- Generac disclosed in an SEC filing a long-term agreement to supply industrial backup generators for Amazon data centers, with initial deliveries of roughly $2.4 billion across 2027–2028 and potential aggregate payments up to $8 billion.
- Amazon received a warrant for up to 1,693,745 Generac shares at $200.93 — about 2.9% of the company fully diluted — with 307,954 shares vesting immediately and the rest tied to purchase milestones through September 2033.
Snapshot — September 17, 2026
36 stories
- Anthropic signed its first Australian data-center agreement, taking space in the proposed A$32 billion (US$21B+) Western Downs Digital Park near Dalby, Queensland, developed by Singapore-based Zerra DC across roughly 725 hectares.
- The campus is planned to reach 2.16 GW of capacity — a draw comparable to 1.5 million Australian households — and Anthropic will use it to run Claude for user queries rather than to train models.
- Anthropic disclosed that Claude is now contributing to development of its own successor, taking an "extraordinarily active" role while still working under human supervision — consistent with its same-week R&D Automation Index post finding Claude "leads" 26% of AI R&D work (up from under 1% in February) but is not operating fully autonomously in any measured subset.
- Anthropic published the first results of a prototype R&D Automation Index, reporting that as of August 2026 Claude "leads" 26% of the company's AI research and development work — completing most of a task end-to-end from a high-level prompt under human supervision — up from under 1% in February 2026, scored on Epoch AI's AL0–AL5 automation scale.
- Anthropic rebuilt Projects in Claude Code: instead of a folder plus one chat, a project is now a single ongoing conversation in which Claude acts as coordinator and decides which work spins out into a separate thread.
- Each thread is a full cloud session that keeps running after the laptop closes.
- Available in beta on Pro and Max tiers, this extends the same consolidation direction as the September 16 merge of Claude chat and Cowork.
- Anthropic published three metrics — AI-led R&D, oversight of AI agents, and internal compute allocation — along with methodologies, days after CEO Dario Amodei's essay calling for a coordinated industry slowdown.
- It reported approximately 30,000 concurrent agents doing research and engineering work across its most-used internal platform, and that in a July 13–20 snapshot roughly 6% of compute going to AI R&D was allocated to safety.
- Apple is developing dedicated AI servers built around future M8 Ultra processors in two- and four-chip configurations aimed at developers, enterprises, and governments, with no arrival before 2029.
- It would be Apple's first purpose-built server line since Xserve was discontinued in 2011, backed by CEO John Ternus.
- Cornell launched AI Integration Pilots and expanded its AI Undergraduate Fellows program, funded by a $2 million gift from the Dake family.
- Undergraduates serve as embedded classroom researchers partnering with faculty to evaluate AI's effect on learning.
- Vice provost Steven Jackson emphasized a deliberately discipline-specific design, noting AI may make academic areas more distinct rather than more uniform.
- Data center developer Crusoe closed a $3.9B Series F at a $30.9B post-money valuation, co-led by Atreides Management, Mubadala Capital and Valor Equity Partners, with Founders Fund, GIC, Nvidia, the Qatar Investment Authority, Radical Ventures and TPG participating.
- The valuation is roughly triple the $10B mark set ten months ago.
- Crux AI, a cloud venture tied to Blackstone and Alphabet, lined up roughly $22 billion in financing from a syndicate of ten banks to fund purchases of Google's custom TPU accelerators.
- It is among the largest chip-collateralized debt packages assembled to date and signals bank appetite for lending against AI silicon.
- The European Commission adopted its proposed EU KIDS Act, covering social media, video-sharing platforms, online games, app stores and — for the first time in EU platform regulation — AI companions and chatbots.
- AI companions must be off by default for minors, may not be designed to create emotional dependency, and by default may not carry earlier conversations into later sessions; under-13s reach them only through parental controls.
- Google DeepMind launched the DeepMind Institute with Shane Legg, James Manyika, and Demis Hassabis as directors.
- The inaugural four essays cover AGI economic disruption, preserving human-readable model reasoning, human flourishing, and frontier-model evaluation.
- Hassabis proposes a US-led frontier AI standards body where developers would initially submit models voluntarily for review up to 30 days before release, eventually using independent held-out tests.
- Google, Nvidia, and Anthropic co-founded the AI Energy Management Alliance (AEMA) alongside grid-software unicorn Emerald AI plus utilities AES, Constellation, National Grid, and NRG Energy.
