AI News for September 18, 2026

Widespread shortages of advanced memory and grid power, plus chip manufacturing concentration in Taiwan, are delaying AI data centers and even pushing next-generation Nvidia and AMD GPUs to 2028, with Intel's CEO warning 2027 will be worse.

Industry leaders detail the data, legacy infrastructure, governance, and explainability barriers holding back AI adoption in financial services despite heavy investment.

A Teradata report finds that healthcare, manufacturing, retail, and financial services are all moving toward agentic AI at different speeds but share the same obstacle of fragmented, poorly contextualized enterprise data.

A vendor guide outlines seven 2026 enterprise AI trends, arguing that agentic AI, multi-agent orchestration via MCP, and AI governance are moving from experimentation into production.

A forward-looking roundup of manufacturing logistics trends through 2027, covering AI-driven planning, smart warehouses, robotics, digital twins, multimodal logistics, and sustainable supply chains.

The Weave paper presents a fine-grained dynamic SM scheduling system within a persistent MoE megakernel that adapts SM partitioning per layer and per GPU to routing results, achieving 2.89x MoE-layer and 1.33x end-to-end speedups over five baselines on 4x H100 GPUs.

Google Cloud's weekly roundup covers new GA features like Storage Intelligence Advisor and Bigtable subscriptions, plus guides for building private WebSocket and agent architectures with Apigee, Cloud Run, and Gemini Enterprise.

An argument that AI-driven economic growth may not translate into military power because what matters is the convertibility of an economy's industrial base, not aggregate GDP.

A Motley Fool investment analysis argues Nvidia will benefit most from PwC's forecast of $31.6 trillion in cumulative global AI infrastructure spending by 2050, driven by chip upgrade cycles.

HPCwire argues that AI's compute scarcity won't be solved by building more data centers alone but by architecting for efficient token usage, with knowledge graphs and precise prompts reducing inference energy by 32-48%.

As AI shopping agents increasingly handle consumer purchases, brands must make influencer and creator content machine-legible through transcripts, schema markup, and structured product data or risk invisibility in an emerging purchase channel.

Google announced that Workspace Studio automation flows now support third-party integrations like Slack and Salesforce plus webhook transmission, with a phased rollout starting September 17, 2026, for Business, Enterprise, and Education plans.

A promotional listing advertises an AI automation course covering AI workflow and agent automation, with hands-on training and a link to enroll.

Alibaba's Qwen team launched Qwen3.8-Omni-Flash, a 1M-context omni-modal model for agentic audio-video understanding and tool use, available via hosted API with reported benchmark gains and steep price cuts.

A former OpenAI researcher who helped create RLHF has launched TypeSafe AI's Jev, a non-language transformer model that outputs calibrated probabilities instead of text, drawing strong developer interest for cheaper, faster, hallucination-free automation tasks.

Salesforce's Dreamforce 2026 recap shows the launch of AIforce as a new interface layer and its Koa CRM reasoning model, but much of the announced stack was previously revealed or remains in pilot.

NVIDIA quietly posted a gated, commercially-licensed CARI4D human-object interaction checkpoint on Hugging Face around August 30, 2026, with a larger 231M-parameter model, a new MHR body rig, and commercial-use terms — but no announcement, benchmarks, or pricing.

PharmExec answers reader questions on a busy day of pharma news, including Novo's rebrand and Anthropic AI drug discovery partnership, Bayer's 30-year CKD first, and Sanofi's portfolio simplification.

Anthropic has opened a physical biology lab in the Bay Area to push its AI drug discovery work beyond computational screening into real-world experiments, just days after its CEO urged slowing frontier AI development.

A weekly analyst tracker argues AI competition is shifting toward end-to-end task delivery via voice and background agents, but claims of cost advantages still await real-world validation.

A vendor tracking page confirms Google has not announced Gemini 3.9 Flash and no official API, pricing, or model card exists, with the latest documented Flash release remaining Gemini 3.8 Flash from September 2, 2026.

A weekend roundup for financial planners highlights Anthropic's new advisor-specific Claude plug-in and surveys showing early AI adoption and client reluctance to replace human advisors.

A Startup Pill analysis of how founders are building AI agents that autonomously execute entire professional workflows, citing recent funding rounds for Harvey, Clay, Cognition, and Ramp as evidence of a $10 trillion opportunity.

Gov. Gavin Newsom ordered California agencies to develop AI safety recommendations—including kill switches, external safety evaluators, and expanded incident reporting—after vetoing a tougher AI safety bill two years ago.

Also reported by: kpbs.org

A Substack analysis argues that AI regulation is becoming central to the AI trade, drawing on historical precedents to frame current debates over safety, concentration, and the pace of diffusion.

