TechnologyNov 16, 20268 min read

Top 10 AI News — November 16, 2026

Washington closes the remote-compute loophole on AI chips, NVIDIA, IBM and Fermilab wire GPUs to quantum processors, and SC26 shows that the next limits on AI are power, cooling and light.

Neural Dispatch

Editorial Desk

Share:

Today's stories have one thing in common: AI is merging with the systems around it. Chips are bound up with geopolitics, GPUs are being paired with quantum processors, models are moving into chemistry labs and clinics, and data centers are being redesigned around power and light rather than raw compute.

1. Washington closes the remote-compute loophole on AI chips

The Commerce Department is rolling out an interim final rule that requires licenses to export, reexport or transfer AI chips and servers to data centers anywhere in the world if those facilities may provide compute to companies in China or other restricted countries. TechCrunch reported the details on Monday. The rule applies to D:5 embargoed destinations, including China, Russia, Iran and North Korea, and to indirect access through brokers and intermediaries.

The threshold is 12 million TPP (total processing performance), which TechCrunch estimates is about the maximum compute of 750 Nvidia H100 or H200 chips. The rule also adds new license requirements for advanced chips going to Malaysia, Singapore, Indonesia, Thailand and Brazil, citing a "higher-than-normal risk of diversion." Existing operators get 90 days after the rule takes effect to obtain licenses. The rule formally rescinds the January 2025 AI Diffusion Rule.

Until now, export policy controlled where chips were shipped. This rule controls who can rent them, which makes every overseas GPU cloud a compliance question. For the background on how Nvidia's market is splitting along these lines, see our analysis of the fracturing AI chip market.

2. NVIDIA, IBM and Fermilab pair GPUs with quantum processors

NVIDIA announced a modular quantum–classical codesign platform with IBM and Fermi National Accelerator Laboratory. NVIDIA DGX B200 systems will sit alongside IBM quantum computers, including the 120-qubit Nighthawk processor, and Fermilab's SQMS Center will lead the error-correction and physics work. The partners have not disclosed a start date or a cost.

NVIDIA also released a new decoding algorithm in CUDA-Q QEC 0.9.0 and says it speeds up error-correction research by up to 13,000x, cutting one computation "from 12 weeks to just 45 minutes." That figure comes from one benchmark comparison, not a measure of general performance. In separate work, IBM and RIKEN ran quantum-classical chemistry on GB200-based systems on complexes of up to 12,635 atoms, more than 40 times the previous 303-atom benchmark. The work is a finalist for the 2026 ACM Gordon Bell Prize.

In this setup the quantum chip does not replace the supercomputer; it works as an accelerator alongside it, and GPUs handle error correction. NVIDIA is also adding QPU scheduling to Slurm, the job scheduler most supercomputing centers already use.

3. NVIDIA launches Nemotron for Science

NVIDIA introduced Nemotron for Science, which packages its open Nemotron 3.5 Super VL model with an open chemistry dataset and post-training recipes for life sciences and chemistry. The dataset, LitmusBench v0.2, has nearly 50,000 chemistry reasoning questions, covering molecular properties, solubility, and how the body processes molecules.

According to NVIDIA, early adopters reported these results. Terray Therapeutics recovered about 76% of annotated reactions on the RxnScribe test. Lila Sciences reached 61% accuracy on quantum-dot protocols, against 54% for GPT-6 Astra, at 10 to 20 times lower inference cost. A KTH agent fixed 81% of test cases in a single overnight run on one DGX Spark. The model and dataset are on Hugging Face.

This is the "programmable biology" bet in its most concrete form: domain agents that run on a desk-side workstation and are trained on open lab data rather than a frontier API.

4. Apple and OpenEvidence launch a 200,000-person study of health AI

Apple released a research app, built with healthcare AI company OpenEvidence, to study how people use AI for health questions. TechCrunch reported that the study is open to U.S. users over 21 with an iPhone 15 or later and Apple Intelligence enabled. It aims to enroll up to 200,000 participants for more than 12 months.

Answers are grounded in clinical sources such as NEJM, JAMA and Nature. Participants rate the answers, take monthly surveys, and can choose to share HealthKit data such as resting heart rate, sleep or lab results. Apple says the data is anonymized and not used to train AI models, and OpenEvidence deletes queries after 24 hours. "Health is deeply personal," said Sumbul Desai, Apple's VP of Health and Fitness.

Consumer health AI has mostly shipped first and been measured later, if at all. A year-long, consented study at this scale could produce the first large dataset on whether people understand and act on these answers correctly.

5. Dell targets agent data and liquid-cooled storage at SC26

Dell expanded its AI Factory with NVIDIA at Supercomputing 2026 in Chicago. The main addition is an Agentic Data Gateway, a single governed route for AI agents to reach enterprise data. In Dell's internal testing, reusing context cut token consumption by up to 76% and raised answer accuracy from 60% to 94%. Dell also announced direct-liquid-cooled storage for rack-scale AI, due in early 2027, and rack management software that handles up to 100,000 devices.