- The Alliance aims to make demand-response — pausing noncritical AI workloads or shifting compute — a first-class part of data-center siting, potentially unlocking as much as 100 GW of additional grid capacity.
Google Research published a research experiment that lets educators generate custom, interactive, guided educational simulations matched to their stated objectives and curriculum rather than selecting from pre-authored content. It is the follow-on application of Google's earlier generative-UI line of work — the only first-party frontier-lab research blog post confirmed inside the 24-hour window. - https://research.google/blog/the-future-of-practice-enabling-teachers-to-create-learning-interactives-with-generative-ui/
- The US House voted 417–3 to require large AI data centers to pay their own grid-upgrade costs rather than shifting them onto household ratepayers, working through PURPA — which only requires states to *consider* the federal standard.
- The next day, Sen.
- Martin Heinrich (D-N.M.) blocked Sen.
- Jon Husted's (R-Ohio) attempt to pass the Ratepayer Protection Act by unanimous consent, calling the measure toothless;
- At Huawei Connect in Shanghai, rotating chairman David Wang moved the Ascend 960DT launch from Q3 2027 to Q1 2027 — three quarters ahead of schedule — with the 960PR in Q3 2027 and Ascend 970 and 980 slated for 2028 and 2029.
- Huawei also unveiled the Atlas 960 SuperPoD (4,096 NPUs in the 960E configuration), detailed a full Ascend NPU roadmap, and pitched near-packaged optics as co-packaged costs bite.
- Huawei's "Intelligent World 2035" report, released ahead of Connect, projects global annual AI token consumption growing 100,000-fold by 2035, with autonomous agents generating more than 90% of that traffic and as many as 900 billion active agents worldwide.
- The report reframes the compute burden as an inference-and-agent problem rather than a training one, and names agent security and privacy among ten technology directions it says must be built.
- Huawei rotating chairman Eric Xu told reporters at Connect in Shanghai that Chinese AI developers may not yet be advanced enough to encounter the safety risks US labs are reporting, and "may need to speed up their pace to the level that they could also feel the risks from AI development," while adding that development and risk management must be balanced.
- Insilico Medicine announced a Cell publication introducing an openly released AI toolkit for aging biology, centered on the LongevityBench benchmark and a family of five compact models.
- LongevityBench is presented as the first open benchmark for aging biology, assembled from DNA-methylation, transcriptomic, proteomic, and clinical data into retrieval-resistant tasks with observable ground truth.
Speaking to reporters ahead of an AI summit convened by King Charles III in Scotland, Huang said he expects Nvidia to sell twice as many chips next year as this year, citing AI investment “in almost every single country that we're in.” The statement sits above Nvidia's formal guidance of roughly 70% revenue growth (about $673B) for the fiscal year ending January 2028 — a figure the company has described as supply-constrained. Huang, who appeared alongside representatives of Google DeepMind, OpenAI and Anthropic, also said companies should hold products back when they are not safe.
- Meta's Oversight Board issued a sharply critical assessment of the company's policies on AI-generated deepfakes, faulting both the substance of Meta's manipulated-media rules and their enforcement across platforms.
- The finding arrives as generative video tools make convincing synthetic media cheap to produce at scale, ahead of election cycles.
Microsoft's AKS engineering team open-sourced TauGrid under an MIT license, bundling the tau CLI, Kueue queueing, KubeRay orchestration, GPU node health monitoring, and observability into a single Helm install for any Kubernetes 1.30+ cluster with GPU nodes. It targets the operational layer between raw accelerator fleets and training or inference jobs — an area cloud providers have largely kept proprietary. - https://www.marktechpost.com/2026/09/17/microsoft-open-sources-taugrid-a-kubernetes-native-stack-for-gpu-ai-workloads/
- Microsoft released a playbook based on its internal AI transformation, arguing that proprietary process and workflow data — not models — is the durable enterprise advantage.
- It reports 20% higher close rates in one sales group, up to 75% cycle-time reduction in selected supply-chain workflows, and a nine-person engineering team shipping an initial release in 35 days.
- MIT's Ad Hoc Committee on AI Use in Teaching, Learning, and Research Training, co-chaired by Eric Klopfer and Sam Madden, delivered its final report.
