AI GDP Clock

AI GDP Clock

How many tokens does it take for AI to become a measurable part of GDP? Change the assumptions and every result updates. Read the essay →

AI crosses 1% of U.S. GDP
June 2028

Realized incremental AI output · share of growing GDP

AI-worker equivalents today
Potential output today
Realized AI GDP today
Realized share today
GDP per 1M tokens
Tokens per $1 GDP
1. Token volumeHow many tokens exist
2. AI workersTokens ÷ tokens needed per worker
3. Potential outputWorkers × value of their work
4. Realized GDPOnly the additional output that reaches the economy

Token efficacy is not another independent assumption. It is the result of the worker-token requirement, worker value, productive-use share, and realization rate.

Choose a year and token forecast

Pick a target year, then drag either the multiple or the absolute annual token volume. All three controls stay linked and the calculator back-solves the growth rate.

Measured on August 24, matching the model's baseline date.
Global annual token volume relative to the Aug. 24, 2026 baseline.
Annualized quadrillion tokens, matching the model's baseline unit.
AI-worker equivalents
Tokens per AI worker
Potential output
Realized AI GDP
Share of GDP
Annual GDP-growth lift
Inference GPUs
Inference power

The 2,030× value shown when Author's case loads is derived by extending its 19-week doubling assumption from August 2026 to August 2030. It is not a separate published forecast.

Goldman Sachs Research forecasts 120Q tokens/month in 2030, or 1,440Q/year. Its 24× multiple uses Goldman's own smaller 2026 baseline, so this button uses the absolute endpoint rather than incorrectly applying 24× to this calculator's 130Q baseline. In Dell's Sept. 1, 2026 earnings call, Jeff Clarke said inference-driven tokens would rise 87× to 3,600Q by 2030, but the transcript does not state a monthly or annual cadence for 3,600Q. Dell's two buttons deliberately show the relative and annualized interpretations separately.

Core assumptions

These four ideas determine the result: token supply, where the tokens work, how many tokens make one accepted worker-equivalent, and what that work is worth.

Everything AI generates worldwide, annualized in a consistent token unit.
The fraction of global tokens used in work attributed to the U.S. economy.
Forecast driving this rate:
Goldman and Dell are included when you select their buttons above: the calculator back-solves a constant doubling time from today's annualized baseline to that endpoint. Moving a token-path control creates a custom path.
All tokens required for one fixed year of accepted work, including reasoning, failures, retries, and verification.
Annual value of that fixed work bundle before the realization discount below.
Advanced: turn potential work into realized GDP
The original worker assumptions are preserved. They start at 100% productive use and 100% realization, meaning the $20k worker value is already treated as incremental GDP. Lower either control for a more conservative realized case; do not discount the $20k separately as well. The GDP denominator now grows independently at the selected real rate.
Of the tokens attributed to this economy, the share used in workflows intended to create measured output. Casual use and entertainment sit outside; failures inside a work attempt stay in tokens per worker.
The share that becomes genuinely additional real output after oversight, adoption, displacement, and bottlenecks. Use 100% only when worker value is already net incremental GDP.
Let tokens needed per AI worker fall This models technical efficacy directly: the same accepted work bundle uses fewer tokens over time. Off keeps the token requirement fixed.
A 52-week setting turns 100B tokens per worker into 50B after one year and 25B after two.
When AI-worker equivalents double, how much their combined potential value grows.
Annual growth of the rest of the economy, in constant dollars. The 1% threshold rises with this denominator.
GDP denominator “Share of GDP” is a level. “Growth lift” is the extra annual growth generated by the increase in realized AI output.

The physical footprint — inference only

Everything below prices inference: the power to generate the tokens. Training the models is a separate draw on top of it.

Realized GDP / 1M tokens
Inference GPUs at 1%
Inference power at 1%
Inference power today

Inference power for the U.S.-attributed token share · GW, log scale

Blended tokens per second one accelerator sustains across the model mix.
Watts per accelerator including the node, networking, cooling, and facility overhead.
Share of peak throughput a real fleet nets after latency targets, small batches, and idle headroom.

Method. Economy-attributed productive tokens ÷ tokens needed per accepted AI worker = AI-worker equivalents. Worker equivalents × gross value per worker gives potential output; the realization rate converts that capacity into incremental GDP. Diminishing returns reduce the value of later worker-equivalents as the supply expands. The target controls translate between year, multiple, and annualized token flow using a smooth exponential path. Non-AI GDP compounds separately in constant dollars, so the 1% threshold rises over time. “GDP share” is the level of realized AI output divided by that GDP benchmark. “GDP-growth lift” is the increase in annualized AI output versus the previous year, divided by the prior-year GDP benchmark.

Token and compute accounting. Tokens per accepted worker should include input, output, reasoning, failed attempts, retries, and verification under the same counting convention used by the volume forecast. Inference GPUs and power use every economy-attributed token, including non-economic use, because all of them consume compute. A fall in tokens per worker improves economic efficacy, but aggregate compute stays unchanged when the total-token forecast is held fixed. Training remains outside the footprint calculation.

Interpretation. This is a scenario model, not a measured economy-wide dollars-per-token series. Worker value, productive share, and realization must not contain the same discount twice. Cost savings, vendor revenue, infrastructure investment, and user productivity are separate accounting channels and should not simply be added together. Token forecasts also need matching dates, baselines, cadence, geography, and token definitions before their multiples are compared.

Sources. Goldman Sachs token forecast · Dell Q2 FY27 transcript · Google I/O 2026 · tokensperday.com · Menlo on OpenRouter · Epoch AI · METR · GDPval · SWE-Lancer · Generative AI at Work · Menlo enterprise report · McKinsey on data-center power · inference share of AI compute

A napkin model — companion to the essay by Rob Leclerc.