Vol. I, No. 15Monday 16 November 2026Published weekly · 05:00 Irish time

The AI Week Ledger

A weekly briefing on artificial intelligence — from silicon to capital. Every item cited.

The week of 10 – 16 November 2026

The rules for frontier AI are moving from statute to paperwork. New York opened registration under its RAISE Act: developers with more than $500 million in annual revenue must register by 1 January, publish safety frameworks and report critical incidents within 72 hours, with fines of up to $3 million for repeat violations. In Paris, 17 heads of state and government and ten AI companies signed a call for a common frontier-model evaluation framework and rapid incident alerts, though neither the US nor the UK is among the listed governments. In Washington, a bipartisan House bill would give the NSA a mandatory 30-day pre-deployment review of the most powerful models. The harder question is who does the checking: independent auditors are scarce, and thousands of models with their safeguards stripped still sit on Hugging Face.

The week's largest shock came from the hardware side. Nvidia said Jensen Huang is stepping down as chief executive, and CNBC reports the board moved to remove him after reports of extramarital affairs; CFO Colette Kress is interim CEO, five days before Nvidia reports quarterly results. CNBC notes that more than $500 billion of preliminary AI infrastructure financing arranged with Wall Street firms now carries a leadership question. Elsewhere, TSMC declined to comment on a report that it is weighing memory production at Nvidia's request, and Cornelis launched an 800G fabric pitched against InfiniBand.

Below the frontier, the stack is widening. Anthropic released Claude Fable 5.2 at $8 per million input tokens while holding its Mythos twin back for vetted organisations, Google launched a $0.10 decision endpoint, and NVIDIA, AMD, Intel, Google and Meta joined an SGLang consortium as the engine gained a native AWS Trainium backend. Google DeepMind's Gemini Robotics 2.5 reached Mercedes-Benz plants on Apptronik humanoids while OpenAI folded its robotics team into general research. Capital kept moving: Anthropic told banks the pricing in an IPO filing showing up to $69.3 billion was inaccurate, Mercor doubled its valuation to $20 billion, and Naver signed a $1.5 billion compute contract.

I Safety & Governance Reliability, regulation & policy

New York opens RAISE Act registration portal; large frontier developers must register by Jan 1, 2027

New York opened a registration portal for its RAISE Act. Large frontier developers (over $500M in annual revenue) must register by Jan 1, 2027, publish safety frameworks, report critical safety incidents within 72 hours and issue transparency reports. Penalties run up to $1M for a first violation and $3M for later ones. The DIGIT office oversees compliance, and Gov. Hochul launched the New York Center for AI Responsibility and Research at Binghamton University to independently evaluate advanced systems.

New York adds to California and Illinois, building a state-level frontier AI compliance patchwork in the absence of federal law.

Paris Call for AI safety and security signed by 17 heads of state and government, 10 AI companies and civil society groups

Signatories (France, Germany, Canada, Ireland, Kenya and others; Anthropic, OpenAI, Microsoft, Mistral and others) commit to a common frontier-model evaluation and risk-management framework covering misuse and loss of control, with non-sensitive findings published in a standard format. They also commit to supporting independent third-party evaluation, rapid alert systems for serious cyber and AI-control incidents, and confidential whistleblower channels. Endorsements remain open.

A multilateral text with lab signatories pointing toward comparable evaluations and cross-border incident sharing. The US and UK are not among the listed heads of government.

Bipartisan American AI Security Act (H.R. 10981) would require NSA pre-deployment review of the most powerful AI models

Reps. Gottheimer (D-NJ) and Lawler (R-NY) announced H.R. 10981, which would require developers of the most powerful AI models to give the NSA full access for a mandatory pre-deployment review of cyberattack and CBRN risk. Reviews run 30 days with one 30-day extension, and companies can appeal through an expedited process modeled on national-security courts. Its legislative prospects beyond the announcement are untested.

It is a national-security route to federal frontier oversight. The source is an announcement summary, so the bill's prospects are untested.

Hugging Face appears to be hosting a load of models that could be used for cyber attacks

Transformer reports that more than 6,500 models tagged abliterated, uncensored or heretic were on Hugging Face in July, some openly marketed for offensive cyber use, despite a policy that prohibits models designed to gain unauthorised access or generate malicious code. Hugging Face removed one GLM-5.3 derivative marketed for offensive cyber in October but left others from the same uploader. It cites Anthropic testing in which GLM-5.3 wrote working exploits for known browser vulnerabilities in 12% of attempts. Hugging Face did not respond to questions.

