The Control Series, Part 5 of 6 · Chokepoint: Distribution. Part 4 ended on a question — you never owned the model, so will you own the door to it? This is the fight for the door.
In June, SpaceX paid $60 billion for a coding tool. Not a model. Not a data center. An interface — the surface where developers actually type. A company that already owns power, compute, a research lab, and a frontier model looked at its stack, found the expensive missing piece, and the piece it bought was the door.
That should reframe how you think about this whole industry. For three years the assumed prize was the model — the smartest weights win. But the model is sliding toward commodity (Part 2: H100 rates down 60–75%; capable open weights months behind the frontier), and the thing that isn’t commoditizing is the surface the human touches. Whoever owns that surface owns the default, the habit, the data that flows back, and — quietly the most important part — the decision about which model gets called in the first place.
This is the chokepoint that sits above the model layer. Here’s why it may be the most valuable one in the series, and the strange reason it’s also the most winnable.
The Door: Worth More Than the Model
SpaceX paid $60B for a coding tool — not a model. As the model commoditizes, the surface the human touches captures the value: the default, the habit, the data, and the choice of which model gets called.
Perplexity
The most valuable chokepoint — and, strangely, the most winnable. You can’t bootstrap a gigawatt or a 555K-GPU cluster, but a small team can still build the door (Cursor was a few founders on rented models). Own the interface and the user relationship even if you rent everything underneath — and never let a platform’s default be your only door to your users.
The $60-billion proof
Start with Cursor, because it settles the argument. Anysphere built a coding interface on top of other companies’ models, reached roughly $4 billion in annualized revenue, rebuffed OpenAI twice and Microsoft once — and sold to SpaceX for $60 billion. The model underneath was rentable; the surface on top was not. What SpaceX bought was the place developers live all day, the usage data that streams off it, and the power to route that demand to whichever model it chooses.
That is the distribution thesis in one transaction: when the substrate is a commodity and the surface is a habit, the surface is worth more than the substrate. Even the company with the most complete vertical stack in AI agreed — and paid a model-sized premium for an application.
AI Voice Chat Module Type C Interface AI Large Model Support with Technology
- Interface Type: Type C interface for connectivity
- Battery Management: Built-in TP5400 battery management
- Integrated Components: Includes INMP441 microphone and amplifier
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The war above the URL bar
The same logic is detonating the calmest corner of software: the browser, untouched for fifteen years, is suddenly the hottest battleground in tech.
OpenAI shipped Atlas in October 2025 — “ChatGPT with hands,” a browser where the address bar is a prompt and an agent acts on pages rather than just summarizing them. The smart read is that Atlas isn’t trying to beat Chrome; it’s distribution, not displacement — a surface that keeps OpenAI from depending on Google or Apple to reach its own users. Perplexity made Comet free worldwide and shipped it across every platform. The Browser Company folded Arc into Dia and sold itself to Atlassian for $610 million. Brave kept Leo lean and local. And the giants with real distribution are now moving: Google is rolling a Gemini panel and OS-level “auto-browse” into Chrome on flagship phones; Microsoft turned on Copilot Mode across hundreds of millions of Edge seats; Anthropic put browser and desktop control behind Claude. Estimates put Atlas around 10–15 million monthly users, Comet at 3–5 million — small numbers that the OS-level defaults could dwarf the moment they flip on.
Here’s the tell that makes this a chokepoint and not a feature race: the interface decides the model. Atlas users get routed to OpenAI’s models, Comet users to Perplexity’s stack, the Claude surfaces to Claude. Own the door and you own the routing — which model is the default, which gets demoted, which never gets reached. The model layer everyone obsessed over becomes plumbing behind a faucet someone else controls.
AI Deployment Pipelines: Enterprise MLOps Governance | AI Tools and Platforms | Data Privacy in AI | AI Performance Metrics | Sustainable AI Systems | Future of AI in Cloud | AI Deployment Strategies
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why the interface is the lever
Distribution has always beaten the better product; AI just raised the stakes. The surface that sits between a user and the machine captures four things at once:
The default — most people use whatever’s in front of them, and the interface decides what that is. The habit — the tool you open every morning owns your attention and your switching cost. The data feedback loop — every query, click, and correction flows back to the surface owner, not the model maker, compounding into the proprietary corpus Part 3 called the real moat. And the routing — the power to send demand to a favored model and starve a rival’s.
