The Billion-Dollar Pivot: Why the Biggest Bets in AI Are Moving From Software to Sensors

In April, Jeff Bezos’s Project Prometheus closed a $10 billion funding round at a $38 billion valuation. While just five months old at the time, and with no products shipped and no research published, its investors included some of the most prominent in the world, including JPMorgan and BlackRock.

But more attractive to investors that current sales was the bet Prometheus is making: building AI systems that simulate the behavior of materials, machines, and physical processes, a burgeoning category known as physical AI.

Prometheus is not alone in this race. That same month, SoftBank filed to IPO a robotics and AI company called Roze at a target valuation of $100 billion. Eclipse Ventures closed a $1.3 billion fund built entirely around physical AI startups. At NVIDIA’s GTC in March, CEO Jensen Huang put it plainly: “Physical AI has arrived. Every industrial company will become a robotics company.”

Following three-plus years of record-breaking investment into AI software, investors are now pointing their pens toward physical AI.

“The AI-induced SaaSpocalypse is putting a chill on PE deals in software,” Lex Zhao, a partner at One Way Ventures, told the International Business Times.

At the same time, he noted, “it’s creating another lever for PE firms to extract more value from traditional industries. This is highly appealing because they have less fear about AI cannibalizing those businesses since they operate in the physical world.”

Why now?

Over the past 10 years, manufacturing companies have been installing sensors everywhere on their operations: conveyor belts in packaging plants. Pressure valves in refineries. Even CCTV cameras across entire cities. The hardware went in, but most of the data it generates has gone nowhere.

Now, physical AI could help translate that data into real-world business gains, its proponents say.

“Physical AI is really a Rosetta Stone for industrial IT data,” Paul Bloudoff, Senior Director of Edge AI and 5G at NTT DATA, told International Business Times. “There are sensors on absolutely everything in some factories. But all of that data just gets generated and ignored.”

What’s changed is that foundation models can now consume raw sensor data and make sense of it. Anomalies in vibration patterns, sharp temperature changes across production lines, a pedestrian stepping into a crosswalk three seconds too late. In fact, industry estimates put the global AI-in-manufacturing market at roughly $34 billion in 2025, with projections reaching $155 billion in the next four years.

The workforce pressure makes waiting harder. The EU is projected to lose between 1 and 2 million workers annually in the coming decades, according to the Egmont Institute. Germany alone faces a deficit of nearly 5 million workers by 2030. Nicolas Barthalon, a partner at Munich-based venture firm Ventech, told International Business Times: “Automation ceases to be a productivity option and assumes the character of a structural obligation.”

Despite projections, European and US labor groups have argued that leading with automation increases risk for displacing workers before they can effectively be retrained. Some unions have formed alliances aiming to ban the introduction of fully automated equipment in their workplaces.

What deployment looks like on the ground

Several startups are already deploying physical AI outside of the lab. One of them is running in a mid-sized American city. In Bellevue, Washington, the Newton model is on traffic cameras at city intersections. It detects near misses between cars and pedestrians.

“If an accident almost happens, it never happens,” Bloudoff said. “But all of this data can better help us plan cities if we know the types of intersections that almost cause accidents.”

Brandon Barbello, co-founder of Archetype AI, is behind the Newton system. He spent years at Google working on Project Soli, the radar-sensing technology embedded in Pixel devices.

Newton, Archetype’s foundation model, works differently from older machine learning systems that handled one sensor at a time. It processes multiple data streams at once; cameras alongside vibration sensors, alongside temperature readings, alongside acoustic signals. The company calls it sensor fusion.

In manufacturing, three use cases are gaining traction: task verification (confirming a worker completes each step of a procedure correctly), anomaly detection for predictive maintenance, and semantic safety, which is basically identifying fuzzy risks before they become incidents.

Barbello said each of these previously required months of custom machine learning work. With a physics-aware foundation model, he said, they can be deployed in days.

The infrastructure requirements are pushing startups across the category toward partnerships with telecom and industrial giants.

At GTC 2026, AT&T, T-Mobile, and Comcast announced plans to turn their network edges into platforms for running AI workloads closer to where the data is generated. SK Telecom launched a “Sovereign AI Package” at MWC Barcelona 2026, bundling telecom infrastructure with industrial AI services. Robotics startup Serve Robotics is already deploying delivery robots that run AI applications across those telecom edge networks.

“No one entity can do the whole thing,” Barbello said. “As a startup, having partners like NTT Data lets us provide the AI for sensors while they bring the hardware and the networking.”

Is the market ready?

In April, physical AI in robotics, aerospace, drones, and autonomous vehicles attracted roughly $5.3 billion in venture funding, according to Crunchbase.

But while investor capital is flowing to physical AI, it’s not clear whether the bets will pay off. Afterall, enterprise AI software pilot projects have a well-documented failure rate.

“AI pilots fail when the technology isn’t good enough, or when the AI company doesn’t fully understand the incentives and limitations of its customer,” Zhao said. He also cautions that in places like China, we’re already seeing signs of success.

“You already see ‘dark factories,’ fully automated manufacturing plants that can go lights out, coming out of China. For certain types of physical processes, we’re already there. The market potential is going after the rest.”

Barthalon also sees market opportunity for physical AI models coming from investors. Large asset managers can bring a portfolio of industrial clients who can help validate and deploy physical AI in their factories.

“The investment and the distribution channel are, in this case, the same transaction,” he said.

Barbello’s advice for anyone entering the space was specific. “Find the piece that you can be the best at, focus there, and then build the right network of partners around you to make the bigger thing possible,” he told the International Business Times.

Whether the current wave of capital translates into production-scale deployment remains the open question. But with institutional investors, telecom operators, and industrial incumbents all converging on physical AI simultaneously, the category is no longer speculative, it is being built.