NVIDIA Cites Five Sources for AI-Factory Returns

NVIDIA’s 1 October blog post on AI-factory returns cites five outside sources. This publication opened three of the originals, SemiAnalysis, Sprout and Ornn Data, and could not open Silicon Data or Barkr. Where the originals add detail, such as an engine label or a comparison the NVIDIA post does not include, this piece reports it. NVIDIA sells the hardware it discusses, and nothing here is independently verified beyond reading those sources.

What NVIDIA Argues

NVIDIA’s post of 1 October 2026, “Productive, Durable, Fungible: How NVIDIA AI Factories Maximize Return on Investment,” says “Each megawatt factory costs roughly $60 million.” It says three things shape a factory’s return: earning capacity, useful life and demand.

NVIDIA says its platform is productive, durable and fungible, which it says maximizes all three. To support that, the post cites SemiAnalysis, Sprout, Barkr, Silicon Data and Ornn Data. This publication opened three of the five originals and could not open the other two.

The $60 million per megawatt figure is NVIDIA’s, and this publication did not check it.

The Throughput Claim and the Engine Label

The post says SemiAnalysis AgentX data shows Vera Rubin NVL72 systems “deliver over 30x higher throughput per megawatt than NVIDIA GB300 NVL72, and up to 45x lower cost per million tokens on the DeepSeek V4 Pro model.”

NVIDIA’s earlier post on this, dated 24 August and updated on 15 September 2026, worded the throughput line as “up to 30x higher throughput per megawatt.” The 1 October text says “over.” Neither version gives an interactivity target or a serving engine.

SemiAnalysis’s written post of 14 September does give them, and the multiple depends on both. For DeepSeek V4 Pro it reports Rubin at about 2.09x the stronger GB300 engine at a 100 tokens-per-second interactivity target, about 7.2x GB300 SGLang at 150, and 2.72x at 200.

At exactly 170, SemiAnalysis reports 62.9x the throughput per megawatt of GB300 TRTLLM and 5.56x against GB300 SGLang. It adds: “The engine label is therefore essential when quoting the high-interactivity gain.”

SemiAnalysis also says the size of the advantage “depends on the interactivity target and the serving engine used for comparison.” This publication did not check the 45x cost figure against SemiAnalysis, whose result tables appear as images.

How Long a GPU Earns: What Sprout Says

NVIDIA says every major operator has extended server life and cites a Sprout chart. Sprout’s own post of 30 September says Microsoft ran several GPU generations for 6.9 to 8.8 years, and that its V100 fleet ran 8.4 years against a six-year book life.

Sprout also says AWS has never retired an A100 and that rental rates for prior-generation GPUs rose through 2026. It says six-year-old A100 systems are selling at roughly 38% to 43% of their launch price, above the 25% residual now used in lending.

NVIDIA’s post gives a different measure from Silicon Data: a six-year-old A100 “still worth a quarter of what it cost.” This publication could not open the Silicon Data post, so it cannot compare the two.

The 80 Percent: What Ornn Says

NVIDIA says Ornn Data finds “the market paying 80% as much to rent an A100 GPU on a five-year contract as on a one-month contract.” Ornn’s paper, “The Economics of Open-Weight Inference,” reports the A100 figure as 80.2%, from marks published on 13 August.

The same paper gives the other families. At five years the H100 is 59.8%, the H200 43.7%, the B200 53.8% and the B300 53.8%, which Ornn summarises as 44 to 60 percent for Hopper and Blackwell. NVIDIA’s post quotes the A100 figure only.

Ornn describes the marks as “analyst-produced indicators, not executable quotes or verified averages of comparable executed contracts at every tenor.” It also says percentage retention depends on each family’s starting level, so a flatter curve “need not mean a higher absolute rent.”

Ornn’s disclosures say it licenses data commercially and also offers GPU rentals through Ornn Compute, which it says creates “a commercial interest in both the interpretation and adoption of its data.” Its own diagram lists continued earning capacity of older accelerator families as an “economic inference to be tested.”

The Two Not Opened

NVIDIA says Barkr puts useful life at five to six years for an eight-GPU H100 system and nine to 10 years for GB300 NVL72, based on resale. Barkr’s report page lists a 16-page PDF dated 29 September, “The Resale Standard: Forward-Looking GPU Valuations,” which it describes as forward valuation curves for the H100 and GB200 NVL72.

This publication did not open that PDF, so it cannot say where the GB300 figure comes from. The Silicon Data claim above is also unchecked.

What Is Not Established

The $60 million per megawatt figure, the claim that CoreWeave extended bookings on 2020 units through 2029, the Silicon Data figure and the Barkr figures are NVIDIA’s statements and are unchecked here. The 45x cost figure is unchecked too.

Nothing here judges whether an AI factory earns a good return. NVIDIA’s post is a company blog that sells the hardware it discusses, and every figure above is attributed to its source. Nothing in this piece is investment advice.

For this publication’s earlier look at the SemiAnalysis Rubin numbers, see our piece on the 67x and 3x figures. For the accounting terms, see the Business Pill on capex and depreciation.