AI Gets Physical

Markets were volatile this week as investors navigated renewed concerns around AI spending, rising oil prices and a more hawkish Federal Reserve. Equities declined modestly early in the week as semiconductor and AI-related shares came under pressure and the 10yr Treasury yield briefly moved >5%. On Wednesday, the Federal Reserve raised the federal funds rate by 25 bps to 3.75%–4.00%, the first increase since 2023, citing persistent inflation (“Fed Rate Decision September 2026: Rates Rise to 3.75%-4%,” CNBC, September 16, 2026). The announcement initially pressured both stocks and bonds, but markets rebounded sharply Thursday as oil prices and Treasury yields retreated. Through Thursday, the S&P 500/NDX were roughly flat for the week after gaining approximately 1% on Thursday. Despite the week’s volatility, resilient economic activity and continued AI-related capital spending have helped support equities, with the S&P 500/NDX still up approximately 12%/13% YTD (“S&P 500 Is Up More Than 12% This Year — And One Analyst Says the Run Isn’t Over,” Yahoo Finance, August 23, 2026). Our NPM Private Market Tracker, an average of the performance of the top 50 names by valuation, is +90% YTD; this is an internally calculated, non-public composite maintained for illustrative purposes only, is not an investable index or a product offered by NPM, and past performance is not indicative of future results. (NPM; Bloomberg; Reuters)

Physical AI Could Be a Major Growth Opportunity Following Generative AI

We have written a lot about humanoid robotics, but this week we wanted to zoom out to the broader world of “physical AI.” Physical AI is the application of AI to machines that perceive and act in the physical world. Simplistically, it is the intersection of foundation models and robotics. The category includes humanoid robots, but also extends to autonomous vehicles, drones, industrial robots, autonomous machinery and eventually consumer products. The common denominator is not the form factor; it is the common ability to perceive, make decisions and take physical action.

Whereas transformer technology was a watershed moment for generative AI, the critical technological developments for Physical AI have been the emergence of Vision Language Action (“VLA”) models and world models. Rather than simply predicting text or images, these systems predict actions and their consequences. Figure AI’s Helix, for example, is designed as a generalist humanoid VLA model controlling perception, movement and reasoning in real time. NVIDIA’s Isaac GR00T similarly acts as a pretrained robot brain that can take instructions and output a series of robotic movements.

In our view, the physical AI market should be approached as a multi-layer stack spanning compute, components, data, foundation models, software and physical platforms. Within that stack, however, there is a critical strategic divide between companies trying to own the “brain,” companies focused primarily on hardware, and vertically integrated companies trying to own the vertical stack.

Our central thesis is that, over time, the greatest economic value is likely to accrue to intelligence and data, rather than commodity hardware. Ultimately, the companies best positioned may be those combining a general model, proprietary data and a large installed hardware base; this is a general market observation and not a recommendation regarding any particular company or security. In our view they will have the greatest operating leverage, but also the highest upfront capital requirements and risk.

The market is still extremely early. And as there is no standardized definition of “physical AI,” there is therefore no reliable single market size number. Estimates, however, for the TAM of only the humanoid robotics market, a single slice of the pie, are in the trillions of dollars (“The $9 Trillion Opportunity in Humanoid Robotics,” RBC Capital Markets, March 2026). In our discussion below we somewhat arbitrarily divide the physical AI world into 3 key segments: (1) intelligence layer; (2) humanoids; and (3) autonomous vehicles.

Intelligence Layer = Strategic Battleground

In our view, the most important layer is the “brain”, the foundation model layer. Legacy robots have been programmed for defined tasks in defined environments. Physical AI models attempt to generalize by collecting visual data and generating physical actions without explicit programming for every scenario. If a single general model can control many robot embodiments, the addressable market expands from a single robot form to a potentially unlimited number.

Skild AI, based in San Mateo, is one of the clearest “brain” bets. The company raised $1.4bn in January 2026 at a reported valuation of ~$14bn (“Robotics Software Maker Skild AI Hits $14B Valuation,” TechCrunch, January 14, 2026). Its Skild Brain is designed to be form agnostic, including humanoids, quadrupeds, arms and mobile manipulators. Skild is therefore attempting to become a horizontal intelligence layer that can control essentially any machine that moves, rather than a robot manufacturer. The company has begun commercial deployment and reported on 9/10/26 that ARR was >$100mm (“Robotics Startup Skild AI Hits $100 Million in Recurring Revenue Run Rate As Customer List Grows,” Bloomberg, September 10, 2026).

Physical Intelligence (“PI”), headquartered in San Francisco, is pursuing a similarly pure foundation-model strategy. The company was reported to be in talks for a valuation of approximately $11bn in early 2026, with a subsequently reported valuation of $11.7bn (“Ex-DeepMind Staffers’ Robotics Startup in Talks for $11 Billion Valuation,” Bloomberg, March 27, 2026); NPM has not independently confirmed the final terms or closing date of this round. PI is developing general purpose VLA models designed to transfer intelligence across robot platforms.

Generalist AI, headquartered in San Francisco, sits between the pure brain and pure deployment models. Its approach emphasizes models that can generalize across robots and learn new tasks from short demonstrations, while commercial deployments provide a mechanism for generating supplemental physical world training data. In that regard, Generalist AI’s strategy is somewhat similar to Physical Intelligence’s. The company has raised $741mm in total, most recently at a $3bn post-money valuation in 8/26 (“Robotics Startup Generalist Reaches $3B Valuation, Sources Say,” TechCrunch, August 25, 2026).

Field AI, headquartered in Irvine, is an “API first” autonomy company. Field AI focuses on robots operating in unstructured environments. Its Core API is designed to let third parties incorporate autonomy capabilities into their own machines. This makes Field AI an important test of whether the autonomy layer can become a horizontal software business even without owning the robot. The company has raised over $400mm to date at a most recent valuation of $2bn (“Nvidia, Bill Gates-Backed Robotics Startup Field AI Hits $2 Billion Valuation After Recent Raise,” CNBC, August 20, 2025).

Outside of the startup world, we would also note that NVIDIA is attempting to participate across the physical AI stack. To us, NVDA is becoming for physical AI what TSLA is for humanoid robotics: a well-capitalized public-market participant. In June 2026, NVDA released its Cosmos 3 “world action model,” which displayed broad capabilities (“NVIDIA Launches Cosmos 3, the Open Frontier Foundation Model for Physical AI,” NVIDIA Newsroom, June 1, 2026). NVDA appears to be attempting to recreate in physical AI what its “CUDA + GPUs” has become for generative AI.

Humanoids: the Most Visible, but Not Necessarily the Largest, Segment

Humanoids naturally attract the most attention. A human shaped machine can theoretically operate most easily in environments already designed for humanoids. To us, the critical issue is whether humanoid hardware becomes commoditized or remains coupled to proprietary models. If models become sufficiently general, hardware manufacturers could capture less of the long-term economics. However, if each robot requires extensive model customization, vertically integrated companies may retain substantially more value. Key US humanoid players include 1x ($6bn reported valuation), Agility Robotics ($2.3bn), Apptronik ($5.3bn valuation) and Figure AI ($39bn).

Autonomous Vehicles, Drones and Defense Demonstrate a Different Path

Autonomous driving is arguably the most commercially mature form of physical AI. Major pureplays include Wayve ($8.5bn valuation via a 7/26 employee tender offer; “Wayve Launches $85M Employee Tender Offer At $8.5B Valuation,” TechFundingNews, July 1, 2026), Applied Intuition ($15bn at 6/25; “Applied Intuition Hits $15 Billion Valuation for AI Vehicle Tech,” Bloomberg, June 17, 2025) and Nuro ($6bn at 8/25; “Self-Driving Vehicle Startup Nuro Valued at $6 Billion in Late-Stage Funding Round,” Reuters, August 21, 2025). These companies demonstrate the potential for software to become a recurring revenue layer on top of physical assets.

Some might argue that drones and defense fall into this bucket as well, although this space is large and distinct enough that we will not cover it here. Drones and autonomous machinery broaden the opportunity further into inspection, agriculture, logistics, mining and construction.

Where Will the Economic Value Accrue?

We believe the foundation model layer could offer significant theoretical value because a successful general-purpose model could be deployed across millions of machines with very low incremental software cost. In our view, that is the logic behind the >$10bn valuations being assigned to companies such as Skild and Physical Intelligence. However, this is also the highest risk part of the stack. The companies must solve generalization, obtain sufficient physical world data and defend themselves against hyperscalers such as Nvidia.

The API, autonomy and orchestration layer may offer an attractive risk/reward profile relative to other segments of the stack, in our view. This layer handles the practical problems of turning intelligence into machine behavior: perception, safety, simulation, integration and adaptation to customer environments. It can potentially remain valuable even if the underlying foundation models become increasingly commoditized. Field AI is the clearest independent example of this strategy.

Vertical hardware/intelligence platforms have the advantages of both proprietary data and customer ownership. Physical AI cannot rely on an internet of real world text data like generative AI. Therefore, companies like Figure AI, for example, must use real world deployment to create data that improves their systems. This creates a feedback loop in which more machines generate more data, better data improves the models, better models increase machine utilization and improved economics support additional deployment. The vertical strategy, however, is enormously capital intensive.

In our view, pure hardware companies are less likely to capture “software like” economics unless they possess a meaningful proprietary moat. Sensors, actuators, batteries and motion control systems could all see substantial demand, but the risk of commoditization is clearly higher.

What are the Risks?

To us, the principal risk is that physical AI proves substantially harder than expected. A humanoid robot must contend with changing environments and errors can have serious consequences. A second risk is that models remain insufficiently general. If every deployment requires extensive customization, the economics begin to resemble systems integration rather than software. Third, valuations have run. The current marks for Figure, Skild, Physical Intelligence and others already embed expectations for an enormous TAM.

A Quick Note on Oil

WTI crude is currently at $102/bbl, +48% from a low of $69 in late June (“Crude Oil Price Today: September 11, 2026,” Forbes Advisor, September 16, 2026). Oil clearly remains volatile and a key driver of sentiment in both public and private markets.

In early July the market had begun to reprice a normalization of Gulf exports.  Some traffic had resumed through the Strait of Hormuz, and other barrels were finding their way to market in ways we cannot totally explain.  People were also optimistic that the US would find an offramp from the conflict, and oil fell roughly to pre-war levels.

Behind the scenes, and surprisingly, China also cushioned the market.  Pre-war, China imported about 11mm b/d of crude, and they are importing about 7mm b/d now.  Essentially China has removed about 4mm b/d of demand from the seaborne crude market.  The assumption is they are drawing down their 1.2bn barrel SPR.

The downward pressure on oil price was reversed on 9/11 when Saudi Arabia announced its East-West oil pipeline, a 750 mile pipe from Saudi oil fields to the Red Sea, had been damaged in a drone attack launched from southeastern Iraq, which regional officials have linked to Iran-backed militias (“Saudi Arabia Shuts Critical Oil Pipeline After Drone Attack: What It Means,” Al Jazeera, September 12, 2026). Estimates of how long the pipeline will remain offline vary by source, from several days to several weeks, and it was a key outlet for ~4-5mm b/d of oil production (4-5% of global supply). This attack elevated perceived “tail risk” because previously this pipe was viewed as a safe alternative to the Strait of Hormuz. This attack brought crude again to >$100/bbl.

Meanwhile on the other side of Saudi Arabia, Houthis are an increasing threat. On 9/8 they attacked several energy facilities in the southern part of the country (“Saudi Arabia Halts Energy Facilities After Attacks Near Yemen Border,” Bloomberg, September 8, 2026). Houthis are further threatening to cut off access for Saudi Arabian crude through the Red Sea. As a result, Saudi Arabia’s offtake redundancy is at risk of disappearing.

In our view, the oil market is back to the April/May mindset where several million b/d of production is disrupted with no clear solution.  A credible ceasefire could easily get prices back down to normalized levels.  However, it feels like the risk of additional disruption is going up, not down.  Also, be aware that China has been a shock absorber here, and eventually they will need to restock inventories.

This communication is prepared by Nasdaq Private Market Securities LLC (“NPM Securities”), a broker-dealer registered with the SEC and a member of FINRA/SIPC, for informational purposes only. It does not constitute an offer to sell, or a solicitation of an offer to buy, any security, and is not a research report, investment recommendation, or investment, legal, tax or accounting advice. It does not take into account the investment objectives, financial situation or needs of any particular recipient, who should consult their own advisors before making any investment decision.

Statements of opinion herein (including those introduced by phrases such as “we believe,” “in our view,” or “our thesis”) reflect the current views of NPM Securities as of the date of publication, are subject to change without notice, and should not be construed as statements of fact, guarantees, or assurances of future performance. Estimates, valuations and forward-looking statements, including with respect to privately held companies and market size, involve significant assumptions, risks and uncertainties, and actual outcomes may differ materially. Information regarding privately held companies — including funding amounts, valuations and commercial metrics — is drawn from third-party media reports and has not been independently verified by NPM Securities unless otherwise indicated.

Past performance, including the performance of the NPM Private Market Tracker referenced herein, is not indicative of future results. The NPM Private Market Tracker is an internally calculated, non-public composite maintained for illustrative purposes only; it is not an investable index or product and does not reflect fees, expenses, or the performance of any NPM-managed account. Market and pricing data are believed to be reliable but are not guaranteed as to accuracy or completeness.

This material is intended solely for the recipient to whom NPM Securities provided it and may not be forwarded, reproduced or redistributed, in whole or in part, without NPM Securities’ prior written consent.