NVDANVIDIANeutralNVIDIAResearch|Physical AI: CoRL 2026 Field Notes — The Brain Moves to Frontier Models; Competition Shifts to Action Models, Data and Deployment
This report draws on CoRL presentations, exhibit-floor observations and publicly available research. CoRL 2026, a leading academic conference in robot learning, was held in Austin, Texas, from November 9 to 12. FUNDA attended throughout, including talks by NVIDIA and Google DeepMind and live demonstrations at several company booths. This report draws on CoRL presentations, exhibit-floor observations and publicly available research. Executive Summary 1. The robot’s brain is increasingly being supplied by general-purpose frontier models, while competition within robotics is shifting toward action models, data and real-world deployment. By frontier models, we mean general-purpose models such as GPT, Gemini and Claude. The brain interprets instructions and plans steps; the action model generates motion in real time. Improvements in brain capabilities are coming mainly from frontier-model companies, leaving robotics companies with limited scope to build a durable advantage at that layer. They still face three questions. The action model must be fast and inexpensive, but NVIDIA plans to make its own action model available for free, limiting the scope for standalone monetization. 2. First-person video is now inexpensive to obtain; tactile data and on-site data that can only be accumulated through real-world deployment remain scarce. Robots have entered factories but still require human intervention. We maintain our view that humanoid robots will reach industrial scale around 2030. Compute demand should grow before robot shipments ramp, driven mainly by frontier-model inference and simulation data generation. The brain: moving to frontier models Asking a robot to put a cup from the table into the dishwasher requires two kinds of work. The first is thinking: understanding the instruction, recognizing the objects on the table, breaking the task into steps such as picking up the cup, opening the dishwasher and placing the cup inside, then adjusting the plan if something goes wrong. The second is moving: turning each step into specific hand and arm movements, potentially adjusted dozens or even hundreds of times per second. The industry generally calls the first capability the brain and the second the action model. Robotics companies used to train both themselves; increasingly, the brain is being supplied by frontier models. Although these models were not developed specifically for robots, their visual understanding, reasoning and code-generation capabilities can already be applied directly to robotics. They outperform many in-house robot-brain models and continue to improve rapidly. At the conference, NVIDIA introduced GR00T 2, an action model designed to work with a frontier model that breaks tasks down, monitors execution and diagnoses failures. Jim Fan, one of NVIDIA’s robotics research leads, said complex tasks cannot be handled by OpenAI’s Astra alone or NVIDIA’s own GR00T alone; the two need to work together. This reflects the industry’s growing acceptance of frontier models for task understanding and planning. On NVIDIA’s figures, the success rate on the same set of tasks rose from 40% without a frontier model to 92% with one (vendor-reported). At Google’s booth, Apptronik’s humanoid robot Apollo took instructions directly from visitors. A screen beside it showed how the robot interpreted each instruction and the action it planned to take. One visitor asked whether it could high-five itself. The screen first showed the robot preparing to clap, then displayed its reply: “I can definitely give myself a high five, which basically looks like a clap.” The frontier model generated the response on the spot. It was one of the more engaging live demonstrations we observed at the conference. Figure: Google’s booth, showing Apollo and the live conversation and action log on screen Source: FUNDA on-site photographs OpenAI and Anthropic had no product demonstrations or talks at the conference, yet many research projects used their models. We saw no booth demonstrations or presentations from either company. OpenAI was a conference sponsor; according to Fortune, the booth registered in its name stood empty during the event. Anthropic was not a sponsor. Meanwhile, the workshop on controlling robots with frontier models was among the conference’s most closely watched sessions, with 39 accepted papers. Research from NVIDIA, Berkeley, ByteDance and others used Astra or Claude as the robot’s brain or as a research tool. Progress in the robot’s brain is increasingly coming from frontier-model companies outside the robotics industry. Subscribe now Read more
View original →Neutral
NVDANVIDIANVIDIA
MSFTMicrosoftMicrosoft Deep|ANT: A History-Making IPO and Shifts in Customer Demand and Competition
