There is a strange contradiction emerging in the race to build humanoid robots.

On one side, some of the world's most influential technology executives are making enormous bets on a future filled with intelligent machines.

Nvidia CEO Jensen Huang has said the “ChatGPT moment” for robotics has already arrived. Elon Musk has predicted that there could be more than 1 billion humanoid robots within a decade, with each potentially producing several times the output of a human worker.

On the other side is Anthropic.

The AI company behind Claude has just released research suggesting that the technological promise of robotics is real—but the economics remain a huge obstacle.

Anthropic's study says robots can already perform about 74% of physical job tasks in the United States, representing roughly 34% of working hours. Yet today's machines are cost-competitive with human labor for just 0.3% of work tasks.

If robot prices continue declining at roughly their historical pace, Anthropic estimates it could take about 40 years for robots to become cost-competitive for even 10% of U.S. work.

That creates a fascinating collision between Wall Street's robotics enthusiasm and an AI lab's economic reality check.

Anthropic's central finding sounds contradictory until the distinction between capability and economics is understood.

A robot might be technically capable of performing a task.

That does not mean a company will buy one to do the job.

The machine might cost too much.

It might require a carefully controlled environment.

It might need maintenance specialists.

It might operate too slowly.

Or it may fail often enough that a human remains cheaper once the full cost of deployment is included.

Anthropic's research explicitly separates what robots can physically perform from whether those systems are economically competitive with people.

And the difference is enormous.

Robots can already perform a large share of physical tasks.

But there are comparatively few jobs where deploying robots makes economic sense today.

That distinction could become one of the most important variables in the next decade of robotics investing.

Nvidia has built much of its AI empire around supplying the computing layer beneath artificial intelligence.

Now the company is targeting physical AI—the software, simulation, chips and models that allow machines to perceive the physical world and act within it.

Nvidia's robotics stack includes Cosmos world models, Isaac simulation tools and GR00T models designed for humanoid and other robotic systems. The company says major robotics developers including Figure, Boston Dynamics, Agility and others are using its technology.

Nvidia's logic is straightforward.

If robotics becomes a huge new computing market, it could create another enormous source of demand for Nvidia processors and software.

And unlike an individual robot manufacturer, Nvidia does not need one particular humanoid design to dominate.

It can potentially supply the technology underneath many of them.

That is a powerful strategic position.

When Huang says robotics has already reached its ChatGPT moment, he is essentially arguing that the underlying technology has crossed an important development threshold. Nvidia's own materials describe major advances in physical AI, world models and robot-learning systems.

But that is not the same as saying millions of general-purpose humanoids are ready to replace workers.

The commercial breakthrough for robotics has more steps.

A system needs to understand its environment.

It needs to manipulate objects reliably.

It needs to operate safely around humans.

It needs enough battery life.

It needs maintenance.

And, ultimately, the economics have to work.

Anthropic's research suggests that last stage is much farther away than the enthusiasm surrounding robotics might imply.

Elon Musk's forecast is dramatically larger.

During remarks at the G20 Innovation Ministerial in September, Musk predicted that more than 1 billion humanoid robots could exist within 10 years and suggested each could produce approximately five times the output of a human. He also argued that robot manufacturing could create a compounding effect in which robots help manufacture more robots.

That is a forecast, not a measured current production trend.

And recent evidence shows why the path to that scale could be difficult.

Tesla's Optimus program, for example, has faced production challenges, with output remaining far below the company's earlier target of 20,000 units per week. The Verge reported that only several hundred units per week had been produced recently and described manufacturing difficulties involving the robot's small, precision-dependent components.

That contrast is revealing.

The vision calls for exponential manufacturing.

The factory floor still has to solve thousands of mechanical problems.

Anthropic's researchers estimate that robots can already perform three-quarters of physical tasks in the U.S., but the systems are usually operating under significant constraints. Many require structured environments that are difficult to reproduce across ordinary workplaces.

Consider a factory.

A robot can be extremely effective when the environment is designed around it.

The floor is predictable.

The parts arrive at a known location.

The lighting is controlled.

The machine performs repetitive movements.

Now place the same robot into a warehouse, construction site, restaurant or household.

Suddenly the environment changes continuously.

Objects are misplaced.

People move unpredictably.

Weather matters.

Surfaces differ.

The robot needs flexibility rather than repetition.

This is one of the hardest problems in physical AI.

Anthropic says robot prices have declined by around 3% annually since the 1990s. At that pace, the company estimates it would take roughly four decades for the share of tasks that are economically competitive with humans to reach 10%.

That number should not be interpreted as a prediction carved in stone.

Technology does not necessarily move in a straight line.

AI could dramatically improve dexterity.

Manufacturing could become much cheaper.

Battery technology could improve.

Robot hardware could become modular and easier to repair.

Or a breakthrough could make today's cost assumptions obsolete.

Anthropic explicitly acknowledges this uncertainty. The company's methodology relies partly on current robot capabilities and historical price trends, while noting that AI-enabled robots could leapfrog today's technological limitations.

So the 40-year figure is better understood as a scenario based on existing trends—not a timetable for the entire robotics industry.

Anthropic's research does not suggest that robot adoption has no economic consequences until humanoids become ubiquitous.

In fact, its historical analysis found that jobs more exposed to robots over the past 50 years experienced greater subsequent declines in wages and employment.

The potential implication is that automation may arrive unevenly.

The first effects could appear in occupations where robots already have a foothold.

Anthropic identifies driving and warehouse work as highly exposed examples, while nursing and general repair appear much less exposed because current robots struggle with their physical complexity and variability.

That suggests the future may not begin with a Hollywood-style arrival of humanoids everywhere.

It may begin with specific jobs gradually requiring fewer people.

Anthropic estimates that roughly 80% of job tasks by working time are exposed to either robots or large language models.

That figure does not mean 80% of jobs will disappear.

“Exposure” means that technology can potentially perform some portion of the tasks involved.

A job can be exposed to automation while still requiring a human for judgment, communication, supervision or responsibilities that machines do not handle well.

That distinction is essential.

A warehouse employee could spend less time moving boxes but more time overseeing automated systems.

An accountant could use AI for document processing while spending more time communicating with clients.

A driver could eventually supervise autonomous vehicles rather than manually operate every mile.

The future of work may therefore involve task substitution rather than simple job replacement.

The most interesting way to understand the emerging robotics market is to separate three curves.

The first is the capability curve—how quickly robots become able to perform new tasks.

The second is the cost curve—how quickly machines become cheap enough to compete with human labor.

The third is the deployment curve—how quickly businesses can integrate robots into messy real-world environments.

Nvidia is betting these curves will accelerate together.

Musk's forecasts assume an especially steep trajectory.

Anthropic's research suggests the capability curve is already impressive, but the cost and deployment curves remain much slower.

That does not invalidate the bullish robotics thesis.

It defines the challenge.

The robotics industry now has something the early AI industry once lacked: enormous computing power, powerful foundation models, sophisticated simulations and global investment.

Nvidia is providing much of that infrastructure.

Companies such as Tesla are betting on vertically integrated humanoid fleets.

Anthropic is studying what happens when those technologies meet the real economics of employment.

The resulting debate is not really about whether robots are possible.

They clearly are.

The question is how quickly capability turns into affordable, reliable labor.

That distinction could determine whether the coming robotics boom resembles the explosive software-driven growth of generative AI—or follows a much slower industrial curve.

For Nvidia and Musk, the opportunity is measured in extraordinary scale.

For Anthropic, the latest research is a reminder that physical reality has its own timetable.

The chips may be ready.

The models may be improving.

But the factory floor still gets the final vote.