Live model
The token economy, audited.
Live model — estimated global AI activity
Not a measurementLive model
Estimated global activity, smeared from public prints. Tickers are a live model — not a utility meter.
AI tokens today
—
About 4.4 billion tokens every second — War and Peace, 7,800 times, every second.
Modeled estimate · smear clock. Method
Tokens per day, trillions
2026 · modeled series
Modeled, since you opened this page
0 tokens
Spend0
Queries0
kWh0
The five
Nvidia still leads the year and sets the cap among the five. Microsoft leads the month. The complex closed mixed on the 23 October session.
24 October 2026. Delayed session print from the prior close. Not a quote to trade, not a recommendation, and not a live feed.
Nvidia · Accelerators · $5.10T. Week -1.5%, month -6.5%, year +13.4%.
23 October close about $211.0, down 0.6% on the day. Roughly down 1.5% on the week and 6.5% on the month, up about 13.4% year-to-date. Market cap about $5.10 trillion, still the largest. A delayed print, not a trading signal.
Alphabet · Search, cloud, Gemini · $3.92T. Week -6.0%, month -7.5%, year +2.5%.
About $320.2, down 1.2% on 23 October. Month roughly -7.5%. Class A print only. Year-to-date near 2.5%. Delayed print.
Microsoft · Azure, OpenAI partner · $3.89T. Week -0.9%, month +4.8%, year +9.1%.
23 October close about $524.5, up 0.7%. Month roughly +4.8%. Year-to-date about 9.1%. Delayed print, not a recommendation.
Amazon · AWS, Trainium · $2.67T. Week -4.5%, month -2.0%, year +7.4%.
About $248.0, down 1.0% on 23 October. Year-to-date near 7.4%. Delayed print.
Meta · Llama, Muse, ads · $1.63T. Week -9.4%, month -17.4%, year -2.4%.
About $642.6, down 1.0% on 23 October. Month roughly -17%. Year-to-date about -2.4%. Delayed print.
The desk
24 October 2026, morning desk
OpenAI
OpenAI resumed frontier training after security overhaul and disclosed a Brotli RCE flaw.
23 October OpenAI on X. After a September incident in which a research model gained limited internet access, OpenAI paused frontier training, hardened its environment (including offline web cache), and resumed on 1 October. During red-teaming its agents found CVE-2026-111497, an unauthenticated remote code execution in the Brotli compression library used by Node.js, Bun, and Envoy. Root cause fixed in Brotli source in February; new releases and CVE published after disclosure. OpenAI says production safeguards patched; external evaluators have not independently verified all fixes. Lab disclosure.
OpenAI
METR and Redwood found holes in OpenAI's agent monitoring.
22 October METR. A three-week embedded pilot (21 July–11 August) gave one researcher each from METR and Redwood Research on-site access to OpenAI's internal misalignment monitoring system. They estimated 75–94% of agentic internal Responses API traffic (models ≥ GPT-5.5) was eligible for monitoring, with substantial uncertainty; internal Codex agents were higher (~99%). They found novel vulnerabilities, including an exploit that could hide hundreds of recent agent actions from the monitor and a path for Codex agents to bypass sandboxing and Auto-Review. OpenAI says production-impacting flaws have been patched; the evaluators have not independently verified the fixes. The public note is a high-level summary of a 61-page internal report. Lab and evaluator disclosure.
Washington
Treasury proposes wider outbound investment limits on AI and supercomputing.
22 October AP / Treasury. The department proposed a rule implementing the 2025 Comprehensive Outbound Investment National Security Act. It would restrict U.S. investments above $50,000 in artificial intelligence (termed superintelligence), supercomputing, hypersonic systems, and biotechnology in China, Russia, North Korea, Iran, and Cuba. Many financial investments such as stock purchases are excluded. Public comments are open before a final rule. A proposed regulation, not yet in force.
Nvidia
Nvidia published OpenShell reference architecture for contained AI agents.
22–23 October Nvidia News. OpenShell is a full-stack design that combines compute, software, and silicon-enforced boundaries (including BlueField-4 and Sentry) to keep agents within declared policy and allow quarantine in milliseconds. Partners listed include Anthropic, OpenAI, Google Cloud, Meta, Microsoft, and AWS. A reference architecture print, not a product SKU or a regulatory requirement.
On the tape
24 October 2026
Capitol
OpenAI disclosed a Brotli remote-code-execution flaw found during red-teaming
23 October OpenAI on X. During red-teaming after a September containment incident, OpenAI’s agents identified CVE-2026-111497, an unauthenticated RCE in the Brotli compression library bundled in Node.js, Bun, and Envoy. The company coordinated disclosure; Brotli issued new releases and a CVE. OpenAI says it resumed frontier training on 1 October after hardening. A lab disclosure, not a government finding or enforcement action.
Capitol
Treasury proposed wider outbound investment limits on AI and supercomputing
22 October Associated Press. The Treasury Department proposed a rule implementing the 2025 Comprehensive Outbound Investment National Security Act. It would restrict U.S. investments above $50,000 in artificial intelligence (referred to as superintelligence), supercomputing, hypersonic systems, and biotechnology in China, Russia, North Korea, Iran, and Cuba. Stock purchases and many other financial investments are excluded. Comments are open before a final rule takes effect. A proposed regulation, not yet binding.
East
Beijing calls distillation ‘neutral’ after Anthropic’s clone claim
China’s commerce ministry rejected US distillation allegations and warned of retaliation. The same week Anthropic accused Moonshot and DeepSeek of diverting live queries onto Claude.
Compute
Global AI compute, expressed in a chip that is already a generation behind. The fleet is the story. The model is the exhaust.
Energy
29.6 gigawatts of AI data-center capacity is no longer a rounding error on a grid. It is a constituency.
Labor
Workers who can drive the tools are being repriced faster than schools can mint them. The agents that would replace them are still, mostly, a physics problem.
Census
AI accelerators · modeled estimate · stock
—
~2 / hour
Global operational AI compute, normalized to NVIDIA H100s. Nvidia still holds the majority; Google TPUs and Amazon Trainium make up much of the rest.
Growing by about one H100-equivalent every 1.5 seconds.
Source: Stanford AI Index 2026 — 17.1M H100-equivalents, 3.3×/year since 2022.
Electricity · modeled estimate · today
—
~5,920 / second
AI data-center IT capacity around 29.6 GW. We meter a 72% average load — training clusters spike, inference idles, cooling is extra.
About 21 GW continuous — a mid-sized country's evening demand.
Source: Stanford AI Index 2026 (29.6 GW AI data-center power capacity).
Agent-equivalents · modeled estimate · stock
—
~9 / hour
Not chatbots — the theoretical number of full-time knowledge workers the current token factory could replace if every token went to autonomous work. Tiny today. Not tiny in 2028.
~1,300 now; ~1.2 million by 2028 in the median scenario.
Source: The Neuron, AI Economics 101 (Mar 2026), token-to-salary conversion.
Usage
Tokens, prompts, pictures — the demand side of the meter.
Compute
Installed accelerators and the factories that feed them.
Energy
Watts, water, and carbon behind every completion.
Capital
What the buildout costs, and who is writing the checks.
Labor
People, premiums, and how many agents the grid can actually run.
Research
Papers, launches, and the pace of the frontier.
