How Many Digital Workers Could Future AI Deploy?

Epoch AI Estimation Framework

Epoch AI estimates the scale of AI‑generated “digital workers” by comparing the total tokens produced by OpenAI each day with the tokens a human worker can process in a day.

Numerator – Tokens Produced by OpenAI GPT‑5

Path 1: Inference‑compute budget – Reported ~1.1 million H100‑equivalent GPUs, ~40 % allocated to inference, yielding ~480 k H100 GPUs. These deliver ~10²⁵ FLOPs per day. Assuming 1 × 10¹¹–6 × 10¹¹ FLOPs per token, daily token output is estimated between 1 × 10¹³ and 1 × 10¹⁴ tokens (10 trillion–100 trillion).

Path 2: Usage statistics – ChatGPT (GPT‑5) processes ~40 billion messages per day; API calls account for a quarter of token volume, with an average of 4 000 tokens per message, giving ~2 × 10¹³ tokens per day.

Combining both paths and accounting for uncertainty, Epoch selects a median of 1.9 × 10¹³ tokens per day. OpenAI DevDay 2025 disclosed 60 billion tokens per minute, i.e., ~8.6 × 10¹³ tokens per day – over four times the median.

Denominator – Human‑Equivalent Token Processing

Method 1: Thought‑speed anchor – 380 words/min ≈ 240 k tokens per 8‑hour workday. Simple 1:1 token exchange yields 1.9 × 10¹³/2.4 × 10⁵ ≈ 80 million digital workers.

Method 2: Task‑based comparison – METR study shows GPT‑5 needs 1 × 10⁵–1 × 10⁶ tokens to complete a one‑hour task. Scaling to an 8‑hour day gives 0.8 – 8 million tokens, resulting in 2.4 – 24 million digital workers.

Epoch reports a median estimate of 7.43 million digital workers with a 90 % confidence interval from 0.4 million to ~300 million.

Limitations of the Static Framework

Token‑equivalence assumption – Treats all AI‑generated tokens as equal to human “thinking” tokens, ignoring differences in cognitive efficiency and economic value.

Static efficiency assumption – Uses a fixed FLOPs‑per‑token and current model performance, overlooking rapid algorithmic and system‑level gains.

Homogeneous labor assumption – Models a single, generic digital worker, missing future heterogeneity of specialized AI agents.

Neglect of physical and economic constraints – Omits data‑center power, cooling, and the requirement that total cost‑of‑ownership be lower than generated economic value.

Dynamic Extension – Supply‑Side Reconstruction

CapEx → hardware → physical constraints – Model the compute supply chain from capital expenditure through hardware deployment, accounting for power and cooling limits.

Physical‑constraint multiplier – A factor < 1 that reduces the theoretical compute deployment rate based on electricity‑grid expansion, cooling‑technology adoption, and emerging SMR energy sources.

Dynamic Extension – Demand‑Side Reconstruction

Inference‑efficiency multiplier – Captures gains from speculative decoding, Mixture‑of‑Experts, quantization, pruning, distillation, and software‑hardware co‑design (e.g., vLLM) that lower the raw compute needed per task.

Task reliability & risk – Incorporates expected failure cost (failure probability × average loss) especially for high‑risk domains such as finance or healthcare.

Human‑in‑the‑Loop (HITL) cost – Adds supervision, verification, and intervention expenses for high‑value AI deployments.

Global AI Compute Supply (Quantifying the Numerator)

Projected data‑center investment reaches ~US$7 trillion by 2030, with >US$4 trillion earmarked for compute hardware.

Major hyperscalers (AWS, Azure, GCP, Meta) plan >US$3 trillion in AI‑related CapEx for 2025, roughly half allocated to accelerators.

TSMC’s CoWoS advanced‑packaging capacity grows from ~75 k wafers/month (2025‑end) to ~110 k/month (2027‑end), remaining a supply constraint.

Power demand projected at 40 GW for a Virginia “data‑center port”; Jevons’ paradox suggests efficiency gains may spur higher total demand.

Air cooling approaching limits; liquid cooling and immersion becoming necessary.

Companies explore modular SMRs to secure low‑carbon power for future AI farms.

Inference‑Efficiency Evolution (Denominator Dynamics)

Speculative decoding can triple inference speed.

Mixture‑of‑Experts decouples model size from compute cost.

Quantization, pruning, and distillation shrink model footprints.

vLLM and similar engines improve GPU utilization via paged attention and dynamic batching.

Task Reliability and Economic Value

SWE‑bench “Verified” subset shows ~70 % success for top models.

SWE‑bench Pro and CAIA reveal lower success rates (≤25 % on complex code, ~67 % on high‑risk crypto tasks), highlighting reliability gaps.

GDPVal framework evaluates AI contributions along four dimensions:

Task Economic Weight – monetary impact of the task.

Labor Substitution & Augmentation Coefficient – degree of automation vs. enhancement.

Quality & Innovation Multiplier – performance relative to human baseline.

Risk‑Adjusted Cost – failure probability multiplied by expected loss.

Integrating GDPVal yields an “economic‑feasibility filter” that discards digital‑worker estimates whose risk‑adjusted ROI is negative.

Scenario Modeling (2030‑2035)

Worst‑case “Physical Brake” (linear growth) – Slow grid expansion and modest inference‑efficiency gains limit growth to tens of millions of digital workers.

Best‑case “Algorithmic Leap” (explosive growth) – Breakthrough inference algorithms cause exponential efficiency gains, pushing the count to tens of billions despite hardware limits.

Average “Balanced Acceleration” (baseline) – Steady but non‑breakthrough advances yield hundreds of millions of digital workers.

Key Insights

Insight 1: Economic value decouples from raw FLOPs; advances on the denominator side become the dominant lever.

Insight 2: Strategic bottlenecks shift in three phases – (2024‑26) CoWoS capacity, (2027‑35) power‑grid and data‑center approvals, (post‑2035) proprietary data, workflows, and talent.

Insight 3: The homogeneous “digital worker” notion is obsolete; a heterogeneous AI‑agent economy will require new market mechanisms and governance.

Insight 4: Jevons’ paradox may turn energy into the ultimate limiting factor for AI scaling.

Insight 5: Automation vs. augmentation paths will diverge across economies, shaping long‑term competitive dynamics.

References

How many digital workers could OpenAI deploy?, Epoch AI

GDPVAL: Evaluating AI Model Performance on Real‑World Economically Valuable Tasks, OpenAI

Measuring AI Ability to Complete Long Tasks, METR