Ecosystems, Not Specs
Like the doomed microcomputers of the 1980s, bespoke AI gadgets prioritize novel specs over the dominant software ecosystems that control the value chain.
By Ben Thompson
Sparked by Apricot Computers: An underrated British brand · discussion
The first British microcomputer to feature a 3.5-inch floppy drive wasn't made by IBM, nor did it run standard PC software out of the box. It was the Apricot PC, launched in 1983, and it was a marvel of technical specification. If you look at the company's portable model released a year later, it featured a remarkable 640x256 pixel LCD screen, speech recognition software, and an astonishing industry first: an infrared trackball/mouse.
Apricot was an underrated British brand that consistently shipped better, sleeker hardware than its American counterparts. And yet, the company was utterly doomed.
To achieve this tight integration of premium components, Apricot used a custom BIOS. This allowed them to optimize their display and peripherals, but it broke deep hardware compatibility with the IBM PC standard. You could run a tailored, ported version of MS-DOS, but standard off-the-shelf software built for the IBM ecosystem would often fail.
That is the problem: Apricot attempted to differentiate through integrated, premium hardware in a market where the software layer was actively dictating modularity.
In the early 1980s, the value chain was rapidly shifting. IBM's original PC had commoditized the hardware layer by using off-the-shelf components. This shift pushed all the strategic value up the stack, allowing the operating system — Microsoft's DOS — to become the actual platform. Apricot was fighting a structural war with superior industrial design, oblivious to the fact that the battleground had moved. Reading through a recent Hacker News discussion about the company's history, the retroactive consensus is stark. Commenters rightly note that Apricot’s machines were technically superior in almost every way to the IBM PC XT, but that advantage was rendered entirely moot by the software ecosystem. Superior hardware, when disconnected from the dominant software layer, is an isolated island. It doesn't matter how good your infrared mouse is if you cannot run Lotus 1-2-3 natively.
The tech industry, though, is remarkably adept at repeating this exact structural error.
Over the past year, a wave of bespoke AI hardware has flooded the market. The two highest-profile examples are the Humane Ai Pin and the Rabbit R1, alongside a smattering of custom desktop AI wrappers and pendants. The prevailing narrative in tech circles is that these are simply failed early iterations in a new computing epoch — misguided, perhaps, but necessary first drafts of a post-smartphone future.
In reality, they are modern incarnations of the Apricot mistake. These companies are attempting to sell isolated, vertically integrated hardware into a paradigm where the intelligence layer is already fiercely modularized.
Before reaching for any grand theories about computing platforms, it is necessary to follow the actual economics of an edge AI query.
Consider the Humane Ai Pin. The device demands a $699 upfront cost, plus a mandatory $24 monthly subscription to use it. Why a subscription for a standalone device? Because bespoke AI hardware operators are caught in a brutal, negative cash flow margin trap. The hardware maker collects a one-time markup on a physical piece of aluminum and glass, but they owe perpetual, marginal API costs to foundation models like OpenAI or Anthropic at the bottom of the stack.
Think about the value chain of a single query. A user taps the pin and asks for a restaurant recommendation. Humane captures the audio (costing battery and local compute), sends it to the cloud (costing bandwidth), routes it to a large language model (incurring a per-token API fee), and synthesizes a voice response (incurring a text-to-speech API fee). Every time a user interacts with the device, the hardware maker bleeds cash. The company is effectively acting as an arbitrageur between a fixed hardware sale and infinite variable computing outflows. To put it another way, their business model requires charging you a monthly fee just to prevent their most active users from bankrupting them via usage costs.
This margin squeeze at the bottom is compounded at the top of the stack, where these devices lack any structural differentiation.
Take the Rabbit R1. It was pitched at launch as a completely new computing interface, driven by a proprietary "Large Action Model" that would bypass traditional apps entirely. The structural reality, revealed by brutal device teardowns weeks later, is much more pedestrian: the R1 interface is essentially powered by a single Android app running on a heavily modified version of AOSP (Android Open Source Project).
They are not building a new computing paradigm. They are renting server infrastructure from Google, licensing intelligence from OpenAI, wrapping it in AOSP, and trying to capture a premium on a bright orange plastic shell. The operating system layer has already been decisively won by iOS and Android, who possess the ultimate moat: the default customer relationship, the authenticated payment credentials, and the distribution channels to every third-party service on earth.
This dynamic forms a clear 1:1 historical map to 1983.
In the PC era, IBM modularized hardware, pushing all the strategic value up to Microsoft's DOS. Today, the foundational models are modularizing the intelligence layer, while the incumbent mobile OS monopolies retain a stranglehold on the user endpoint.
If we diagram this out in prose, the AI endpoint market sits on two axes. At the bottom are the capital-intensive compute providers and foundation models, commanding the margin for raw, modularized intelligence. At the top are the smartphone operating systems, commanding the margin for distribution, identity, and user attention.
There is simply no money in the undifferentiated middle.
The framework at play here is Integration versus Modularity. Apple can successfully integrate AI into the iPhone because they already own the modularized app ecosystem and the user's pocket. They are integrating new intelligence into a mature, high-margin distribution node. Humane and Rabbit, conversely, are attempting vertical integration from scratch. They are trying to build a new hardware form factor, a new operating system wrapper, and a new user habit all at once, in a market where the underlying intelligence APIs are already standardized commodities. They are Apricot Computers, boasting about an infrared trackball while the rest of the world standardizes on someone else's software layer.
This is why the romantic framing of a "post-smartphone future" is so fundamentally flawed.
Consumers are not looking for new form factors or voice-first cameras to pin to their lapels; they are looking for AI access embedded into the devices that already house their digital lives. The solipsism required to assume a user will abandon the smartphone ecosystem — which contains their iMessage groups, their photo libraries, and their Apple Pay wallets — for a standalone AI gadget is remarkable.
The endgame for consumer AI at the edge is not a proliferation of bespoke hardware widgets. It is an API call made by iOS or Android, routed to a foundation model running in an Azure or AWS data center, and returned directly to the glass screen you already stare at for six hours a day. The hardware companies that survive this epoch will do so by understanding the structural reality of the value chain, not by fighting it with clever industrial design.
Disruption doesn't come from a new plastic shell. Ecosystems, not specs.