AI Daily Digest November 19, 2026: NVIDIA Posts $115.8B Quarter, OpenAI Turns GPT-6 on Physics
NVIDIA Q3 FY27: a $115.8B quarter and the fastest ramp in company history NVIDIA filed its first quarterly report since Jensen Huang stepped down, and the numbers did the talking: $115.8 billion in Q3 revenue, up 103 p
NVIDIA Q3 FY27: a $115.8B quarter and the fastest ramp in company history
NVIDIA filed its first quarterly report since Jensen Huang stepped down, and the numbers did the talking: $115.8 billion in Q3 revenue, up 103 percent year over year, with data center revenue at $107.6 billion. The Vera Rubin platform generated nearly $20 billion in its first quarter of production shipments, which interim CEO Colette Kress called the fastest product ramp in company history. Rack-scale systems have already landed with CoreWeave, Google Cloud, Azure, Oracle, and Nebius, with SpaceX SI coming online next.
Guidance for Q4 sits at $130 billion against a consensus near $125 billion, and the fiscal 2028 growth outlook was raised from 70 percent to roughly 75 percent. Cloud industry backlog reached $2.5 trillion, up from $2 trillion three months ago. On the earnings call, Kress reaffirmed the company's open model strategy and said the planned acquisition of Hugging Face, announced in September, is proceeding as expected. She also flagged a more selective posture toward neocloud customers and strong demand for Vera CPU, with CPU revenue expected to more than double in fiscal 2028.
The ecosystem numbers are arguably the most useful part of the call. Weekly token volume on OpenRouter is up 35x year over year, agentic workloads now account for roughly 80 percent of tokens, and Anthropic's annualized revenue hit $81 billion at the end of October. Frontier labs are expected to contribute about 15 percent of data center revenue this year, and token prices have fallen nearly 60 percent since May. Chair Harvey Jones said the board will not rush the permanent CEO search. For developers, the signal is that open-model inference is becoming the fastest-growing slice of NVIDIA's ecosystem rather than a side bet. β NVIDIA Β· Reuters
π NVIDIA IR
Isomorphic Labs raises $5B Series C
Isomorphic Labs, the AI drug design company founded by Nobel laureate Demis Hassabis, announced a $5 billion Series C co-led by Greenoaks and Sequoia Capital, with Andreessen Horowitz joining as a new investor and Alphabet continuing to put in significant capital. The round was oversubscribed and comes just six months after a $2.1 billion Series B, bringing total disclosed funding to $7.7 billion. It is the largest single financing round recorded in AI-driven drug discovery.
The company's core asset is IsoDDE, a drug design engine built from foundational simulation models that work across multiple therapeutic areas and drug modalities, paired with proprietary agents that reason over complex biological systems. Isomorphic reports positive preclinical results in its internal oncology and immunology programs as well as partner programs, and the new capital is earmarked for foundational biology models, pipeline expansion, and the infrastructure needed to carry its first candidates into clinical trials.
One fact worth holding onto: no Isomorphic-designed molecule has entered human trials yet. The company has research deals with Eli Lilly, Novartis, and Johnson & Johnson, but the clinical test of the platform still lies ahead. Investors are clearly betting on the trajectory rather than the endpoints. β Isomorphic Labs Β· Reuters
π Isomorphic Labs
OpenAI opens a science blog with a lattice Standard Model result
OpenAI launched an official science blog this week, and the first guest post is not a demo video or a leaderboard. It is a write-up by Nathaniel Craig, a visiting scientist at OpenAI who leads the particle theory group at UC Santa Barbara, describing how he worked with GPT-6 Astra on a candidate nonperturbative lattice definition of the Standard Model. The problem has sat open for over fifty years: the Standard Model is a chiral theory, and putting chiral fermions on a lattice triggers the Nielsen-Ninomiya doubling problem, while constructing a gauge-invariant measure for overlap fermions stalled for decades.
The division of labor is the interesting part. Craig supplied the physical intuition, a refinement strategy that compares progressively finer lattices, but could not prove the construction converged. Astra turned the strategy into a concrete construction and led the development of the detailed proof across a 123-page preprint, including a Lean4 formalization. Lattice field theory experts Maarten Golterman, Andreas Kronfeld, and Yigal Shamir reviewed the derivation and pressed on key details.
The caveats are real and OpenAI keeps them visible. The result remains a candidate definition, reflection positivity at finite lattice spacing is unresolved, and Edward Witten has argued the chiral sign problem makes direct Monte Carlo simulation impractical regardless. Craig's own manual audit ran to 2,488 line items. Whatever you think of the framing, this is a different genre of model output than code autocomplete. β OpenAI
π OpenAI Science Blog
Claude Code Projects: the coordinator speaks first
Anthropic updated Claude Code Projects so the coordinator now reports progress without being asked. In a project, Claude splits work across parallel threads, and until this update you had to open each thread to see where things stood. Now the coordinator answers questions in the project chat and posts status updates on running threads, and you can set notification rules such as only pinging you on completion or blockers. Threads can be parked with a reminder and will come back on schedule.
