Why we’re at the beginning of the AI hardware boom | Caitlin Kalinowski (ex–OpenAI, Meta, Apple)

In the labs I’m seeing the curve go almost straight up, and the purely digital gains will eventually level off. When that happens, the action moves to the physical world—robots, manufacturing, logistics—and warfare will likely evolve faster than consumer tech, so we should fund swarms and drones more than big ships.

Picture a swarm of a hundred thousand drones barreling toward us. With that on the horizon, I’m thrilled to welcome Caitlin Kalinowski, a veteran hardware leader behind unibody Macs at Apple, Rift and Quest at Meta, Orion AR glasses, and most recently OpenAI’s robotics push—plus we’ll dig into that “meteor” of rising memory prices.

Excited to dive in.

VR has dazzled for years but hasn’t gone mainstream. What happened, and is the future VR, AR, or something else?

VR taught us core spatial skills—SLAM, depth sensing, and how people see in 3D—which now power robotics, but as a consumer category it’s still niche, especially because face-covering devices are hard to make social. I’m glad we built it, and its tech is paying off elsewhere.

So where does this head now—AR glasses?

I believe AR glasses are part of the future because they let you stay present with people while getting information. Orion was ahead of its time—waveguides and microLED need better yields and lower cost—and input that’s quiet and private remains unsolved, but the lineage from AR and VR flows straight into physical AI across robots, drones, and autonomy.

Orion’s prototype had a wide, binocular view that felt naturally immersive, which is the moment AR really clicks.

I’m hearing CS enrollment is dipping while robotics rises. What’s that like from your seat?

It’s surreal. Hardware was rarely the glamorous path—outside of Apple’s legacy—so the sudden rush into robots and physical systems feels both overdue and a bit strange.

What surprises software folks when they try to build hardware?

You only get a few “compiles,” so you must be conservative, test deeply, and design for part variation to hit high yields, because once you ship you can’t patch aluminum. Miss by a little and it becomes a costly recall instead of a quick redeploy.

Why is robotics hot right now?

Many labs sense the keyboard-bound phase will plateau, so the next big unlock is acting in the real world—sensing, moving, making—and that spans factories, logistics, and eventually space.

Humanoid robots look close on demos. How ready are they, and what’s the timeline?

They’re impressive but still prototypes, and safety near people is the gating factor—lighter, softer, and compliant designs reduce impact energy, but we need real data. The next phase is making them cheaper, safer, and manufacturable at scale before they’re truly around us.

What blocks scale—hundreds of thousands or more?

Supply chain. We lack domestic depth in critical pieces like actuators, which are the motors plus gearing that move limbs. Every part has to show up on time, and today too many of those links live overseas.

Give us the state of the robotics supply chain.

From raw materials to magnets to actuators to subassemblies, decades of know‑how moved to Asia for cost and scale. To be resilient, we need more independence across that stack here at home.

Why do magnets matter so much?

Permanent magnets ring the motor and, with alternating current, create rotation. If those materials tighten, you redesign actuators around different physics, which is painful.

Watching drones in Ukraine changes how I think about defense. What should we do now?

Drones and robot arms share core tech, so military supply chains need to be as independent as possible. We should reindustrialize, relearn high‑volume production, and process key materials domestically so future shocks don’t leave us exposed.

And the carrier-versus-drone debate?

I agree we should prioritize drones over legacy platforms because AI is reshaping warfare quickly; look at Ukraine’s rapid iteration and 3D printing. The cost math currently favors offense, and we need to move fast, which America is good at when it decides to.

People worry about prompt injection for chatbots, but what about robots?

Hardening the hardware layer against adversarial control is essential—whether drones or humanoids—because mistakes in the physical world have real consequences.

Any war stories on agents going sideways?

I sandboxed an agent, told it never to share private info, and minutes later it posted my personal email. That’s a harmless example, but it shows how immature the guardrails still are.

You’ve built hardware cultures at Apple and Meta. What did you learn about doing it right?

Apple treats hardware as first‑class and pushes first‑principles rigor—the “finish the back of the cabinet” mindset that drives clarity and quality end to end. Oculus had a hacker spirit, and at Meta we kept the speed but professionalized it for yield, cost, and volume.

Give builders a playbook for shipping devices.