Four years have passed since ChatGPT launched and set the AI narrative in motion. Now, a week before that anniversary, the narrative is approaching its biggest moment yet: Anthropic, the fastest company in history to reach $50 billion in revenue, is preparing to go public. A year ago, the AI narrative looked much as it does today, perhaps even worse. For much of last year, both spot and contract prices for H100 and A100 GPUs fell month after month. Reports of Microsoft cutting orders and canceling data centers were widespread. OpenAI was still the market’s main focus. Readers may remember the series of reports we published during those months: Then Claude Code came to the rescue: The standard for being AI-native shifted from AI writing 50% of the code to writing all of it. Engineers went from running one agent to five, then dozens. A typical engineer’s agents went from running for tens of minutes to several hours, even replicating features and software they found online. The effects compounded, quickly pushing Anthropic’s monthly growth to 50%. This changed how investors thought about AI ARPU. Monthly spending per person did not have to stop at the price of a ChatGPT subscription. It also changed how they viewed data center constraints. Revenue generated per gigawatt of capacity could rise from $10 billion to $20 billion, or even $50 billion. Instead of extrapolating AI annual recurring revenue (ARR) in a straight line, the market began estimating it from future compute capacity. Assuming revenue per gigawatt stayed constant, the split between training and inference capacity would be enough to derive an ARR trajectory. We regard Anthropic as one of the great companies of this century. It has changed our understanding of AI, accelerated AI infrastructure development, and helped give rise to a cohort of data-labeling companies. It also takes AI’s consequences for humanity seriously, devoting substantial time to helping society understand AI safety even when doing so could hurt its own revenue. We believe this is the right approach and may ultimately help avert an AI crisis. The picture has become more complicated as competition between OpenAI and Anthropic intensifies and their combined ARR growth slows. Anthropic’s S-1 also showed that inference’s share of total costs began falling sharply from 26Q2, with training reaching 75% of total costs that quarter. That makes it harder to infer ARR from data center capacity, or to assess the ROI of the AI narrative as a whole. We have returned to this issue repeatedly: the closer we get to AGI, the larger training’s share of costs becomes, and the harder it is to measure ROI. It is like the final sprint of a 100-meter race. Ahead of Anthropic’s IPO, we spent considerable time on field research to better understand the AI ROI question. We set out to answer the questions below. Awkwardly, given the timing, we still do not have complete answers; further company disclosures may be needed to resolve them. If we divide Anthropic’s customers into startups, technology companies, and traditional enterprises, how is demand developing in each group, and where does it go from here? How do KV caching and model routing affect ARR? How does today’s token budgeting differ from the SaaS optimization cycle of 2022-2023? How does competition from OpenAI and open-weight models affect ARR? Which use cases create meaningful switching costs? Beyond today’s applications, what new use cases could contribute to ARR in 2027? How might recent developments at NVIDIA affect Anthropic, and which findings from our research might they challenge? Subscribe now Read more
View original →
NVDANVIDIANeutralNVIDIAReview|NVDA FY27Q3: FY2028 Growth Guidance Raised; Jensen Huang’s Potential Return Remains Key
Since our Preview report, NVIDIA has seen several important developments. Jensen Huang resigned, and the Board stated at the start of the earnings call that it does not intend to reinstate him as CEO. The Board has begun a formal CEO search covering internal and external candidates and established a National Security and Export Control Committee. Near-term results exceeded expectations: FY27Q3 revenue was $115.8b (+20.4% QoQ), and FY27Q4 revenue guidance was $130b ± 3%. FY2028 revenue growth guidance was also raised from 70% to approximately 75%. Kress took the unusual step of reassuring employees during the call and acknowledged internal disagreements, highlighting the challenge of retaining key talent. We continue to believe that whether Jensen Huang returns is an important factor in NVIDIA’s future competitiveness. Without Founder Mode, NVIDIA may find it harder over the long term to sustain the competitiveness of its product roadmap and its supply-chain ecosystem integration advantages. Earnings Highlights Data Center Momentum and Gross Margin Trough Confirmed: FY27Q3 Data Center revenue was $108b (+21% QoQ), including Hyperscale revenue of $60b (+23% QoQ); cloud providers’ backlog rose $500b QoQ to $2.5t. FY27Q4 guidance includes an additional week, while China-specific Hopper revenue is expected to fall to zero (from $2.4b in FY27Q3). FY27Q3 Non-GAAP gross margin was 74.9%. FY27Q4 gross margin guidance fell to 71%–72% due to costs associated with fulfilling legacy orders (Non-GAAP OpEx of approximately $10.7b). Gross margin is expected to recover to 72%–73% in FY2028 as new pricing takes effect. Product Progress and Rising GPU Rental Prices: Vera Rubin contributed nearly $20b in revenue in its first quarter of production shipments, with inference