To separate price from capability in revenue growth, start with a price, volume and mix bridge built from company disclosures and analyst models. Normalised performance per watt and performance per dollar come second, and their only job is to test whether the capability story is technically plausible. A benchmark cannot size the revenue figure: LynxBenchAI measures throughput, latency, memory, energy and cost per unit of work, not shipments or selling prices.
LynxBenchAI covers the capability half of that check. The 26Q3 release publishes measured scores for the same training, inference and compute tests across 48 devices, so the ratio between two hardware generations comes from measurement, not from a spec sheet. The first step is to find both devices in the AI Hardware Directory of all LynxBenchAI results and read their per-test scores side by side.
Build the price, volume and mix bridge before touching a benchmark
Price, volume and mix (PVM) analysis splits the revenue change between two periods into three drivers: what was charged per unit, how many units were sold, and how the product composition shifted. The standard formulas are Price Effect = (Actual Price − Base Price) × Actual Volume, Volume Effect = (Actual Volume − Base Volume) × Base Price, and Mix Effect = (Actual Mix % − Base Mix %) × Total Base Revenue. Add the three effects together: if they do not equal the total revenue variance, the bridge is wrong. Some guides calculate mix as the residual left after price and volume are taken out, so mix always closes the gap by construction.
The split matters for more than presentation. In industrial revenue analysis, price-driven growth reaches margin at close to 100% incremental contribution. Volume growth carries normal variable costs, with incremental margins typically of 30-50% for manufacturers. A PVM bridge also separates real demand growth from volume that rose only because prices were cut. The capability question needs exactly that distinction. Volume reflects unit demand changes, and mix reflects shifts in the composition of products sold. Pricing power shows up as price.
New products break the standard formulas, because a product sold only in the current period has no base price or base volume. Icypeas recommends running PVM on the comparable portfolio only and reporting separate new-product and discontinued-product effects. For a chip vendor in the middle of a generation change, the new-product bucket is often where the capability story actually sits. Burying it in mix credits the wrong driver.
Are there disclosures that separate ASP from unit shipments?
Sometimes, but rarely completely. In the United States, Regulation S-K Item 303 requires management’s discussion to describe how far material revenue changes come from price changes, from volume, or from new products and services. That makes the MD&A section of 10-K and 10-Q filings the first place to look for a company’s own price-versus-volume attribution. Earnings calls often use the same framing, with statements such as “price increased 8%, adding $120 million”.
Exact unit and price splits are rarely disclosed. Epoch AI’s chip-sales methodology notes that most companies do not disclose exact chip sales for most chip types and periods. Its figures are estimates built from revenue disclosures, earnings commentary and analyst price reports, often with significant uncertainty. One rare direct disclosure shows how much uncertainty remains even then. Nvidia’s CEO said the company had made $100B from 4 million Hopper GPUs through October 2025 (excluding China), which works out to $25k per GPU. Epoch’s own revenue model points to an average closer to ~$30k, which Epoch says could be consistent with rounding. For AMD, Epoch has meaningful uncertainty about Instinct revenue because the company does not regularly disclose the split between EPYC and Instinct. It also notes that contract prices vary by customer and deal structure in ways a reported ASP range cannot fully capture.
A decomposition built on these inputs is a range, not a point. Treat any capability share derived from it as approximate as the least-disclosed input.
Premium-SKU mix shifts get mistaken for capability
A falling average price can mean either a real price cut or more sales of SKUs that were already cheaper. Without a decomposition, the two cannot be told apart. The same ambiguity runs in the other direction. A shift toward higher-priced, higher-margin products raises revenue even when total units are flat. Datama’s worked example shows average price falling by nearly €4 while no list price changed, because the cheaper product’s share grew from 20% to 45% of units.
If buyers move to the higher-priced flagship part, average revenue per unit rises. That is a mix effect, even when the flagship is also the more capable part. The fix is segmentation. A true price effect is a change inside each segment with composition held constant, and mix is the change in weights at unchanged segment prices. A capability argument that only works at the blended average has not yet been tested.
Does the benchmark gain match the revenue growth claimed on the earnings call?
It is not supposed to match, and a close numerical match should make you suspicious, not reassured. Performance-per-watt and performance-per-dollar gains measure how much work hardware does for a unit of energy or money. Revenue reflects pricing, volume, mix, competition and market conditions on top of that. The useful question is narrower. Is the measured capability gain large enough to make the claimed demand shift believable, and is it in line with ordinary generational progress or well beyond it?
Worked example: reading a capability ratio from the 26Q3 results
Assumption: we want the inference capability ratio between two generations on the same workloads at the same precision. Here we use FP16 inference for the NVIDIA A100-SXM4-80GB and the NVIDIA H100 80GB HBM3 in the 26Q3 results.
| Test | A100-SXM4-80GB FP16 | H100 80GB HBM3 FP16 | Ratio |
|---|---|---|---|
| MobileNetV3 | 1709.3 | 2833.0 | 1.66 |
| Flan-T5 Small | 1645.7 | 3919.6 | 2.38 |
| DistilGPT2 | 1607.9 | 3797.7 | 2.36 |
| DistilBERT | 1267.4 | 2611.7 | 2.06 |
| GraphGPS (Peptides) | 548.4 | 1060.0 | 1.93 |
The scores in the table are read from the 26Q3 release, and the ratios are simple division. Averaging those ratios calls for the geometric mean. The arithmetic mean of normalised numbers is meaningless, while the geometric mean stays meaningful however the numbers were normalised. The product of the five ratios is about 37.1, and its fifth root gives a geometric mean of about 2.06. On this test set, the H100 80GB HBM3 scores about twice as high as the A100-SXM4-80GB in FP16 inference within 26Q3. Each test is one timed window after a discarded warm-up, so the ratio supports a plausibility check but is not a precise multiplier.
