Research brief · AI economics · October 2026 · Epoch, a16z, Altman + sceptics

The collapsing cost
of intelligence

The price to use a fixed unit of AI intelligence is falling about 10x a year, faster than Moore's law ever moved. At the same time, the cost to build the frontier is rising about 2.4x a year. We read Epoch, a16z, the labs and the sceptics, and mapped where two opposite curves are taking the whole business.

By Christos Papadimitriou · House of 47 Research
Last updated:

See the fall
10x/yrPrice of a fixed capability falls each year
1,000xCost to reach GPT-3's level, in 3 years, $60 to $0.06
2.4x/yrFrontier training cost rising, past $1B by 2027
3.2QTokens a month Google now processes, up ~330x in 24mo
A computer chip in close-up
The price of GPT-3-level intelligence · 2021 → 2025 · per million tokens

The floor falls out

In November 2021, one model could reach a given level of general knowledge, and it cost sixty dollars per million tokens. Three years later the same capability cost six cents. This is not a discount, it is a regime: a fixed unit of intelligence gets roughly ten times cheaper every year, and the floor keeps dropping toward zero.

Nov 2021 · GPT-3

The only option

$60

One model could clear the bar, at $60 per million tokens. Intelligence was scarce, metered, and expensive. The starting line for everything that followed.

Mar 2023 · GPT-3.5

30x in a year

$2

The ChatGPT API lands at two dollars per million tokens. Same capability, a thirtieth of the price, in sixteen months. The slide has begun in earnest.

2024 · Open weights

Near the floor

$0.07

Llama, Gemini Flash and others reach the same level for about seven cents. Open weights turn pricing brutal once many labs match a capability.

2025+ · Too cheap

Near-free

~0

The cost to use a fixed level of AI approaches zero. Intelligence stops being the scarce input. What you point it at, and who pays the frontier bill, become the questions.

The two curves of the barbell

Cost to use a fixed capability

Cross-check · Epoch AI, a16z, OpenAI, NVIDIA, Sequoia

Where they agree

  • All
    A fixed capability gets ~10x cheaper a year

    a16z and Altman put it at 10x every 12 months. Epoch measures 9x to 900x a year across benchmarks, median near 50x. Faster than Moore's law at its peak.

  • All
    Two engines compound

    Algorithmic efficiency improves ~3x a year, hardware price-performance roughly doubles every 2 to 2.5 years. Multiplied together, they crush the unit cost.

  • All
    Open weights are the accelerant

    Once Llama, Qwen or DeepSeek match a level, many providers serve it and pricing turns brutal. Competition, not charity, drives the floor down.

  • All
    Jevons paradox: cheaper means more

    Falling token prices grow usage, not shrink spend. Google went from ~10 trillion tokens a month to over 3 quadrillion. Nadella: AI use will "skyrocket."

  • Most
    The frontier gets more expensive

    Training a frontier model rises ~2.4x a year, heading past $1B by 2027. Data-center capex hit ~$420B in 2025 and is heading over $1 trillion in 2026.

  • Most
    The model layer commoditizes

    When capability is near-free and replicable, model access is no moat. Value moves to distribution, data and workflow. The moat is not the model.

Where they disagree

How long does ~10x a year last?

Does the deflation keep compounding, or do scaling returns and physical limits stall it? Placed by how long each camp expects it to run.

Plateaus soonContinues for years

Does the frontier capex ever pay off?

The build bill is real and rising. Does exploding usage absorb it, or is it a bubble resting on loss-making labs?

Never closesAbsorbed by demand

Where does the value end up?

If the model is a commodity, who captures the margin as intelligence goes near-free?

The modelWhat you build around it

Epoch AI 2024-25 (inference 9x-900x/yr; training 2.4x/yr; GPU FLOP/$ ~2x per 2.3yr) · a16z 2024 (LLMflation, 10x/yr, 1,000x in 3yr) · Altman 2025 (10x/12mo, "too cheap to meter") · Nadella 2025 (Jevons paradox) · Sequoia / Cahn 2024 ($600B question) · Marcus 2025 (diminishing returns) · Zitron 2025 (OpenAI loss figures).

Probability of outcome · House of 47 synthesis

The odds

Synthesis estimates across the sources below, scenario estimates, not market forecasts. The through-line: the cost to use intelligence keeps falling and usage keeps exploding, while the frontier bill keeps climbing, and the two stories have not yet been reconciled.

The barbell · two prices moving in opposite directions

Two speeds

The cost to use a fixed capability deflates toward zero. The cost to build the frontier inflates toward the hundreds of billions. The squeezed middle is undifferentiated model access, exactly the thing that looks most like a business today.

Cost to use (inference)Cost to build (frontier)

Directional, synthesised from Epoch, a16z, Altman, Dell'Oro and Sequoia. "Use" tracks the cheapest price to reach a fixed capability; "build" tracks frontier training runs and the capex behind them.

House of 47 read · What to do about it

Own what you point it at

Intelligence is not the moat

If a fixed capability is near-free and the model layer is a commodity, buying access to a model is no edge. Everyone has the same intelligence on tap. The edge is what you do with it, and what only you can do with it.

Own the agents, not the tokens

When inference is near-free, value accrues to the systems that turn intelligence into finished work: the workflows, the proprietary data, the client relationship. Owning the workforce that does the work is the business of theagency47.

Spend the savings on usage

Jevons is the strategy. Do not bank the price cut, redeploy it. The winners run intelligence everywhere, on every task, not sparingly on a few. Cheap tokens reward the ones who use the most.

Stay portable on the frontier

The build bill is rising and sits with a handful of labs. Do not hard-wire your business to one model you do not control. Keep workloads portable across providers so the commodity stays a commodity for you.

Base case

Plan for intelligence as a near-free utility. The cost to use a fixed capability keeps falling about 10x a year for the next several years, usage explodes to match, and the durable margin sits in distribution, data and workflow, not in model access.

Tail

Stress-test for the reconciliation. If scaling returns plateau or the capex never earns out, the frontier bill forces a repricing. The deflation on the use side is real either way, so build to benefit from cheap intelligence, not to depend on any one supplier of it.

Intelligence is getting too cheap to meter. The scarce thing is knowing what to do with it.

Sources · figures checked against primary and named reporting

Sources

Epoch AI, LLM inference price trends (9x-900x/yr), analysis

Epoch AI, how the cost of LLM inference changed, analysis

a16z 2024, LLMflation, 10x/yr and 1,000x in 3 years, essay

Epoch AI, cost to train frontier models (~2.4x/yr, $1B+ by 2027), analysis

Epoch AI, trends in GPU price-performance, analysis

OpenAI 2020, AI and Efficiency (44x in 7 years), report

Sam Altman 2025, Three Observations (10x/12mo), essay

Sam Altman 2025, "too cheap to meter", via Fortune

Satya Nadella 2025, Jevons paradox, via Fortune

Alphabet 2025, token volume (980T+/mo), Q2 2025 earnings

Tom Tunguz 2025, OpenAI ~6B tokens/min, analysis

DeepSeek 2024, V3 ~$5.6M training-run claim, via IEEE Spectrum

NVIDIA 2025, Blackwell cost-per-token claims, benchmarks

Stargate 2025, $500B AI infrastructure, SoftBank

Dell'Oro Group 2025, data-center capex +57% to $420B, report

Sequoia / David Cahn 2024, AI's $600B question, essay

Gary Marcus 2025, diminishing returns to scaling, analysis

Ed Zitron 2025, OpenAI loss figures, analysis

Changelog · what changed in this brief, newest first

Changelog