The Physical World is the Next Frontier for AI — and Nobody’s Talking About It Enough

The Setup: Why Now

The digital economy got its AI upgrade first. Now the physical world is next — and the numbers are staggering.

The macro pressures are real and converging:

Aging populations in every major economy (Japan, Germany, US, China) are shrinking the workforce — particularly in physically demanding jobs
US manufacturing wages hit $34/hr in 2025 and are rising. Robot costs are falling
$1.2 trillion in US reshoring investments announced in 2025 alone — but there aren’t enough workers to build it
Climate infrastructure (wind farms, grid upgrades, data centers) needs to be built fast — with an industry that already can’t find enough skilled operators
Geopolitical pressure is forcing nations to rebuild domestic manufacturing capacity regardless of cost
Construction alone needs 454,000+ additional workers in 2025 — and the pipeline isn’t coming

The bottleneck isn’t capital. It’s human bodies willing to do hard, dangerous, repetitive physical work.

The Prize: Physical GDP

Here’s the number that should stop you cold:

More than half of the $30 trillion US economy is directly tied to physical human labor.

Breaking it down:

Industry (manufacturing, construction, mining, utilities): ~20% of US GDP
Transportation & logistics: ~6%
Healthcare with physical labor (nursing, elder care, surgery): ~8%
Agriculture, extraction: ~2%
Services that require physical presence (hospitality, security, maintenance): ~10%+

Global industry share is ~26% of a $110T world economy — roughly $28 trillion in industries that, until now, have resisted meaningful automation. This is the territory that software AI barely touches.

Compare that to knowledge work — the main target of LLMs — which represents the other half of GDP but has already seen 20+ years of software-driven productivity gains. Physical work is where the productivity deficit is deepest.

The Technology Unlock

For decades, “Moravec’s Paradox” held: it was easier to build a machine that plays chess than one that folds laundry. That’s now breaking.

Three simultaneous shifts are converging:

Vision-Language-Action (VLA) models — AI that sees a scene, understands a spoken command, and outputs physical movements. Think of it as giving robots a brain that was pre-trained on the entire internet. Key models: Physical Intelligence’s π0, Figure AI’s Helix, Google’s Gemini Robotics, NVIDIA’s GR00T N1.6 (open source), Hugging Face’s SmolVLA (runs on a laptop)
World models — AI that simulates physics. Instead of needing millions of real-world robot hours, you can train in simulation. NVIDIA Cosmos, World Labs’ Marble, and Waymo’s internal models are early leaders. This is the data bottleneck-breaker
Cheaper, better hardware — Humanoid robot material costs projected to fall from ~$35K to $13–17K per unit by ~2030. Cobots (human-collaborative robots) are growing 27.5% CAGR and expected to hit $7B+ by 2030. Industrial robot installations hit 542,000 in 2024 — double the rate a decade ago

Where the Money Is Going

VC funding in robotics:

2024: ~$15B globally in robotics/physical AI
2025: $34B+ — more than double year-over-year
Pace continues accelerating into 2026

Big rounds that define the narrative:

Physical Intelligence (foundation model lab): $1B total raised, $600M Series B led by Alphabet’s CapitalG
Figure AI (humanoid): valuation approaching $39.5B; partners include BMW, NVIDIA, Microsoft
Skild AI: raised at multi-billion valuation, building a “general brain” for robots

But here’s the catch: Most of the headline companies are still pre-revenue at scale. High valuations reflect the size of the opportunity, not current cash flows. The funding is concentrated in humanoids and foundation model labs — arguably the riskiest, longest-timeline bets.

The Revenue Reality Check

Who’s actually making money today in physical AI?

Intuitive Surgical (da Vinci surgical robot): ~$6B annual revenue. 50% of surgeons now perform some robotic surgery, up from 9% in 2012. The floor-to-commercial journey took 20+ years
FANUC, ABB, Yaskawa (legacy industrial robots): Billions in annual revenue, now integrating AI
Amazon’s robotics fleet: 1 million+ robots deployed internally — not sold, but a massive internal cost center proving the model
Serve Robotics, Richtech: Small but real revenue in last-mile delivery and hospitality

Most physical AI startups in 2025–2026 are cost centers, not profit centers. This is normal for the phase we’re in — but it means time horizons matter enormously for where you place bets.

