Three Boring Words That Explain the Entire AI Buildout

Buying faster chips is not what makes an AI data center pay for itself. That sounds wrong, especially if you have spent the past few years reading headlines about speed records and benchmark wins. But NVIDIA’s own framing of the problem puts performance inside a much less glamorous trio of ideas: productive, durable, fungible. Those three words do more to explain the economics of AI than any throughput chart.

I want to walk through what each one actually means, because once you have them, a lot of confusing industry news suddenly makes sense.

Productive means output per watt, not output per chip

An AI factory is a building full of computers that turns electricity into tokens. Tokens are the small chunks of text, code, or images that AI models read and write. Every answer an AI agent gives you is a pile of tokens.

Here is the constraint most people miss: these facilities are limited by power, not by floor space or chip count. You cannot plug in more hardware than your electrical supply allows. So the number that matters is tokens per megawatt — how much useful AI work you squeeze out of each unit of electricity.

NVIDIA’s claims on its Vera Rubin NVL72 system are built around exactly that measure. The company says it delivers more than 30x higher throughput per megawatt than the GB300 NVL72, and up to 45x lower cost per million tokens running DeepSeek V4 Pro.

Notice what those two numbers have in common. Neither is “this chip is faster.” One is output divided by power. The other is cost divided by work delivered. That is the vocabulary of someone selling to a person who signs an electricity bill, not someone chasing a benchmark leaderboard.

Why this matters if you use AI agents

Cost per million tokens is the hidden variable behind every AI product you touch. AI agents are token-hungry by nature — they think in steps, re-read their own notes, call tools, and check their work. A single agent task can consume far more tokens than a simple question-and-answer exchange.

When the cost of generating tokens drops, agents that were previously too expensive to run become ordinary features. That is the quiet mechanism behind why AI assistants keep getting more capable without getting more expensive to use.

Durable means the hardware keeps earning long after it stops being new

The common assumption about AI hardware is that it ages like milk. New generation arrives, old generation becomes scrap. NVIDIA points to evidence that this is not what happens.

The A100 GPU shipped in 2020 and is still in commercial service. Six years later, it is still doing paid work. Meanwhile CoreWeave has extended bookings through 2029 — customers committing to capacity years out.

The most telling signal is an accounting one. Every major operator has extended the depreciation schedule on its servers. Depreciation schedules are how companies spread the cost of equipment across the years they expect it to be useful. Stretching that schedule is a formal statement, reviewed by accountants and disclosed to investors, that the hardware will earn for longer than previously assumed.

Accountants are not known for optimism. When they extend a schedule, they are responding to observed reality.

Fungible means one machine, many jobs

Fungible is the least familiar word of the three. It means interchangeable — able to be swapped into a different role without losing value. A $20 bill is fungible. A concert ticket for a specific date is not.

Applied to AI infrastructure, fungible means the same hardware can run whatever work shows up. Training a new model, serving chat requests, running agents that plan and call tools, generating video, handling scientific simulation. NVIDIA’s claim is broader workload support across its systems.

This is risk management dressed up as a technical feature. Nobody building a facility today knows what AI work will look like in three years. Agents barely existed as a serious workload a few years ago. Hardware that only does one thing well is a bet on a specific future. Hardware that handles many things is a bet that something will be in demand.

Reading the industry through this lens

These three ideas work as a filter. Next time you see an AI infrastructure announcement, the useful questions are: how much work per megawatt, how many years will it keep earning, and how many different jobs can it take on.

There is a fair amount of self-interest in NVIDIA framing the conversation this way. A company whose hardware scores well on durability and flexibility benefits from making those the terms of debate. The claims are NVIDIA’s own, measured on NVIDIA’s own comparisons.

Still, the framing holds up independently. Depreciation schedules extended by multiple operators and multi-year bookings are decisions made by other companies putting their own money at stake. Productive, durable, fungible is an unglamorous way to describe the most expensive construction boom in computing. It is also, probably, the accurate one.

🕒 Published: October 1, 2026