I’ve spent two decades building consumer hardware: as an executive at traditional hardware companies, inside large internet firms, and as a startup operator. Most recently at Dreamer, an AI startup, I was a one-person hardware team. In practice that meant taking a consumer product from definition through manufacturing, logistics, Amazon fulfillment, and the DTC store in four months. We were days from launch when the company was acquired. What follows is what I learned doing that at that scale and against that clock, and what I would now tell anyone building an AI hardware company.

The Apple Playbook Is Too Heavy Now

For twenty years, “build hardware like Apple” has been the default ambition for anyone building a hardware company: the founder pitching investors, the executive standing up a new org, the program manager looking for a template. It is the gold standard, the safe answer in the board meeting, the benchmark everyone measures themselves against. And it should be questioned, not because Apple was ever wrong, but because the two conditions that made its playbook correct have both quietly collapsed.

It is easy to dismiss the Apple playbook as bureaucracy. It was not. It was a rational response to two facts about the world.

The first fact: the supply chain wasn’t strong, so the knowledge had to live inside Apple. When Apple built its machine, there was no mature ecosystem you could simply hand a hard problem to. The metallurgy, the display tuning, the thermal architecture, the radio stack, the assembly tolerances: none of that expertise existed as a purchasable service. So Apple internalized all of it. It hired the world’s best people for every layer of the stack, integrated vertically, and ran a heavy, gated process (EVT, DVT, PVT, hundreds of people orbiting a single SKU) precisely because the knowledge had nowhere else to live.

The second fact: everyone else was slow, so perfection beat speed. In a world where every competitor moved slowly, “take three years and ship something flawless” was a winning strategy. You were slow, but you were perfect, and your rivals were slow and imperfect. Time was not the binding constraint; execution quality was.

Both of those facts have reversed.

The supply chain is no longer downstream of the innovation. It has become the source of it. The clearest evidence is in the one category Apple is supposed to own: the phone. Foldable displays, the most significant smartphone form-factor shift in a decade, were pioneered not in Cupertino but by Royole, a Chinese company that shipped the first commercial foldable in 2018. The entire category that followed (Huawei, Xiaomi, Oppo, Vivo, Honor) was built out in China while Apple watched, eight years behind, with a foldable iPhone finally rumored for late 2026. Fast charging, full-screen industrial design, ceramic bodies, periscope optics: the pattern repeats across the spec sheet. These were not innovations that leaked out of Cupertino into the supply base. They were born in the supply base.

And speed has overtaken perfection as the thing that decides outcomes. Chinese manufacturers don’t just innovate. They iterate at a cadence that the gated, perfection-first process simply cannot match. “Slow but flawless” now loses to “fast and continuously improving,” because by the time the flawless thing ships, the fast competitor has shipped, learned, and shipped again three times.

Take the iPhone display. Every generation is locked years in advance: the panel supplier chosen, a dedicated production line built, yields ramped, two to three years before the phone reaches a customer’s hand. In a world where a screen’s value (brighter, more efficient, higher refresh) is a safe bet that still holds three years later, that is fine. The thing you’re betting on doesn’t expire.

AI breaks that. In the AI era, the question you actually need to answer is: is this idea even real? And you need to answer it in three months, because the model capability your hardware was designed around may have shifted underneath you in that window. The feature that justifies the device today may be something the model just does natively four weeks later, and now the hardware you committed to isn’t useful anymore. When the half-life of your core assumption is three to six months but your hardware planning horizon is two to three years, you are not polishing a product. You are pouring capital and organizational energy into a bet that is likely to be invalidated while it is still being poured. The heavy process treats “locked two to three years in advance” as a mark of rigor. Today it is the single largest liability you can carry.

Someone will point at Apple Silicon and say: Apple bet big on vertical integration, and won. They did. But M-series is not a counterexample. It is the perfect illustration. Apple has had chip design capability for thirty years; they were a founding investor in ARM in 1990. They iterated quietly through eleven A-series generations on iPhone before bringing the architecture to Mac. They waited until A14 had outperformed Intel’s desktop silicon, until 20 million Macs a year could amortize the investment, until Intel was clearly failing them. Then they moved: cautiously, MacBook Air first. That is not the Apple playbook beating the AI era. That is the Apple playbook doing what it was always good at: patience, accumulation, scale, forced moves. None of those conditions are available to a founder.

