Dark still-life of a black leather briefing binder with a blank metal nameplate beside a blank tent card on a slate desk. No people, no logos.

The Tuesday AI brief: Newsom bills the halls, DeepSeek joins Altman at the UN, and Alibaba sketches a bigger Qwen

September 22, 2026 Aaron Harme

In this Brief

California signed seven bills so AI data centers pay for the grid and water they strain
OpenAI asked the U.S. to lead global AI incident standards ahead of Altman at the UN
Reuters sources say DeepSeek was invited to brief the Security Council the same day
Nvidia bought another $1.5 billion of SoftBank’s SB Energy at a 10% IPO discount
Alibaba put a 5-10 trillion-parameter Qwen plan and a Zhenwu chip on the Hangzhou stage

California makes AI data centers pay for the grid and water they strain

Setup: California Gov. Gavin Newsom signed seven bills Monday designed to stop AI data centers from shifting utility costs onto residents, The Verge reports, citing earlier Los Angeles Times reporting. The package requires the California Public Utilities Commission to create a new rate classification for data centers, a separate utility billing category, and forces them to pay for upgrades to local power grids and water systems.

Proposed centers must disclose estimated water use to local governments, plus energy-efficiency and drought-planning information. Meeting energy, water, and fuel gates is also required before a streamlined approval path opens.

The rate class is the quiet lever. Once the halls sit in their own utility bucket, neighbors stop eating the upgrade cost for new lines and substations. Water disclosure and drought planning turn consumption into a permit fact before concrete. Streamlined approval only opens after those gates clear, so speed becomes a reward for paying for the strain. California is writing the cost shift into statute while other states are still arguing over whether to host the boom at all.

Longer cut: Newsom forces California AI data centers to pay for grid and water upgrades.

OpenAI wants Washington to write the AI incident rulebook before Altman speaks at the UN

Setup: OpenAI released international AI safety standards proposals on Monday as the U.S. and China weigh risk mitigation, Axios reports. Altman is set to cover the proposals in his address at the United Nations Security Council in New York on Wednesday. The White House hosts Xi on Thursday.

OpenAI recommends routing work through AI safety institutes worldwide via the U.S. Center for AI Security and Innovation at Commerce, plus shared global measurements for classifying incidents and tracking alignment issues, when a model’s behavior drifts from what the lab intended. The Next Web, citing Bloomberg, adds that the push covers which incidents must be reported, how countries cooperate, and access to compute. OpenAI’s own blog says fully autonomous recursive self-improvement, systems that increasingly improve the next generation of AI with less human oversight, should not be pursued unless and until it can be done safely.

Incident reporting is the paperwork fight dressed as safety leadership. Shared measurements and national institutes sound cooperative until someone decides what counts as serious enough to phone home. The self-improvement pause line sits in the same memo OpenAI wrote for itself. Altman carries that package into the Security Council on Wednesday, with Xi due at the White House a day later. Washington gets asked to lead the forms for a frontier fight it has not yet agreed how to police. The ask is timed for the week both capitals share a New York frame.

DeepSeek is walking into the same Security Council room as Sam Altman

Setup: Chinese AI startup DeepSeek will be among companies briefing the United Nations Security Council this week on AI risks, two sources familiar with the matter told Yahoo Finance, carrying Reuters. The 15-member Council meets Wednesday on AI and international security. Diplomats also expect senior Anthropic representatives. DeepSeek and other Chinese firms including Moonshot have been invited to make statements, one source said. Founder Liang Wenfeng does not plan to attend, and plans remain fluid.

An invitation buys a seat under the dome. Putting DeepSeek and Moonshot on the same Wednesday roster as Altman and Anthropic turns the Security Council into a staged mirror of the lab race. Liang staying home keeps the founder off the mic while the company still shows up in the room. Plans remain fluid, so the seating chart can still move. The body that usually manages wars now hosts the firms shipping the models. That is the governance fight in one chamber, on one day.

Nvidia is buying another $1.5 billion of SoftBank’s data-center landlord at a discount

Setup: Nvidia is buying a further $1.5 billion of shares in SoftBank-owned SB Energy at 90% of the price of the developer’s coming U.S. listing, The Next Web reports, citing a Bloomberg regulatory filing. That takes Nvidia’s total backing to $3 billion in new N-class nonvoting shares, ownership economics without board votes.