The update also ports and repurposes two terminal commands. /btw opens a fresh thread for a side issue so the current one stays on topic, and /fork branches a new thread from the current state so you can trial a different approach in parallel. Single threads can now be shared as read-only links with any signed-in Claude user, with the sensible restriction that threads which touched your local machine cannot be shared. The threads panel gained a recent and pinned row plus artifact thumbnails.
The fine print matters for anyone planning to run a small fleet. Every thread draws on your existing plan limits, new projects default to Opus at high effort, and accounts are capped at 200 new threads per day. The public beta covers Pro, Max, and Ultra on web, desktop, and the latest mobile app; Team and Enterprise are not included yet. β Anthropic
π Anthropic
FlipTest: video world models score below random
A paper from Google DeepMind, UT Austin, and UC Berkeley asks video world models a blunt question: if the present changes slightly enough to flip the future, does your prediction flip too? FlipTest builds 1,680 starting frames from 10 simulated and 15 real-world tasks around visual minimal pairs, for example a box resting a few millimetres before a table edge versus a few millimetres past it. One stays put, the other falls.
All 11 models evaluated failed. The best averaged 18.3 percent pair accuracy across real and simulated scenes, below the 25 percent of random guessing, and no model exceeded 5.3 percent on simulated scenes. Scale did not help: the physical-AI specialist Cosmos3-Super scored 13.7 percent, essentially level with the general-purpose Wan2.2 at 14.0 percent. The sharpest control in the paper: when a vision-language model was allowed to measure the frame with code instead of eyeballing it, accuracy on boundary-crossing steps jumped from 31.3 to 93.3 percent, which locates the failure in state reading rather than missing information.
For anyone wiring world models into robot or driving planners, the implication is uncomfortable. Plausible pixels are not a held state, and models that predict the same future for both sides of a tipping point cannot be trusted with rules. The authors frame next-frame prediction as needing causal state identification, and the paper is a preprint with judge-dependent scoring, so treat the exact numbers as provisional while taking the design seriously. β Google DeepMind Β· arXiv
π arXiv:2611.19822 Β· FlipTest
Gemini Robotics 2.5 now runs natively on Franka FR3 Duo
Franka Robotics, an Agile Robots company, announced that Gemini Robotics 2.5 runs natively on its FR3 Duo dual-arm robot. Because DeepMind used FR3 Duo as one of the primary training embodiments, many policies run on the robot with no additional training, which is the kind of boring operational detail that actually saves research teams weeks. The integration was shown live at CoRL 2026 in Austin, and Franka open sourced its tooling for teleoperation, data collection, and inference.
The company also took orders for the next-generation Mobile FR3 Duo with whole-body control, where the arms and the base run in a single 1 kHz control and data acquisition loop under one safety concept, so the mobile system behaves as one robot instead of an arm bolted to a cart. Availability is mid-2027, with EU and US certification to follow. A new Franka Gripper with real-time force, position, and speed control arrives in Q1 2027.
Context from DeepMind's own announcement: Gemini Robotics 2.5 lifted success rates by 30 to 100 percent across more than 200 evaluated tasks and cut completion times by 20 to 30 percent, and an Apollo 2 humanoid running the model achieved a 96 percent success rate on parts kitting on active Mercedes-Benz production lines. Independent validation at scale remains the open question, as with most robotics benchmarks. β Franka Robotics Β· Google DeepMind
Voyage Labs raises $12M for Atlas, a travel agent that books instead of linking
Voyage Labs announced a $12 million seed round led by Kleiner Perkins with SignalFire and SV Angel participating, plus angels including Brian McClendon, co-creator of Google Earth, Lenny Rachitsky, Notion cofounder Akshay Kothari, and Skyflow CEO Anshu Sharma. The founders are Manik Gupta, who led the rebuild of Google Maps and ran product at Uber, and Yogesh Jain, who built generative AI features at AWS and ServiceNow.
The product, Atlas, is a personal agent for travel and local plans that learns who you travel with, what you like, and what you will spend, then books flights, hotels, and experiences through supplier partnerships rather than handing back links. It runs as three agents: a Live Agent you talk to, a Proactive Agent that watches prices, availability, and check-in times after you log off, and a Computer Agent that operates across the web where no API exists. It lives in iMessage and on the web, currently in private beta.
The pitch leans on a familiar statistic, that 71 percent of US travelers already use AI for trip inspiration, and the gap the team is attacking is everything that happens between inspiration and a booked trip. Consumer travel agents have a long graveyard, so the interesting question is whether supplier partnerships plus an agentic booking layer closes that gap better than search did. β Voyage Labs Β· PR Newswire
π Voyage Labs
KD Agentic Β· AI Daily Digest
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β full credit and traffic to the original publisher.