"AI outcomes don't come from models alone," said Arthur Lewis, president of Dell's Infrastructure Solutions Group. Dell also launched services to "align power, cooling and space requirements" with AI workloads. Enterprise AI planning now includes facilities planning: data access and floor space are as much the bottleneck as GPUs.

6. Lambda deploys co-packaged optics across 10,000 GPUs

Lambda and NVIDIA reported a deployment of co-packaged optics (CPO) networking across more than 10,000 Blackwell Ultra GPUs at Lambda's 142 MW Quebec campus, using NVIDIA Quantum-X Photonics switches. Industry coverage describes it as the first verified commercial CPO deployment in an AI cloud.

The reported gains: latency 10.1% lower at pod scale, optics running 15°C cooler than pluggable transceivers, and 3.5x to 5x better optical power efficiency. That matters because retimers in pluggable optics can consume 20% to 30% of a switch's power. Instances are being phased into production from late 2026.

Moving data between chips is now a large share of a cluster's power budget. Building optics into the switch package is how operators cut it.

7. Intel and SambaNova split inference across two kinds of silicon

At SC26, Intel and SambaNova demonstrated disaggregated inference for the GLM-5.3 model. Four Intel D270 GPUs handle prefill and a SambaNova SN50 rack handles decode, connected over standard Ethernet. Both systems are air-cooled. The D270 runs at 350 W per card, and an SN50 rack draws about 20 kW during inference.

The pitch is aimed at the many data centers that cannot host 100–130 kW liquid-cooled GPU racks without multimillion-dollar retrofits. Audited MLPerf results are still pending, so the throughput claims are the vendors' own. Production configurations are planned for early-access clouds in Q1 2027.

8. Noeri launches with $110 million to build AI that companies own

Stanford professors Yejin Choi and Carlos Guestrin launched Noeri, a Palo Alto and Zurich lab, with $110 million led by Menlo Ventures. NVIDIA, Cisco Investments and Snowflake joined as strategic investors, alongside Madrona, Kindred, Wing and NEA. No valuation was disclosed.

Noeri describes itself as building self-improving AI for enterprises and scientific discovery: models and agents that learn from a company's live traffic and train specialist models the customer owns. It frames this against "a single superintelligence, owned by a few." Guestrin co-created XGBoost and co-founded Turi, which Apple acquired in 2016. The company has not disclosed customers or revenue.

9. AI-linked job cuts more than double, but the economy is now the bigger reason

U.S. employers announced 121,401 job cuts attributed to AI from January through October, according to Challenger, Gray & Christmas data reported this weekend. That is about 20% of all announced cuts and more than double the 54,836 recorded for all of 2025. October itself had only 1,265 AI-attributed cuts, the lowest month of the year.

Market and economic conditions have overtaken AI as the most-cited reason, at 130,531 cuts. "The reasons employers give have shifted from technology to the economy," said Andy Challenger. The next Challenger report is due December 3, a day before the BLS jobs data.

The labeling shift matters as much as the totals. Companies that blamed AI for cuts earlier this year are now citing the economy, and workforce plans should read both labels skeptically.

10. A humanoid robot that takes cues from lab-grown human neurons

OpenMind, founded by Stanford bioengineer Jan Liphardt, showed Iris, a humanoid whose decisions are partly shaped by lab-grown human cortical neurons working alongside conventional computing. TechTarget reported that across nearly 500 stimulations, 2D neuron cultures decoded signals with up to 94% accuracy within 150–240 milliseconds.

Liphardt called the result "less like a brain and more like a reflex," and the company describes it as basic research, not a product. Keeping the cultures alive is the main barrier. It is still a clear example of robotics, biology and computing converging in a single system, which no single field could have produced.

What to watch

The comment period and the 90-day compliance window for the chip rule will show how much overseas GPU capacity has to be relicensed or shut down. SC26 runs through November 20, and Slurm 26.11 with QPU scheduling is due later this month, which will put the quantum-HPC claims in front of the centers expected to use it. Challenger's December 3 report will show whether companies keep shifting from citing AI to citing the economy.

#ai-news#daily-brief#export-controls#quantum-computing#sc26#ai-infrastructure#health-ai#ai-jobs

Neural Dispatch

Editorial Desk · The Neural Dispatch

Covering the intersection of AI, engineering, and the future of building. We dig into what the tools actually do, how builders are using them, and what it means for the industry.

Keep reading

Related dispatches

Technology
11 min read

Headline Top 10 Neural News: Q1 to Q3 2026

An AI model that finds zero-days in every major operating system, a $965 billion Anthropic, IBM's contested quantum advantage, and Washington redrawing the rules for crypto. The ten advanced-technology stories that defined the first nine months of 2026.

Oct 17, 2026Read more
Technology
7 min read

Top 10 AI News — October 16, 2026

China proposes export controls on 23 AI technologies, the White House superintelligence task force starts a 120-day clock, and the 20 largest AI stocks shed more than $1 trillion in a week.

Oct 16, 2026Read more
Technology
8 min read

Top 10 AI News — October 7, 2026

OpenAI publishes 372 math results from an unreleased model, the Pentagon says it has stopped using Anthropic, and Google signs a 3.6 GW power deal with Constellation.

Oct 7, 2026Read more