Robotics has the money. Deployment is still the product.

The money, and the gap

Scale Ventures describes a robotics market awash in capital and still short on deployments. The firm says approximately $60 billion has been invested in robotics model labs since 2023, with companies raising at valuations from $5 billion to $30 billion and beyond. Forecasts for how many robots will exist a decade from now range, in Scale's account, from Goldman Sachs forecasting 6.5 million humanoids by 2035 to Elon Musk forecasting one billion robots within ten years. That spread is the story: the capital has arrived before anyone agrees on the volume. Scale's own thesis is that success depends on closing the gap between demo performance and reliable, scalable operation. m:…7059

The cost curves, on Scale's numbers, favor the optimists. An industrial arm that cost an average of around $46,000 in 2010 was $27,000 by 2017 and can be purchased for $10,000 today, and the firm says the average cost of an industrial arm has fallen by more than 80% over the last 20 years. A hobbyist compute board that did about 0.5T operations per second in 2019 now has a successor doing 67T operations per second. And the firm cites DHL going from 78 units an hour to around 150 at its Locus sites. Those are deployment results, not demo videos, and Scale uses them to argue that the economics of buying a robot have already changed even if the reliability of running one at scale has not. m:…7059

Scale's framing for who captures the value is also worth noting. Businesses buy outcomes, not robots, the firm says, and the new robotics companies are functioning as neo systems integrators that will likely capture a larger share of market value than traditional integrators. The industry, in its account, sorts into three camps: model labs, vertical integrators, and classical automation providers. Toyota's historical success, Scale argues, came from process optimization rather than technology adoption alone. A demo can win a pilot, the firm says, but a multi-year contract waits for human intervention to go away and the economics to work. m:…7059

Safety first, literally

Agility Robotics, answering questions about its Digit 5 humanoid, puts the same bottleneck in engineering terms. Safety is a hard blocker for deploying robots until it is solved, the company says, and Digit 5's safety decisions are made onboard the robot, using an islanded computer separate from the main control computer for safety functions. The safety strategy includes AI-based human detection, and the robot uses lights and sounds to signal intent to humans. Agility says it targets 99% reliability for bin and tote manipulation tasks. That number is the deployment contract in miniature: a customer does not buy a humanoid, it buys a tote moved reliably, thousands of times, without hurting anyone. m:…2483

The other Digit 5 details fill out the same picture. The robot weighs 129 kg, has a 50 lb lifting capacity, and its battery charging time is 9 minutes, according to Agility. It uses a scaled impact approach to manage reflected inertia in locomotion, and its compute is modularized in a box at the top of the torso. Agility also says data collection from deployed fleets is the primary driver for scaling production. That is a loop worth watching: deployment produces data, data improves reliability, reliability wins the next deployment. It is also a loop that favors whoever deploys first, which is why the capital race and the reliability race are the same race. m:…2483

Pudu Robotics offers a quieter version of the same lesson. Its elevator control solution has achieved CSA certification, confirming assessments for electrical safety, structural design, and operational reliability, the company says. The solution lets robots call elevators and travel between floors autonomously, with multi-robot and multi-elevator scheduling. Pudu says it has been deployed in multiple markets, supporting more than 10,000 robots. Nobody films a robot waiting for a lift, but a robot that cannot ride one cannot work in a multi-floor building. m:…3307

The energy ceiling

Naveen Rao, founder of Unconventional AI, told Twenty Minute VC that the inference market is projected to reach $10 trillion within 10 years, and that robotics will be 100 times larger than the market for coding agents. He argues that energy supply is the primary bottleneck for AI scaling, with current infrastructure growth rates insufficient to meet demand, and that building one gigawatt of energy infrastructure costs approximately $50 billion. Those are forecasts and opinions, not reported results, and they should be read as a founder's positioning. But they set a ceiling on the robotics thesis: even if deployment works, the power has to come from somewhere. m:…6264

Rao also argues that pure software investing is becoming significantly less viable compared with hardware-centric AI investment, that the transformer architecture is not the final paradigm, and that development is shifting from pre-training to post-training reinforcement learning and specialization. He said verification of AI-generated artifacts is becoming a larger bottleneck than producing them. The through-line is that value moves from generating things to verifying, powering, and embodying them. m:…6264

What to diligence

For a VC or a buyer, the diligence question that follows from all four sources is the same: show the deployment, not the demo. How many hours has the system run without human intervention, at what reliability? Scale's cost curves say the hardware is cheap enough. Agility's safety architecture and Pudu's certification say the gates are known. Rao's energy numbers say the constraint may sit outside the robot entirely. m:…7059 m:…2483 m:…6264