SB Energy holds 8.8 gigawatts of data-centre capacity in contract or under construction, including Texas and Ohio. OpenAI is also an investor. SB Energy is targeting $5 billion to $7 billion of proceeds. Techmeme’s evening cluster said sources report the IPO, originally planned this month, has been delayed as investors question a sought $50 billion-plus valuation.

Ninety cents on the IPO dollar buys landlord economics before the public book opens. Nvidia already creates the power demand that fills those halls. SoftBank gets capital for the U.S. vehicle. Nvidia gets more exposure to 8.8 gigawatts of contracted or under-construction capacity without taking board control. Delay talk around a $50 billion-plus ask only sharpens why a discounted private placement is useful now.

Alibaba stacked a bigger Qwen, a homemade chip, and a 20-gigawatt cloud path in Hangzhou

Setup: Alibaba rolled out Zhenwu V900, its most powerful AI chip yet, at Alibaba Cloud’s annual conference in Hangzhou, Morningstar / Dow Jones Newswires reports. CEO Eddie Wu said performance is three times better than the chip Alibaba released four months earlier, with larger GPU memory and bandwidth. A single V900 cluster can support up to 500,000 cards. Mass production and commercial release are scheduled for Q1 2027.

Alibaba also announced plans to train a new AI model with 5 trillion to 10 trillion parameters, the adjustable knobs a model learns from data. Flagship Qwen 3.8 Max, from July, has 2.4 trillion. Finimize pairs the chip with a plan to push Alibaba Cloud data-center capacity past 20 gigawatts by 2032, and notes Wu acknowledged supply-chain shortages mean Alibaba cannot add servers as fast as AI demand.

Jumping from 2.4 trillion toward 5 to 10 trillion is a size bet on the whiteboard, still short of a shipped product. Zhenwu V900 is the homemade accelerator meant to train and run that stack in half-million-card clusters, with commercial release aimed at early 2027. Twenty gigawatts by 2032 is the power yardstick for how many halls Alibaba Cloud wants to light. Wu’s own supply-chain admission is the check on the keynote: the ambition is written, the silicon schedule is the bottleneck.

## Also noted

– The FAA rolled SMART, its $875 million AI routing helper, live at Reagan National, Dulles, and BWI. Controllers still get the final say on suggested routes. Rep. Don Beyer (D-VA), whose district includes Reagan National, called for immediate suspension, saying controllers were not consulted or trained before adoption.
– Xiaomi open-sourced MiMo-V2.6 Pro and Flash, natively omnimodal models that handle text, images, and related agent tasks together. TestingCatalog says Pro scored 46.32 on Artificial Analysis Intelligence Index v4.3, ahead of Kimi K3 and Qwen3.8 Max among open-source models in that comparison, and ties Grok 4.7 (xHigh) on the Techmeme/VentureBeat paraphrase.

## Sources

– The Verge: California AI data center bills
– Axios: OpenAI AI safety standards / U.S.-China
– The Next Web: OpenAI global AI standards push
– OpenAI: Building standards for the next phase of AI
– Yahoo Finance / Reuters: DeepSeek to brief UN Security Council
– The Next Web: Nvidia SB Energy pre-IPO stake
– Morningstar / Dow Jones: Alibaba Zhenwu chip and larger model plan
– Finimize: Alibaba 5-10 trillion-parameter AI model
– Rep. Don Beyer: SMART suspension call
– Morning Brew: FAA launches AI tool to reduce flight delays
– Futurism: FAA AI air traffic control / SMART contract
– TestingCatalog: Xiaomi MiMo-V2.6 Pro and Flash
– Live BF: Newsom forces California AI data centers to pay for grid and water upgrades

Also noted

The FAA rolled SMART, its $875 million AI routing helper, live at Reagan National, Dulles, and BWI. Controllers still get the final say on suggested routes. Rep. Don Beyer (D-VA), whose district includes Reagan National, called for immediate suspension, saying controllers were not consulted or trained before adoption.

Xiaomi open-sourced MiMo-V2.6 Pro and Flash, natively omnimodal models that handle text, images, and related agent tasks together. TestingCatalog says Pro scored 46.32 on Artificial Analysis Intelligence Index v4.3, ahead of Kimi K3 and Qwen3.8 Max among open-source models in that comparison, and ties Grok 4.7 (xHigh) on the Techmeme/VentureBeat paraphrase.