Highlights a platform-governance gap for open-weight models with stripped safeguards.

Independent AI auditors have a recruitment problem

Independent verification organizations are central to proposed AI oversight, but model-evaluation talent is scarce and mostly employed by frontier labs, and training pipelines are at least a year from producing graduates. A Massachusetts bill section still in conference, Illinois's July auditing law and the federal FRONTIER Act (unlikely to pass this Congress) would all involve audits. The EU Code of Practice lets companies skip auditing if they can show no qualified auditor is available.

Auditor capacity, not legal mandate, may be the binding constraint on credible third-party AI oversight.

Trust but Verify: Implementing Frontier AI Policy in California and Beyond

Carnegie describes California's SB 53 as a 'trust but verify' model: frontier developers design and disclose their own risk-management practices while government and third parties check the claims. It maps how New York's RAISE Act, with a 72-hour incident window and a new oversight office, and Illinois's SB 315, which adds annual third-party audits, diverge from California. It warns that policies often fail at implementation and urges the state to build permanent in-house technical expertise rather than rely on fellowships.

A practical map of how state frontier AI laws will operate and where they may fail.

EU AI Act Newsletter #114: Who evaluates the evaluators?

The AI Office presented proposed qualification and independence requirements for external evaluators of systemic-risk general-purpose models at a workshop (reported as 15 July 2026), and plans a scoring system. The issue cites a Cambridge law professor saying information requests have gone to over 30 providers. This is his account in an interview, not an AI Office statement. It also relays a Lawfare argument that the Act's serious-incident definition covers AI systems but not clearly models.

Shows EU enforcement moving from rules to operational detail on evaluators and incident reporting.

Despite Huang scandal, US-China AI safety dialogue still seems on track

Opinion piece by Mark MacCarthy. After Nvidia CEO Jensen Huang's reported resignation, the Senate NDAA includes the Chip Security Act (chip tracking), which the House has not yet acted on, and Commerce says it is finalizing rules on Chinese access to restricted chips in servers abroad. The author argues the US-China AI dialogue agreed in September, which includes an incident line, remains on track ahead of the Trump-Xi meeting on Nov 18-19 in Shenzhen.

Export controls and bilateral AI safety diplomacy are entangled, with implications for compute access. It is a perspective piece, and the measures cited are pending.

II Training Infrastructure Silicon, storage & the physical layer

Nvidia says Jensen Huang is stepping down as CEO; CFO Colette Kress named interim chief

Nvidia announced on November 12 that Jensen Huang is stepping down as chief executive, without giving a reason. CNBC reports, citing a person familiar with the matter, that the board decided to remove him after reports of extramarital affairs, one of them linked to a businesswoman US intelligence reportedly believes is a Chinese spy. CFO Colette Kress becomes interim CEO and stays CFO while the board searches for a permanent successor. Nvidia still plans to report fiscal third-quarter results on November 17.

The dominant AI accelerator vendor loses its founder-CEO days before earnings, putting roadmap execution and hyperscaler commitments under unusual scrutiny.

TSMC 3D DRAM report linked to Nvidia draws response

Following a report that TSMC is weighing advanced memory production at Nvidia's request, TSMC said it does not comment on market rumours and remains focused on its core foundry business. The statement does not explicitly rule out an internal study, and neither company has announced a memory manufacturing program.

A foundry move into DRAM would reshape the HBM supply chain that currently depends on SK hynix, Samsung and Micron.

Cornelis launches CN6000 800G networking to rival Nvidia InfiniBand

Cornelis announced CN6000 800G Omni-Path cards and switches, claiming 1.6 billion messages per second and 21% faster small-message passing than InfiniBand, both from its own tests. An Ethernet card with Ultra Ethernet support is due later. Dell, Lenovo and Supermicro systems are expected in early 2027.

Another non-Nvidia fabric option adds open-interconnect pressure on InfiniBand in AI and HPC clusters.

Dell's 157 PB all-flash object storage rack

Dell built an all-flash ObjectScale rack of 16 PowerEdge R7725xd servers holding 640 Kioxia LC9 245.76 TB SSDs in 32 RU, for about 157 PB raw. It targets AI and data lake workloads.

