General AI News (Sep 22): Grok 4.7, BC Lawsuit & Alibaba V900
British Columbia sued OpenAI and Sam Altman, xAI shipped Grok 4.7 at an unchanged price, and Alibaba unveiled the Zhenwu V900 chip on September 22, 2026.
Two very different things landed on September 21 and 22. A frontier lab shipped a new flagship model at an unchanged price, and a provincial government took a frontier lab to court over a school shooting. In between, Alibaba put a number on China's data center ambitions and a university lab showed a 550 billion parameter model running on a laptop.
Here is what actually changed in the 24 hours to September 22, 2026, with sources for every claim.
Models & Products
Grok 4.7 arrives at the same price as Grok 4.6
xAI, which now brands itself SpaceXAI on its own release page, launched Grok 4.7 on September 21 as its strongest model for coding and knowledge work. The company describes a new and larger base model than Grok 4.6, trained with a longer reinforcement learning run weighted toward problems that take many hours to finish, plus better self-verification and native understanding of the Grok Bot harness. Source: xAI
The pricing is the story. Grok 4.7 is served at $2 per million input tokens and $6 per million output tokens, unchanged from Grok 4.6. On the company's own comparison table it scores 46.3% on CursorBench 4.0, against 40.4% for Grok 4.6 High, 41.7% for GPT-5.6 Sol Max and 51.8% for Fable 5.1 Max. Fable 5.1 charges $10 in and $50 out. Source: xAI
Read the chart the way the company frames it: Grok 4.7 is positioned at the frontier of price to performance, not at the top of raw capability. The Hacker News thread on the launch passed 600 points, which is where the community argument started over how much the CursorBench delta is worth. Source: Hacker News
Policy & Regulation
British Columbia sues OpenAI and Sam Altman
British Columbia filed suit against OpenAI in California on September 21, arguing that the February mass shooting at Tumbler Ridge Secondary School could have been prevented if the company had warned law enforcement that the shooter had used ChatGPT to plan the attack. The filing in San Francisco federal court names both OpenAI and chief executive Sam Altman as defendants. Source: Reuters
Provincial attorney general Niki Sharma announced the action in Victoria. The province says it is seeking answers and accountability for the community. The Guardian reports the claim goes further, alleging Altman promised reforms after the shooting and did not deliver them. Source: The Guardian Source: Government of British Columbia
This matters beyond one case because the plaintiff is a government, not a family. The province is asking a court to define what a lab owes the public when its platform is used to plan violence. The New York Times reported that the province's core argument is a duty to warn. Source: The New York Times
DeepSeek, OpenAI and Anthropic head to the UN Security Council
The three labs are set to brief the UN Security Council on AI safety and risk, at a moment when US and Chinese officials are fighting over chip exports and model provenance. That is an unusual venue for commercial labs, and it follows the UN panel brief we covered on September 21, which argued that agent safeguards are unravelling faster than oversight. Source: Tech Startups
If you missed that one, the September 21 edition has the panel detail: General AI News (Sep 21).
Chips, Data Centers & Energy
Alibaba's T-Head unveils the Zhenwu V900 and a 20GW cloud target
At the Apsara Conference in Hangzhou on September 22, Alibaba chief executive Eddie Wu said the company plans to train a successor to Qwen in the 5 to 10 trillion parameter range, up from roughly 2.4 trillion today, and to grow Alibaba Cloud's global data center capacity past 20 gigawatts by 2032. Source: The Next Web
Its chip unit T-Head, also called Pingtouge, showed the Zhenwu V900 accelerator for training and inference, claiming about triple the compute of its predecessor and cluster support scaling to 500,000 cards. Mass production of the full Panjiu supernode server is targeted for the first quarter of 2027, with a Zhenwu J900 generation previewed for March 2028. Alibaba shares rose roughly 3% in Hong Kong on the day, with some reports citing gains above 5%. Source: TechNode
Treat the chip claims as vendor numbers until independent benchmarks land. The roadmap dates are the verifiable part.
