Two Doors: Qwen3.8-Max Opens Wide as Washington Shuts Chinese Robots Out

Capability & Integration

• Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter mixture-of-experts model with only 95 billion parameters active per query; Alibaba’s own benchmarks claim it edges out GPT-5.6 Sol and Claude Fable 5 on several coding and research-reproduction evals, per Bloomberg — an internal comparison, not yet reproduced by a third party. Weights are due on Hugging Face next week.
• Update: Friday’s brief flagged OpenAI’s “$2,000 in compute” claim for Astra’s math proofs as single-sourced. That figure is still unverified, but Thomas Bloom, who curates the Erdős problem catalogue at the University of Manchester, gave the results themselves independent mathematical weight, calling three new Erdős-problem solutions “big news” — “in terms of constructions, this is big” — while rejecting the idea that Astra replaces mathematicians rather than extending work built on their proofs, per Quanta Magazine.

• Microsoft’s Q4 FY26 earnings booked a $3.2 billion mark-to-market gain on its Anthropic stake — nearly matching what it made on OpenAI for the entire year — against a $600 million markdown on that OpenAI position, per TechCrunch.

Robotics

• Update: The BYD humanoid teased in Friday’s brief has a name and specs: “Xiao Di,” 1.61 meters, 58.5 kg, 31 degrees of freedom, real-time translation across six Chinese dialects and six foreign languages, unveiled at BYD’s Di Space center in Zhengzhou, per Notebookcheck. It’s a working prototype greeting showroom visitors — BYD’s stated goal is two or three per dealership — not a manufacturing or logistics deployment.
• The FCC updated its Covered List on July 28 to bar new imports of Chinese-made humanoid and quadruped robots on national-security grounds; previously purchased units and government use are exempt. China, which supplies roughly 85% of the global humanoid market, called the move protectionism dressed up as security policy, per the Washington Post.

Hardware & Supply Chain

• SK hynix and SanDisk published the first Open Compute Project standard for High Bandwidth Flash at Flash Memory Summit 2026 — a new memory tier between HBM and SSDs, up to 512GB per device across three bandwidth grades topping 3TB/s, using the UCIe interconnect, per Tom’s Hardware. It’s an open spec, not a single vendor’s proprietary lock-in.
• SK hynix, Samsung, and Micron are all now qualified to supply HBM4 for Nvidia’s Vera Rubin platform shipping in Q3 2026; analyst estimates from UBS, relayed by TrendForce, put SK hynix at roughly 60–70% of Rubin’s HBM4 allocation — an estimate, not a disclosed contract figure.

Environmental & Cultural Impact

• Global data centers consumed an estimated 448 trillion watt-hours of electricity and 1.2 trillion gallons of water over the past year, producing roughly 208 million tons of CO2 — comparable to Argentina’s annual emissions — per UN University, which projects these figures could double within four years.

• More than 300 data-center-related bills have been introduced across 30-plus states in 2026, shifting from incentive competition toward moratoriums and grid-cost-allocation rules; 16 states have already enacted at least one such law, per MultiState.

AI in the Wild

A trend called “Costanza AI Edits” has taken over X since early July: user @e_cdalton started it by AI-editing Seinfeld’s George Costanza into a bald-and-bearded makeover, and it’s since spawned into a full genre of AI face-swaps and reimaginings, catalogued in Know Your Meme’s weekly roundup. Nobody involved is claiming it’s real — it’s the opposite of Friday’s Taiwan deepfake story, and a reminder that most of AI’s cultural footprint right now is closer to a Photoshop gag than a fraud.

Takeaway