Interactive Fermi model

How many tokens will analytics consume?

A scenario model for steady-state global LLM usage in data analytics. Adjust the market, workload share, and token price to see the infrastructure spend implied by your assumptions.

Public anchors

The updated model replaces the original Google run rate with May–July 2026 disclosures. These are directional anchors, not a complete census of global usage.

Google · all surfaces3.2Q+tokens/month · May 2026
Google · first-party APIs22Btokens/minute · Q2 2026
Microsoft Foundry300+customers tracking to 1T+ tokens/year
Visible annual floor>38QGoogle all-surfaces alone, annualized

Sources: Google I/O 2026, Alphabet Q2 2026, and Microsoft FY26 Q3.

Workload mix

Today’s observed developer-heavy mix is separated from the 2029 steady-state hypothesis. Data analytics is the highlighted share.

Build your case

The defaults reproduce the original 2029 base case. Move any assumption to update the analytics token volume and infrastructure spend.

650 Q/yr
250Q1,600Q
20%
10%30%
$1.00
$0.25$3.00
650Q × 20% × $1.00/M$130Banalytics infrastructure spend / year

130Q analytics tokens per year

Forecast envelope

The original global token scenarios, retained as the model’s uncertainty range.

YearBear caseBase caseBull caseKey anchor
20258–12 Q/yr15–20 Q/yr25 Q/yrEarly API and hyperscaler disclosures
2026E35–50 Q/yr60–90 Q/yr130 Q/yrGoogle alone annualizes above 38Q
2027E80–110 Q/yr120–180 Q/yr280 Q/yrAgent rollout and inference capacity growth
2028E160–220 Q/yr280–420 Q/yr700 Q/yrFalling unit costs unlock more workloads
2029E280–380 Q/yr500–800 Q/yr1,500+ Q/yrSteady-state enterprise agent adoption

Why 20%

The estimate triangulates workforce size, SQL penetration, and the extra token intensity of agentic analytics loops.

Workforce ratio

Roughly 12–18M data workers versus 28.7M software engineers, with data workflows estimated at 50–80% of SWE token intensity. That implies 10–26% of all tokens.

SQL penetration

SQL reaches 54% of professional developers, while Python is deeply used for data work. The proxy implies roughly 25–35% of coding tokens come from analytics.

Agentic multiplier

Plan → inspect schema → generate SQL → execute → validate → iterate creates repeated context-heavy calls, estimated at 5–15× a single-pass interaction.

Demand expansion

Natural-language analytics removes the SQL barrier and can expand the active user base by 3–5×, while enterprise schemas add 10K–50K context tokens per query.

Assumptions & caveats

The model is deliberately adjustable because both usage and token efficiency remain unusually uncertain.

Adoption lag

Data analytics trails software engineering adoption today, but the model assumes the gap narrows materially by 2029.

High context intensity

Enterprise schemas, tool calls, result inspection, and iteration make analytical sessions structurally heavier than casual chat.

Specialized-model risk

Smaller NL-to-SQL models, caching, and reusable query plans could compress token consumption without shrinking the software opportunity.

Measurement gap

Public disclosures mix API use, consumer surfaces, internal products, and different token-counting conventions. Treat every top-down total as approximate.