I've spent a long time working with technology companies at moments when a platform shift becomes real, when something moves from impressive to inevitable. What I'm watching in AI right now doesn't feel like inevitable. It feels like a technology with genuine capability trapped behind an infrastructure that wasn't designed for it. And I think that gap is where the most significant opportunity in a generation is sitting, largely unaddressed.
I'm currently in the middle of this problem, designing AI hardware that has to earn trust in practice, not just demonstrate capability in a lab. That proximity has sharpened one conviction above all others: the reason AI keeps disappointing people in practice has nothing to do with the models. It's the hardware they're running on.
The excitement is real. The benchmarks are improving. The costs are falling. But when I talk to people actually using AI tools day to day, the experience rarely matches what the demos promised. The instinct inside companies, in the press, across the industry, is to blame the models. Assume the next version will close the gap. Keep improving the intelligence.
I don't think that's right. And I think it's sending a lot of smart people in the wrong direction.
The next stage of AI, systems that take action, manage complexity, operate with real autonomy in the world, needs something no model improvement can provide: context. Genuine, continuous, real-world context. Better reasoning is irrelevant if the system has no meaningful signal to reason about. And right now, it largely doesn't. Your smartphone can access the most capable language model ever built, but it cannot tell whether you're stressed. Your smart speaker can query the entire internet, but it cannot understand the context of your afternoon. The devices we carry and live with were designed for a different era of computing, one where humans did the sensing and machines did the processing.
No sensing, no context. No context, no useful action. That's the gap I keep coming back to. And it points directly at hardware.
When I work through what agentic AI genuinely requires from the physical layer, it comes down to three things.
The ability to sense continuously, accurately, across multiple modalities. Not a microphone that activates on command, but an environment that understands what's happening without being asked.
The ability to process locally. Round-trips to the cloud introduce latency, create privacy exposure, and break the coherence that makes ambient intelligence actually useful. Real-time intelligence has to happen close to where the sensing does.
The ability to close the loop by translating understanding into action through interfaces that feel considered rather than bolted on. Not a feature set. A grounded, physical presence that people can locate, understand, and trust.
That last point matters more than the industry currently acknowledges. Hardware isn't just the delivery mechanism for AI capability. It's the thing that makes AI legible to the people living alongside it. A physical object has edges. It has a state that can be read. It can signal when it's active, what it's doing, whether it's listening. Software alone cannot do this. And without it, capability and trust remain decoupled, which is precisely where we are now.
Most implementations I see are missing at least one of these three things. Usually two. The demos look impressive. The products disappoint. Consumers, rightly, start to lose faith in the whole category. This is the implementation gap and in my view, it is not a software problem.
Every platform shift I've worked through has come down to the same thing: not which company had the best software stack, but which product people actually wanted to live with. The physical moment of use is what determines adoption, and adoption is what determines who wins.
The obvious objection deserves a direct answer. Don't Apple, Google, and Amazon already have exactly the hardware expertise this requires? On paper, they should own this transition.
But each of these companies built their hardware in service of something else. Apple's devices exist to sell software within a closed ecosystem. Google's are surface area for advertising. Amazon's deepen e-commerce relationships. In each case the hardware is the means and the platform is the end. That's a rational business model, but it creates a structural problem when the hardware itself needs to become the relationship.
All three have tried to solve ambient AI by distributing intelligence across existing device categories: phones, speakers, watches. The result is a sensing layer shaped by product decisions made years before ambient AI was the design brief. You cannot retrofit genuine context-awareness onto hardware designed to sell subscriptions. The seams always show.
What the agentic era requires isn't better versions of existing devices. It's hardware where trust is the primary design requirement from the start, not a feature added at the end. The companies best positioned to build that are not the ones dominating the current AI conversation. The question is whether they recognise it before the window closes.
I've become increasingly convinced that the industry is asking the wrong question about AI hardware. The question isn't how to make technology perform better. It's how to make it more accountable. How to give people something they can locate, read, and control. How to design the point at which a consequential system enters someone's life in a way they can actually understand.
This matters because trust is not a feature you can add at the end. It has to be built into the physical form from the beginning, into the way a device signals its state, declares its presence, and gives people something real to hold onto. The brands that understand this, and design for it with the same rigour they bring to capability, will define what living with ambient AI actually feels like. That's a bigger prize than a better benchmark.
The instinct, when AI products underdeliver, is to wait for better models. It won't work. The ceiling on real-world AI value isn't intelligence. It's the physical layer that intelligence is trying to operate through. A more capable model with no meaningful sensing, no local processing, and no physical presence that people can orient to will still disappoint in exactly the same ways.
The companies that will define the next platform era are the ones that recognise their hardware expertise as the strategic asset the moment actually requires, not just for capability, but for trust. The knowledge of how people live with technology, what accountability feels like in practice, where relationships with products get built and broken: that is the advantage. The question is whether it gets deployed while it still is one.
This is the first in a series of essays on the physical layer of AI. The second, Threshold Objects, explores how physical artefacts can make the presence and boundaries of ambient AI systems legible to the people who live alongside them.
