Researchers at Stanford University, with co-authors at UC Berkeley and UT Austin, ran the same AI model on four chips and measured the economic value each produced per unit of energy. In their paper, Joules-to-GDP, one NVIDIA B200 delivered $535.5 of net value per megajoule, against $124.1 for one AMD Instinct MI350X. By our arithmetic that is about 4.3 times as much.
On expert time saved, the B200 recorded 11.684 hours per megajoule against 1.260 for the MI350X, about 9.3 times, by our arithmetic. The comparison holds the model fixed: Qwen3.6-27B, run with the OpenHands harness on vLLM, on 220 GDPval professional tasks.
Business Pill · GPUs Explained
An evergreen Business Engineer explainer on GPUs and why AI runs on them. It teaches the general idea only and does not discuss this paper, NVIDIA or AMD.
The key insight: As we read it, the paper turns chip choice into a question of economic return per unit of power: with the model fixed, one B200 returned about 4.3 times the MI350X’s net value per joule, by our arithmetic, while the gap between NVIDIA’s own generations was about 1.33 times.
The paper describes Joules-to-GDP as “a framework that measures output dollar value minus AI execution cost per unit of energy, which we call Net Value per Joule (NVJ), and expert time saved per unit of energy, or Time Saving per Joule (TSJ)”.
It applies the framework to “220 GDPval tasks across 44 occupations”, weighted by sector GDP and employment. The authors include John Hennessy, Christopher Re and Azalia Mirhoseini. The arXiv listing shows version 2 dated 2 November 2026.
Net value per megajoule for Qwen3.6-27B on one accelerator across 220 GDPval tasks, from Table 25 of the Joules-to-GDP paper (Stanford, UC Berkeley, UT Austin). Point estimates; the paper gives 95% intervals. Ratios are our arithmetic.
What Was Posted on X
Stanford’s new research: The NVIDIA B200 delivered over 4x the dollar value and over 9x the expert time saved per unit of energy compared with the AMD MI350X.
i.e. chip alone, with the model unchanged, can decide how much economic return you get from your power budget.
Here, Stanford researchers ran the same AI model, Qwen3.6-27B, on 4 different single AI chips across 220 real professional tasks and measured how much useful work each produced for the energy it used.
Rohan Paul posted this summary of the Stanford paper at 21:34 UTC on 4 November 2026, saying the NVIDIA B200 “delivered over 4x the dollar value and over 9x the expert time saved per unit of energy compared with the AMD MI350X”. The figures in this piece are taken from the paper’s own Table 25; the ratios are our arithmetic.
The Four Chips
The authors ran Qwen3.6-27B “with OpenHands on one NVIDIA B200, H200, or H100 and one AMD Instinct MI350X, using the same expert-only reference and judge”. They write: “B200 delivers the highest GDP-weighted NVJ and TSJ among these deployments”.
In the paper’s Table 25, net value per megajoule is $535.5 for the B200, $402.2 for the H100, $391.3 for the H200 and $124.1 for the MI350X. Expert hours saved per megajoule are 11.684, 7.243, 7.293 and 1.260.
The gap between NVIDIA generations is smaller. By our arithmetic the B200 delivered about 1.33 times the H100’s net value per joule and about 1.61 times its time saving per joule. The H100 and H200 results are close, and their 95% intervals overlap.
Why the Chips Differ, per the Paper
With the model held fixed, the authors write, “differences in accelerator architecture, memory bandwidth, and power draw help explain differences in inference time and energy use, and hence the spread in NVJ and TSJ across devices”.
Energy covers metered GPU, CPU and DRAM. The paper also finds that adding chips can cost more energy than it saves: “eight H100s use 3.2x the energy of one to cut completion time by only 22%, at similar net value”.
The Paper’s Own Caveats
The intervals are wide. The MI350X’s 95% interval for net value per megajoule runs from $81.7 to $177.8, and the B200’s from $388.4 to $721.6. Execution cost, which is subtracted from value, uses different price sources: Modal’s published per-second prices for the NVIDIA chips and CloudRift’s on-demand rate for the MI350X. The authors write: “These provider prices do not guarantee identical host specifications or performance across providers.”
They also write that output dollar value “uses expert completion time and occupation median wage, not measured GDP”, and that future comparisons “could vary model, harness, backend, precision, batching, and retries individually to isolate their effects”.
As we read the table, the widest gap is between vendors rather than generations: B200, H100 and H200 sit between $391.3 and $535.5 of net value per megajoule, and the MI350X at $124.1.
As we read the caveats, the result is one model on one software stack with different price sources per vendor, so it measures these deployments rather than the chips in general.
Zeinali et al., Joules-to-GDP, arXiv:2610.30858
“More capable models and more hardware must therefore earn their extra energy with enough gains in output value.”
Three Implications
FOR BUYERS OF AI COMPUTE The paper measures return per joule as net value after execution cost, so the chip, the model and the price paid all enter the same number.
FOR CAPACITY PLANNING The paper finds eight H100s used 3.2x the energy of one to cut completion time by only 22%, at similar net value.
FOR CHIP COMPARISONS The B200 recorded 11.684 expert hours saved per megajoule against 1.260 for the MI350X, about 9.3 times by our arithmetic, with wide intervals on the MI350X.
The framework puts it this way: “Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”
As we read it, Joules-to-GDP adds a measure to that power equation: value delivered per joule, which in this test varied about fourfold across chips running the same model. The framework does not discuss this paper or these chips.
The comparison uses one model, one harness and one inference engine on one chip of each type, so it does not show how the chips compare on other models or software. The paper does not say whether the MI350X result reflects the chip or its software stack. NVIDIA and AMD were not asked to comment by us, and the paper does not report their views.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
In Stanford-led Joules-to-GDP research, the same model on one NVIDIA B200 returned about 4.3 times the net value per joule and about 9.3 times the expert time saved per joule of one AMD MI350X, by our arithmetic from the paper’s table, with wide intervals and different price sources behind the numbers.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. We read the Joules-to-GDP paper on arXiv (version 2, 2 November 2026); every figure and quotation comes from it. X posts by Rohan Paul pointed us to it. Ratios between chips are our arithmetic from the paper’s Table 25. We did not contact NVIDIA, AMD or the authors.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