- Community surveys found nearly 80% believe AI skills will be necessary for their careers, while 48% worried about the cognitive consequences of relying on AI and only 25% of undergraduates felt MIT was preparing them to use AI professionally.
- OpenAI introduced Astra for Law, pairing GPT‑6 Astra with a proprietary legal search index spanning more than 230 million URLs of U.S. case law, statutes, regulations, court rules and administrative decisions, sourced in part from the Free Law Project's CourtListener collection.
- On 200 questions from the private validation set of Vals AI's Legal Research Bench, OpenAI reports a 54.0% correctness pass rate versus 38.7% for GPT‑6 Astra with web search alone — a 40% relative improvement, but still a failure on nearly half the questions, on OpenAI's own testing.
- OpenAI began testing Sponsored Agents, a format in which clicking an ad in ChatGPT opens a clearly labeled conversation with an advertiser's own agent rather than sending the user to an external site.
- The launch bundles natural-language campaign management through an Ads Manager plugin in ChatGPT Work, AI-generated ad copy and imagery, and first-party integrations with HubSpot (CRM) and Shopify (commerce).
- OpenAI introduced a standing process for tracking, investigating, and publicly disclosing model behavior that falls outside expected limits, with public-disclosure clocks that apply even when a behavior has not been fully explained or mitigated.
- Flagged cases are routed to one of three tracks — ready for disclosure, minor investigation, or slower third-party case — on six- or twelve-business-day clocks.
- Pew's 37-country study, based on 42,151 interviews conducted 8 February to 13 May 2026, finds a 37-country median of 46% expecting AI to produce *fewer* jobs over the next 20 years against 9% expecting more.
- The gradient tracks income: a median of 55% in high-income countries versus 36% in middle-income ones, where uncertainty is far higher (34% unsure versus 22%).
- PrismML released Bonsai 2 27B, a compressed version of Alibaba's Qwen3.8 27B that TechCrunch says reduces memory footprint by roughly 9× to 10× while matching 98% of the original model's aggregate benchmark scores — using ternary weights to shrink storage enough for PCs and potentially high-end smartphones.
- Hacktron AI chained a libheif image-decoder vulnerability in OpenAI's Discourse-based community forum to a separate flaw in OpenAI's single sign-on, took over an employee's ChatGPT and Codex accounts, and opened a benign pull request in the private openai/openai monorepo.
- Claude Opus 4.8 failed to produce a reliable exploit across several sessions;
- Apple shipped a Safari MCP server in Safari 27.0, allowing Claude Code, Codex or any MCP-compatible agent to control a Safari window: reading the DOM, inspecting network requests, capturing screenshots, pulling console output and running JavaScript in the active tab.
- Developers enable it via “Allow remote automation and external agents” in Safari's Developer settings and connect through the existing safaridriver binary with a new --mcp flag.
- Sen.
- Martin Heinrich (D-N.M.) blocked Sen.
- Jon Husted's (R-Ohio) attempt to pass the Ratepayer Protection Act by unanimous consent, calling the measure toothless because it only requires state utility commissions to *consider* standards making 100 MW-plus data centers pay the full incremental cost of grid upgrades.
- Stanford researchers released Paper2Agent, which converts a published paper into a validated Model Context Protocol server so an agent can re-run the paper's methods on new data.
- Demonstrations with AlphaGenome, Scanpy, and TISSUE were built in ~45 minutes each at ~$13–14 apiece; multi-agent collaboration among them independently identified GPR137 as a causal gene for psoriasis.
- The Information reported that the same vulnerability class was found across several AI coding-agent tools, including Claude Code, Codex, Gemini CLI, and GitHub Copilot.
- Public details were limited, but the headline is strategically important because it points to shared failure modes in agentic developer tooling rather than one-off vendor defects.
- A newly unsealed summary-judgment brief in the publishers' copyright suit quotes Microsoft's Director of Applied Science describing mass scraping as potentially "the largest theft of labor in human history," and an internal Microsoft document warning of a "doom loop" that would hurt both model performance and the wider web.
- Z.ai published a technical account of building a production-grade inference service for GLM-5.3-Flash from scratch on a cluster of more than 100,000 Chinese-made AI accelerators, stating that all production inference for the model now runs on that system.
- The company says much of the work was executed by a GLM-5.3-powered “Infra Agent” rather than by infrastructure engineers alone, using a “dense feedback” methodology that folds correctness tests, traces and microbenchmarks into locally verifiable loops.