And this is the trap. Anyone building a new hardware team in 2026 instinctively reaches for the Apple template (more people, more gates, more vertical integration, more polish before launch), and in doing so they build an organization superbly optimized for a world that no longer exists. The weight isn’t a sign of seriousness. It’s a sign that the team is solving the 2007 problem in 2026.

The harder question is what you build instead. Because “lighter” is not an answer. It’s a direction.

The Capability Axis Has Flipped

So what does lighter actually mean? It does not mean fewer people. A 30-person team running the old playbook is just as broken as a 300-person team running it. What changes is the axis the team’s capability is organized along.

The old axis was vertical depth. Display, acoustics, RF, thermal, mechanical, battery, electrical: each a deep well, staffed with the world’s best specialists, owned end-to-end inside the company. This was correct when the relevant expertise didn’t exist anywhere else. You had to internalize it because there was no alternative.

The new axis is judgment and trade-off. The team is organized around a different question: can we validate this assumption in three months? Or even two? The core competence is no longer “we can build anything.” It is “we know what is already mature in the ecosystem, what we have to attack ourselves, and what we should not be attacking at all right now.”

The capability axis has rotated 90 degrees. From depth to judgment.

Two examples should make any hardware leader uncomfortable, because they are exactly the things instinct says you have to own. FCC compliance, once handled by internal regulatory teams running their own pre-compliance labs, is today done faster, cheaper, and better by specialized third-party labs in Taiwan, Shenzhen, and California, because it is their entire business. CMF (color, material, finish), once owned by a famous in-house industrial design group, now sits as a complete ecosystem in Shenzhen and Dongguan, ready to sample within weeks. You do not need internal teams for either of these. You need someone with the taste to pick the right vendor and the instinct to know which finish will survive at scale.

These are not edge cases. The same pattern repeats across the whole hardware stack: radio modules, antenna design, battery packs, thermal solutions, significant portions of mechanical and electrical engineering. The ecosystem has gotten so good that almost everything that used to demand an internal deep well is now something you can buy, partner with, or commission. The companies still trying to internalize these capabilities aren’t building a moat. They are rebuilding wheels the ecosystem has already perfected, and they are paying for it twice: once in headcount, once in the speed they are losing to teams that didn’t bother.

So where did the moat go? For an AI hardware company, it is not in the device. The moat is the experience that emerges when the device meets the AI behind it, and that experience is a function of how fast you can iterate the integration between the two. Hardware that doesn’t change isn’t a moat. AI that improves but isn’t well-integrated into the device isn’t a moat either. The moat is the gap between a team that can iterate hardware and model together at the AI’s cadence and a team that cannot. And that gap compounds, because every cycle the model moves and the slower team falls further behind on what the product actually does. This is what makes the cadence question (three months, two months, can you go faster) load-bearing rather than stylistic. It is the moat itself.

Inside that moat, two things are actually scarce, and they don’t look like traditional hardware expertise on a résumé. Which is exactly why most companies miss them.

The first is product taste. Taste in the artistic sense: aesthetic judgment about what a device should look like, feel like, behave like; which experiences are worth offering and which ones cheapen the product; which 5% of details matter and which 95% don’t. In a world where any competent Shenzhen supplier can build whatever you spec, the difference between a product worth using and a product that gets returned is taste. This is also the capability that current AI tools cannot replace, because taste is the kind of judgment that has no training set. It is accumulated, idiosyncratic, often contrarian.

The second is product instinct. Instinct in the technical sense: engineering judgment about which trade-offs to make. Which subsystem belongs on-device and which belongs in the cloud. Which spec to push and which to accept. When an imminent model capability shift will invalidate your hardware bet, and when it will not. Which Chinese chipset family to commit to given the geopolitical trajectory. Whether to do a custom board or use a reference design. In the old playbook this instinct lived implicitly inside a chief engineer with twenty years of experience. In the new playbook it has to be explicit and distributed across the team, because the cadence the AI era demands cannot bottleneck on one person.

Taste and instinct (artistic judgment about what to make, technical judgment about how) are what is scarce now. Almost everything else is increasingly buyable.

The capability axis flip is the actual reorganization. Everything else is cosmetic.

What the AI-era Hardware Team Looks Like

So what does this team actually look like?

