may rely on remote human operators for unfamiliar tasks, and Tesla has announced no Optimus price or sale date, per Botinfo. Treat all of this as demo or limited-pilot stage, not broad deployment.
Hardware & Supply Chain
- Tesla changed its planned AI5 chip memory twice in early October, per Tech-Insider: from 144GB of LPDDR5 to 72GB, then to 96GB, with the rationale tied to securing memory supply for high-volume Optimus production. The reversal suggests memory availability is constraining a robot that has not yet reached volume.
Unverified
Single source, and the reported Musk statements are not linked to a primary post.
- Nvidia’s 10-Q warns that new products may face unilateral export controls by the time they reach market, citing the H20 write-down. That is a standing risk disclosure, not a new event, and I found no new BIS action since the May guidance covered Friday.
Environmental & Cultural Impact
- Update: Friday’s 448 TWh figure for 2025 data center electricity use, from the June UN University report (Insurance Journal), is not the only baseline. An IEA-based tracker uses 485 TWh for 2025, per Gradually.ai, a gap of about 8%. Both put 2030 well above 900 TWh in most projections, so the estimate range is narrower than the growth range.
Review
“Both put 2030 well above 900 TWh in most projections” — no source is linked for the 2030 figure and neither cited page is shown to say it; the 448 vs 485 TWh gap rests on a tracker (Gradually.ai) rather than the IEA itself — cite the projection or cut it.
- I found no dated October layoff figures attributed to AI, so I am not reporting any.
AI in the Wild
A viral clip of a “Korean Fetty Wap” was exposed as AI-generated after a community note flagged it, per Complex. The coverage is about three weeks old, not from this week, and the creator is not identified in what I could read. That a crowd-sourced label rather than the platform or the poster settled its origin shows verification is still being done by volunteers.
Takeaway
Takeaway
The humanoid story is currently being decided in memory budgets and pilot contracts rather than demo reels, which is why a chip-spec change tells you more about Optimus than any video does.