Sanjay Mehrotra has spent more than four decades in a business that repeatedly punishes confidence. Memory manufacturers enjoy periods when demand outruns supply, prices rise faster than costs and factories become cash machines. The prosperity encourages new capacity. Customers over-order to protect themselves. Competitors chase market share. Then supply catches demand, inventories rise and the same chips that appeared strategically scarce become brutally interchangeable.
Micron Technology’s latest results belong to the prosperous side of that cycle, but on a scale the industry has never seen. In the third quarter of fiscal 2026, revenue reached $41.5 billion, more than four times the level a year earlier. Gross margin rose to 84.6 per cent. Operating cash flow was $25.4 billion, and management expects fourth-quarter revenue of about $50 billion with a gross margin near 86 per cent.
The immediate cause is artificial intelligence. Accelerators cannot operate efficiently without enormous amounts of fast memory; models cannot preserve growing context without storage; autonomous systems push more data between compute, memory and the network. DRAM and NAND have moved from supporting components to architectural constraints.
Yet the most important event in Micron’s quarter was not the price increase that produced those extraordinary margins. It was Mehrotra’s attempt to make the economics last after scarcity eventually eases.
Micron has signed 16 strategic customer agreements covering data centres, consumer devices, automobiles and industrial markets. Most extend for five years. They commit customers to purchase specified volumes on a take-or-pay basis, often within price floors and ceilings. Fourteen of the agreements represent roughly $100 billion of minimum contracted revenue over their remaining terms. The company expects $22 billion of customer deposits and related financial commitments.
This is an effort to rewrite the memory industry’s commercial constitution. Instead of selling much of its output into a market where price can collapse within quarters, Micron wants customers to reserve supply years in advance and share the risk of building the factories required to produce it. In return, buyers receive assurance that critical memory will be available for product roadmaps whose value can be measured in hundreds of billions of dollars.
Mehrotra is not claiming to abolish the cycle. The laws of semiconductor manufacturing remain intact. Capacity still arrives late, technology transitions still change output per wafer and demand forecasts still fail. What he is trying to change is who carries the uncertainty.
The wager is audacious. If the agreements endure, Micron could become a more predictable, higher-return company whose customers treat memory supply as a strategic partnership rather than a quarterly procurement exercise. If the current shortage encourages too much construction, the same contracts could lock in customers at high prices while Micron commits more than $250 billion to US manufacturing and technology investment through 2035.
Mehrotra has converted scarcity into contractual power. His defining test is whether he can use that power to prevent the boom from financing the next glut.
The memory engineer was waiting for memory to become strategic
Mehrotra’s authority at Micron begins with the technology. He is not a generalist executive who arrived when artificial intelligence made memory fashionable. He designed non-volatile memory, holds more than 70 patents and co-founded SanDisk in 1988, when flash storage was still a technical proposition rather than an everyday utility.
Over nearly three decades, he helped turn SanDisk from a start-up into a Fortune 500 company with $6.6 billion of annual revenue. The business built removable storage, embedded products, solid-state drives and a long manufacturing partnership with Toshiba. Western Digital acquired it for $16 billion in 2016.
Micron appointed Mehrotra chief executive in May 2017. He inherited one of the world’s three major DRAM producers and a leading NAND business, but also an industry whose products were often valued primarily through cost per bit. Technology leadership mattered because it lowered manufacturing cost; differentiation was difficult to sustain when customers could qualify multiple suppliers and shift purchases as prices changed.
Mehrotra spent his early years improving Micron’s process cadence, portfolio and operating discipline. The company exited weaker businesses, reorganised around markets and invested in advanced DRAM and NAND nodes. It also endured the recurrent reality of memory: every improvement in technology could be overwhelmed by a few percentage points of excess industry supply.
The 2023 downturn was a severe reminder. Annual revenue fell to $15.5 billion from $30.8 billion. Micron reported a GAAP net loss of $5.8 billion and negative adjusted free cash flow of $5.5 billion. At the low point, quarterly gross margin was negative. Products that required some of the most advanced manufacturing on earth were being sold for less than their cost.
Mehrotra cut wafer starts, reduced capital expenditure and slowed supply growth, while preserving spending on the technologies required for the recovery. That distinction is central to memory management. A company must reduce near-term output without falling behind on the next process node, because technological delay raises cost precisely when the market recovers.
The arrival of generative AI changed the demand curve more quickly than the industry could change its factories. Micron had spent years arguing that data growth would elevate memory’s value. The argument was correct, but the decisive shift came when computing performance became visibly constrained by memory bandwidth, capacity and power.
Mehrotra’s career prepared him to recognise the change. Flash taught him that a component can become a strategic system when architecture begins to depend on it. His challenge is to make sure Micron captures that value through the full cycle rather than surrendering it when supply returns.
AI moved the bottleneck from compute to memory bandwidth
The first phase of the artificial-intelligence investment boom was described largely through processors. Graphics chips became the scarce asset, and the market rewarded companies able to supply accelerators, networking and the software required to train large models.
