2026: Physical AI Becomes a Capital Allocation Problem

Every January, a big chunk of the tech world flies to Las Vegas for CES, the Consumer Electronics Show.

Companies go there to do three things: show prototypes, announce partnerships, and recruit belief. Belief from customers, from talent, and from capital. And when the incentives are to recruit belief, the whole environment naturally optimizes for possibility.

That’s why CES can make almost any technology feel inevitable.

Not because it’s fake.

Because the stage is built to answer one question: “What could this become?”It’s not built to answer the harder question: “What does it take to deploy this at scale without blowing up the economics?”

So when people come back saying “this is the year,” I translate it into a more practical sequence:

• First, it works.

• Then, it runs.

• Then, it scales.

“Works” is a demo.

“Runs” means it survives Tuesday night at 2am with real users, real variance, and real failures.

“Scales” means it can be repeated - city after city, facility after facility - without renegotiating the entire risk and cost structure every time.

That’s where the real world shows up: permission (regulators, cities, safety rules), liability (who eats the downside), unit economics (cost per hour of useful work), and operations (maintenance, uptime, parts, service networks).

This is why Deutsche Bank’s post-CES framing matters. They’re basically saying 2026 is when autonomous driving starts moving beyond pilots into commercial rollouts, and humanoid robots start moving from labs into small, paid deployments.

If that’s even directionally right, 2026 won’t be remembered as the year robots got “smarter.”

It’ll be remembered as the year the market stopped buying demos and started pricing execution.

And execution, in physical AI, is where incentives get exposed.

Software scales when the marginal copy is cheap.

Physical AI scales when the marginal hour of useful uptime is cheap.

That one sentence makes the rest of the reasoning easier, because it tells you what decides winners:

• Permission (rules, approvals, operating boundaries)

• Liability (who carries the downside)

• Service (maintenance, parts, time-to-repair)

• Supply chain (cost curve, component bottlenecks)

• Infrastructure (energy, charging, depots, facility constraints)

This is why I keep repeating “permission is an economic input.”

Not a legal footnote.

An input.

If you want proof that “permission” is a real economic variable, look at two industries that already went through the same transition: dangerous tech → public trust → scale.

Early U.S. commercial aviation didn’t start with passengers.

It started with mail.

The Contract Air Mail Act of 1925 (the Kelly Act) allowed the Post Office to contract private companies to carry mail. That created steady revenue, forced operational discipline, and, most importantly, made the activity legible to institutions that matter (insurers, investors, regulators).

The mechanism is the lesson:

Mail contracts didn’t just “support” aviation. They reduced uncertainty enough for capital to fund repetition.

That’s what turns a risky invention into an industry.

Elevators existed before they were trusted.

The market didn’t unlock because people suddenly loved height.

It unlocked when the downside became containable.

Elisha Otis’ 1854 safety-brake demonstration made safety visible and believable to the public.

Then the industry scaled through something even more important than the demo: standards.

ASME issued its first elevator safety code in 1921, creating a shared baseline for safety provisions and making adoption repeatable across buildings and cities.

Again, the mechanism matters:

Standards compress uncertainty. Compressed uncertainty increases permission. Permission unlocks scaling.

Now apply that directly to autonomy and humanoids.

Most autonomy conversations are still framed around “levels.”

That’s not the axis that determines scaling.

The axis is: who owns the downside when the system makes a mistake.

Because that drives:

• how regulators respond

• how insurers price risk

• how cities behave

• how fast fleets expand

Once you see it that way, “why robotaxis scale first in constrained places” becomes obvious. Not because engineers lack ambition. Because permission and liability are easier to bound.

And bounded downside is what lets utilization rise.

A robotaxi is not “a car that drives itself.”

It’s a fleet that must be dispatched, cleaned, maintained, repaired quickly, and insured - while staying on the road as many hours as possible.

So the core economic object becomes:

revenue per vehicle-hour minus cost per vehicle-hour

That’s why the real scaling work looks unsexy: simplifying hardware, reducing maintenance burden, improving reliability in weather, tightening operating playbooks.

Waymo’s sixth-generation approach is explicitly framed around reducing system cost while enabling more capability - and even the small details (like sensor cleaning mechanisms) are really about uptime in real conditions.

The lesson is simple:

In fleets, demos don’t compound. Operations compound.

Humanoids trigger a lot of sci-fi thinking.

Companies don’t buy sci-fi.

