← Business

Who serves the tokens?

Tokens are the AI economy's unit of account — every chat, coding agent and pipeline consumes them. This stacks estimated LLM inference volume by model family since 2019, the way forms of rule stacks humanity by regime: hover for a breakdown, and switch between relative share and absolute tokens per day.

OthersOpen-weights (Llama, Mistral…)Chinese models (DeepSeek, Qwen…)Anthropic (Claude)Google (Gemini/PaLM)OpenAI (GPT)
0%25%50%75%100%2019202020212022202320242025Time →

Editorial, order-of-magnitude estimates — no official industry-wide token statistics exist. Anchors include vendor disclosures (e.g. Google's tokens-processed-per-month figures, Microsoft earnings mentions) and public router rankings; the split between families is inferred and approximate, and “tokens served” conflates consumer, API and internal workloads. Hover for a breakdown; switch between share and absolute volume above.

How to read it

Three stories sit in the stack. First, the explosion: from research-scale volumes in 2019 to tens of trillions of tokens a day by 2025 — the fastest adoption curve of any computing platform. Second, the erosion of the OpenAI monopoly: near-total dominance after ChatGPT's launch gives way to a contested oligopoly as Gemini rides Google's distribution, Claude takes a deep position in code and enterprise work, and DeepSeek and Qwen make frontier-adjacent capability nearly free. Third, the open-weights floor: Llama, Mistral and their descendants guarantee that some share of the world's tokens can never be enclosed — the “non-state societies” band of this chart.

Status of these figures. There are no official statistics for industry-wide token volumes: these are editorial, order-of-magnitude estimates anchored on documented public disclosures (vendor tokens-processed figures, earnings mentions, public router rankings) and interpolation, current as of early 2026. The split between families is inferred; “tokens served” mixes consumer, API and internal workloads; and self-hosted open-weights usage is genuinely unmeasurable. Treat the shape as informative and the decimals as illustration — corrections welcome, the data lives in one editable file.