Ultra-dense QLC flash is displacing HDD for AI data lakes, tightening NAND supply and changing training-data tier economics.

III Inference Models, serving software & the KV-cache economy

Anthropic releases Claude Fable 5.2, with Mythos 5.2 limited to vetted organisations

Fable 5.2 and Mythos 5.2 share one underlying model; Fable carries stronger cybersecurity and biology safeguards and is generally available, while Mythos is restricted to organisations in Anthropic's cyber and life-sciences verification programmes. Pricing is $8 per million input tokens and $40 per million output tokens, with an optional fast mode at $16 and $80 that Anthropic says is up to 2.5 times faster. Anthropic claims a step up from Opus 5.5 on most benchmarks, including 49.1% on FrontierCode 2.0; these figures are self-reported.

Splitting one frontier model into a safeguarded general release and a vetted dual-use tier is becoming the template for shipping the most capable systems.

NVIDIA, AMD, Intel, Google, Meta and others form SGLang Consortium

Launched at the first SGLang Summit in San Francisco, the consortium's founding members are NVIDIA, AMD, Intel, Google, Meta, Cloudflare, Modal and RadixArk. Each names a technical representative and joins two roadmap meetings a year, while SGLang's maintainers keep responsibility for technical decisions and code acceptance. Joint work is planned on frontier model support, reasoning and multimodal workloads, distributed inference and hardware optimisation.

Cross-vendor backing positions SGLang as a vendor-neutral serving engine alongside vLLM.

Google launches Gemini Decisions API, a single-pass decision model at $0.10 per million input tokens

Gemini Decision 1 scores closed-set choices (yes/no, up to 255 options, rubrics) in one forward pass at $0.10 per million input tokens and $0 output, with confidence scores. Artificial Analysis measured about 211 ms per task and a Decision Index of 66, tied with Elephant 9B and Jev 1.14 and 7 points behind Drex Max v1.1 and GLiDE v1.2. Its per-task cost ($345 per 1M tasks) is higher than Jev's ($156) in AA's figures.

Non-generative decision endpoints are becoming a distinct, cheaper inference product category.

SGLang-Neuron v0.1.0 open-sourced: SGLang runs natively on AWS Trainium

The SGLang team and AWS Annapurna Labs released SGLang-Neuron v0.1.0 under Apache 2.0 as a public preview at the SGLang Summit. It runs LLMs on Trn3 (eight architectures, including Qwen3 family, GLM-5.2 and Kimi-K3 with prefill/decode disaggregation across three nodes). Features, APIs and supported models may change, and no third-party GPU comparison has been published.

A mainstream engine gaining a first-class non-NVIDIA backend widens hardware choice for production serving.

Meituan open-sources LongCat-2.5, a 1.6T-parameter MoE trained without Nvidia GPUs

Weights were released Nov 15 under MIT for a MoE of about 1.6T total and 48B active parameters, announced Nov 2. Meituan describes a sample deployment on 194 Huawei Ascend Atlas A2 accelerators. Launch promo API pricing is roughly 0.27 EUR input and 1.07 EUR output per million tokens and may change. All figures come from Meituan and a press report and are not independently verified.

A frontier-scale open model trained and served off-NVIDIA is a data point for hardware diversification.

Thinking Machines publishes Inkling 2 Small open weights under Apache 2.0

A week after promising open weights, Thinking Machines Lab published Inkling 2 Small on Hugging Face on November 12 under Apache 2.0. It is a 594B-parameter mixture-of-experts model with 17B active per token, routing each token to 8 of 384 experts, with a 1M-token context window. The lab recommends its native NVFP4 checkpoint, which needs at least about 500 GB of GPU memory, such as two B300s or four H200s.

A large, permissively licensed MoE with an FP4 checkpoint raises the bar for open-weight models.

IV Edge AI Industrial, robotics, client machines & telco

Google DeepMind unveils Gemini Robotics 2.5, tested on Apptronik Apollo 2 in Mercedes plants

Google introduced Gemini Robotics 2.5, a vision-language-action model reported to deliver 30% to 100% higher success rates than prior baselines depending on task. In a Mercedes-Benz proof of concept at plants in Kecskemet, Hungary, and Sindelfingen, Germany, Apptronik's Apollo 2 achieved a reported 96% success rate on automotive part picking and packing. It is not stated whether inference runs on the robot.

