Technology · Reference

The price went up.
The price of the work
fell five and a half times.

Eight datacenter AI accelerator generations, 2017 to 2029, with launch price, cost per petaflop, and what each is worth one, two and three years later. Both halves of the answer are in the same table.

By Marcin Firmuga·2026-09-01·6 min read·Reference

Somebody comparing sticker prices concludes AI hardware is getting more expensive every year. Somebody comparing cost per unit of work concludes the opposite. Both are reading the same catalogue.

Where these numbers come from. Specifications and launch prices are public vendor figures, rounded. Residual values are computed by the depreciation model described below, which is the same code that runs inside the simulation game this table was built for. Two rows are marked as projections: they are roadmap extrapolation, not shipped product.

The table

Throughput is dense BF16 petaflop/s per unit. The FP8 and FP4 numbers vendors quote are a different measurement and are deliberately not used here. Residual value assumes purchase at launch and is expressed as a share of what was paid.

AcceleratorLaunchPF/sLaunch price $ per PFAfter 1 yrAfter 2 yrAfter 3 yr
V100 SXM2 32GB2017-06-210.125$10,000$80,00071%50%35%
A100 SXM4 80GB2020-11-160.312$15,000$48,07771%47%27%
H100 SXM52022-10-010.989$30,000$30,33471%37%20%
H200 SXM2024-02-010.989$32,000$32,35669%37%20%
B2002025-01-152.250$40,000$17,77870%36%19%
GB3002026-03-013.300$48,000$14,54570%36%23%
VR200 projection2027-06-016.500$55,000$8,46268%37%24%
Next generation projection2029-01-0113.000$62,000$4,76968%46%32%
5.5x The fall in cost per petaflop between the V100 in 2017 and the GB300 in 2026, from about $80,000 to about $14,500, while the price of a single unit rose from $10,000 to $48,000.

Every generation loses about 30 per cent in year one

This is the most consistent number in the table. V100, A100, H100, H200, B200 and GB300 all sit between 68 and 71 per cent of purchase price twelve months after launch, despite nine years and a fivefold change in what a unit costs.

Year two is where they separate. The V100 still held 50 per cent, because little shipped against it. The H100, H200 and B200 land near 36 per cent, because the launch cadence tightened and each of them faced more successors inside the same window.

The model, in three rules

It is deliberately simple, and every part of it is a number you can argue with.

value = price * 0.5 ^ (days_held / half_life_days) // time * (1 - 0.25 * ramp) per newer generation // successors, phased over 180 days value = max(value, price * 0.08) // scrap floor

The half-life is 730 days for parts up to the H200, 700 for the B200 and GB300, and 660 for the projected generations. It shortens because the cadence is tightening.

The successor penalty is the one that hurts

Time alone would be survivable. What actually removes capital is that every newer part of the same class takes a further quarter of what is left, and it does so over 180 days rather than on launch day, because a second-hand market reprices over a couple of quarters rather than overnight.

A launch that had already happened when you bought is not counted again. It was priced into what you paid.

What this means if you are actually buying

The decision is almost never what to buy. Within a generation the parts are close enough that the answer is obvious. The decision is when, and the table gives it a shape.

Capital committed shortly before a successor ships loses a quarter of its remaining value inside six months on top of ordinary time decay. Capital committed early in a generation and run at high utilisation converts into work before that happens. The same hardware, the same price, and a materially different outcome depending on the month.

Renting inverts the risk. It costs more per unit of work and carries no residual exposure at all, which makes it the correct answer when you cannot predict your own utilisation, and the expensive answer when you can.

What this table is not

It is a model, not a market report. Real second-hand prices are set by supply, export controls, hyperscaler dumping and how many units a buyer needs at once, none of which are here. The half-lives are chosen figures, not measured ones.

What it is good for is the shape of the decision: that the first year costs about thirty per cent whatever you buy, that successors matter more than age, and that cost per unit of work has moved in the opposite direction to cost per unit.

This table is the premise of a game. Scaling Laws is an AI company tycoon where the depreciation above is charged against your balance every day, whether the cluster is busy or idle. It is free for Windows and the source is open, including the valuation code quoted here.

The game page and the source

Questions people ask about this

How fast does an AI accelerator lose value?

About 30 per cent in the first year, consistently across every generation here. Near 36 per cent remaining by year two for modern parts, and close to 20 per cent by year three.

Has AI hardware got more expensive or cheaper?

Per unit, more expensive: $10,000 to $48,000 in nine years. Per petaflop, far cheaper: about $80,000 down to about $14,500, roughly 5.5 times better.

Rent or buy?

Renting costs more per unit of work and carries no residual risk. Buying is around a third of the cost per FLOP and puts depreciation on your balance sheet whether the cluster is busy or not. Utilisation and timing decide it, not price.

What is the floor on a used accelerator?

8 per cent of purchase price in this model. Value decays toward it and never below, because somebody always wants cheap silicon.

MF

Marcin Firmuga

Solo developer · HCK_Labs · building in public

I write about what I actually shipped, with real numbers and real code, including the parts that did nothing. More: my story.