Congress extended most Defense Production Act authorities through Dec. 11, 2026, preserving a national-security tool that experts say could be used to compel reporting from frontier AI developers as safety and cyberattack concerns mount.

Legal experts say standing requirements, speculative damages, and difficulty proving existential harm make it nearly impossible to sue AI companies preemptively, even as executives warn of catastrophe.

Legal and national security experts say the Cold War-era Defense Production Act, extended through December 2026, could give the Trump administration leverage over frontier AI companies, particularly through information-gathering and reporting requirements, though its powers remain largely untested for this purpose.

A new South Carolina Senate special committee held its first meeting to begin a two-year review of how the state should regulate artificial intelligence across government, schools, law enforcement, businesses, and consumer markets, with no legislation yet produced.

A Modern War Institute analysis argues that the Pentagon's rapid GenAI.mil rollout has outpaced its accountability mechanisms, leaving governance to be bypassed by workforce attrition, operational urgency, and top-down adoption mandates.

A Think Digital Partners webinar with Defra, GDS, and SAS offers five practical takeaways for how government bodies can move responsible AI principles into everyday public service delivery.

A reported AI-assisted intelligence error that nearly triggered a US-China naval incident is prompting investor and regulator attention toward stricter human-validation and auditing requirements in defense AI procurement.

Bellwether's education newsletter analyzes the trend of school districts adopting screen time and AI bans or limits, arguing leaders should ask better questions about current usage, goals, and what replaces restricted tools before enacting blunt policies.

A RefAssured survey of 646 U.S. jobseekers and hiring professionals finds that generative AI is fueling resume misrepresentation and eroding trust in automated candidate screening.

A University of Konstanz survey of 1,105 employees finds workplace AI use rose only slightly to 38%, with adoption largely employee-driven and uneven across occupations, education levels, and organization sizes.

A finance executive argues that trust and data governance, not budget, are the real gating factors for AI adoption in accounting and reporting workflows.

A thin, AI-generated commentary piece speculating about an OpenAI report on how workers are using AI to change their workflows, without access to the report's actual findings.

A short commentary argues that accountants' core superpower is trust, which AI automation of bookkeeping tasks does not displace, drawing on a decade-old Xero survey showing accountants as the most trusted profession in the UK.

A Frontiers study of 106 Portuguese professionals finds that perceived organizational support buffers AI-driven job-security worries but fails to protect long-term career satisfaction, highlighting limits of generalized support amid AI adoption.

An analysis of research on women changing industries argues that as AI reshapes work, employers must become more flexible in recognizing transferable skills rather than requiring prior industry experience.

Potential Democratic presidential candidates are racing to respond to AI risk warnings while Trump dismisses them, with governors Newsom and Shapiro taking state-level action and senators proposing federal legislation ahead of the midterms and 2028 race.

US officials warn that Russia has authorized an AI-driven covert influence campaign to spread fabricated political content and deepen divisions ahead of the November midterms, with China also being monitored for possible targeted interference.

Two election-security experts assess how foreign and domestic actors are using increasingly sophisticated AI tools in influence operations and argue that AI companies, lawmakers, and civil society must strengthen safeguards before the November midterms and 2028.

UNESCO published "The Algorithm in the Room," a collection of 26 think pieces examining how AI and algorithmic systems are reshaping education and framing the "alignment problem" between AI goals and education's deeper purposes.

AI healthcare stocks surged after Novo Nordisk's Anthropic collaboration and a clearer Tempus AI reimbursement outlook renewed investor conviction that AI is becoming a revenue-generating part of pharma and diagnostics.

A daily funding roundup highlights Crusoe's $3.9 billion Series F at a $30.9 billion valuation alongside major Chinese deep-tech raises and smaller rounds in payments and security, showing capital concentrating on scarce AI compute and infrastructure.

Cory Doctorow argues that the real fear driving AI executives' superintelligence warnings is not existential risk but competition and falling unit economics, and that invoking superintelligence could justify antitrust exemptions against rivals and Chinese open-weight models.

Creators are beginning to pitch brands on their ability to influence LLM-generated answers, marking a shift in creator marketing toward answer engine optimization.

A teacher-focused review recommends six AI tools for classroom paperwork and preparation while warning against using AI for judgment calls about students.

At the CIEX conference in Indianapolis, software vendors and chemical executives discussed how to overcome scientist skepticism and build trust in AI tools for laboratory research and materials discovery.

RAYA is a hybrid learned-analytic control framework that embeds a learned recoverability margin into a hard-constrained optimal controller and shifts task priorities during recovery, validated across thousands of simulations and on-board Crazyflie quadrotor flights.

Anthropic's September 2026 Threat Intelligence Report details disrupted malicious uses of Claude, including AI-generated dating personas, mass surveillance platform development, and hotel WiFi compromise in espionage campaigns.