| 1 | # THE AI INDUSTRY 2026 — Annual Report · Speaker Notes |
| 2 | |
| 3 | # 01_cover |
| 4 | |
| 5 | Welcome to Volume One, Number One of The AI Industry annual report. The year we are about to walk through is the year the league table changed — the revenue leader is no longer the user-count leader, training costs cleared the half-billion-dollar line per run, and benchmark half-lives dropped from years to months. Ten pages, one front page, four stories, one chronicle, one closing read. Let's begin. |
| 6 | |
| 7 | --- |
| 8 | |
| 9 | # 02_issue_at_a_glance |
| 10 | |
| 11 | Four numbers frame the year. Worldwide AI spending reaches two trillion dollars in 2026, up thirty-six percent. Three point eight billion people now use a large-language-model product each month. The frontier labs collectively booked twenty point seven billion dollars of revenue in the first quarter alone. And across seventy-two countries, more than a thousand distinct AI policy initiatives are active. From those four facts, three observations: revenue concentration is breaking from user-count rankings, training costs have crossed the half-billion-dollar line per run, and inference cost for the same capability is falling roughly ten times per year. Each is picked up by a story later in this issue. |
| 12 | |
| 13 | --- |
| 14 | |
| 15 | # 03_revenue_league |
| 16 | |
| 17 | The Q1 2026 revenue league table is the headline. Anthropic captures thirty-one point four percent of global LLM revenue, narrowly ahead of OpenAI at twenty-nine percent. This is the first time a non-OpenAI lab holds the number-one revenue position since the modern LLM era began. Google trails on twelve percent, Microsoft on seven, Tencent on five. The China cohort — Tencent, Baidu, Alibaba combined — sits at roughly eleven percent, comparable to Google's solo share. And despite extensive open-weights attention through 2025, DeepSeek and Zhipu together hold under a third of one percent of revenue. Distribution does not translate to direct revenue capture. |
| 18 | |
| 19 | --- |
| 20 | |
| 21 | # 04_arpu_divide |
| 22 | |
| 23 | The story behind the leaderboard is per-user revenue. Anthropic earns sixteen dollars twenty cents per monthly user. OpenAI earns two dollars twenty. Google earns one dollar ten. Meta earns ten cents. The rank flip at the extremes is the punch: Anthropic monetises a one-hundred-and-thirty-four-million user base one hundred and sixty-two times harder than Meta monetises a billion. Scale is not the moat. Enterprise-API workloads, coding agents, and per-seat enterprise pricing compound faster on a per-user basis than consumer free-tier scale ever has. |
| 24 | |
| 25 | --- |
| 26 | |
| 27 | # 05_training_cost |
| 28 | |
| 29 | Frontier training costs have moved into a new band. GPT-4 cost roughly seventy-eight million dollars in compute. Llama three point one was about one hundred and seventy million. Gemini Ultra reached one hundred and ninety-one million. The 2026 frontier sits between two hundred and five hundred million per run, at one to ten times ten-to-the-twenty-sixth FLOPs. The late 2027 projection — based on announced data-center buildouts — is one to three billion dollars per run. The constraint, importantly, is no longer chips or capital. The constraint is multi-gigawatt power siting and grid interconnect. The next cohort is being built around substations, not around silicon. |
| 30 | |
| 31 | --- |
| 32 | |
| 33 | # 06_inference_watts |
| 34 | |
| 35 | Inference is the only frontier metric falling faster than capability is climbing. For the same medium-length prompt, Claude four Opus consumes about five watts; DeepSeek V Three consumes about twenty-three. That is a four point six times spread inside the frontier class for identical work. Carbon emissions follow: the least-efficient inference path emits more than ten times the most-efficient. Cost at constant capability falls roughly ten times per year. The implication is that model selection has become a cost-of-goods decision, not just a capability decision. Two models a single benchmark point apart can differ ten times in compute footprint at production scale. |
| 36 | |
| 37 | --- |
| 38 | |
| 39 | # 07_benchmarks_saturated |
| 40 | |
| 41 | Capability benchmarks designed to last years are now lasting months. On SWE-bench Verified — the canonical coding benchmark — frontier models rose from sixty percent to roughly one hundred percent of the human baseline in a single year. Humanity's Last Exam, built explicitly to resist saturation, gained thirty percentage points in twelve months. PhD-level science question-answering, multimodal reasoning, and competition mathematics all crossed human-baseline territory in the same window. The bench-design community is now permanently behind the model-release cycle, and the conversation is shifting from static benchmarks to continuous evaluation harnesses that regenerate problems on every run. |
| 42 | |
| 43 | --- |
| 44 | |
| 45 | # 08_three_rulebooks |
| 46 | |
| 47 | Compliance is no longer a single rulebook. Europe regulates the model class with the AI Act, in full force since August 2025, carrying a maximum penalty of thirty-five million euros or seven percent of global turnover. The United States has no federal horizontal statute; California, Colorado, and Texas are doing the load-bearing work. China regulates the output — synthetic-content watermarking has been mandatory since September the first, 2025, enforced by the cyberspace administration with administrative penalties and licence revocation. Convergence on principle, divergence on implementation. A model deployed globally now ships three compliance bundles, not one. |
| 48 | |
| 49 | --- |
| 50 | |
| 51 | # 09_timeline |
| 52 | |
| 53 | Sixteen months changed the industry's shape. January 2025: the United States signed Executive Order fourteen one seventy-nine, reversing the federal posture. August 2025: the European AI Act took full effect, the first binding horizontal AI statute on a major economy. September 2025: China's synthetic-content rules became effective. Quarter four 2025: DeepSeek R one shipped, reframing the inference-cost conversation through open-weights reasoning at frontier-adjacent capability. Quarter one 2026: Anthropic overtook OpenAI on LLM revenue. April 2026: the Stanford HAI AI Index 2026 reported the first year SWE-bench Verified was effectively saturated. Three structural shifts — compliance, compute, capability — and none of them unwinds on the present trajectory. |
| 54 | |
| 55 | --- |
| 56 | |
| 57 | # 10_closing_read |
| 58 | |
| 59 | To close, eight observations to carry forward. What changed: revenue leadership decoupled from user count, training cost crossed the half-billion line, inference became a four point six times discriminator inside the frontier class, and benchmark half-life collapsed from years to months. What did not change: the frontier cohort is still roughly five labs, China frontier labs still have zero EU-market revenue, open-weights distribution still does not translate to direct revenue capture, and power and grid interconnect remain the single hardest input to source. The test of the next twelve months is whether a second wave of labs can land the half-billion-dollar runs without breaking the three-rulebook trade barrier. That is the question this report leaves open. Thank you for reading. |
| 60 |