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

What It Covers

Caitlin Kalinowski — hardware leader with tenures at Apple, Meta (Oculus/Orion AR glasses), and OpenAI's robotics division — maps the convergence of AI and physical hardware. She covers why digital AI capabilities will plateau and push innovation into robotics, the fragility of global supply chains for actuators and memory, humanoid robot safety, and what it takes to build hardware programs from scratch.

Key Questions Answered

• •Hardware compilation constraint: Unlike software engineers who can redeploy code daily, hardware teams get roughly four to five total design iterations across an entire product's lifetime before mass production locks everything in. This forces a fundamentally different discipline: define KPIs upfront, change them as rarely as possible, and treat every build cycle as expensive and irreversible. Companies transitioning from software to hardware routinely underestimate this constraint and burn months on avoidable redesigns.
• •Design the hardest part first: Experienced hardware architects identify the highest-risk physical constraint — the pinch point where the design is most likely to fail — and resolve it before touching familiar components. Kalinowski cites routing cables through a laptop hinge as an example: the architect started with cable diameter and hinge clearance, not the display or chassis. Most teams do the opposite, defaulting to what they already know how to build, which delays discovering fatal constraints until late in the program.
• •Memory price shock incoming: AI data center demand is consuming DRAM supply at a rate that consumer hardware and robotics companies cannot compete with on price sensitivity. Prices have already risen significantly, with estimates suggesting further doubling on an uncertain timeline. Kalinowski advises hardware startups to pre-buy memory inventory now to buffer against supply spikes, accepting the risk that prices could fall, because the alternative — halting production — is more damaging than overstocking.
• •Actuators are the robotics supply chain bottleneck: The motor-and-gearing assemblies that power robot limbs depend on rare-earth magnets processed predominantly in China and Japan. This supply chain was deliberately offshored over 25 years, and rebuilding domestic actuator manufacturing capability does not yet exist at scale in the US. Even prototype-stage robotics teams face one-to-two month lead times just to source actuators for testing, making supply chain independence a prerequisite for any serious robotics program.
• •Humanoid robots require softness and mass reduction for safety: Current humanoid robots capable of meaningful physical work carry mandatory warnings prohibiting humans within three feet. Kalinowski points to One X Neo as a design that addresses this by pulling mass inward, reducing arm weight and using compliant materials to lower impact energy. Two separate factors determine injury risk: the kinetic energy of the moving limb and the impulse delivered on contact, both of which must be engineered down before humanoids operate alongside people.
• •Robot social design borrows from Pixar, not engineering: Researcher Leila Takayama's work shows that robots must signal intent before moving — looking in a direction before turning, for example — to avoid triggering threat responses in nearby humans. Stationary or unresponsive robots read as creepy. Kalinowski argues Pixar and Disney hold the deepest institutional knowledge for designing approachable, emotionally legible characters, and that robotics teams should study animation principles for conveying attention, softness, and non-threatening intent through physical movement.
• •AI-native junior engineers are a hiring priority: Engineers in their early twenties who have built their entire problem-solving workflow around AI tools operate fundamentally differently from experienced engineers who adopted AI later. Kalinowski actively recruits these individuals to teach senior team members how to think AI-first, not just use AI as a productivity add-on. She frames this as analogous to how internet-native engineers outpaced predecessors in the early web era — the cognitive model, not just the tool proficiency, is what transfers.