How to Analyze AI Lab Economics Without Bad Comparisons

How to Analyze AI Lab Economics

Analyze an AI lab by keeping valuation, financing, revenue, revenue quality, serving cost, training expense, infrastructure capacity, capital commitments, cash use, and demand in separate dated measures. Do not place a valuation, annualized revenue figure, gigawatt target, and reported loss in one comparison as if they describe the same thing.

The safest tool is a metric ledger. If a number has no date, definition, source type, status, and limitation, it is not ready for analysis.

Begin with a ledger, not a narrative

The AI market produces dramatic numbers because the companies are raising capital, building infrastructure, serving large audiences, and disclosing selectively. Private companies do not provide the same standardized record as a public issuer's audited annual report.

Record the evidence before deciding whether the company is thriving, under pressure, efficient, or overextended.

This structure prevents a current financing from silently replacing an episode-era condition.

Funding, valuation, and cash are different

Funding is capital raised or committed under a transaction. Post-money valuation is the implied equity value after the financing. Neither is revenue.

OpenAI's March 31, 2025 announcement reported $40 billion in new funding at a $300 billion post-money valuation and 500 million weekly ChatGPT users. The company did not say that the $300 billion was cash available or annual sales.

One year later, OpenAI's March 31, 2026 announcement reported $122 billion in committed capital at an $852 billion post-money valuation. "Committed" belongs in the status field. The announcement should not be read as an audited cash balance.

Valuation can reflect expected future cash flows, strategic scarcity, investor demand, control terms, market conditions, or optionality. It is evidence of a financing price, not proof that the business already earns that value.

Revenue needs a period and quality

Annual revenue for a completed year, monthly revenue, annualized revenue, contracted revenue, bookings, and annual recurring revenue are not interchangeable.

OpenAI's March 2026 announcement says the company was generating $2 billion in revenue per month and that enterprise contributed more than 40 percent of revenue. Those are dated company disclosures. The page does not supply a full audited income statement, gross-margin bridge, retention cohort, or segment cost record.

Anthropic's April 2026 Amazon collaboration announcement says run-rate revenue had surpassed $30 billion, up from approximately $9 billion at the end of 2025. A run rate annualizes a current pace. It is not the same as recognized revenue for a completed twelve-month period.

Revenue quality asks where demand comes from, whether it recurs, how concentrated it is, what discounts or credits apply, how usage changes after pilots, and how much cost is required to serve it.

User count is not monetization

Weekly active users, paid subscribers, enterprise seats, API customers, and developers describe different populations.

OpenAI's March 2026 update reports more than 900 million weekly active users and more than 50 million subscribers. Dividing those numbers to produce a "monetization rate" would still be risky because the denominators, plan types, organizations, and timing may not align.

Anthropic reports more than 100,000 customers running Claude on Amazon Bedrock. That does not reveal equal usage, retention, revenue, or production depth across those customers.

A useful demand record defines the account, person, organization, active event, period, paid state, and product surface.

Compute must be separated into capacity, use, and cost

Gigawatts measure power capacity, not spending. Chip counts measure hardware units, not utilization or useful output. A contract ceiling is not deployment.

OpenAI's April 2026 infrastructure post says the company surpassed ten gigawatts of secured infrastructure. It also says more than three gigawatts were added in the prior ninety days. These are company capacity statements.

Anthropic's April 2026 announcement describes up to five gigawatts of new capacity and a commitment of more than $100 billion over ten years to AWS technologies. The ledger needs separate rows for commitment value, term, capacity ceiling, and deployment milestones.

OpenAI's original Stargate announcement said the project intended to invest $500 billion over four years and begin deploying $100 billion. An intention or plan should not be described as fully spent.

flowchart TD
    A["Capital raised or committed"] --> B["Infrastructure contracts and build"]
    B --> C["Available capacity"]
    C --> D["Utilized training and inference"]
    D --> E["Models and product service"]
    E --> F["Usage and revenue"]
    F --> G["Cash generation or further capital need"]

Each arrow requires execution. No one number proves the full flywheel.

Serving economics need missing denominators

Gross margin requires revenue and directly associated cost under a consistent accounting definition. Unit economics require a unit, such as a token, task, active user, seat, or successful workflow.

Token price is not serving cost. The provider may use several chips, routing systems, caches, model sizes, regions, and capacity contracts. A customer's price may reflect strategy as much as current marginal cost.

Training cost also needs a boundary. Hardware, cloud, power, data, labor, failed experiments, research, and allocated infrastructure can be counted differently. A reported model-training estimate cannot be compared casually with company-wide cash use.

When the denominator is missing, state that the ratio cannot be calculated.

Work through one date at a time

E092 recorded a late-2025 mood after Gemini 3. Several transcript numbers came from reporting or estimates and were not preserved with complete definitions. The 2026 disclosures do not prove that the episode was wrong. They show why the historical and current records must remain separate.

The table is not a company ranking. It demonstrates the ledger discipline.

The conclusions a public record can support

A reader can compare the scale and direction of company disclosures, the size and status of financings, the pace at which reported demand changed, the diversity of distribution, and the magnitude of infrastructure ambition.

Without audited or consistently defined data, the reader usually cannot establish comparative profitability, burn, gross margin, utilization, unit cost, customer retention, or return on invested capital.

That limitation is not a reason to stop analyzing. It is a reason to stop one sentence earlier than the headline wants.

[[Rough Vibes - Dalton Anderson on the AI Race After Gemini 3]] applies the date boundary to the episode. [[What Makes an AI Product Moat Beyond the Model]] asks whether capital and distribution create durable product advantage.

This guide reflects current first-party OpenAI and Anthropic disclosures, the recovered E092 catalyst, and the preserved transcript. It is analytical education, not investment advice. AI assistance was used for research organization, drafting, and validation. Publication remains unauthorized.