Treat memory as a priced, volatile input with a public supplier side and a private buyer side, and plan on the private side.
Researched by an agentic pipeline. Reviewed and approved by the editor before publication.
Memory is the one server cost line where the seller's accounts are public and the buyer's price is not. Micron's annual report for the fiscal year ended 3 September 2026 shows DRAM revenue of $100.68bn, against $28.58bn the year before, and NAND revenue of $31.79bn against $8.50bn1. These are reported figures from an audited filing. They are also revenue, not price: the year ran 53 weeks against 52, and the filing's business-unit table does not separate price from volume. The only price evidence in the sections of the filing we could read is a range: over five fiscal years, annual changes in DRAM and NAND average selling prices have run from an increase of about 200% to a decrease of about 50%1. The cycle is violent in both directions.
On the buying side, the vendors' own results moved the same way. Dell's quarterly filing for the period ended 31 July 2026 reports revenue of $46.97bn against $29.78bn a year earlier, gross margin of $9.83bn against $5.45bn and operating income of $5.39bn against $1.77bn2. Those are reported. The ratios are ours: gross margin of about 20.9% against 18.3%, operating margin of about 11.5% against 6.0%. Dell's own release, as reproduced by a trade site, puts infrastructure revenue at $31.8bn, up 89%, with operating income of $4.8bn, about 15%; traditional servers and networking, excluding AI-optimised systems, grew 122% to $10.5bn3. HPE's third quarter to 31 July reports server revenue of $6.8bn, up 35.3%, a GAAP gross margin of 40.1%, up 1,090 basis points, and a Cloud & AI segment operating margin of 17.0% against 7.0%4.
Both ends of the chain reported wider margins in the same period. That is consistent with memory cost being passed to buyers. It is not evidence of how much, because AI mix lifts server margins on its own.
HPE's chief executive said on the March 2026 earnings call that DRAM and NAND now make up more than 50% of the bill-of-materials cost of a traditional server, and that the share would keep rising5. This is an operator statement. It names no configuration, date or denominator, and it is HPE's characterisation of its own products. The same report records that HPE raised prices in November, December and January, cut quote validity from 30 days to 14, and can reprice orders until shipment5. These are reported actions; they show direction, not size.
The size appears only in reported plans and forecasts. A trade report in December 2025, citing unnamed channel sources, said Dell, Lenovo, HP and HPE were planning server price rises of around 15%6. That is unconfirmed, and no vendor filing we read repeats it. TrendForce, an independent research house, forecast on 30 September 2026 that conventional DRAM contract prices would rise 10–15% quarter on quarter in the fourth quarter and NAND 15–20%, with server DRAM still undersupplied and some supplier increases delayed by long-term agreements7. This is a forecast, not a measurement.
| Question | Basis | Status |
|---|---|---|
| Memory suppliers' revenue has multiplied | Micron 10-K, reported | Established, but includes a 53rd week and does not separate price from volume |
| Server vendors' margins rose | Dell 10-Q and HPE release, reported; ratios derived for Dell | Established |
| Memory exceeds half of server BOM | HPE CEO statement | Operator claim, no basis |
| Realised server price rise | None published | Not established |
| Fourth-quarter contract rise of 10–15% | TrendForce forecast | Projection |
Three things are missing. No vendor publishes the realised change in server price, or the share of the memory increase it absorbed. The memory share of cost has no stated basis. And the vendors' margin gains cannot be separated from the shift towards AI systems. We did not read the management discussion sections of Dell's quarterly filing or Micron's annual report, which may contain the pass-through language this analysis lacks. Independent confirmation is therefore partial: TrendForce corroborates the direction of contract prices, and nobody independent measures what an enterprise paid for a server.
Treat memory as a priced, volatile input with a public supplier side and a private buyer side, and plan on the private side. Quotes that expire in 14 days and can be repriced before shipment are not budget commitments. Ask each vendor to itemise memory in the configuration, state the quote's price-protection terms, and name the date after which the price can move. Budget server refreshes against the contract price you can lock, not against list price or the 50% figure, and treat any forecast of a further 10–15% as a scenario to test, not a number to book. The evidence supports caution about timing. It does not support a claim about how large the increase is for your configuration.
Do not scale the national 52 GW, or Cleanview's 2 GW of gas, into a figure for data-centre storage.
Oversubscription shows demand for allocations; it does not show how much of the allocated time ran jobs.
A fab cannot be more than full, so the denominator is an estimate that output can exceed.