They buy throughput.

So humanoids will win by taking one workflow where ROI is obvious and failure is survivable.

That’s consistent with the Deutsche Bank framing: humanoids moving from lab experiments into small deployments, with “general purpose” effectively being postponed while narrow use cases prove viability.

There’s a nuance people miss here:

Humanoids aren’t competing only with labor. They’re competing with every other form of automation and process redesign.

Sometimes the best “robot” is still a conveyor belt and a better layout.

So the only way humanoids scale is by being the best option for specific bottlenecks.

In every big tech wave, a platform layer emerges.

Platform layers do two things at once:

• make it easier for others to build

• concentrate margins and leverage at the platform

At CES 2026, Nvidia explicitly pushed “physical AI” and highlighted humanoid partners integrating its edge compute stack (Jetson Thor).

If that stack becomes default, it changes industry structure.

Not because of hype.

Because standards, de facto or de jure, shape who captures economics.

You can’t talk about autonomy and humanoids seriously without Tesla.

If you believe, as I do, the same perception + planning stack can generalize across vehicles and humanoids, Tesla has a structural advantage: speed of iteration and a manufacturing machine that can push cost curves down.

Tesla is indeed the most aggressive attempt to integrate physical AI into an industrial machine:

• a massive deployed vehicle base

• autonomy software iteration in the wild

• manufacturing culture built around cost-down

• a humanoid program (Optimus) inside the same company

That integration is Tesla’s advantage.

But it’s also Tesla’s exposure: every constraint shows up earlier.

Tesla said it will offer Full Self-Driving (Supervised) only via monthly subscription starting February 14, 2026.

Most people will interpret that as “recurring revenue.”

The other possible read is incentive-based:

If a product is already durable, universally trusted, and clearly worth it, upfront pricing is easy.

Subscription is what you choose when you want lower adoption friction and you want cash flow to match a product that is still evolving and still supervised.

It’s Tesla telling you where it thinks the adoption constraint really is.

Reuters tied the change to increased scrutiny, including an NHTSA investigation into 2.88 million vehicles after reports of safety violations and crashes involving FSD.

This is the heart of the “permission as an input” argument:

In autonomy, incidents are not just bugs. They become evidence. Evidence becomes policy. Policy becomes timeline.

Tesla can ship fast.

But speed doesn’t remove the permission constraint. It often activates it.

Robots don’t scale on model weights.

They scale on actuators, motors, magnets, power electronics, and service parts.

And Tesla has already run into that reality: Musk said Optimus production was affected by export curbs on rare-earth magnets, and Tesla was working through licensing.

That single detail tells you where the market will bottleneck.

Hyundai’s move to appoint Milan Kovac (former head of Tesla’s humanoid robot program) as an adviser tied to Boston Dynamics’ strategy and commercialization is another signal that this is becoming an industrial execution race.

Serious companies don’t hire for vibes.

They hire because they expect the category to demand real execution.

Most people will watch announcements.

I would watch four boring variables.

Can they deploy the next site / city with less friction than the last one?

If not, it’s not scaling. It’s touring.

Who carries the downside when something goes wrong?

Ambiguity slows scaling by default, because insurers and regulators fill the vacuum.

Who fixes it, how fast, with what parts, under what contract?

This is where impressive tech usually dies.

Not cost per unit. Not capabilities. Not demos.

Cost per hour of useful work, at acceptable safety.

That’s the variable that compounds.

CES is useful because it shows what companies want you to believe.

But physical AI becomes real only when the market can price it.

Airlines scaled when mail contracts turned risk into a repeatable business. Elevators scaled when safety became credible and standardized.

Autonomy and humanoids are walking the same path.

The winners won’t be the best storytellers.

They’ll be the operators who turn permission, liability, service, and cost curves into a repeatable deployment machine.

𝘈𝘯𝘺 𝘷𝘪𝘦𝘸𝘴 𝘰𝘳 𝘴𝘵𝘢𝘵𝘦𝘮𝘦𝘯𝘵𝘴 𝘦𝘹𝘱𝘳𝘦𝘴𝘴𝘦𝘥 𝘢𝘳𝘦 𝘮𝘪𝘯𝘦 𝘢𝘯𝘥 𝘯𝘰𝘵 𝘵𝘩𝘰𝘴𝘦 𝘰𝘧 𝘮𝘺 𝘦𝘮𝘱𝘭𝘰𝘺𝘦𝘳