Stack those and you get classic aggregation: the front-end captures the value and the user relationship, while the supplier underneath — here, the frontier model — commoditizes. The open web becomes back-end infrastructure to a handful of synthesizing, acting front-ends. Ten blue links collapse into one answer, and the answer’s owner decides what’s in it.
Platform Engineering for Artificial Intelligence: Designing scalable infrastructure, data pipelines, and model lifecycle management for generative AI and agentic protocols (English Edition)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The friction
You can already see the control fight breaking into the open, because an interface that acts on the web has to push through everyone else’s. Agent traffic has exploded — one security firm measured a roughly 6,900% rise in agent requests since mid-2025, and agentic shopping surged about 145% over a single Black Friday weekend. The first real legal test arrived when Amazon sued Perplexity in January over Comet’s automated shopping, fighting to control whether someone else’s agent can transact inside its store. Publishers are weighing whether to block agent access the way they’re now licensing training data. New standards are appearing to let agents act on pages at all.
Every one of those is the same question in a different venue: when the interface owner sits between every user and every website, who controls the transaction — the store, the agent, or the surface that sent it? The door isn’t just where users enter. It’s where everyone else has to ask permission.
Learning OpenCV 4 Computer Vision with Python 3: Get to grips with tools, techniques, and algorithms for computer vision and machine learning, 3rd Edition
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Who holds the door
Three sets of hands, and they overlap. The device and OS owners — Apple, Google, Microsoft — hold the ultimate distribution, because the default browser and assistant on a phone or laptop start with a near-insurmountable lead the instant they ship. The AI-native interfaces — OpenAI, Perplexity, and others — are racing to build a habit deep enough to survive those defaults. And the application layer — Cursor and the tools people live inside all day — owns specific, high-value surfaces worth more than the models beneath them.
Access at the interface is gated, defaulted, and routable — the same gated-repriceable-revocable signature as the rest of the stack, expressed as whose surface you reach the model through, and what that surface decides on your behalf.
My take
I think distribution is the most valuable chokepoint in this series, for the reason it’s the one ordinary people actually touch: it owns the default, the habit, the data, and the routing, while the model it sits on top of slowly commoditizes. The $60-billion Cursor deal isn’t an outlier — it’s the market pricing this in.
But here’s the turn, and it’s the most hopeful note in the whole series. Distribution is also the most winnable chokepoint — the only one a small team can still take without owning power, compute, data, or a frontier model. Cursor was a handful of founders building on other people’s models, and it became worth $60 billion by owning a surface people loved. You cannot bootstrap a gigawatt or a 555,000-GPU cluster. You can still build the door. For anyone constructing on top of AI, that’s the strategic center of gravity: own the interface and the user relationship even if you rent everything underneath it. Be present on more than one surface. Never let a platform’s default be your only door to your own users.
Be fair to the optimists too: agentic interfaces deliver real value — collapsing dozens of clicks into one instruction is not a trick — and no one has won. Atlas and Comet are small; Chrome and Apple haven’t fully swung; the field is genuinely contested. But the direction is the worry. If a few front-ends come to intermediate how everyone reaches information, buys things, and gets work done, then the open web becomes their supply chain, and the surface owners inherit the leverage that search engines spent two decades accumulating — faster, and with the power to act, not just rank.
Watch items for the rest of 2026: whether OS-level defaults on phones and laptops smother the standalone AI browsers through sheer distribution; how Amazon v. Perplexity resolves, because it sets the rules for who controls agentic commerce; whether agent-action standards open the surface to everyone or entrench the incumbents; and whether an interface owner routing users to its own model draws the self-preferencing scrutiny that logic invites.
Next in the series
One chokepoint left, and it’s the one underneath all the others — because power plants, GPU clusters, exclusive data, frontier models, and category-defining interfaces share a single prerequisite: someone has to pay for them. Part 6, the finale, takes up capital — the circular, intra-industry financing and the coming IPO wave holding the entire structure aloft, and what happens to every other chokepoint if that money ever stops circulating.
You never owned the model. You might still own the door. The last question is who’s funding the building.
Sources: SpaceX SEC filings; WSJ; Reuters; CBS; TechCrunch; reporting on AI browsers from multiple outlets; HUMAN Security; Anthropic’s State of AI Agents (2026); and coverage of Amazon v. Perplexity (Oct 2025–June 2026). User and revenue figures are as reported; MAU estimates are approximate. Analysis and opinions are the author’s.