That ratio answers one question only: whether a generational capability story is credible. It does not say how much of a vendor’s revenue that capability earned. For normalising across tests, precisions and stacks, see our normalisation framework for comparable GPU benchmarks.
Price-adjusting the gain against a performance-per-dollar baseline
To judge whether a gain is exceptional, compare it with the trend in price-performance. Epoch AI (2022), using 470 GPUs released between 2006 and 2021, found that FLOP/s per dollar doubles roughly every 2.5 years. A 2.5-year doubling is about 32% a year, since 2^(1/2.5) ≈ 1.32. A later Epoch AI estimate of the AI chips actually purchased each quarter found performance per dollar growing about 49% a year from early 2023 to the end of 2025 (90% CI: 36% to 66%). The growth came in spurts: nearly flat until mid-2024, then rising as Blackwell-generation chips took over spending. The two figures measure different things over different periods (listed GPUs against purchase-weighted AI chips). For a recent datacentre generation, we would take the purchase-weighted figure as the closer baseline.
The purchase-weighted baseline rests on spec-sheet peak performance. Epoch AI itself warns that realised performance per dollar on a given workload can differ substantially from on-paper specs, depending on software, workload shape, system design and memory and interconnect bandwidth. That warning is the reason to read measured ratios like the one above next to the trend, not to swap one for the other.
The price-adjustment itself is simple to state. Epoch AI puts the GB300 at about 5.5 times the P100’s 2016 launch price in 2025 dollars, delivering roughly 200 times the performance, which makes it about 37 times more cost-effective. The higher unit price is a price effect in the revenue bridge, even though it buys far more capability per dollar. Capability shows up in the bridge only indirectly: as volume, as mix toward the newer part, or as the pricing power that lets the vendor charge more. Comparing the per-dollar gain with the trend tells you whether that pricing power is plausible. For why cost, efficiency and value are different metrics, see our page on GPU performance per dollar.
Performance per watt has to be measured at full-system power
Performance per watt should count the watts of the whole system. SPEC’s power methodology recommends requiring entire system-under-test power to be reported even when a subsystem is highlighted. Subsystem watts are lower than total watts, so a subsystem performance-per-watt figure comes out larger than the system figure. MLPerf Power calls component-only measurement a common misconception: accelerators, networking and CPUs are active at different stages of an ML workload, so it measures full system power and reports throughput efficiency as samples per joule.
Rated power is not measured power either. Intel’s advice is to measure at the wall with a power meter. Power on its own is not efficiency: a lower-power system that takes longer to finish a task may use more energy overall.
Two more limits apply to any revenue argument built on efficiency figures. SPEC states that performance-per-power ratings from different benchmarks are not comparable and should be labelled to discourage cross-benchmark comparison. It also says such figures indicate relative efficiency but are not good sizing tools for other environments. An efficiency gain quoted from one benchmark and set against revenue from a product line running many workloads stretches both limits at the same time. The 26Q3 figures quoted above are scores, not watts, so a performance-per-watt claim needs its own full-system power measurement.
Which evidence answers which question
| Question | Evidence that answers it | What it cannot do |
|---|---|---|
| How much of revenue growth was price, volume or mix? | PVM bridge from filings (Item 303 MD&A), earnings commentary, analyst ASP and unit models | Prove that demand came from capability |
| Did a product launch drive growth? | Separate new-product and discontinued-product buckets | Explain why the new product sold |
| Is the claimed capability gain real? | Measured scores on matched workloads and precision, averaged with a geometric mean | Size or validate the revenue figure |
| Is the gain exceptional for its price? | Measured ratio compared with a performance-per-dollar trend baseline | Replace workload measurement, since the baselines use spec-sheet peaks |
| Is the efficiency claim real? | Full-system power within one benchmark | Compare across benchmarks or size another deployment |
Who can check whether a claimed capability gain is real on the hardware?
LynxBenchAI gives the capability side of this analysis a measured basis: published scores for training, inference and compute tests on the devices a revenue story depends on, all comparable within one release. To check a claimed generational gain, read both devices’ per-test scores in the 26Q3 results through the directory and take the geometric mean of the ratios on the tests that match the workload in question. If the hardware or workload is not covered, the LynxBenchAI Personal Edition runs the suite on a machine you control.
Revenue, ASPs and unit shipments come from filings, earnings commentary and analyst estimates. Scores are comparable only within one release, here 26Q3, and a published score is not a measurement of your own production workload. For a figure that sits in a business case, run the Personal Edition on representative hardware or set the score next to your own measurements.
What stays uncertain once both halves are done is the link between them. A clean bridge and a measured capability ratio can both be right, and the share of demand caused by capability can still be a judgement, because analyst ASP ranges and unit estimates carry significant uncertainty. State that range openly instead of collapsing it into one capability percentage.
