Is compute getting more expensive?

Two claims hold up the AI infrastructure trade. I checked them against a year of public price data.

Dan Bressner · Updated October 14, 2026 · Part 1 of a series on the economics of AI (part 2: who's winning open weights)

In October 2026, Gavin Baker of Atreides Management argued that the case for AI infrastructure comes down to two claims. First, GPU rental prices are rising, not falling. Second, citing 3Fourteen Research, a B200 cluster earns roughly a 37% internal rate of return even if rental prices fall 40% a year. Most bullish positions in the trade lean on some version of them.

Both can be checked against public data, so I checked them. I built a pipeline that collects GPU rental prices from public marketplaces, loaded a year of weekly history into SQLite, and rebuilt the cluster return model from scratch. This page is generated by that pipeline, and the methods section covers how it works. Highlighted terms are in the glossary; hover or tap.

Claim 1: rental prices are rising

For the year of history I have, yes. The median on-demand price rose for every datacenter GPU in the dataset, old and new. Spot prices, the tier most sensitive to demand, rose fastest.

H100 on-demand, trailing yr

+15%

$2.98 → $3.42

H200 on-demand, trailing yr

+33%

$3.50 → $4.67

B200 on-demand, trailing yr

+31%

$5.39 → $7.08

H100 spot, trailing yr

+149%

$0.90 → $2.24

On-demand rental price, weekly median

USD per GPU-hour across cloud providers · Oct 2025 to Oct 2026

Trailing-year price change by GPU

Change in median on-demand price, first 4 weeks vs last 4 weeks

A few details stand out. The A100 is a six-year-old GPU and its price still rose, so this isn't enthusiasm for the newest chip. The gap between spot and on-demand collapsed: H100 spot went from $0.90 to $2.24 an hour, which is what it looks like when the cheap, interruptible slack in a market gets bought up. AMD's parts rose alongside Nvidia's, so the demand is for compute generally, not one vendor's product. And prices rose even though supply kept broadening. The number of providers listing an H100 went from 36 to 41 over the year, and H200 listings went from 21 providers to 33. More sellers usually means lower prices. Not this year.

Context, before anyone gets carried away. The two years before this window were rough for GPU landlords: H100s that rented for $7 to $8 an hour in 2023 were down to $2 to $3 by mid-2025 as supply caught up (Silicon Data tracks the full history, and Introl covered the collapse). So this year is a reversal, not a continuation. The rise is real and broad, and it also started from a low base. Both things are true.

The standard assumption is that a GPU is a depreciating asset, with each new generation pushing the rents of older chips down something like 30% a year. From 2023 to 2025 that assumption looked conservative. This year it didn't just miss, it got the sign wrong.

Full table: all GPUs, on-demand and spot

GPU	billing	Oct 2025 (4-wk avg)	Oct 2026 (4-wk avg)	change	providers
MI300X	on demand	$1.99	$2.81	+41.4%	5
MI300X	spot	$1.85	$1.68	-9.4%	1
MI355X	on demand	$4.09	$8.04	+96.5%	2
MI355X	spot	$2.59	$4.84	+86.9%	2
A100	on demand	$1.60	$1.67	+4.3%	40
A100	spot	$1.08	$0.83	-22.9%	7
B200	on demand	$5.39	$7.08	+31.3%	20
B200	spot	$2.69	$4.99	+85.6%	6
B300	on demand	$5.39	$8.84	+64.0%	14
B300	spot	$6.94	$5.40	-22.2%	6
GB200	on demand	$13.25	$16.00	+20.8%	3
GB200	spot	$27.04	$27.04	+0.0%	1
H100	on demand	$2.98	$3.42	+14.8%	41
H100	spot	$0.90	$2.24	+149.1%	9
H200	on demand	$3.50	$4.67	+33.4%	33
H200	spot	$1.63	$2.89	+77.1%	8

Claim 2: a 37% IRR even if prices fall 40%/yr

This claim needs a model, because the result depends on assumptions the claim doesn't state. I modeled a single GPU's slice of a B200 cluster: all-in capex for the GPU, server, networking and datacenter fit-out; power and fixed opex; a five-year life; 10% residual value; and monthly rental revenue that declines at a chosen annual rate. The IRR comes out of those monthly cash flows. Every parameter is listed in the methods section, and the two tables below show what happens as the big ones move.

