OpenAI, Anthropic and Google DeepMind coordinate on AI safety; Micron shows a 512GB DDR5 module

OpenAI, Anthropic and Google DeepMind spend weeks coordinating on AI safety

OpenAI's global policy chief Chris Lehane said on Tuesday that the company has been working with Anthropic and Alphabet's Google DeepMind on AI safety for several weeks. Speaking at a briefing in Washington, Lehane said the three labs do not need an antitrust waiver to coordinate on safety and pointed to other industries that share safety work. "It's better to try to work together to prioritize safety," he said. Lehane was in Washington to meet lawmakers and said OpenAI would back bipartisan legislation aimed at catastrophic AI risks, including a provision in the FRONTIER Act that would require leading labs to let independent organizations verify model safety (Reuters, Sep 15).

OpenAI weighs another raise at a reported $1.2 trillion valuation

OpenAI has held early talks with large investors about a new private funding round that would value the company at about $1.2 trillion, according to reports on Tuesday, up from $852 billion in its March round. The contacts were started by investors, and whether the round proceeds depends on the company's listing plans. OpenAI's annualized revenue passed $40 billion last month, up 20% from the prior month. The company filed confidentially for an IPO in June, though CEO Sam Altman said last week that now is not the right time to list in the US (National Business Daily, Sep 16).

Musk floats cross-testing between AI labs; Huang rejects new rules

At the All-In Summit in Los Angeles on Monday, SpaceX CEO Elon Musk said leading AI companies could open their APIs to each other before a model ships, so rivals can run their own safety tests and look for risks such as bioweapons, nuclear weapons and deception. He said the arrangement would not need new legislation, and that competitors could publicly question a release if problems went unresolved. Musk also said China should take part for the mechanism to work. On the same stage, Nvidia CEO Jensen Huang said the AI industry does not need new safety regulation, arguing that market pressure will push companies to ship safe products and that safety and speed can both be met at once (National Business Daily, Sep 16; ai0.news, Sep 15).

Microsoft ships an emergency fix after its largest patch ever

Microsoft has started rolling out an out-of-band Windows 11 update to repair problems introduced by its September Patch Tuesday release, which fixed close to 1,000 flaws in the largest monthly patch in the company's history. The defects broke folder shares on Hyper-V Linux virtual machines and Remote Desktop Services sessions, and affected some USB audio devices. The emergency update, released a day earlier, covers Windows 11 versions 26H1, 25H2 and 24H2. Microsoft also shipped fixes for Windows Server 2025 and 2022, Windows 10 LTSC builds, and Windows Server 2012 and 2012 R2. Out-of-band releases used to be rare for Windows but have become routine in 2026: four in January, two in March and one in July (The Verge, Sep 15).

Micron shows the first 512GB DDR5 server module

Micron said on Tuesday it had demonstrated the first 512GB DDR5 RDIMM across several server platforms. The 3DS RDIMM stacks DRAM dies with through-silicon vias to fit more capacity into a single package. It runs at up to 9,200 MT/s, and in a dual-socket server with 24 memory slots it allows up to 12TB of DDR5. Micron says the module uses more than 60% less operating power than four 128GB RDIMMs. AMD and Intel are validating it on next-generation server platforms, and volume production is expected in the second half of 2027 (Micron, Sep 15).

Apple accepts Samsung memory prices for the first quarter of 2027

Apple has accepted Samsung Electronics' memory prices for the first quarter of 2027, according to supply chain sources cited by Jiemian News on Tuesday. DRAM is priced near $2.0 per Gb and NAND near $0.33 per Gb, 30% to 40% above quotes for the third quarter of 2026. Samsung's third-quarter quotes were about $1.4 per Gb for LPDDR4X, $1.5 for LPDDR5X and $0.26 for NAND. TrendForce says Samsung, SK Hynix and Micron have largely allocated their 2027 DRAM and HBM capacity, with customers receiving about 60% to 70% of what they asked for. Apple did not respond to a request for comment (Jiemian News, Sep 15).

Huawei opens a "3D data center" in Wuhu

Huawei used an industry conference in Wuhu, Anhui province, on Tuesday to show what it calls the world's first 3D data center and to release a set of outline design drawings for the architecture. The layout stacks the cooling, IT equipment, power supply and battery layers vertically rather than spreading them out, which the company says allows zoned cooling and 168MW of capacity in a single building, aimed at clusters of more than 100,000 Ascend super-nodes. Deputy chairman and rotating chairman Wang Tao said daily token consumption in China rose from about 100 billion at the start of 2024 to roughly 500 trillion in 2026. Huawei Cloud CEO Zhou Yuefeng said the immediate priorities are cutting end-to-end cost and getting chips into service faster (The Paper, Sep 15).

Salesforce and Nvidia launch Koa, a CRM reasoning model

At Dreamforce in San Francisco on Tuesday, Salesforce and Nvidia introduced Koa, Salesforce's first reasoning model for CRM work. Koa is a post-trained version of Nvidia's Nemotron 3 Super, built on a synthetic dataset modeled on nearly 27 years of CRM deployments and covering more than 14 industries, including manufacturing, financial services, healthcare and travel. Salesforce says no customer data was used. On the company's own CRM benchmark, which includes tasks such as updating an opportunity and routing a case, Salesforce says Koa matches or beats leading models with about three times fewer errors. Salesforce controls the weights, and training and inference run inside its own infrastructure. Koa is already used internally and in a Slack employee agent, with customer pilots starting in October and general availability expected in US regions in winter 2026 (Salesforce, Sep 15; Nvidia, Sep 15).