Tokens and Operating Systems

2026-04-26

In 1956, Malcom McLean loaded 58 aluminum truck bodies onto the SS Ideal-X in Newark and sailed them to Houston. He had been thinking about it since the 1930s, when he sat in his truck at a port in Hoboken watching stevedores unload cargo one crate at a time: “I was looking at a lot of wasted time and money.” The standardized shipping container hollowed out the middle of the freight industry within two decades. The Port of New York’s commission agents, the Brooklyn waterfront where longshoremen had unloaded mixed cargo by hand for a century, the piers that once employed tens of thousands: gone. Below the box, capex and port concessions concentrated shipping into a handful of global carriers; a modern Triple-E carries 18,000 TEU at roughly $190M per ship. Above the box, Walmart, Toyota, and Sony built supply chains regional rivals could not match: store networks, supplier relationships, brands, things the container could not synthesize. The middle had been thousands of forwarders, brokers, and regional shippers. By 1980, it was a graveyard.

AI is the same shift on a steeper curve. Two company shapes survive. Token distributors sell raw intelligence by the token, win on capex and distribution, and consolidate to a handful of global players. Operating systems sell intelligence applied to something the model cannot synthesize: atoms, relationships, attestation, or proprietary context. The middle, what we currently call SaaS, gets hollowed out.

Below the Box, Above the Box

A token distributor sells inference. The product is a number of tokens at a latency and a price. Hyperscaler economics rule: training runs in the tens of billions, talent fungible only at the margins, distribution that compounds with usage. Margins compress toward zero the way commodity prices do. Expect roughly the same population as cloud hyperscalers; the inputs to scale are the same.

An operating system is a company that owns one or more of four things the model cannot synthesize: atoms, relationships, attestation, or context.

Tesla’s factories, a utility’s wires, a port operator’s berths, a regional dialysis chain’s clinics: these are atoms. They exist in space. They have to be paid for, permitted, maintained, and operated. No prompt produces them, and no one builds them overnight.

The bank holding trillions in client assets, the Fortune-100 CRM, the primary care physician at the center of a patient’s chart: these are relationships. Trust, brand, decades of compliance posture, fiduciary standing, all accumulated slowly and transferred even more slowly.

Auditors signing 10-Ks, ratings agencies, SOC-2 firms, state licensing boards, UL Labs: these are attestation. They sell signed truth about someone else’s substrate, and the signature is regulated, insured, and not synthesizable from a model.

Bloomberg’s terminal, Epic’s installed base, an ERP with thirty years of joins no model has seen: these are context. Proprietary data assembled by people, in a specific place, over a specific time, with assumptions baked into every column that no one fully remembers but everyone depends on.

A company is an OS if it owns at least one of these and the others run through it. If it owns none, it is a SaaS interface to commodity intelligence, sitting inside the box rather than above it. The structural test: hand a competent in-house team a frontier coding agent and a quarter of runway. Can they reproduce the product? If yes, the product was the middle. Try the test on your own company. What could they rebuild, and what would they miss?

Scorched Adjacencies

The mechanism is older than AI: platforms with a profitable core commoditize their adjacencies, and anything charging margin between the user and the core dies. In 2008, Andy Rubin’s team at Google released Android as an open-source operating system and gave it to every handset manufacturer who wanted it. Google did not need to profit from the OS; it needed the OS on every phone so that Google Search stayed the default, which protected the advertising core. By 2013, Android ran on 80% of smartphones sold worldwide, and the margins of every mobile software company that sat between user and search compressed toward zero. Amazon ran the same play with AWS: price infrastructure below profitability thresholds competitors could not match, and let the cheap infrastructure feed e-commerce traffic back to the core.

The token distributors are running this play one floor up. The core is token consumption. They price everything that sits between the user and the model toward zero: coding assistants, search, translation, document analysis, customer support tooling. They do not need to win those categories; they need to make them cheap enough that no SaaS company can charge a margin on top of inference. The margin destruction is deliberate, driven by the distributors themselves. Generation cost falls toward zero, switching cost follows, and the middlemen vanish the way the freight brokers did.

Some categories survive this pressure because no amount of scale lets a distributor reach through the screen and into the clinic, the courtroom, or the port. Atoms, relationships, attestation, and context resist commoditization not because the distributor cannot ship a feature, but because it cannot ship a network of physical operators, a thirty-year compliance posture, a signed audit trail, or a longitudinal proprietary dataset.

The deeper reason is taste. An OS accumulates a reward function tuned to its domain: which edge cases matter, which workflows users actually run, which failures are catastrophic versus cosmetic. Consider a dialysis chain at 3 AM: an alarm sounds, and the nurse reaches for the silence button because she’s heard this one a thousand times. But another alarm, with a slightly different cadence, means a patient is crashing. She knows the difference. No training manual taught it to her; ten thousand shifts did. A freight broker knows which carriers ghost loads on Fridays and which ones answer the phone. That knowledge accrues through thousands of shifts and seasons, encoding in product decisions no distributor will ever see. Taste cannot be shipped from outside. The distributor sends features; the OS sends judgment about which features matter.