Cluster IRR: starting price × annual price decline

$55k capex/GPU, 80% utilization, 5-yr life · outlined cell ≈ the claim's scenario

Cluster IRR by starting rental price and annual price decline
start price	−0%/yr	−20%/yr	−40%/yr	−60%/yr
$3.00/hr	24%	-2%	-28%	-45%
$4.00/hr	46%	15%	-16%	-41%
$5.00/hr	69%	34%	-2%	-35%
$6.00/hr	94%	54%	13%	-28%
$7.75/hr	141%	93%	43%	-10%

Blue cells are positive returns, red cells are losses. At the current $7.75/hr B200 on-demand median, a 40%/yr price decline still yields about 43%, close to the claimed 37%. At committed-contract prices of $4 to $5 an hour, the same decline produces losses.

Same stress case, by capex and utilization

$5/hr start, −40%/yr price decline

Cluster IRR by capex per GPU and utilization, at $5/hr start and 40% annual price decline
capex/GPU	60% util	70% util	80% util	90% util
$40,000	0%	12%	26%	41%
$55,000	-18%	-10%	-2%	6%
$70,000	-27%	-21%	-16%	-10%

Even at $40k per GPU, the stress case needs utilization of 70% or better, sustained for five years, to break even.

So the claim reproduces, but only in one corner of the assumption space. It needs on-demand marketplace rates held for the full life of the hardware, or unusually cheap capex with high sustained utilization. Large clusters mostly sell capacity through committed contracts priced well below on-demand, and at those prices a 40% annual decline wipes the return out. My base case, $5 an hour and $55k of capex per GPU, comes out at -2%.

There's a better version of the argument, and it's the one the data supports. Cluster economics look excellent right now because prices rose. The business isn't robust to a 40% annual price decline, and it doesn't currently need to be. The return claim and the price trend are the same bet stated two ways: that demand keeps outrunning supply.

Verdict

Supported by the data

"Compute prices are rising." Every datacenter GPU in the dataset rented for more on-demand than it did a year earlier, and spot prices rose between 77% and 149%. The caveat: this is marketplace and listed-provider pricing. Hyperscaler contract prices are negotiated privately and may behave differently.

True only under aggressive assumptions

"About 37% IRR even if prices fall 40%/yr." Reproducible if a cluster earns on-demand rates for five years. On-demand is the most expensive tier, and big clusters mostly sell below it. At realistic contract prices the stress case fails, and the tables above show where.

Three things would change this verdict. Falling prices for two consecutive quarters, with spot leading the way down as it led on the way up. Provider counts rising while prices fall, which would mean new supply landing faster than demand. Or disclosed contract repricings at the big neoclouds coming in below their 2024 and 2025 rates.

Objections worth taking seriously

Fair challenges to this analysis, and where each one lands.

"These are listed marketplace prices, not what anyone actually pays."

Partly fair. The median listed price isn't a transaction price, and the biggest deals in this market are private. But listed prices are the market's visible edge, they move with the private market's supply and demand, and a year of them rising across every GPU and both public tiers is not nothing. The private-contract claim cuts both ways, too: it's the same reason nobody quoting hyperscaler economics can fully back their numbers either.

"The median moved because the provider mix changed, not because the market did."

The data argues against this. Provider counts grew for every major GPU over the year (H100 from 36 to 41, H200 from 21 to 33, B200 from 10 to 20), and new entrants typically list below the incumbents to win business. A broadening, more competitive supply base alongside rising medians is the opposite of a mix artifact.

"You started measuring at the bottom."

True, and the page says so. H100 prices collapsed from 2023 to mid-2025 before this window opens. One year of rising prices after a two-year slide is a reversal, not a regime. That's why the verdict names the specific signals that would flip it rather than claiming a trend.

"Your model isn't 3Fourteen's model."

Correct. Theirs isn't public, so nobody outside can check it, which is part of the point. What I can do is show which assumptions are required to produce their number. The answer: on-demand rates held for five years, or cheap capex with high utilization. If their model gets to 37% another way, publishing the assumptions would settle it.

"Rising prices might just be Blackwell scarcity or datacenter power constraints, not durable demand."

Could be, and this data can't fully distinguish those stories. Two observations push against the pure-scarcity reading: six-year-old A100s and AMD's parts rose too, which scarce Blackwell supply doesn't explain, and prices rose while provider counts grew. But if the constraint is power rather than chips, the price strength could outlast chip supply and still end abruptly when new capacity energizes. That's a real risk and the model's downside rows show what it does to returns.

Methods & data

The pipeline is about 300 lines of dependency-free Python, standard library only: urllib, csv, sqlite3. Charts render with Apache ECharts 5.6.0, pinned from cdnjs, from data embedded in the page.

This research was built working with Claude, Anthropic's AI. It wrote most of the pipeline code and drafted prose under my direction; the questions, method, assumptions and conclusions are mine, and so are any errors.

