In early 2022, large language models were a niche research interest. GPT-3 had been described by academic reviewers as “comparable to an old typewriter” — useful for novelty text, frustrating for anything multi-step. Only 26% of Americans said they had heard “a lot” about AI. Most enterprise software roadmaps for 2023 did not contain the words “generative” or “LLM” at all. The technology existed; the product did not.

ChatGPT changed the distribution problem. It put a frontier model behind a chat box and made the chat box free. The growth curve that followed has no good historical analog in consumer software — and the more interesting fact is that nobody at OpenAI expected it. Internal projections, later disclosed, were for tens of thousands of weekly users in the first month. The actual number was three orders of magnitude larger.

Speed of adoption is the obvious story. The less obvious one is the capability leap that was running underneath it. The model behind ChatGPT in November 2022 (GPT-3.5) scored 43% on MMLU, the standard general-knowledge benchmark — roughly an attentive undergraduate. The frontier model in May 2026 scores 92%, above the human expert ceiling on most subtests. The benchmarks that academic reviewers in 2022 used to argue LLMs were a dead end have all been saturated.

Weekly active users are a more honest number than the cumulative figure. They show whether the product retained, not just whether it landed. ChatGPT did both. The curve below has two interesting features: a year-long plateau through 2023 and into 2024 (the “is this just a toy?” period), and a near-vertical re-acceleration from mid-2024 onward — driven by GPT-4o, voice mode, mobile, and the simple fact that the model stopped hallucinating about as often as a tired colleague.

Between 2022 and the end of 2025, US tech employers cut roughly 666,000 jobs across 3,500 employers. Disentangling “AI displacement” from the broader macro story is difficult — but by late 2025, IEEE estimated that about 27% of tech layoffs were explicitly framed as AI substitution.

Three distinct phases are visible in the headcount data. The first, through 2022 and most of 2023, was a post-ZIRP correction — cheap-money hiring being unwound after interest rates rose. The second, through 2024, was a margin story: large incumbents using AI as cover for the kind of mid-cycle efficiency drive they would have run anyway. The third, from late 2024 onward, is the one that is genuinely new. Companies began cutting roles whose tasks they had measurably automated — not because the budget required it, but because the work no longer required the person.

The clearest signal in the data is not displacement of mid-career workers. It is the collapse of junior hiring across nearly every function. SignalFire’s 2025 analysis found that across the largest public tech firms and mature VC-backed startups, the number of new roles filled by people with less than one year of experience fell by roughly half between 2019 and 2024.

The drop is broad-based, not isolated to a single function. Recruiting fell hardest (down 58%) — the people who hire other people were among the first to be replaced by AI-assisted workflows. Engineering and marketing both more than halved. Even legal and finance, traditionally the most resistant to disruption from above, posted double-digit declines in entry-level seats.

The flip side of task displacement is the productivity gift to those who remain. Every major workplace study published between 2023 and 2025 found the same asymmetry: novices and low-performers gain dramatically from AI assistance, experienced workers gain a little or sometimes nothing. AI compresses the experience curve. Whether that is good for the workforce depends on whether you already have experience.

Three things that should not normally happen at the same time happened at the same time. Per-unit cost collapsed by more than two orders of magnitude. Capability went up by every benchmark anyone could write down. And revenue — the thing that is supposed to be hardest to grow once cost is falling — grew faster than any prior software category on record. Pick any two of the three and you have a normal technology cycle. All three at once is what makes this period structurally weird.

The cost line is the more important of the two. A single token of GPT-3.5-equivalent output cost roughly 300× more in November 2022 than it does today. The implication is that almost every product idea that was "too expensive to run at scale" in 2022 is now economically trivial. Customer-support bots, document-review assistants, agentic coding loops that re-prompt themselves dozens of times — all of these were blocked on price, not on capability.

The headline figure — roughly $1 trillion of AI-related capital deployed between 2023 and Q1 2026 — is large, but the more interesting number is how unevenly it has landed. The labs themselves have absorbed less than a fifth of it. The bulk has gone to chips, and to the data centres and electrical infrastructure that house them.

The most-discussed feature of this capital story is its circularity. Microsoft's equity in OpenAI flows back to Microsoft as Azure spend. Nvidia's chip allocation to OpenAI is paired with a $100B compute commitment from Nvidia to OpenAI. CoreWeave buys GPUs from Nvidia, in which Nvidia is also an equity holder, then resells the compute to Microsoft. Each leg looks rational in isolation; aggregated, the system has a small number of participants exchanging very large numbers with each other.

Anthropic spent 2022 and most of 2023 as the quieter of the two leading labs — a safety-first spin-out from OpenAI, with a smaller user base, a smaller revenue line and a model family (Claude 1, then Claude 2) widely regarded as a half-step behind GPT-4. By late 2023, the conventional wisdom was that the LLM market would consolidate around OpenAI the way search consolidated around Google. Within eighteen months, that consensus had inverted.

The inflection arrived with Claude 3 Opus in March 2024 — the first non-OpenAI model to credibly beat GPT-4 on MMLU — and accelerated through the 3.5 Sonnet release in June, which delivered Opus-class quality at a fraction of the price. From there the cadence held: 3.7 Sonnet, Claude 4, 4.5 Sonnet, 4.5 Opus, each release pushing further into the workload Anthropic had quietly bet the company on.

The bet was coding. Where OpenAI optimised for the consumer chat interface, Anthropic optimised for the developer plugged into an editor at 2am. Claude 3.5 Sonnet held the top spot on every public coding benchmark for most of a year. Computer-use, shipped quietly in October 2024, gave the model an arms-and-legs interface to the operating system. By the time Claude Code launched as a first-party CLI in 2025, the workflow was already entrenched in Cursor, Replit, Windsurf and GitHub Copilot's Claude option.

The quadrant below places each frontier lab against the two metrics that used to move together — weekly active consumers and enterprise spend share. In every prior platform era, the same firm sat in the top-right corner. In May 2026 the corner is empty.