Micron earns more from each dollar of memory sales

Revenue grows much faster than cost of goods sold

Micron announced its fiscal fourth-quarter and full-year results on September 30, 2026, for the period ended September 3. Quarterly revenue reached $54.229 billion. Cost of goods sold was $7.182 billion, leaving $47.047 billion in gross profit and an 86.8% GAAP gross margin. These are reported, unaudited results.¹

Compared with the thirteen-week year-earlier quarter, revenue increased by $42.914 billion while cost of goods sold increased by $921 million. Of each additional dollar of revenue, about 98 cents remained as additional gross profit. That calculation describes the reported totals. It does not measure the profitability of manufacturing one more chip.¹²

In June, Micron had forecast $49 billion to $51 billion in fourth-quarter revenue and $30 to $32 in adjusted earnings per share. Actual revenue exceeded the top of that range by $3.229 billion. Adjusted EPS of $33.42 exceeded its upper bound by $1.42.³¹ Durability still depends on how much of today's selling price survives future deliveries.

Prices outpace bit growth

Micron's fourth quarter contained fourteen weeks, against thirteen in the preceding quarter.² Revenue rose 30.8% sequentially. Comparing weekly revenue gives 21.5% growth: $3.874 billion per week against $3.189 billion.¹

The expansion began before Q4. First-quarter revenue was $13.643 billion and GAAP gross margin was 56.0%; second-quarter revenue reached $23.860 billion with a 74.4% margin.⁴ Across the full fiscal year, revenue increased 256.3%, while cost of goods sold increased 14.1%. Annual gross margin rose from 39.8% to 80.7%.¹ FY2026 contained fifty-three weeks, against fifty-two in FY2025.²

Micron reports DRAM bits growing sequentially by mid-single digits, prices by high teens; NAND bits approximately 10%, prices approximately 30%. Conventional DRAM had higher gross margins than HBM. Increased HBM mix offset higher prices, keeping cloud gross margin flat. Management forecasts potentially double-digit calendar-2026 PC/mobile unit declines.⁵

Bits measure memory quantity. Revenue reflects quantity multiplied by realized prices, with product mix affecting that comparison. Price increases can produce much higher revenue without a similar increase in cost of goods sold.

Management attributes the margin dip to manufacturing incentive compensation moving from Q4 inventory into Q1 expense.⁵ Even with that lower percentage, the revenue-guidance midpoint implies more gross-profit dollars. The midpoint remains a forecast.

AI uses more than accelerator memory

HBM, or high-bandwidth memory, supplies accelerators with working memory. Server DRAM supplies the host processors. NAND supplies persistent storage. Treating these as one interchangeable AI product obscures the computing constraints that make customers willing to pay for each.

Micron's agent-workflow study used Llama 3.1-8B and CrewAI, with GPU inference and CPU tool execution. At 10,000 concurrent requests, the key-value cache exhausted its available memory, and inference duration reached roughly 105.6 times the 100-request baseline. Spare compute and bandwidth did not eliminate that capacity limit. The paper's approximately 1.4-times HBM4 throughput gain holds compute constant and comes from a model, not an HBM4 benchmark.⁶

The key-value cache stores intermediate attention data used during generation. Running more requests together requires room for more of that data. Buying a faster accelerator cannot, by itself, ensure that the cache fits. This is a reason memory capacity can remain valuable even when an inference service is not consuming all its available compute. The experiment establishes that mechanism within one constructed workload; it supplies no estimate of Micron's future sales.

A separate Micron/Meta study compared low-power DRAM (LPDDR5X) server systems with 256GB per socket at 8,533MT/s and 512GB at 6,400MT/s. The faster system improved tensor-rebatching throughput by 11%. The larger system delivered two to three times the Spark throughput under studied conditions by avoiding data spilling to solid-state storage. Capacity, speed and rank differed, so the tests do not isolate one variable.⁷

The two experiments explain different purchasing decisions. An operator can need more accelerator memory for concurrent generation and more host memory to keep a data-processing job off disk. Neither result licenses a universal performance multiplier. Both studies include Micron authors, and the host-memory paper also includes Meta authors. They explain technical demand more directly than they validate an industry-wide spending forecast.