Weekly price history: the GetDeploying GPU rental price dataset (CC BY 4.0), covering 103 GPU models over the trailing 53 weeks. Prices are the median of listed prices across providers.
Live marketplace snapshots: the Vast.ai public offers API, captured on each run for spot-level detail.
Storage and analysis: raw files are kept immutable and date-stamped. Loads are idempotent upserts into SQLite. Trend figures compare 4-week averages at each end of the window to smooth single-week noise.
Return model: monthly cash flows, solved by bisection. Every assumption is a named parameter, and both sensitivity tables above come directly from analysis/irr.json.

Model assumptions and where they come from

Parameter	Base value	Basis
Capex per GPU	$55,000	8x B200 HGX servers listed at $360k to $394k in late 2026, about $45k to $49k per GPU, plus networking, power distribution and fit-out. The grid tests $40k to $70k.
Starting rental price	$5.00/hr	Below the $7.75 on-demand median in this data, reflecting committed-contract discounts. The grid tests $3.00 to $7.75.
Price decline	40%/yr	The claim's own stress case. The grid tests 0% to 60%.
Utilization	80%	Assumption. The grid tests 60% to 90%.
Power draw	1.4 kW/GPU	Roughly 1 kW of board power plus a share of server and cooling load. Estimate.
Electricity	$0.07/kWh	Just under the EIA US industrial average of about 8.7 cents in 2026; large operators site where power is cheaper.
Fixed opex	5% of capex/yr	Estimate covering staff, space, maintenance and insurance.
Hardware life	5 years	Matches common accounting treatment of AI hardware; the real number is debated.
Residual value	10% of capex	Estimate. Prior-generation GPUs still resell six years on.

Where a value says estimate, that's exactly what it is. The sensitivity grids exist so you can judge how much any single one matters.

Limitations

Marketplace and listed pricing, not negotiated hyperscaler contracts. The prices Baker's larger argument turns on are mostly private.
The history covers 53 weeks. One year of rising prices is a fact, not a regime.
A median of listed prices can move when the mix of providers changes, not only when the market does. Provider counts are in the data table as a partial control, and they grew alongside prices.
The model omits taxes and financing structure, and it treats price and utilization as independent. In a glut they'd fall together, so the real downside is worse than the tables suggest.

Glossary

Capex: Capital expenditure: the upfront cost of the asset. For a cluster this means the GPUs plus the servers, networking and datacenter fit-out around them.
Committed contract: A negotiated deal for capacity over months or years, priced below on-demand in exchange for the commitment. Most large cluster capacity is sold this way, and the terms are rarely public.
Depreciation: The decline in an asset's value as it ages. For GPUs, the standard assumption is that each new generation pushes the rental rates of older chips down until they no longer earn their running costs.
GPU generations: Nvidia's datacenter GPUs arrive in generations: A100 (Ampere, 2020), H100 and H200 (Hopper, 2022 to 2024), then B200, B300 and GB200 (Blackwell, 2024 to 2026). Each generation is much faster than the last, which is why older chips are expected to rent for less over time. MI300X and MI355X are AMD's competing parts.
GPU-hour: One GPU rented for one hour. The standard billing unit for rented compute.
Hyperscaler: One of the largest cloud providers: Amazon Web Services, Microsoft Azure, Google Cloud. They build datacenters at enormous scale and negotiate their compute deals privately.
IRR (internal rate of return): The annualized return implied by a stream of cash flows. Formally, the discount rate at which the investment's net present value is zero. An IRR above your cost of capital means the investment creates value; 37% would be exceptional for infrastructure.
Median listed price: The midpoint of the prices providers advertise for a GPU. It tracks the market, but it is not a transaction price. What large buyers actually pay is usually negotiated below it.
Neocloud: A newer cloud provider built around GPU compute rather than general cloud services. CoreWeave, Lambda, Nebius and Crusoe are examples.
On-demand: Pay-as-you-go rental with no commitment, billed per GPU-hour. The most expensive tier, and the one with public prices, which makes it the easiest to measure.
Opex: Operating expense: the recurring cost of running the asset. For a cluster, mainly electricity, staff, maintenance and space.
Payback period: How long an investment takes to earn back its upfront cost. A rough, intuitive check that ignores timing and terminal value; IRR is the stricter version of the same question.
Residual value: What the hardware is worth at the end of the investment period, expressed here as a share of its original cost. Old GPUs keep some resale value; A100s from 2020 still trade today.
Spot: Discounted capacity the provider can reclaim at short notice. Spot prices float with whatever supply is sitting idle, so they react to shifts in demand faster than any other tier.
Utilization: The share of hours a GPU is actually rented and earning revenue. Fixed costs run whether or not the hardware is busy, so returns are extremely sensitive to this number.