Software Still Needs Hands

The intelligence is no longer the bottleneck. The hands are.

That’s the uncomfortable shape of the next few years, and it inverts a decade of intuition. We spent the 2010s assuming the scarce thing was smarts — the model, the algorithm, the IQ in the room. Now you can rent a near-frontier model by the token and point a thousand of them at a problem overnight. So the constraint moves. If a system can think but can’t make — etch a wafer, pipette a reagent, bond a die — then making is where the world slows down. As one builder put it on a recent conversation, “the software still needs hands,” and “if it can’t make things, then those are real, real boundaries.” This piece was sparked by The AI Industrial Revolution with Naval Ravikant; the line stuck because it’s a builder’s claim, not a philosopher’s, and it’s testable against what’s actually shipping in 2026.

The thesis is sharper than “AI can’t do physical stuff yet.” It’s directional: intelligence will keep getting cheaper and better, so the leverage is in instrumenting the physical world — wiring up foundries and labs so a smarter model can show up immediately as a real-world output instead of a suggestion. Build the hands, and every future model upgrade lands in atoms the next morning.

The deployment gap is measured, not vibes #

The cleanest evidence is robotics. a16z’s analysis of physical AI found that “the vast majority of robots in production environments remain narrowly preprogrammed, executing fixed routines in carefully controlled conditions” — and that a policy hitting 95% success in the lab can fall to roughly 60% in real deployment once lighting, textures, and angles drift off the training distribution (a16z). That 35-point cliff is the whole argument in one statistic. Every percentage point of that gap is, today, a human hand reaching in to unstick the thing. The intelligence isn’t what’s missing in that moment. The robustness of the embodiment is.

This is why the interesting work right now isn’t a smarter planner. It’s the boring layer underneath: fixtures, sensors, transport, calibration — the stuff that turns a 95%-in-a-clean-room policy into something that survives contact with reality.

Foundries are being wired up in real time #

The semiconductor industry is running the experiment live. At GTC Taipei, NVIDIA and TSMC announced a program to push AI into the fab itself: computational lithography (cuLitho) targeting 20–50% improvement in cost-effectiveness or cycle time, and process simulation (cuEST) running 50x faster chemistry modeling (NVIDIA). This is exactly the “instrument the foundry” move — and notably, the same announcement is candid about the limit. It states plainly that a fab still requires “precise coordination across tools, materials, robots, humans and facility systems.” The most advanced AI-in-the-loop fab on earth still lists robots and humans as load-bearing.

The practitioners say it louder. At the 2025 Advanced Semiconductor Manufacturing Conference in Albany, executives from Intel, GlobalFoundries, and EMD Electronics named the hard constraints on AI in chipmaking: black-box decisions you can’t explain, data scarcity behind proprietary walls, model-validation gaps, and the failure to standardize across companies (Manufacturing Dive). Every one of those resolves to a human standing at the physical layer, verifying. Intelligence doesn’t dissolve those walls. Plumbing does.

There’s also a downstream chokepoint worth flagging carefully. Advanced packaging — the physical step of bonding many dies together (CoWoS) — has reportedly become the binding supply constraint for AI silicon, with packaging capacity oversubscribed and steps spilling over to outside assemblers. The most-cited account of this (CNBC, April 2026) is paywalled and blocks automated verification, so treat the specific figures as secondhand. But the structure of the claim is exactly the thesis: the scarce thing isn’t compute design, it’s a physical bonding process that doesn’t yet exist at scale outside a few buildings in Taiwan. You can’t prompt your way out of a missing factory.

Labs are the purest case #

If foundries make the point, labs prove it. A CHI ‘26 study, “Beyond the Desk,” catalogues why scientists can’t simply hand physical experiments to AI: the experiments are too high-stakes to risk model errors, the environments are too constrained for general-purpose automation, and — the deepest one — AI can’t access the tacit knowledge of hands-on bench practice (arXiv). That third barrier is the killer. A lot of what a good experimentalist knows isn’t written down anywhere a model can read it; it lives in their hands.

So the response isn’t a smarter chatbot. It’s hardware. Ginkgo Bioworks’ Cloud Lab is the proof-of-concept the thesis is reaching for: a browser interface onto an autonomous Boston facility built from Reconfigurable Automation Carts — modular units with robotic arms, maglev sample transport, and 70+ instruments spanning liquid handling, readouts, and incubation (Ginkgo Bioworks). A scientist submits a protocol in plain English; the physical fleet runs it. Strip out the carts and the AI has nothing to act on — it’s a brilliant disembodied advisor. The intelligence was never the hard part of that system. The carts were.

The honest counter-case #

The thesis is directionally right, but a builder should know where it’s softest.

Lights-out manufacturing already exists — just narrowly. FANUC has run near-unsupervised robotic production in Japan for over two decades, with cells operating for days without a human. The catch is that these are high-repetition, tightly-bounded processes, not the adaptive general-purpose making the thesis imagines. The hands aren’t gone; the problem was just made rigid enough not to need them. Full lights-out chipmaking remains aspirational, with industry watchers putting broad adoption around 2035.

The physical envelope is moving faster than “always needs hands” implies. Autonomous reactors running Bayesian-optimization loops can explore chemical space far faster than manual benchwork, and Ginkgo is decommissioning traditional benches outright in 2026. The “self-driving labs” framing has its own caveats — one widely-shared Substack analysis describes today’s systems as “artisanal server farms” still wrestling with hardware heterogeneity, calibration drift, and data standardization (it proposes an autonomy-level framework, though I’d be cautious citing any precise “most systems are at Level N” figure — that specific prevalence claim isn’t well-substantiated). But the trendline is real: the physical surface is expanding, not frozen.

And the binding constraint may not stay physical. MIT Technology Review argues that as physical AI scales, “trust becomes the limiting factor” — the gate becomes governance, safety, and observability, not raw capability (MIT Tech Review). That’s a friendly amendment, not a refutation. It sharpens the thesis: the wall isn’t only “can’t make things,” it’s increasingly “won’t be allowed to make things unsupervised” until the trust scaffolding catches up.

What to actually do with this #

Strip away the futurism and there’s a concrete instruction for builders. If intelligence is getting cheap and embodiment is the bottleneck, then the highest-leverage thing you can build is a physical surface a model can act through — and the durable moat is the instrumentation, not the model riding on top of it.

That reframes the work. The valuable asset isn’t the cleverest agent; it’s the wired-up fab, the cart fleet, the calibrated rig, the sensor mesh — anything that converts “the model knows what to do” into “the thing got done.” Build that layer, and you’ve created a place where every future intelligence upgrade lands as a real-world output for free. Skip it, and you’re stuck doing what most of the industry does today: generating excellent advice that still needs a human to walk over and execute it.

The slogan is right. Software still needs hands. The opportunity is to be the one who builds them.