NewPhysical AI: When Software Meets Steel post

Your clients don’t care how many databases you searched. They care about credible appearances, visible momentum, and results they can show leadership.

PodPitch finds the right shows for every client, studies what each host cares about, writes personalized outreach, sends it from your team’s inbox, and follows up until interviews land.

PR and communications teams use PodPitch to create more authority, more client content, stronger search visibility, and better renewals without adding researchers or coordinators.

Only 12 PR team demo spots are available this month. When they’re gone, registration closes.

Book more interviews. Prove your value. Give every client another powerful reason to renew.

Appear on 3,853,234 Podcasts

The conventional wisdom says software eats the world. That idea is now eating itself.

The next wave of AI is not running on your laptop. It is running on robot arms in Shenzhen factories, inside waveguide lenses in a pair of glasses, and through the neural networks of autonomous systems navigating warehouse floors in Jakarta and Stuttgart.

This is physical AI, and the capital flowing into it in 2026 is not incremental. It is structural. By the end of H1 2026, robotics and physical AI startups had raised $55.8 billion globally, nearly double the prior full-year record. This is not hype. It is infrastructure.

What follows is a map of where that capital is going, why one Chinese glasses startup just became a unicorn on the back of privacy hardware, how Southeast Asian factory floors are being rewired in real time, and three under-the-radar companies you should be watching before everyone else is.

Most people still think of AI as a software problem. You train a model. You serve it through an API. You measure outputs in tokens. That framing misses what is actually being built right now.

Physical AI refers to systems where the model is fused directly with sensors, actuators, and real-world feedback loops. The AI does not just predict. It perceives the environment through cameras, LiDAR, and tactile sensors. It makes decisions. It acts through motors, grippers, and mobility systems. And crucially, it learns from what actually happens in the physical world, not just from data scraped from the internet.

NVIDIA described it precisely at GTC 2026: physical AI is models that understand the real world, reason about it, and plan actions within it. The company then launched its Physical AI Data Factory Blueprint at the same event, partnering with ABB, FANUC, and KUKA to solve the real bottleneck in robotics development, which is not algorithms but training data.

The gap between language AI and physical AI is this: a language model fails silently. A physical AI system that fails can drop a car part on a factory floor, crash a delivery drone into a pedestrian, or damage expensive manufacturing equipment. That raises the stakes for everything: the hardware, the sensors, the simulation layer, the safety cases.

It also raises the returns. Because when physical AI works, it compounds. A warehouse robot that handles 800 SKUs today can handle 5,000 next quarter, without hiring a single human.

The numbers from 2026 require a second read.

Global robotics and physical AI VC funding reached $27.6 billion in 2025, already more than double 2024's $13.8 billion. Then 2026 arrived. By mid-year, the sector had already surpassed $55.8 billion, nearly double the prior full-year record.

The concentration is notable. The top 10 physical AI startups have collectively raised approximately $11 billion. Robot foundation models alone attracted more than $2.2 billion in 2025, led by Physical Intelligence's $400 million raise. That same company is now in discussions for a further $1 billion round at an $11 billion valuation, despite having no commercial product or revenue. Skild AI raised close to $1.4 billion in January 2026 at a valuation above $14 billion.

The pattern investors are following: fund the brain layer first. Robot foundation models, simulation infrastructure, and perception software attract the largest bets because they are the reusable components that travel across every physical AI application. The hardware companies are next in line.

On July 6, 2026, Even Realities, a three-year-old Shenzhen startup founded by ex-Apple engineers, closed a $150 million pre-Series B led by Meituan and Tencent. The round valued the company at $1 billion.

On the surface, this looks like a China tech giant backing a local hardware startup. Look closer and it reveals something more interesting.