A Thousand Local Monopolies

The token layer concentrates because its inputs are fungible globally: GPUs, electricity, transformer architectures, web-scale text. The OS layer fragments because its inputs are not. Healthcare alone produces hundreds of vertical OSes (dialysis chains, radiology networks, regional EHR integrators, pediatric specialty groups), and the same is true in industrials, logistics, regulated finance, education, government procurement. Each one looks more like a regional bank than a horizontal SaaS company. But the long tail is selective, not democratic: a vertical AI workflow tool that owns none of atoms, relationships, attestation, or context is a prompt waiting to be written, and the first time the customer’s internal team writes that prompt, the wrapper evaporates. Selling a slim layer over a frontier model and a Postgres database is a doomed middle in vertical clothing.

What the Agent Can’t Build

If every business can spin up a coding agent, why do they need OSes at all? Why not build their own end-to-end?

For some things, they will. Anything that terminates inside one company collapses to a prompt: internal CRMs, dashboards, project trackers, ETL glue, ops runbooks, even significant slices of what used to be horizontal SaaS, all regenerated by whichever team needs them, often nightly. The interface is ephemeral, the company owns the data, and the “vendor” is its own engineering org augmented by an agent.

The dividing line is whether the system terminates inside one firm or spans outward. Spans outward means: payment rails that touch banks the firm does not run, regulated reporting to authorities the firm does not control, supply chains across operators the firm does not own, multi-party trust where no single party can unilaterally certify, datasets assembled across decades and parties.

But a coding agent does not give a hospital a thirty-year longitudinal patient record. It does not produce a Class-A trucking license, a SOC-2 history with named auditors, a relationship with state insurance regulators, or the trust of a primary care physician. It does not turn an enterprise’s intern into Stripe. It does not move atoms.

Disintermediation is real but bounded: the inside of the firm collapses to prompts, while the outside, where the firm meets other firms, atoms, regulators, and history, does not.

The Survival Test

Pick any company. Ask four questions. Does it own something physical that has to be permitted, maintained, and operated? Does it sit at the center of relationships that took decades to build and cannot be transferred by API? Does it produce a signed attestation that is regulated, insured, or legally required? Does it hold proprietary context assembled over years by people in a specific place, with assumptions no one fully documented? If the answer to all four is no, the company is a prompt waiting to be written. It may have brand, distribution, a hundred million users, a beloved product. None of those are in the four.

A payments company that owns banking relationships and regulatory standing passes. A commerce platform that owns the merchant relationships and the payment rails passes. A code-hosting tool that a coding agent and a quarter of runway can reproduce fails. A project management suite fails for the same reason. Most analytics companies, most customer support platforms, most document editors: prompts. The question is not whether the product is good. The question is whether the product owns something a model cannot synthesize.

Some companies will discover they pass for a reason they never marketed. A design tool survives not because of its AI features but because it became the design-system-of-record for thousands of enterprises, and that is context. A payroll company survives not because of its interface but because it holds decades of tax-filing history and state-by-state compliance relationships. The survivors will rebrand around their unsynthesizable asset, not their software. The ones that can’t find such an asset will have their answer.

Full Vertical

The adjacency-scorching will not stop at coding assistants and search. A distributor has no reason to leave any token-generating surface in someone else’s hands. Chrome exists because Google needed every browser’s default search to be Google. The lab browser will exist for the same reason: whoever owns the surface where reasoning happens owns the billing meter for every token that flows through it. Expect at least one major distributor to ship a browser by 2027 and offer git hosting at or below cost in the same window. The distributor does not need to profit from browsers or repos the way Google never needed to profit from Chrome. It needs to own the surfaces, and it will price every incumbent out of the way to get them.

Follow the logic one step further and the distributor becomes the literal operating system, not metaphorically but architecturally. Apple already ships on-device models. Google already runs Gemini inside Android. Within a decade the traditional OS (files, apps, windows) is a legacy compatibility layer, and the primary interface is a model runtime that holds the full context of everything you do on the device. The token is not just the new shipping container. The token is the new instruction set. The lab is not just the new shipping line. It is the new Intel, the new Microsoft, the platform on which everything else runs.

Apply the taxonomy to that endgame and it gets uncomfortable. Brand is not in the four. Distribution is not in the four. User base is not in the four. Network effects, unless they are relationships in the essay’s specific sense (fiduciary standing, decades of trust, regulated ties), are not in the four. Companies that think they are safe because they have 200 million users are in the middle unless those users are connected by obligations no model can intermediate. The framework, taken seriously, is more destructive than the essay has admitted until now.

The Endgame

Token distributors win at the bottom, OSes win above them, and the top layer is disposable, rebuilt per-firm, often overnight; the regenerated applications run on the OSes, and the OSes run on the tokens. The stack is literal, not metaphorical. Two or three lab-platforms will own the runtime the way Microsoft owned the desktop in 1995, except with the ability to observe and intermediate every cognitive task, not just every file operation. Alongside them, a local-sovereignty ecosystem will function the way Linux does today: powerful, principled, and used by a small minority who care enough to maintain it. Everyone else will choose the most capable distributor and never look at the weights.

McLean’s container killed the middle of the freight industry because it standardized the unit of transport. The forwarders, brokers, and regional shippers who had charged margin for handling mixed cargo lost their reason to exist. Below the box, a handful of global carriers. Above it, companies whose advantage was in what the box carried, not in the carrying. AI does the same thing to software, with the token as the container, but the container is eating further than McLean’s ever did. The middle has the same choice it had in 1956: own something the box cannot carry, or be carried away by it. The difference is that this time the box is learning what it carries.