Skip to content
i-think-(locally),-therefore-i-am-…-–-embedded-computing-design

I think (locally), therefore I am … – Embedded Computing Design

November 18, 2026

Blog

A snowman has stepped out of a film and into our world.

Disney’s robotic Olaf walks, moves his eyes, and holds conversations. His movements were learned in simulation, then brought to life in a physical body. A character made of imaginary snow now has to contend with gravity.

Four centuries ago, Descartes wrote, “I think, therefore I am.” He found certainty in thought. Robotics brings us to a more practical problem: how much intelligence can a machine carry inside the body that needs it?

For years, the most capable AI lived in data centers. Now more of that capability can travel with the machine. A robot can respond to the world without sending every decision somewhere else.

That is also a labor story. As intelligence becomes cheaper to carry, more physical work becomes economical to automate. I suspect some of the earliest opportunities will be jobs employers struggle to fill.

More intelligence in less space

Two trends are easy to conflate.

  • Model size measures how much space an AI needs.
  • Intelligence density measures how much capability it delivers for the resources it consumes.

For this definition, “footprint” means the resources needed to run the model. Comparisons need a consistent measure, such as memory or energy. Capability per dollar is useful too, but it measures something different from capability per gigabyte

The Qwen team reports that Qwen3-4B can rival Qwen2.5-72B-Instruct on its evaluations, despite having eighteen times fewer parameters, the numerical settings a model learns during training. Microsoft’s Phi-3-mini, with 3.8 billion parameters, also rivals much larger models on selected benchmarks. These comparisons show how much capability can fit into a smaller package.

And, this is not from even last month. The trend is what I wanted to shed light on.

One way to achieve this is distillation: a smaller “student” model learns from a larger “teacher.” The teacher still needs substantial computing power to be trained. Running a trained model, called inference, is a separate expense.

That distinction helps explain why enormous data centers and small robot brains can grow together. We can spend heavily to develop intelligence, then deploy versions that cost much less to run. A robot needs enough capability for its task; it does not need every capability of the model that helped teach it.

The Catch: this is only an inference time story

One caveat to keep in mind – most of this compression comes from techniques such as distillation – a smaller “student” model trained to mimic a larger “teacher.” The teacher still has to exist, and training it still requires data-center-scale compute, data, and energy. Size-density is falling at inference time, not training time. Data centers don’t disappear in this story. However, they specialize into teaching, while compressed models handle acting. That’s a more precise claim than “AI is getting smaller”.

A brain the body can carry

This intelligence lands on an inference machine in deployment – a moving robot – AI gets a physical embodiment. NVIDIA’s Jetson Thor is the clearest example of what “compressed intelligence, deployed” looks like in hardware: 2,070 FP4 TFLOPS, 128GB of memory, 40–130 watts, small enough to sit inside a robot’s torso. Boston Dynamics is putting it in its humanoid Atlas; Galbot is putting it in its G1 humanoid. These are just two example but the list is fairly long.

Last week’s GTC Berlin keynote is a useful snapshot of what that compressed-intelligence stack now buys, economically. NVIDIA unveiled Agile ONE, a humanoid reference robot built with Agile Robots, meant to cut the time it takes a new robot platform to learn a skill.

Disclaimer: NVIDIA provides compute for both the training-scale chips and the edge chips, so a narrative where intelligence gets both bigger (training) and smaller (inference) expands its addressable market either way. That’s a real incentive – the density gains show up independently in open-weight models from Microsoft, Alibaba, and DeepSeek, which have no stake in NVIDIA’s edge business.

The canonical architectures of the future

A robot brain maps surprisingly well onto Kahneman’s split, with one addition.

  • Planning resembles Kahneman’s System 2: deliberate reasoning about what to do next.
  • Action resembles System 1: turning a plan into a grasp or a step.
  • Beneath both sits what I would call System 0: the reflexes that maintain balance, control motors, and protect people nearby.

Those reflexes cannot wait for a reply from a remote server. The cloud can help with training, shared learning across robots, and some difficult decisions. Immediate control must stay close to the body.