CEO Will Wang spent time at Apple working on the Apple Watch and iPhone before founding Even Realities in 2023. The company's founding thesis is a direct rejection of the dominant product strategy in wearable AI: no cameras. While Meta and Snap build glasses around content capture and ambient recording, Even builds around display-first interaction. The G2 model uses a proprietary optical system called Even HAO (Holistic Adaptive Optics) that integrates the microchip, waveguide, and prescription support from the ground up. The glasses retail at $599, with the average order landing near $1,000 when including prescription lenses and the companion Even R1 smart ring.

The growth numbers tell the real story:

•  Even sold more than 10,000 pairs of the G1, becoming the first company in the category to reach that milestone

•  The company grew from 30 to 40 staff in 2024 to 300 to 400 employees today

•  More than half of its users are in the United States

•  The G2 offers real-time translation across 35 languages and a voice-controlled teleprompter

Why did Meituan and Tencent invest? Not just because the hardware works. Because Even has cracked something structural: a reason for a consumer to wear AI on their face for 8 to 10 hours a day. That is the metric that matters. The glasses are not a novelty. They are a utility.

The privacy-first angle is not a marketing choice. It is a moat. As AI wearables proliferate, regulators in Europe and enterprise security teams globally will increasingly push back against ambient cameras. Even has positioned itself on the right side of that line before the line is drawn.

Previous backers include Hillhouse, Sequoia China, and Northern Light Venture Capital. The company does not sell in China. It is a Chinese-built product targeting the US, Japan, and Europe, which tells you exactly what the investors are betting on.

Here is the uncomfortable truth about factory automation: most of it, until very recently, was not intelligent. It was scripted.

Legacy industrial robots operate on fixed programs. You tell them exactly where to reach, exactly what torque to apply, exactly how to move. They are fast. They are precise. They break the moment something in the environment changes, because they cannot perceive or adapt.

Physical AI changes that equation. The new generation of industrial systems combines computer vision, on-device inference, and learned manipulation policies. A robot powered by physical AI can locate a part in a random orientation. It can handle variability in packaging, SKU mix, and workflow without a technician reprogramming it every time.

Asia Pacific is the fastest-growing region for physical AI deployment. The market for physical AI logistics robots alone is projected to grow at a 25.8% CAGR from 2026 to 2034, driven by China, Japan, South Korea, and accelerating adoption across Southeast Asia. China alone accounted for roughly 14.3% of global physical AI logistics revenue in 2025.

In Southeast Asia specifically, a pattern that plagued the industry for years is breaking. Manufacturers would run a pilot, see results, issue a press release, and then leave the technology exactly where it started. That is changing in 2026. The pilots are turning into production deployments, driven by three forces: labor cost increases, global supply chain restructuring, and the availability of AI systems that actually generalize across tasks.

Grab's acquisition of Infermove in January 2026 is a clean signal. Infermove builds autonomous delivery robots that operated with Meituan, Alibaba's Ele.me, and JD.com's Dada in China. Grab brought them in to extend their logistics infrastructure across Southeast Asia. That is not a pilot. That is a strategic asset acquisition.

Deloitte opened its Asia Pacific Physical AI Centre of Excellence in Shanghai in early 2026, describing it as a hub for helping clients move beyond proofs of concept to scaled deployments. NVIDIA launched its Physical AI Data Factory Blueprint with ABB, FANUC, and KUKA at GTC 2026, providing the simulation infrastructure to train robots before they ever touch a factory floor.

The shift in language from 'automation' to 'physical AI' reflects something real: the systems can now handle environments that were previously too unstructured for robots to navigate. That is the unlock.

1. Linkerbot (Beijing)

Most humanoid robot companies focus on the body. Linkerbot focuses on the hands. The Beijing-based startup mass-produces dexterous robotic hands across a range spanning 11 to 42 degrees of freedom, with precision reaching sub-millimeter tolerance. The company recently raised nearly $217 million and claims to be the first worldwide capable of mass-producing more than 1,000 dexterous units per month. Dexterous manipulation is one of the hardest unsolved problems in physical AI. Linkerbot is attacking it through manufacturing scale, not just research. Watch this one.