SoftBank's $11 billion bond sale shows where the money is coming from
SoftBank launched an $11 billion bond sale to fund its OpenAI investment, which is a reminder that a growing share of AI capacity is financed with debt rather than operating cash. Source: Tech Startups
The rate environment is not easy. The Federal Reserve raised its target range by 25 basis points on September 16 to 3.75% to 4.00%, so nine-figure private rounds are happening against a real cost of capital. On the same day as the Alibaba news, data center infrastructure provider Accelevation set US IPO terms at $20 to $24 per share, implying a market capitalization near $4.9 billion at the midpoint. Source: Tech Startups
Research & Open Source
dlab puts a 550 billion parameter model on a MacBook
Tim Dettmers and his dlab group at the University of Washington opened their open source week on September 21 with a claim that reframes local inference. Their framework runs DeepSeek V4.1, a 550 billion parameter model, on a MacBook with 128GB of memory, an AMD Strix machine or an NVIDIA DGX Spark. Qwen 3.8 Flash Next at 125 billion parameters runs on a single 24GB GPU. Source: Tim Dettmers
The efficiency numbers come from an agent they turned loose on their own inference kernels. They report quantized inference of Qwen 3.6 35B-A3B at 450 tokens per second at roughly 1.5 bits per weight, against 16 bits for half precision. They also describe a fully local deep research system that they say outperforms frontier lab research tools and comparable academic systems, with no internet access at all. Source: Tim Dettmers
Two caveats. These are the lab's own numbers on the week of its own launch, and no third party has replicated them. The Hacker News discussion, which ran to 179 points, focused on exactly that question. Source: Hacker News
Quick Comparison
| Entity | Key Metric | What Changed | Source |
|---|---|---|---|
| xAI Grok 4.7 | $2 in / $6 out per million tokens | New larger base model at unchanged price | xAI |
| British Columbia v OpenAI | Filed in San Francisco federal court | Government sues OpenAI and Sam Altman | Reuters |
| Alibaba Zhenwu V900 | 20GW cloud by 2032 | Triple predecessor compute, Q1 2027 production | TechNode |
| dlab | 550B model on a 128GB MacBook | Local frontier inference at about 1.5 bits | Tim Dettmers |
| SoftBank | $11B bond sale | Funds its OpenAI investment | Tech Startups |
Frequently Asked Questions
What is the top AI news today?
The top AI news for September 22, 2026 is British Columbia suing OpenAI and Sam Altman over the Tumbler Ridge school shooting, filed in San Francisco federal court on September 21. It is the first government action of its kind against a frontier lab over a specific harm, and it turns on an alleged duty to warn law enforcement.
Why did Grok 4.7 matter if the price did not change?
Grok 4.7 matters because capability moved while the price stayed flat. xAI kept the $2 input and $6 output pricing from Grok 4.6, so the model improves price to performance rather than raising the price of the frontier. That is the same dynamic OpenAI and Anthropic pushed on September 23, when both cut or repriced their tiers.
How big is Alibaba's data center plan?
Alibaba said its cloud capacity will pass 20 gigawatts by 2032, with Qwen successors in the 5 to 10 trillion parameter range. Its T-Head unit also showed the Zhenwu V900 chip for training and inference, with mass production of the Panjiu supernode server slated for the first quarter of 2027.
Bottom Line
The day's two loudest signals point in opposite directions. Alibaba, SoftBank and a stack of private rounds are betting that more compute, more debt and more gigawatts convert into capability. Meanwhile a province in Canada is testing whether the lab that builds the models owes anything to the people harmed downstream. Grok 4.7 at an unchanged price says the capability race is still on. The Tumbler Ridge filing says the accountability race just started. More AI news and the rest of this week's coverage are on the blog.
Related reading
Background worth having before you act on the above:
- Alibaba Raises $10B for AI, Nvidia Price Hikes 15%: General AI News for August 24, 2026
- General AI News (Sep 19): Gemini Hack, Newsom Order & Alibaba
- Anthropic Targets Biggest IPO Ever, Alibaba -75%: General AI News for August 21, 2026
- Grok Bot Shops, fal's Live AI TV & Tencent 770B Compresses: Viral AI News for August 30, 2026
Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →
