Physical AI: The Brain and the Body Have Become Their Own Markets

Disclaimer: This article is for educational purposes only and is not investment advice. Trading and investing are subject to volatility and market risk, including partial or entire loss of capital. Please consult your financial advisor before making investment decisions. Financial Gurkha and its writers are not liable for your losses, nor do we take credit for your gains. FinancialGurkha.com and its authors assume no liability for financial outcomes. The author does not hold any position in any of the assets mentioned in the article.

AI disclosure This article was researched and drafted with the help of Muse, an AI model made by Meta. All figures were checked against the cited sources, and the analysis and conclusions are Financial Gurkha's. Financial Gurkha ensures the article is objective and factual.

Financial Gurkha | New York City | October 4, 2026

"The brain is indeed the weakest link in the complete robotics stack."

That sentence, from Spirit AI co-founder and chief scientist Gao Yang to Reuters in September, is the cleanest summary we have seen of where the robotics industry stands — and where the value is going. Chinese humanoids can already sprint, dance, and do backflips on command. The hardware, increasingly, works. What does not work yet is the software that tells the hardware what to do in a messy, unstructured world. Gao expects a breakthrough comparable to OpenAI's GPT-3 — the model that powered ChatGPT — around mid-2027.

This is the first installment of a new Financial Gurkha research section on Physical AI: the application of artificial intelligence to the physical world through robots, machines, and embedded systems. Our organizing thesis is simple, and it comes in two parts:

1. The body and the brain have become their own markets. Robotic intelligence (foundation models for robots) and robotic hardware are now separable, independently investable layers — with different economics, different leaders, and different clocks.
2. Value is accruing to the enabling layer of the stack. As with every computing platform shift before it, the scarcest inputs — not the most visible products — capture the durable economics.

What follows is the evidence for both claims, the bottlenecks that will decide the pace of deployment, and the proof that Physical AI already earns its keep in at least one industry.

The brain: foundation models for robots#

The most important architectural idea in robotics right now is borrowed directly from large language models: a general-purpose action model that transfers across multiple robot bodies, instead of software optimized around a single machine.

This is the bet that the best-capitalized private robotics companies are making, and private markets are pricing it aggressively:

• Physical Intelligence (San Francisco) is reportedly raising $1 billion at an $11 billion valuation, doubling its $5.6 billion valuation from 2025. Investors in prior rounds include Jeff Bezos, OpenAI, Thrive Capital, Lux Capital, Sequoia, and CapitalG. The company builds foundation models for robots and no hardware of its own.
• Skild AI is valued at $14 billion or more — software only, no robot — according to press-reported valuation rankings from August 2026.
• Spirit AI (Beijing), founded in 2024, has raised over $670 million and is valued at 20 billion yuan (about $2.9 billion), Reuters reported in September. Its approach is instructive: rather than training in simulation like many competitors, Spirit AI employs roughly 1,000 contractors wearing motion-capture equipment in homes and on production lines, collecting real-world movement data. The company found that "dirty data" — varied, imperfect human motions — trains models faster than repeating a single "clean" movement fifty times. Tens of its Moz1 wheeled humanoids are already deployed on production lines at battery maker CATL and at JD.com, which is also an investor.

Note what the valuations are rewarding: not robots shipped, but brains built. The company shipping 44% of the world's humanoids is worth a fraction of the software-only firms still running pilots. Markets are pricing the intelligence layer as the scarce asset — which is exactly what the "weakest link" framing predicts.

The demand roof: why the largest companies cannot opt out#

None of this happens without demand, and the demand structure for AI has a distinctive shape: frontier versus open source, plus a compute arms race.

AI capability at the frontier has become a strategic competitive input for the largest companies in the world. Staying on top of it is not an R&D luxury; it is treated as a strategic imperative — which is why capital expenditure on AI infrastructure keeps printing records even as individual quarters get debated. The dynamic is self-reinforcing: open-source models diffuse capability downward, which forces frontier labs to push further, which demands more compute, which funds the infrastructure buildout.

For Physical AI specifically, this matters because embodied models are downstream of the same scaling inputs — data, compute, talent — and will inherit the same competitive intensity. The robot brain does not get built in a calmer market than the chatbot brain. It gets built in the same furnace.

The bottlenecks: power, thermal, infrastructure#

Here is the part of the Physical AI story that gets the least attention and may matter the most for timing: the constraints on deployment are physical, not algorithmic.

Two scaling vectors are colliding:

• Deployment scale is rising. Tens of robots on production lines today; thousands and eventually millions across factories, warehouses, and commercial settings if the technology works.
• Model size is rising. More capable embodied models need more onboard and edge compute per unit.

The product of those two vectors lands on three bottlenecks: power delivery, thermal management, and electrical infrastructure. A robot that cannot be powered, cooled, and charged inside existing facilities cannot deploy at scale no matter how good its brain is. As models get bigger and fleets get larger, power delivery, thermal management, and electric infrastructure stop being facilities trivia and become the binding constraint on the industry's growth rate.

This is worth internalizing because it inverts the usual way technology stories are told. The exciting layer (the brain) gets the valuations; the binding layer (power and heat) decides the timeline. Investors who understand both can think clearly about sequencing — which problems get solved first, and which companies sit on the critical path.

The enabling layer: content density is rising#

Which brings us to the body — or more precisely, to what goes inside the body. Every robot, every intelligent edge device, is a bundle of sensing, signal processing, connectivity, and position-sensing content. And content per system is rising: each generation of machine carries more sensors, more compute, more connectivity than the last. Content density — the dollar value of electronics inside each unit — is the quiet compounder of the whole Physical AI buildout.

This is the "enabling layer of the stack" from our thesis, and it is where established industrial technology companies live. Consider the shape of the opportunity through one example:

Jabil (NYSE: JBL) has made its Intelligent Infrastructure segment the growth engine of the company — roughly $5.8 billion in revenue with expanding margins, tied directly to hyperscaler AI demand. The company is investing $500 million in U.S. manufacturing for cloud and AI infrastructure, acquired Mikros Technologies for liquid-cooling capability (thermal management — the bottleneck above), and announced a collaboration with Endeavour Energy targeting up to 2 gigawatts per year of modular AI infrastructure capacity. Jabil's story in one line: the AI buildout needs someone to physically engineer and manufacture it, and content per system keeps rising in that someone's favor.

The general principle extends across the bill of materials: sensors that see, chips that process signals, radios that connect, power systems that deliver, thermal systems that cool. None of these are the robot. All of them are in every robot.

Proof it already works: the farm#

Skepticism about humanoid timelines is healthy. So it is worth noting that Physical AI already earns its keep — at scale, today — in at least one industry: agriculture.

John Deere's See & Spray technology uses computer vision on sprayer booms to identify weeds in real time and apply herbicide only where needed, instead of broadcasting it across entire fields. In September 2026, Deere's agriculture division president said customers are seeing herbicide savings of more than 50 percent. The company's own product materials claim 77% average savings for See & Spray Select on fallow ground, and across 10 million acres treated in Australia, growers averaged 74% savings, Deere reported in August 2026.

This is the template to watch: cameras plus models plus actuators, deployed on machines that already exist, producing a measured ROI in input costs. No humanoid required. The farm is where Physical AI's economics are already proven — the factory and the warehouse are next in line, and the home comes last, exactly in the order Spirit AI's Gao described: industrial applications in the next one to two years, commercial services after that, homes "far harder than both."

What we are watching#

A few open questions will decide how this section develops:

1. The GPT-3 moment. Gao's mid-2027 call for a robot-brain breakthrough is the industry's most specific public prediction. If it lands, the deployment timeline compresses; if it slips, the hardware lead extends and the brain-layer valuations get tested.
2. The data bottleneck. Spirit AI's bet on real-world "dirty data" over simulation is a genuine methodological fork. Which approach scales to general-purpose reliability first matters enormously — data collection at that scale is its own industrial operation.
3. China versus the West. The pattern so far: America raises the capital, China ships the units. Whether that divergence persists or converges will shape every competitive assumption in the space.
4. The bottleneck trades. Power delivery, thermal management, and electrical infrastructure are the least glamorous and possibly most timing-sensitive part of the thesis. We will be tracking who owns them.

We will continue building this section with company-level research in the same format as the rest of Financial Gurkha's coverage: primary sources, shown arithmetic, stated assumptions — and no recommendations.

This is the first installment of Financial Gurkha's Physical AI research section. Corrections or additions: contact@kanchanksharma.com.