The Underfunded, Underhyped Zones

The hype clusters around humanoids and foundation models. The opportunity may be elsewhere:

Construction automation — $13 trillion global industry, 0% productivity growth over 50 years, massive skilled labor shortage, high injury rates. Investment is tiny relative to size. Gravis Robotics ($23M) is turning excavators into autonomous machines. Bedrock Robotics is building autonomous construction equipment. The autonomous construction equipment market is only $8.8B today but growing fast

Why it’s underfunded: Messy, unstructured outdoor environments; long sales cycles; conservative buyers

Agriculture robotics — Herbicide use, harvest labor, crop monitoring. Carbon Robotics uses laser-equipped robots to eliminate weeds without chemicals. Sunday AI does AI-driven lawn care. The talent that usually goes to this space goes to sexier categories

Why it’s underfunded: Seasonal demand, thin margins, fragmented customers

Elder care and home assistance — Japan and Germany are building national strategies around this. The US is not. By 2030, the 65+ population in the US will reach 73 million. Humanoid robots for rehabilitation and mobility support are in early pilots

Why it’s underfunded: Regulatory uncertainty, emotional sensitivity, harder to monetize than B2B

Simulation and testing infrastructure — Every robot company needs to validate systems safely before real-world deployment. This is the “picks and shovels” play. Zeromatter (simulation infrastructure), Nominal and Flow Engineering (hardware observability and testing) are early — and mostly missed by generalist VCs

Why it’s exciting: Sells to every robotics company; no single hardware bet needed

Mining and energy field automation — High injury, high cost, remote sites. Autonomous trucks already in Australian mines. Huge demand for inspection drones, autonomous drilling, pipeline monitoring. Relatively little startup activity vs. the market size

Small Bets Already Showing Promise

These are replicable proof points — companies doing specific hard things with real customer traction, not billion-dollar moonshots:

Company	What they do	Why it works
Gravis Robotics	AI-controlled construction machines	Operator shortage is acute; retrofits existing equipment
Carbon Robotics	Laser weeder for farms	Replaces herbicides; clear ROI for farmers
Serve Robotics	Sidewalk delivery robots	Addresses last-mile labor costs; already operating in LA
Zeromatter	Simulation infra for autonomy	Infra layer; needed by every hardware company
Dexory	Warehouse inventory scanning robots	Real revenue; solves known pain point

The pattern: narrow problem, clear ROI, existing customer with the pain.

Key Opportunities — Ranked by Bang for Buck

🥇 Highest conviction, near-term

Physical AI infrastructure (simulation, data pipelines, robot observability tools)

Why: Every robot company needs this. No hardware risk. Scales like software
Experiment: Build a data collection or labeling tool specifically for robot training data; sell to the 50+ funded robotics companies as customers

AI-augmented cobots for SMB manufacturers

Why: Reshoring creates demand; cobots are proven; SMBs can’t afford custom integration
Experiment: Build the “plug-and-play” integration layer that lets a small shop deploy a cobot without a robotics engineer

🥈 Medium conviction, 3–5 year horizon

Construction site automation (specific tasks, not whole sites)

Rebar tying, concrete inspection, bricklaying, surveying — all can be done now
Experiment: Partner with one contractor; deploy a single-task robot; measure time savings

Precision agriculture robots

Focus on harvest labor or targeted spraying; clear ROI vs. $18/hr seasonal workers
Experiment: License an open VLA model (SmolVLA/pi0), fine-tune on a specific crop task, partner with one farm co-op

🥉 Longer horizon, high upside

Elder care robots — Huge TAM, underserved, geopolitical tailwind; but regulatory and adoption cycles are long

Autonomous mining/energy field robots — High willingness to pay, but entrenched incumbents and complex deployment

The Bottom Line

Is it worth investing time and money here? Yes — but the where matters enormously.

Foundation model labs and humanoid companies are absorbing billions of dollars and years of time. The winners will be massive. But most won’t win, and timelines are uncertain
The asymmetric bets are in the infrastructure and application layers, not the foundation models themselves
The physical world’s “ChatGPT moment” hasn’t happened yet — but the components are assembled. The question is what triggers mass commercial deployment
The analog: in 2010, everyone wanted to build the smartphone OS. The money was made building apps and selling picks-and-shovels (Stripe, Twilio, AWS)
In physical AI: the picks-and-shovels are undersupplied, and the narrow application bets in construction, agriculture, and elder care are undercapitalized relative to their TAM

The GDP that physical AI can unlock dwarfs what software AI touches. The builders who show up early — with small, focused experiments in unsexy but critical industries — will define the next industrial era.

Sources: Generation Investment Management Roadmap Series (March 2026), Bessemer Venture Partners Physical AI Atlas, ABI Research Global Robotics Market Outlook, Stanford HAI 2025 AI Index, Crunchbase robotics funding data, IMF Global AI Impact Working Paper 2025, Deloitte Tech Trends 2026, Goldman Sachs AI research, Gravis Robotics / Fortune (Nov 2025), PitchBook via Generation IM.

Physical World as the Next AI Frontier