Welcome Back to XcessAI

Robotics companies have become very good at producing impressive videos.

Robots fold clothes. Move boxes. Weld components. Walk through factories. Load dishwashers. Recover when pushed.

Increasingly, the demonstrations create a natural conclusion:

If the robot can do the task, automation must be close.

A fascinating new study from Anthropic suggests the reality is considerably more complicated. Researchers analysed roughly 19,000 tasks across the US economy and compared them with the capabilities of robots available today. Their conclusion was striking.

Robots can already perform around 74% of physical work tasks in at least some environment. But they are currently cost-competitive with humans for only around 0.3% of work. That enormous difference may be one of the most important numbers in robotics.

Because there is a very large gap between:

Can a robot do this? and Should a company pay a robot to do this?

Call it the Robot Gap.

Almost a year ago, in Humanoid Breakpoint, we explored the idea that humanoid robots were moving beyond laboratory demonstrations and towards genuine deployment.

Earlier this year, Humanoid Breakpoint, Part II looked at another acceleration: better AI was allowing robots to move away from rigidly programmed movements towards systems capable of learning, adapting and performing useful work for longer periods.

Those trends haven't reversed. If anything, they have strengthened.

But Anthropic's new research adds an important piece to the puzzle.

Robot adoption has at least three gates:

Capability → Environment → Economics

Passing the first does not automatically open the other two.

A robot may technically be capable of picking up an object. But can it find that object when somebody leaves it in the wrong place? Can it perform the task safely beside people? Can it recover when something unexpected happens? Does the factory need to be redesigned around it? How much supervision does it require?

And after buying, installing, maintaining and powering the robot, is it actually cheaper than paying a human?

These questions rarely appear in demonstration videos. They dominate investment decisions.

One of the most revealing findings in Anthropic's research is where today's robots can actually perform their tasks.

Roughly half of physical work can be performed by robots only in environments specifically designed around the machine, such as traditional automated production lines.

Another 22% can be performed in structured human environments such as warehouses.

Only around 2% of physical tasks can currently be performed robotically in genuinely unstructured environments, such as public roads.

That distinction is fundamental.

Humans are extraordinarily good at dealing with disorder. We walk into a room we have never seen before, recognise objects, understand roughly what needs to happen and continuously adapt.

Robots have traditionally solved the opposite problem. Rather than making the robot adapt to the environment, engineers make the environment predictable for the robot.

That is why industrial automation has been so successful inside factories.

The floor is flat. The objects are known. The components arrive in predictable orientations. Safety zones can be defined. Connectivity can be installed. The workflow can be redesigned. The robot may look general-purpose.

The business case often isn't.

There is another reason robotics may progress differently from software AI.

Software is extraordinarily easy to replicate.

Once an AI model has been trained, another million people can use it without manufacturing another million models.

Robots live in the world of atoms.

Every additional robot requires motors, sensors, actuators, batteries, chips, materials, assembly, shipping, installation and maintenance.

Then it needs electricity. It breaks. Its gripper wears out. Someone needs to fix it. Sometimes a human still needs to supervise it.

This makes the economics fundamentally different from an LLM.

Anthropic estimates that robot prices have historically declined by roughly 3% per year. If that trend continued, its model suggests it could take around 40 years for robots to become cost-competitive for 10% of work.

That is not a forecast. The researchers explicitly model different scenarios, and rapid advances in AI, hardware or manufacturing could accelerate the curve dramatically.

But it demonstrates the scale of the economic hurdle. The robot doesn't merely have to become capable. It has to become cheap enough.

One example from Anthropic's modelling makes this clearer.

Packers and packagers are among the few occupations where today's robotics can already approach economic parity.

Automating the exposed tasks requires multiple machines costing more than $2 million to purchase and install.

That sounds terrible.

But those machines can replace the output of roughly 14 workers.

Spread the capital cost across approximately ten years, then add maintenance, energy and human supervision, and Anthropic estimates annual robot costs at around $45,000 per equivalent worker.

The comparable human compensation is about $49,000.

Suddenly the numbers almost work.

Compare that with welding.

Robots can already weld extremely well.

But welders do much more than make a weld. They move components, position material, climb, inspect, grind, finish and deal with irregular situations.

Automating the entire bundle of exposed tasks would, according to Anthropic's estimates, cost roughly five times as much as employing human welders.

Same technological story. Very different investment case.

This creates an important trap for executives and investors.

We naturally evaluate robotics through capability demonstrations.

Can it fold the shirt? Can it assemble the component? Can it cook the meal? Can it unload the dishwasher?

Those are useful engineering milestones. They are poor capital-budgeting metrics.

The more useful questions are economic.

How many times per hour can it perform the task? What percentage of attempts succeed? What happens when it fails? How much human intervention is required? What modifications does the facility need? How many hours per year is the machine productive? And what is the fully loaded cost per useful unit of output?

A robot performing a task once in a demonstration tells us something important about capability. It tells us surprisingly little about ROI.

I don't think Anthropic's findings weaken the humanoid thesis. They make it more precise.

The first Humanoid Breakpoint argued that automation would begin with tasks rather than entire jobs, particularly repetitive work in manufacturing, logistics, packaging and other controlled environments.

That is remarkably close to what this research finds.

Warehouses and factories are attractive precisely because companies can control the environment. And repetitive tasks are attractive because robot utilisation can be high enough to amortise the expensive hardware.

The breakthrough therefore isn't the moment a humanoid becomes physically capable of doing everything a human can do.

That threshold may not even be economically necessary. The meaningful breakpoint occurs when enough tasks cross three thresholds simultaneously:

That intersection may initially be narrow. But every improvement in hardware cost, reliability, dexterity and AI makes it wider.

There is an important reason not to interpret today's 0.3% as a reason to dismiss robotics.

These variables interact. Better AI can reduce the amount of engineering required to deploy a robot. Better dexterity expands the number of tasks one machine can perform.

Higher utilisation spreads the cost of the hardware over more productive hours. Larger production volumes reduce manufacturing costs.

Fleet learning means improvements discovered by one robot can potentially improve thousands of others. And once a robot can perform several adjacent tasks rather than just one, the economics of the entire installation can change.

This is why the relevant question isn't simply:

How big is the Robot Gap today?

It is:

How quickly is it closing?

Anthropic's historical analysis finds that robots have become capable of performing around 2% of the previously inaccessible physical tasks each year over the past several decades. AI-powered robotics could potentially change that rate, although nobody yet knows by how much. That is the number worth watching.

There is something counterintuitive about Anthropic's numbers.

Robots can apparently perform 74% of physical work tasks under at least some circumstances.

Yet almost none of that work is currently economical to automate.

Both statements can be true. And together they reveal where the robotics industry really is.

The capability revolution may already be well underway. The economic revolution is much earlier.

For the companies building humanoids, the next battle therefore isn't simply about making robots more intelligent. It is about reducing the distance between possible and profitable.

That distance is the Robot Gap.

And when it begins closing quickly, the impressive videos will stop being the interesting part. The balance sheets will be.

Until next time,
Stay adaptive. Stay strategic.
And keep exploring the frontier of AI.

Fabio Lopes
XcessAI

💡Next week: I’m breaking down one of the most misunderstood AI shifts happening right now. Stay tuned. Subscribe above.

Read our previous episodes online!