Micron's cloud and core-data-center units together produced $34.285 billion, or 63.2% of quarterly revenue.¹ That is a segment calculation, not a measurement of revenue caused by AI. The distinction matters for forecasting: a technical reason to buy memory, a customer's realized purchase and a supplier's reported revenue are separate quantities.

Contracts cover part of future sales

Micron reports 26 take-or-pay strategic customer agreements covering over 35% of estimated revenue through 2030. Three-quarters of estimated revenue under those agreements has defined pricing, mostly floor/ceiling bands; one-quarter remains periodically market-priced. Remaining performance obligations total $150 billion at minimum defined prices. Financial commitments total $32 billion.⁵

A price band limits both parties' exposure. Below its floor, the supplier has contractual support for a minimum price. Above its ceiling, the customer has protection against further increases. The value to Micron depends on future prices relative to those bounds and on the costs of making the promised memory. Undisclosed floor and ceiling prices prevent a calculation of either future gross margin or the amount of upside surrendered.

The May-quarter SEC filing describes binding, enforceable volumes and excludes agreements without pricing frameworks, along with certain short contracts, from remaining performance obligations. It records deposits as liabilities before revenue recognition.² The $150 billion cannot be treated as the entire commercial opportunity or as already earned revenue. Nor can the $32 billion be added to it to produce a larger sales total. The measures have different definitions.

Micron received approximately $12.3 billion in Q4 deposits, with returns conditional on minimum purchases.⁵ The earnings release classifies customer-deposit inflows as financing cash flow. Q4 operating cash flow was $43.973 billion, and adjusted free cash flow after net capital expenditure was $33.199 billion.¹ Customer financing adds liquidity separately from that operating cash generation.

Micron's accounting policy recognizes most product revenue when control transfers to the customer.⁸ Signing a future-delivery agreement changes sales visibility; receiving a deposit changes cash and liabilities. Neither event automatically records the future shipment's revenue or profit. Describing contracted sales as booked earnings before delivery would conflate all three events.

Take-or-pay terms strengthen Micron's ability to enforce purchasing obligations. The sources disclose no individual termination clauses, default protections or cancellation remedies. The stated coverage of over 35% does not establish protection for all projected revenue. Micron has better contractual protection for part of future sales, with no disclosed company-wide floor under today's margin.

New fabs must earn their cost

Micron forecasts scarcity through calendar 2028, about $25 billion in first-half FY2027 net capital expenditure, and higher second-half spending. That spending forecast is net of anticipated government incentives. Idaho ID1 initial output is scheduled for mid-calendar-2027; ID2 and Japan expansion for late 2028. Meaningful production follows several quarters later.⁵

That schedule separates construction spending from sellable output. A customer agreement can help justify starting a factory, but cash leaves before the factory reaches useful production. The forecast half-year spending implies full-year net capital expenditure above roughly $50 billion if the second half exceeds the first. This is arithmetic on approximate guidance, not a separately announced annual budget.

Micron's FY2025 annual filing describes large fixed manufacturing costs, past sales below manufacturing cost and the risk that added capacity earns inadequate returns. It warns that competitors' expansion and demand forecasting errors can impair those returns.⁸ The same fixed-cost structure that permits revenue to rise much faster than expense also allows lower prices to reduce profit quickly.

For scale, holding Q4 cost of goods sold constant, a 10% reduction in quarterly revenue would reduce gross profit by $5.423 billion, or 11.5%. This is a sensitivity calculation, not a forecast: volumes, mix and manufacturing expense would change in an actual downturn.¹ It illustrates why greater contractual price protection matters, while leaving its undisclosed value unresolved.

Q4 establishes that Micron converted higher memory prices into operating cash and gross profit on a much larger scale. The next test is whether deliveries under the new agreements preserve profitable prices as manufacturing capacity expands. Until contract floors, production costs and realized deliveries can be compared, an 86.8% gross margin remains a reported quarter's result rather than an established long-term earning rate.