Small. Ten people, not three hundred. In a hardware org of ten there is no room for an “RF antenna group” or a “firmware team.” What you need is a few PMs and engineers with ambition that doesn’t quit and the ability to learn fast across disciplines.

A good firmware engineer, by walking the supplier floor and asking the right questions, can combine the right RF reference design with software-side power management to hit your power budget. One person like that is worth more than an RF team plus a firmware team plus an antenna design team. Most hiring rubrics will filter them out.

This is more possible now than ever, because AI tools have compressed the cost of climbing into adjacent disciplines. A firmware engineer can learn in an afternoon what used to take a month. At Dreamer I ran the hardware program as one person by building the workflow around AI: reviewing ODM plans, cross-checking what suppliers said across meetings, managing development, automating the Amazon launch. I handled the execution. My cofounder confirmed. That was the loop.

Once you have that team, three things matter more than anything else.

Don’t over-design. Hardware is expensive right now: every component, every tooling run, every prototype iteration. Use cheap parts wherever they work, and off-the-shelf modules wherever they exist. The graveyard of hardware startups is full of beautifully over-engineered first versions whose founders ran out of money before anyone knew whether the thing was worth building. Get to PMF first. There is plenty of time to be elegant after you have customers.

Design the cloud-edge boundary the same way: as something that lives, not something that freezes. The model improves every six weeks; the feature that needs to run on-device for latency today may be cloud-native next quarter. Build the BOM and connectivity envelope so the boundary can move without a hardware respin.

Find real partners, not vendors. Real partners are teams whose people light up when they talk about finding the weird component your spec needs, clawing yield up point by point on the line, running color sample after color sample until your finish is right. They care about the work. The best Shenzhen partners I have worked with weren’t vendors I paid; they were people I built with. The kind whose family you ask about, who call you at midnight about a yield problem, who you call back at midnight when your launch is on fire.

Most founders never get there. They go looking for vendors who do exactly what they are told, and they find them, and the vendors are mediocre. The Shenzhen teams whose core competence is catering to Western masters who don’t know what they want gave up on excellence a long time ago.

These teams are extraordinary at managing you. They will make you feel great. You will be convinced you are getting exactly what you asked for. Then, when the timing is right for them, they will deliver the bad news, and somehow you will end up paying for it. Eventually they will hand you a twenty-page report explaining why it could not be done, even though the actual solution lives one block away.

The genuinely interesting hardware is being done by the teams who insist on being peers, not contractors.

Pack up. Go to Shenzhen and the Yangtze River Delta. Bring a translation app. Walk the shop floors, eat the meals, see what gets people excited. You will be surprised, repeatedly, by what a fifteen-person shop in Dongguan can do, and care about, that a two-hundred-person team at a Tier-1 contract manufacturer cannot.

Break the validation theater. The EVT/DVT/PVT stack most hardware founders inherit was designed for products that took five years to develop and stayed in the field for thirty: IBM-era mainframes, defense systems, things where a missed validation meant catastrophic recall. You are not building mainframes. Calibrate your process to what you are actually shipping. A PoC needs almost no validation; you just need a working unit. A dogfood batch needs basic reliability. A 1,000-unit pilot needs more. Mass production needs the full stack. Almost every founder makes this mistake: running mass-production-grade validation on a PoC. That is how you spend six months and a million dollars proving nothing.

There is a Chinese phrase from the martial-arts tradition: 天下武功，唯快不破 — among all the martial arts, only speed cannot be defeated. That is the team you are building.

This is uncomfortable for executives whose intuition was built in the heavy era. It looks under-resourced. It looks risky. It looks like giving up control of things you used to own. And most importantly: it will be challenged as having no moat, by boards and investors whose strategic frameworks were built in the heavy era. All of that is true in some narrow sense, and all of it misses the point: the heavy team is not actually doing the things you think it is doing. It is solving the wrong problem at high cost, on the wrong timeline, with the wrong moat. The light team gives those things up because they were never the source of value to begin with. The real moat is iteration speed against an AI that is itself iterating. It is exactly what the heavy team cannot produce. Which is why building a team that can is no longer a technical problem. It is a problem of conviction.

The team that wins the AI hardware era will be the one whose leadership has the conviction to build for the world that is, not the world that produced the playbook. The capabilities are buyable. The processes are designable. The people are recruitable. The conviction is the bottleneck.