2. Inbolt (France)

Inbolt builds AI-powered 3D vision software that gives existing industrial robots the ability to perceive and respond to dynamic environments in real time. Their GuideNOW system enables robots to locate and manipulate parts in random orientations, dramatically reducing setup time and integration costs. This is a software approach to upgrading the installed base of legacy industrial robots globally, without requiring factories to replace hardware. The addressable market is enormous and the sales cycle is shorter than selling new robots.

3. Field AI (United States)

Field AI raised $405 million at a $2 billion valuation in August 2025, backed by Bezos Expeditions, Khosla, Temasek, and NVIDIA's NVentures. The company builds what it calls Field Foundation Models: risk-aware universal robot brains designed for unstructured, high-stakes environments such as oil rigs, mines, and construction sites. This is the opposite of warehouse automation. Field AI is going after the hardest physical environments, where the cost of a failed deployment is catastrophic. High risk, high defensibility.

Whether you are investing, building, or partnering, the same filters apply:

1.  Identify the feedback loop. Does the system learn from real-world deployment? If it only improves in simulation, ask why.

2.  Check the data moat. Physical AI models improve with real-world data. Who owns the data generated by deployment? That is the actual asset.

3.  Map the failure mode. Software fails silently. Physical systems fail visibly and expensively. What happens when this system makes a mistake?

4.  Assess generalization. A robot that handles one task is an appliance. A robot that handles 50 tasks is a platform. Ask where the company sits on that spectrum.

5.  Follow the deployment signal over the funding signal. In a market where $55.8 billion flooded in during six months, capital is a lagging indicator. Deployment contracts are the leading one.

The physical AI wave is real. So are the risks. Here are the ones nobody is talking about loudly enough.

Deployment economics are not settled.

Most of the largest valuations in physical AI exist before commercial deployment at scale. Physical Intelligence is seeking an $11 billion valuation with no revenue. The gap between what a robot foundation model can do in a lab and what it can do reliably across 10,000 different factory environments is not small. Investors are pricing in generalization that has not yet been proven.

The training data problem is structural.

Language models trained on internet text. Physical AI models need real-world interaction data from robots moving through physical environments. That data is scarce, expensive to collect, and often proprietary. The companies that control large deployment bases will have an enormous advantage. Newcomers entering the space now face a data cold start problem that money alone cannot solve quickly.

Hardware cycles are brutal.

Software startups can ship a new version on a Tuesday. Hardware startups carry inventory risk, manufacturing dependencies, and product cycles measured in years, not sprints. A single major supplier failure or chip shortage can stall a physical AI company regardless of how good the model is.

Geopolitical friction is rising.

Even Realities is a Chinese-founded company deliberately not selling in China, targeting US and European markets with Chinese-manufactured hardware. That is a deliberate positioning choice in a world where geopolitical scrutiny of hardware supply chains is increasing. This tension will affect multiple companies in the physical AI stack.

The question is not whether physical AI is real. The $55.8 billion raised in the first half of 2026 answers that. The question is which layer of the stack captures durable value.

The history of previous technology waves says the same thing every time. The infrastructure layer (compute, connectivity, data) wins first. The application layer (companies solving specific problems for specific industries with real deployments) wins second. The middle layer, the 'platform' companies trying to be general-purpose before the market is ready, is where most capital is destroyed.

In physical AI terms: NVIDIA owns the infrastructure. Companies like Linkerbot, Inbolt, and Field AI are building specific, defensible applications with real deployment signals. The robot foundation model companies in the middle are the most funded and carry the most risk.

This week, pick one physical AI sector relevant to your work, whether manufacturing, logistics, wearables, or something else, and find out which companies are actually deployed at scale in it. Not which companies have raised the most money. Which companies have robots or devices working in the real world today. That is where the next decade of value is being built.

Brief Stak

Insights for builders, founders, and future innovators.