The Starting Gun Was a Text Box
On November 30, 2022, the most consequential product of the decade shipped as a text box. ChatGPT reached one million users in five days and roughly 100 million monthly by January.
This month, the same company switched off their web browser. Atlas (launched in October 2025 as the future of the internet) died on August 9, nine months old, demoted to a Chrome extension.
A record launch and a quiet retreat, four years apart. We could try to explain it by analyzing the models: parameters, benchmarks, capabilities. The more predictive variable was the shape of the product: how close it sat to your actual work.
On the surface, this is a race between models. Beneath it, value is migrating from raw capability to the control points where data, workflow, and action converge. That migration, not the leaderboard, is the story.
A Race With No Fixed Track
The AI race has run through five main shapes so far: a chat box, an assistant inside tools you already use, a search engine that answers instead of linking, a browser, and an agent that does the work itself. Each is a bid to sit closer to your context, each following the same rhythm: breakthrough, clones within weeks, commoditization, a scramble up the stack.
The race was never about intelligence, which kept getting copied. It was about context: the data the model reasons over, the workflow it lives inside, the gravity that makes leaving expensive.
Five Battlegrounds in Four Years
The Magical Box: everyone builds a brain (2022–23).
GPT-4 claimed a top-10% bar exam score (later research pegged it nearer the 62nd percentile); Claude 2 opened a 100,000-token context window; Bard flubbed a James Webb question and Alphabet shed roughly $100 billion in a day. Meta’s LLaMA weights leaked; Mistral shipped frontier models via magnet link. Parity arrived faster than budgeted: a brain was not a business.
The Copilot: the incumbents strike back (2023).
Microsoft understood that distribution beats capability. Bing Chat became Microsoft 365 Copilot at $30 per seat, then a key on the Windows keyboard. GitHub Copilot, shipping since 2021, supplied the template: AI as a sidebar in the tool already open. Incumbents didn’t need the best model. They needed the surface you were already staring at.
The Answer Engine: the assault on the query (2024).
Google shipped AI Overviews over its ten blue links; OpenAI launched SearchGPT; Perplexity grew from roughly 230 million monthly queries to about 780 million. The query is the highest-intent context a user ever volunteers, which is why this shape moved directly at Google’s core control point.
The Browser: if you can’t buy it, copy it (2025–26).
When regulators declined to force a Chrome divestiture (Perplexity had floated a $34.5 billion bid), the labs built their own: Comet, Atlas, Dia, Gemini folded into Chrome. The prize: own the browser and you own the context and the ability to act on it. It lasted nine months, ending in a Chrome extension. The browser was a feature after all, not a destination, and not just for OpenAI. Owning the surface is not the same as owning the work that happens on it.
The Agent: from tool to workforce (2025–).
This shape stops assisting and starts doing. Anthropic’s computer use begat Operator and Codex; Claude Code became the breakout at a reported ~$2.5 billion run-rate. Agentforce, Copilot Studio, and Gemini Enterprise made agents an enterprise category, and MCP (open-sourced by Anthropic, adopted by its rivals, donated to the Linux Foundation) made it a protocol. That’s the full escalation: feature, product, platform, protocol. Once a shape reaches protocol, the category stops differentiating.
Every Lead Lasts Six Weeks
OpenAI announced Deep Research on February 2, 2025; Perplexity shipped its version twelve days later. Within six weeks, Google, xAI, and Anthropic had near-identical products — several under the same name. Reasoning models followed the same curve: o1, then DeepSeek’s R1, then Gemini, Claude, and Qwen.
This isn’t an anecdote; it’s a law. Copy latency keeps compressing, so a feature moat now has a half-life measured in weeks. Its corollary is the wrapper steamroller: Plugins became GPTs became deprecated in favor of agents; the survivors (Cursor, Claude Code, Agentforce) never had defensible features, just workflow lock-in.
The Race’s Casualties (and Its First Fortunes)
Humane raised roughly $241 million for the AI Pin and sold to HP for $116 million; Rabbit’s R1 followed.
The gold column reads differently: Microsoft 365 Copilot at 20 million paid seats; GitHub Copilot at 4.7 million subscribers; Agentforce at $800 million ARR; Cursor from zero to ~$2 billion ARR in three years. One line divides the columns: the gold owns a workflow; the graveyard chased a novelty.
The Finish Line Was the Enterprise All Along
Now the uncomfortable stat. MIT’s Project NANDA found roughly 95% of 300-plus enterprise AI deployments produced no measurable P&L impact — against $30–40 billion invested. How does that coexist with 20 million paid Copilot seats?
It doesn’t contradict the thesis; it confirms it. Pilots fail when AI is bolted beside a workflow. Money appears when the shape is the workflow: the IDE, the open sidebar, the CRM with the data already in it.
The clearest illustration is a structural fork across the field. OpenAI won consumer mindshare: 800 million weekly users. Anthropic won the wallets: 40% of enterprise LLM spend by late 2025, per Menlo Ventures, against OpenAI’s 27%, down from 50%. In coding, 54% versus 21%. Enterprise AI spend tripled in a year, to $37 billion.
The Rules the Race Keeps Enforcing
Four Laws of the AI Race
1. Every launch is fast-followed within weeks.
2. Wrappers get steamrolled; workflow-owners survive.
3. Commoditization pushes value up the stack.
4. Context, not intelligence, is the moat.
What This Means If You Are Building
For founders, the implication is practical. Don’t wrap a capability that will be free in six weeks. Build a control point: a workflow you own end to end, a data flow that improves with use, and a credible path from insight to action. We believe that is the difference between getting steamrolled and compounding.
The Race Isn’t Over: It’s Changing Shape Again
By mid-2026 the fight had moved down a layer, to the “harness,” the scaffolding that turns a model into a reliable agent.
This is where data and money flows get real. Agents are bounded by the data they can access and the actions they’re trusted to take. The winners will control both: defensible data flows and the money flows those decisions trigger: approvals, payouts, procurement, reconciliation.
Every shape has moved closer to where you already are: beside your code, in your browser, then acting without you.
Four years ago, this all began with a text box. This month, it buried a browser. The sixth shape is already forming, and if the first five are any guide, the shape itself won’t stay a moat for six weeks.
What endures is underneath it: the flows the shape controls.