Frontier VLA models are entering factory pilots, raising the question of where robot inference will run.

OpenAI disbands stand-alone robotics team, folds work into general research

OpenAI confirmed to Fortune that it dissolved its stand-alone robotics team; research staff now report to the Chief Research Officer and applied robotics and hardware staff moved to Applied. The company says its general-purpose models can be applied directly to robotics and that robotics remains core to its long-term vision.

Signals a bet that general foundation models, rather than specialized teams, will drive embodied AI.

Agile Robots raises $500M at $8.5B valuation led by SoftBank

Agile Robots closed a $500M Series F led by SoftBank, with Nvidia, Hitachi and T Capital participating, bringing total funding to $2B. Its Agile One SN humanoid is preconfigured for NVIDIA Isaac GR00T models.

Heavy capital continues to flow to industrial physical-AI players aligned with NVIDIA's robot stack.

NXP launches unified robotics platform with open SDK and drone reference designs

NXP announced on Nov 9 a platform spanning sensing, real-time control, AI and comms, with an open-source SDK integrating PX4, Zephyr and ROS 2 and four drone reference designs (NavQ95 companion computer, Tropic flight management unit, 100BASE-T1 sensor node, AM32 ESC). NXP says subsystem integration drops from months to days or weeks; service robot and AMR enablement is expected in 2H 2027.

Embedded incumbents are moving from components to full-stack, self-serve robotics systems.

MIPS adds motor control, edge AI and TSN to RISC-V Actus platform for humanoid joints

At electronica in Munich, MIPS (owned by GlobalFoundries) showed Actus, a reference platform with two Atlas M8500 real-time RISC-V cores, a 512-bit vector unit for local sensor fusion and predictive maintenance, and three-port TSN Ethernet. No customers, pricing or timelines were disclosed and performance figures are vendor claims.

Pushes inference to individual robot joints, a distributed-compute opening for RISC-V against Arm and NVIDIA.

Swiss startup Embodied AI unveils AIR wheeled humanoid that learns factory tasks from demonstrations

At Humanoid Europe 2026, Embodied AI showed AIR, a tendon-driven wheeled humanoid designed under ISO/TS 15066 (a design basis, not a certification), with CE marking planned for 2027. It starts under operator control and learns from demonstrations, typically operating autonomously after about a month. The 98% reliability figure in a cable-handling test is company-reported with no independent verification.

Illustrates a European safety-first, demonstration-learning approach to industrial humanoids.

V Academia Preprints, benchmarks & research signals

SO-Arena: Benchmarking the Capability Spectrum and Reliability of System One Models

SO-Arena is a benchmark for fast-decision 'System One' models, spanning 11 domains and five capability families with sandboxed interactive tasks. It also tests decision stability when the option set changes and whether a model recognizes that no valid answer is offered. Across 21 models, gains from scaling and added reasoning vary by domain, and the most accurate model ranks only seventh on reliability.

Shows that accuracy leaderboards can hide reliability gaps that matter for deployment decisions.

Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents

The authors argue that lessons distilled from past agent trajectories can be wrong or misapplied, so agents with memory sometimes fail tasks they would otherwise solve. Their WikiTrace stores each lesson linked to a summary of its source task and the raw trajectory, so the agent can verify or correct it before acting. It beats existing memory approaches on five agentic benchmarks and reduces this degradation, and the code is public.

Agent memory can hurt performance, so teams building persistent agents should add provenance and verification.

Scaling public input on AI policy with pluralistic agents

The paper proposes 'pluralistic agents', AI representatives that deliberate on behalf of individuals to feed diverse views into AI policy. In a preregistered study of 746 participants, people preferred them to giving input to a single AI model, though most still preferred participating directly. The agents recovered 68% of the baseline-to-direct gap in satisfaction and 82% of the gap in fairness.

An early empirical test of AI-mediated public consultation; the authors frame it as amplifying rather than replacing human input.

Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery

Sage-Taiyi is an AI co-scientist architecture that monitors and revises its assessments of evidence and assumptions to choose what to investigate. It ranks first on all eight metrics of the authors' own Taiyi-Bench (265 tasks), scoring 54.0 against 47.9 for AI Scientist-v2. In five expert-guided chemistry collaborations, one reaction yield rose from trace amounts to 84% and another from 13% to 69%.