Intelligence Density doesn’t need to reach every layer at once. It needs to reach the layer that is the most latency sensitive. Another way frame the observation is that a higher density AI model does not directly equate to a fact that robots go fully offline.

SemiAnalysis did a great coverage on how these architectures may evolve.

A smaller model does not automatically make a competent robot. The machine must still turn what it sees into a reliable movement. Intelligence density helps make that possible within a robot’s limits of memory, power, and response time. The task still has to be tested in the physical world.

A robot phones home to get smarter; it doesn’t phone home to take a step.

The Economics of the Automation Frontier

The economics of where this gets deployed follow directly from the density curve. The basic economic test is familiar.

A task gets automated when the value it produces exceeds the cost of doing it

Observations from this model

As ID rises, the Cost falls for every task at once, without redesigning the task.

The second equation sketches how intelligence density affects the computing part of that cost, ceteris paribus. A task’s footprint represents its computing demand. As density rises, the same task requires less spending on intelligence. Hardware, maintenance, energy, integration, and supervision still belong in the full deployment bill.

Imagine an inspection task whose value stays the same. Nobody makes it easier, and nobody pays more for it. If the cost falls far enough, it becomes worth doing with a robot.

The “automation frontier” is the boundary between tasks that pay and tasks that don’t. Higher intelligence density can move that boundary. Several tasks may become economical together because they benefit from the same improvement in computing.

A cheaper brain can therefore matter more than a more impressive demonstration. It changes the calculation for work that was previously too expensive to automate.

The automation frontier moves as intelligence density rises

The chart above is illustrative, not measured, although the mechanism it shows is real. A task doesn’t need to get easier or more valuable to become automatable. It just needs Cost(task) to fall below Value(task), and Cost falls on its own as ID rises.

Tasks don’t jump the line one at a time because someone redesigned them. Whole clusters of tasks drift across it together as the underlying compute gets denser, i.e., the dotted line is the frontier shifts to the right, and as cost drops the points also move while the task value and challenge stays the same.

Why the job-loss fear is pointed at the wrong jobs?

This is the part worth being precise about, because the popular fear and the actual arithmetic don’t point at the same jobs.

Most of the fear around AI and robots assumes one simple inequality: if a robot is cheaper than a worker, the worker loses. That isn’t how the economics actually work. Displacing someone who already has a job clears a far higher hurdle than filling a job nobody is doing, and the popular fear conflates the two.

If someone is already in the seat, deploying a robot only makes sense when:

Switching cost covers retraining, workflow redesign, integration downtime, and the organizational friction of removing an incumbent. Risk premium covers the real cost of getting it wrong, from a botched rollout to a safety incident. Together these two terms set a high hurdle rate, and for an ordinary job with a reasonably productive worker already in place, that hurdle usually sits well above any price advantage a robot brings today. This is why ordinary jobs, a retail cashier or a line cook doing fine work, are not where this goes first. The hurdle is too high, and current robot economics don’t come close to clearing it.

A vacant role is a different equation entirely:

No incumbent, no disruption, and no switching cost. The bar is simply “it beats the job not getting done.” That’s why hazardous, filthy, or chronically understaffed work is exposed long before anyone’s actual job is.

A robot clears a much lower bar in a vacant role

Consider an illustrative example. Suppose a worker costs $25 an hour, while switching costs and risk add up to $14 an hour over the period being considered. The robot must cost less than $11 an hour to justify replacement:

$25 − $14 = $11

A $12/hour robot fails that test. Its advantage after those costs is negative:

$25 − $14 − $12 = −$1 per hour

Put the same robot into an unfilled role that creates $25 of value an hour. Assuming its $12 hourly cost includes deployment expenses and risk, the surplus becomes:

$25 − $12 = $13 per hour

The robot’s capabilities have not changed. But, the alternative has. In one case, a capable worker is already doing the job. In the other, the work is going undone.

That makes hazardous, unpleasant, or persistently understaffed work a plausible starting point. Where turnover is high, employers already pay for repeated recruiting, training, and overtime. A robot may compete against that recurring disruption.

Vacancies still need to be examined task by task. A role may be empty because the work is difficult to staff, yet remain difficult to automate. An employer needs evidence that a machine can deliver the required output before counting the value of a filled shift.