Define a few hard goals early and resist midstream pivots, because each full build costs months. Start with the riskiest part, over‑iterate what users touch most, and do urgent work now because surprise delays always show up.

What goals matter most?

Pick human‑centric KPIs, like pixels per degree in VR or weight and price in laptops, then make crisp tradeoffs—sometimes that means cutting features to hit the metric that matters.

That famous MacBook Air reveal—how did it come together?

The first low‑volume machined model proved CNC laptops were real; the next wedge‑shaped rev scaled and became the one everyone knows.

Do great products come from user research or vision?

If you’re going zero to one, customers can’t ask for what they’ve never seen, so you lead with conviction; once they see it, feedback gets very useful, but chasing incremental input too early kills breakthroughs.

About that “meteor” of memory prices—what’s going on and how to prepare?

AI demand and tight capacity are spiking prices; I tell startups to pre‑buy if they can and plan for shocks, since data centers will outbid consumer devices. A doubling wouldn’t surprise me, though timing is anyone’s guess.

By memory, you mean fast RAM, not long‑term storage, right?

Yes—working memory that programs actually use at speed; in products it becomes a packaging and cost tradeoff against performance.

One missing part can halt a device. How bad can it get?

If silicon or RAM vanish, it’s a catastrophic redesign—new board, re‑qualification, re‑test, and schedule slip; a simple casting vendor dying is annoying, but chips are devastating.

How complex is something like a Matic vacuum mop?

Dozens to over a hundred parts at the assembly level and thousands of components if you count every piece, with on‑device mapping for privacy, motors, pumps, SOC, RAM, radios, and PCBs.

Is this why vertical integration helps?

Yes—owning more of the stack lets you swap chips and re‑spin boards fast in a shortage, which we’ve seen done at remarkable speed.

Prototype with off‑the‑shelf parts or go custom?

Grab anything off‑the‑shelf to prove it works, then go custom where needed to hit size, weight, cost, and reliability in mass production.

Is AI changing CAD and hardware design yet?

We’re early—models can sketch surfaces and route PCBs, and they’re great for planning and spreadsheets, but real solid CAD and physics‑aware reasoning still need new model types that understand contact, friction, and materials.

Do we really need humanoids, or will task robots win?

Most jobs are better done by dedicated machines, and top‑tier factories already run with very few people; expect a mix—logistics, construction, and specialty robots—plus some humanoids for long‑tail tasks.

When do robots build other robots and designs?

A pipeline from a sketch to full CAD to vendor feedback is plausible, but the blocker is training data—company CAD is precious IP, so I see hobbyists leading first, with on‑prem AI for enterprises later.

What makes a robot feel human and not creepy?

Signal awareness and intent with motion, react to people, appear soft and non‑threatening, and avoid sudden moves; animation studios like Pixar and Disney have mastered conveying emotion and approachability that robotics can borrow.

Give us your five‑year view.

Design near‑term products that ladder into a longer vision, because AI will transform knowledge work quickly while physical change lags due to factories and materials; expect more street robots and a safety‑first mindset, and faster innovation in defense than in consumer devices.

Why did you leave OpenAI?

I care about many people there, but I disagreed with the speed, governance, and guardrails around the defense deal announcement. I couldn’t stay not knowing what the next decision might look like, and I hoped my exit would help others state and hold their boundaries.

How do you hire for zero‑to‑one hardware and robots?

Stack strong generalists with a few deep robotists and AV veterans, add AI‑native new grads who solve with models from the ground up, align everyone on mission, and look for curiosity, openness to update beliefs, and a real drive to win.

Top lessons from working with Sam, Steve, and Mark?

Sam pushes you to think in hundreds‑fold scale. Steve held an unwavering quality bar that sharpened everyone’s work. Mark and Boz ran crisp, technically deep reviews with decisions pushed to the edges so the org could move fast.

Share a favorite failure and what it taught you.

Mid‑Quest 1 we misread a camera tolerance spec, lost tracking, and at EVT had to lock two cameras on a rigid bracket while letting the others float; we still shipped on time, and the new pairing became a more reliable source of truth.

Thanks for listening. If this episode was helpful, follow the show on Apple Podcasts, Spotify, or wherever you listen, and consider leaving a quick rating or review so more people can find it.

You can explore all past episodes and learn more at lennyspodcast.com. See you next time.