JMM RESEARCH Machina Scientiae Search K Pricing
Menu
PredictionsEventsCompaniesToolsTrending stocksPeopleGuides Pricing CorrectionsTermsPrivacy

Search runs in your browser, across every published page on this site.

Sam Altman's AI compute capability model

Model Altman's compute-first thesis: hold a steady compute budget while the cost of compute falls, and see the capability multiple that same budget buys each year under your own cost-decline assumption.

Sam Altman of OpenAI, photographed in February 2025 Sam Altman Reviewed Aug 6, 2026 Rule located in Sam Altman (personal blog), 2025

Portrait: Office of the Prime Minister of Japan, CC BY 4.0, via Wikimedia Commons. Self-hosted by JMM.

Interactive model

Put your numbers through the rule

The source establishes the rule. The values below belong to you, and the output is JMM's deterministic calculation.

Make the assumptions yours

Every field recalculates immediately. Changed values can be copied into a shareable URL.

Compute per dollar in year 512.9×
Year-one capability multiple
1.7×
Cumulative capability bought
29.65
Total budget spent
$5,000,000

The model prices capability as units bought per dollar at an entered cost-decline rate. It does not claim how much capability converts into revenue or returns.

Inspect the calculationThe same result in a readable record view

Year 1

Capability multiple
1.7×

Year 2

Capability multiple
2.8×

Year 3

Capability multiple
4.6×

Year 4

Capability multiple
7.7×

Year 5

Capability multiple
12.9×
Source and translation

What the source says, and what the calculator adds

Source-supported ruleAltman states the rate rather than the direction: the cost to use a given level of AI falls about 10x every 12 months, which he contrasts with Moore’s Law at 2x every 18 months, and he treats driving that cost down as the precondition for putting AI in as many hands as possible.
JMM calculationJMM prices the thesis: a constant budget buys a growing number of capability units at an entered annual cost-decline rate, and the model reports per-year and cumulative capability multiples.
Formula and methodCapability per dollar in year n = (1 ÷ (1 − cost decline))n. Cumulative capability = the sum of yearly multiples times the constant budget.
Sam Altman (personal blog) Feb 9, 2025Three ObservationsSecond observation: "The cost to use a given level of AI falls about 10x every 12 months, and lower prices lead to much more use," with the GPT-4 to GPT-4o token price cited as roughly 150x over that stretch. Sam Altman (personal blog) Sep 23, 2024The Intelligence AgeClosing section: "If we want to put AI into the hands of as many people as possible, we need to drive down the cost of compute and make it abundant (which requires lots of energy and chips)."
Decision notes

What changes the answer

A compounding tailwind

A 40 percent annual cost decline makes the same budget buy nearly 13 times the capability in five years. The model makes that exponent visible instead of rhetorical.

Capability is not returns

The thesis connects compute to capability, not capability to profit. The model stops where Altman stops, and the limitation is stated rather than smoothed over.

The cost curve is real and the inference people draw from it is not.

Altman’s claim is unusually falsifiable for a technology prediction, and so far the direction has been right: the price of a given level of model capability has collapsed. What does not follow is the conclusion attached to it. Falling cost per unit of capability is a gift to buyers of AI and a problem for sellers of it, because a price that drops tenfold a year is a margin that drops with it unless volume grows faster. The same observation supports a bullish case for AI adoption and a bearish case for AI pricing power, and almost every use of it in an investment argument picks one and hides the other. Enter 90 in the cost-decline field to price his stated rate exactly, then notice that the model returns capability, not profit, because that is where the claim stops.

Written by the JMM editor. This is a view, not a recommendation for your circumstances. Who writes this

Not the calculator’s limits. The rule’s.

  1. Cost of a capability level is not the price of compute

    The 10x figure describes reaching a fixed quality bar more cheaply, achieved largely through better models and distillation. The market price of a GPU-hour has not moved anything like that way, and the two get quoted interchangeably.

  2. A trend stated as a rate is not a commitment to that rate

    The observation covers a short window in an industry with heavy subsidy and land-grab pricing. Extrapolating it five years compounds an assumption about competition, not a law of physics.

Limits

What this model does not know

  • Altman states a rate; he does not guarantee it, and the input here is your scenario rather than a vendor price list.
  • The cited rate is for the cost of a given level of AI capability, which is not the same thing as the price of a GPU-hour.
  • Capability is modeled as units per dollar; conversion into revenue, margins, or returns is not asserted.
  • The budget is held constant: no scaling capex, no competition, no supply constraints.
Questions people ask

Before you use the result

Why does the page default to 40% when Altman says 10x?

Altman’s stated rate is 10x every 12 months, which is a 90% annual decline. Enter 90 to price his claim exactly. The 40% first load is a deliberately conservative scenario, because 90% compounds to roughly 100,000x capability per dollar in five years.

Is this an OpenAI valuation model?

No. It prices capability per dollar and deliberately stops before any company or return claim.

Take the answer further

The next question this page cannot answer

Keep researching Sam Altman

Person fileSam Altman

Open the verified role and source record behind this current strategy.