This arithmetic does not prove that occupied jobs aren’t replaceable. It explains why filling a vacancy can make economic sense before replacing an existing worker does. As robots improve and deployment gets cheaper, both calculations can change.

Look at opportunities where the switching-cost term is already close to zero, because there’s no stable incumbent to disrupt in the first place. From public data (BLS JOLTS):

  • BLS’s Job Openings and Labor Turnover Survey (JOLTS) for August 2026 puts the monthly quits rate (a direct measure of voluntary turnover) at 3.5% in accommodation and food services and 3.0% in retail trade, versus 1.9% in transportation and warehousing and 1.8% in construction.
  • Annualized at twelve times the monthly rate (a rough, standard approximation, not a BLS figure itself), that works out to roughly 42% a year in accommodation and food services and 36% in retail.
  • Nationally, the same report showed 7.08 million job openings against 5.19 million hires in August, a gap of about 1.9 million.

In the highest-churn sectors, this cost already shows up in the budget: constant recruiting, training new hires, overtime to cover gaps or at worst ‘lost output’, all while the vacancies remain open. Note that a robot deployed into one of these roles isn’t competing against a long-tenured employee. It is only competing against a position that already turns over every few months on its own.

Ordinary jobs, the ones most people are actually afraid for, stay protected ‘for now’.

Not by policy, not by sentiment, but because the hurdle rate is simply too high for the economics to clear yet.

Four ways to put this to work

Policymakers should back trials in dangerous or persistently understaffed work. Require operators to report safety incidents, human interventions, and cost per completed task. Match approval requirements to the task’s actual risks; a narrow task can still be dangerous. Fund worker training alongside deployment. Expand support when trials demonstrate safe, useful output, and publish failures so others can learn.

Technologists should choose one task and prove that it pays. Start with work customers struggle to staff. Define acceptable quality, speed, and safety before building. Use existing models and shared platforms where they reduce integration work. Test across full shifts and count every human intervention. Price the result per useful task, with maintenance and supervision included.

Economists should measure the cost of useful work. Track training and inference costs separately, then calculate cost per completed physical task, including downtime and human supervision. Separate deployments that fill vacancies from those that replace workers. Follow output, wages, and employment after deployment. That would let us test whether cheaper intelligence is expanding production or mainly cutting payroll.

Everyone else should examine how these machines would change their own work. Identify tasks that are repetitive, hazardous, or regularly left undone. Ask employers which tasks a robot would take over, who would supervise it, and what happens to affected workers. Seek training in operating, maintaining, or checking the systems your workplace plans to adopt. Judge promises by working deployments and published results.

After the applause

Olaf belongs under the lights. A talking snowman is theater, and theater makes an audience believe, for a moment, that something impossible has become real.

After the applause, the machine faces a less forgiving audience. Somewhere, an employer has an unfilled shift, a dangerous inspection, or work that keeps falling behind. The robot must show up, do the task safely, and earn its cost.

Descartes could begin with “I think.” A working machine will have to go further. When the lights come down, it still has a shift to finish.

What to watch next

A few concrete signals worth tracking, not predictions.

  • The density indices from Artificial Analysis and AIMultiple could keep climbing at this pace, or the easy architectural gains could run out and the curve could flatten.
  • Policymakers could actually scale certification to task narrowness, the way the policy argument above calls for, or default to one-size-fits-all review that slows deployment in exactly the vacant, dangerous roles that need it most.
  • And humanoid pricing could start compressing the way inference pricing did, which, more than any single stage demo, is what would turn this from a pilot story into a labor-market one.

Aayush Pathak is a Product Manager at NVIDIA working on edge AI and robotics platforms, including Jetson and IGX. The views expressed in this newsletter are purely his own and not of his current or previous employers and customers or partners.

Aayush has a graduate degree in Electrical Engineering from the University of Southern California and an MBA from The University of Chicago Booth School of Business. He writes from Santa Clara, California.

More from Aayush

Categories
AI & Machine Learning
Industrial – Automation & Robotics

colind88

Back To Top