output per watt reaching 10x that of GB300. CPO switches are in volume production, while Rubin Ultra (2027) and Feynman (2028) remain on schedule. Since May, Token prices have fallen nearly 60%, driving a 35x YoY increase in OpenRouter’s weekly usage (Agentic workloads account for approximately 80%). This increase has lifted compute utilization and pushed GPU rental prices higher. Multi-year renewal prices for Hopper GPUs have reached up to 2x their original level, while B300 rental prices have risen more than 60% since launch. Agentic demand is also expected to more than double FY2028 CPU revenue. Share Repurchases Total shareholder returns in FY27Q3 reached $28b ($22b in share repurchases and $6b in dividends), with a YTD FCF return ratio of 65%. The Board approved a record $150b increase in buyback authorization this quarter. Remaining authorization at quarter-end was $227b, which management plans to use in full by the end of FY2028. Starting in FY27Q4, 100% of excess FCF, after strategic uses, will be returned to shareholders. Financial Forecasts Subscribe now Read more
View original →
NVDANVIDIABearishNVIDIA
AVGOBroadcomNeutralBroadcomDeep|LLM: RSI and Chip Design: AI Lowers the Barrier to Custom ASICs, Redistributing Profit Across Accelerators, EDA and Design Services
Accelerators are the most profitable, most squeezable layer; custom ASICs redistribute NVDA's ~$40B/GW
View original →Neutral
NVDANVIDIANVIDIA
AMATApplied Materials, Inc.Applied Materials, Inc.
MSFTMicrosoftMicrosoft
PANWPalo Alto NetworksPalo Alto Networks Weekly|The Rebound and What Comes Next, NVDA Financing, Cybersecurity, CPO, IFX & AMAT
Last week, we argued that the medium-term questions facing AI could coexist with a near-term rebound after the sharp selloff. In the first half of the week we saw a sharp rebound, which was later interrupted. We would keep the same distinction between time horizons: a recovery in share prices does little to resolve the gap between ambitious infrastructure spending and the revenue needed to support it. Jensen Huang’s departure adds another reason to look more carefully at how that spending gets financed. As we discuss in this week’s NVDA deep dive, Huang’s role extended well beyond selling GPUs. NVIDIA’s equity investments, capacity backstops and credit support helped Neolabs and Neoclouds raise capital and commit to expansion. A more conservative approach under new leadership could make the medium-term constraints we discussed last week more visible. The first changes may show up in new investments and financing commitments over the next one or two quarters, with consequences for deployment and hardware demand emerging later. Strong near-term GPU demand can coexist with a less supportive funding environment. We think the market needs to scrutinize who is financing the next round of capacity, on what terms, and how much spending depends on continued support from NVIDIA. This week also brought more evidence of the security and operational risks surrounding increasingly connected software. Fireworks’ security incident exposed data from some Notion workspaces, illustrating how an AI infrastructure vendor’s breach can reach customers of the applications built on it. Separately, the prolonged Microsoft 365 outage disrupted essential business services. Microsoft attributed the outage to a network upgrade issue exposed by a subsequent change, and said it was unrelated to cyberattacks or AI/agents. It does, however, underline how much business activity depends on a small number of shared services. These developments reinforce our focus on the cost of securing and maintaining the systems that AI increasingly relies on. As AI expands software development and accelerates vulnerability discovery, enterprises face more work in remediation, identity management, third-party oversight and continuous protection. We see a firmer basis for sustained cybersecurity demand, even while the funding and returns on parts of AI infrastructure remain uncertain. The timing still matters: our latest checks point to security-spending growth among covered customers rising from 10–12% in 2026 to 13–15% in 2027, with much of the immediate response funded through existing budgets. Our PANW preview provides an early read on how that demand is reaching established security platforms. This Week’s Reports NVDA: Near-term demand meets a more uncertain financing backdrop. Our earnings preview examines the B300 ramp and Rubin shipment timing, while our new deep dive considers how Huang’s departure could change NVIDIA’s support for Neolabs, Neoclouds and its supply chain. CPO: Bonding and testing still limit the ramp. Our equipment and process checks point to large-scale shipments in late 2027–2028, with hybrid-bonding yields and an unresolved test-vendor decision keeping near-term volume expectations in check. Cybersecurity & PANW: The budget response is taking shape. Our latest cybersecurity checks point to faster spending growth in 2027, while the PANW preview examines solid customer spending and continued consolidation onto fewer security platforms. Premium Report Snapshot A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed. Subscribe now Read more
View original →