Part of the AI-for-science push toward wet-lab results, though the main benchmark is self-introduced and the claims are unreplicated.

Learning Heterogeneous Tactile Representation for General Dexterous Manipulation

LingBot-Touch is a tactile foundation model that maps piezoresistive, magnetic and visuotactile sensors into one representation and can be extended to new sensor types. Its tokens plug into policies from diffusion policies to vision-language-action models such as pi0.5 and LingBot-VLA. The authors report gains on complex manipulation tasks in simulation and on real robots across several hardware platforms.

A step toward sensor-agnostic touch pretraining that could reduce per-hardware retraining in robotics.

Few-Step Streaming Video Generation via Condition Shifting

The paper extends APEX to streaming video generation. It drops the separate fake-score network used in Distribution Matching Distillation by estimating the velocity field inside the generator, using a shifted text condition. On 180-second videos it beats S2-Avatar on instruction fulfillment and aesthetics, while raters favor S2-Avatar for motion quality. Versus DMD it uses 4.4% less peak GPU memory and 9.2% shorter iteration time.

Cheaper few-step distillation for long real-time video generation lowers serving cost.

VI Investment Financials, trends & notable moves

Anthropic tells banks the pricing in an IPO filing showing a raise of up to $69.3 billion was inaccurate

Reuters reported that a filing sent to banks showed about 867 million Class A shares at $70 to $80, which would raise up to $69.3 billion at the top of the range. Late on Monday, November 9, Anthropic told banks the pricing details were inaccurate and would be updated, without saying what was wrong. Investors had earlier discussed a valuation of up to $2 trillion, but Reuters says $1.5 trillion is now seen as the high end. Anthropic declined to comment.

If completed, it would be the largest tech IPO on record and a test of whether public markets will absorb frontier-lab valuations; the retraction leaves terms unsettled.

Jensen Huang's Nvidia exit casts shadow over Wall Street's $500 billion AI financing bet

Huang resigned as Nvidia president, CEO and director on Nov. 12 following reports of extramarital affairs; CFO Colette Kress is interim CEO, with no permanent successor named. Shares fell more than 13% before partly recovering. CNBC notes that in August Nvidia signed preliminary memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to raise over $500B for AI infrastructure; these remain preliminary and subject to final agreements.

Leadership risk at the central chip supplier is now a variable in debt-financed AI infrastructure build-out.

The Week's 10 Biggest Funding Rounds: Agentic AI And Infrastructure Are Still On A Roll

For November 7 to 13, the largest US rounds were Mirendil's $1 billion Series A (Kleiner Perkins, Clearlake), Mercor's $750 million Series D at $20 billion, and Torus's $500 million growth financing led by Magnetar. LangChain ($300 million Series C at $3 billion, CapitalG) and Path Robotics ($300 million Series E at $2.45 billion, Georgian) tied for fourth.

Capital remains concentrated in agentic software, expert-data supply for model training, and infrastructure.

Mercor raises $750M at $20B valuation, doubling its worth in just over a year

Mercor announced a $750M Series D on Nov. 10 at a $20B valuation, up from $10B when it raised $350M in October 2025. Investors include Nvidia, Salesforce, Benchmark, Felicis and General Catalyst; the round includes an employee-share tender and the primary/secondary split is undisclosed.

Investors are paying frontier-lab-adjacent multiples for the human-expert data layer behind post-training.

Naver clinches $1.5 bil. AI computing deal with US firm

Naver signed a roughly 2 trillion won ($1.5B) contract for 20MW of GPU computing, with supply on next-generation GPUs starting in 2027. It is the first deal for its AI Factory business built with Brookfield and Nvidia, and capacity is planned to grow from 55MW next year to 200MW by 2028.

Shows AI compute capacity being exported by non-US operators and monetized through long-term contracts.

Axon puts about $1.2 billion into drone maker Skydio at a $5.2 billion pre-money valuation

Axon will invest about $1.2 billion in cash and stock, combining new capital for Skydio with purchases from existing shareholders, adding roughly 20% to its fully diluted stake. Axon becomes Skydio's exclusive distributor to public-safety customers in the US and Canada. The transactions are expected to close across the fourth quarter of 2026 and the first quarter of 2027, subject to final terms.

Autonomous drones are being pulled into vertically integrated public-safety platforms, a sign of consolidation in defence and security physical AI.