But an accelerator without sufficient memory spends time waiting for data. Training systems need high-bandwidth memory positioned close to the processor. Inference requires fast access to model weights and growing context. Agentic systems add control-plane servers and storage for the record of actions, tools and interactions. As models expand, memory becomes part of the performance equation rather than a peripheral cost.
High-bandwidth memory changes manufacturing economics because it is not ordinary DRAM sold in a different package. Multiple memory die are stacked and connected through advanced packaging. Yield must be managed across the stack, thermal behaviour becomes more complex and each unit consumes materially more wafer capacity than conventional DRAM of equivalent bit volume.
That trade ratio constrains supply beyond the growth of AI demand itself. When Micron allocates more wafers to HBM, fewer bits are available for servers, personal computers, smartphones and automobiles. Each new HBM generation increases performance but also raises complexity and often consumes more manufacturing resources.
Micron is now shipping HBM4 in high volume for a lead customer’s platform. Its 12-high HBM4 ramp is proceeding twice as quickly as the equivalent HBM3E product, and cumulative HBM4 revenue has already exceeded $1 billion. HBM4E, based on the company’s 1-gamma DRAM technology, is planned for volume production in 2027.
The product progress matters because Micron entered the current cycle against two larger Korean competitors with deep customer relationships and manufacturing scale. HBM is not a market where nominal capacity wins. Products are qualified alongside specific accelerators, packaging systems and performance requirements. Missing a customer platform can exclude a supplier for an entire generation.
Mehrotra has placed Micron inside those development cycles earlier. The company is working with customers on memory architecture, power, packaging and capacity years before volume shipments. That collaboration strengthens the case for long-term agreements: customers are no longer buying a standard component after designing a system; they are designing systems around the memory Micron has committed to deliver.
This is the strongest evidence that the industry may be changing structurally. When memory affects the number of accelerators a data centre can use, the performance of the model and the launch timing of a platform, procurement becomes a strategic function.
It remains possible to overstate the shift. HBM will command premium economics only while Micron sustains competitive performance and reliable yields. A qualification win does not guarantee the next generation. The company has moved closer to the centre of the system, but it has also accepted the execution standards of the most demanding customers in computing.
Scarcity has rewritten Micron’s income statement
Micron’s third-quarter results show the force of the current shortage more clearly than any industry forecast. DRAM revenue reached $31.3 billion, up 343 per cent from a year earlier. Sequential bit shipments increased only in the low single digits; average selling prices rose in the low-60 per cent range.
NAND revenue reached $9.9 billion, up 361 per cent from the previous year. Bit shipments rose in the mid-single digits from the preceding quarter, while prices increased in the mid-80 per cent range. The revenue surge was overwhelmingly a pricing event supported by a richer product mix.
The result was a consolidated non-GAAP gross margin of 84.9 per cent, more than double the level a year earlier. Operating expenses of $1.5 billion were small relative to gross profit of $35.2 billion. Non-GAAP operating margin reached 81.2 per cent.
Margins of this scale are not evidence that manufacturing memory has suddenly become inexpensive. They reflect an exceptional imbalance between demand and immediately available supply. Semiconductor fabs take years to permit, construct, equip and ramp. Process transitions are becoming harder. HBM consumes more cleanroom capacity. Customers need memory now, while most new production will not arrive until 2027 and beyond.
Pricing power extends beyond the data centre. Micron’s mobile and client unit produced $11.5 billion of quarterly revenue and an 87 per cent gross margin even as bit shipments declined. Automotive and embedded revenue reached $4.6 billion with a gross margin of 79 per cent. The shortage is spreading the economics of AI across products whose own unit demand is more modest.
This creates both opportunity and danger. Higher memory prices transfer profit from device manufacturers, cloud operators and end consumers to suppliers. Customers may accept that transfer when the alternative is delayed production. Over time, however, expensive memory can reduce system configurations, increase product prices and suppress demand.
There are already signs of adaptation. Server manufacturers are balancing unit growth against the amount of DRAM installed in each machine. Smartphone and PC markets can generate higher revenue despite lower unit shipments if premium devices carry more expensive components, but that model cannot expand indefinitely.
Mehrotra must avoid confusing scarcity rents with permanent product economics. Micron has earned the right to higher margins through technology, supply discipline and investment made during the downturn. Yet an 85 per cent gross margin is also a signal to every competitor, government and customer that additional capacity has enormous apparent value.
The very success of the quarter increases the probability of the response that eventually weakens it.
Five-year contracts could tame the memory cycle
The strategic customer agreements are Mehrotra’s answer to that problem. Sixteen customers have committed to purchase specified volumes of DRAM or NAND over multiple years. Most contracts run from 2026 through 2030, while automotive agreements generally cover three years.
The signed commitments represent roughly 20 per cent of Micron’s DRAM volume and one-third of NAND volume during the term. Management intends eventually to place half or more of company revenue under such arrangements across end markets.
The structure matters. These are take-or-pay contracts: the customer is obligated to purchase the agreed volume or bear the economic consequence. The largest agreements generally use current second-quarter market prices as a ceiling for existing products and include floors that remain well above the peak quarterly margins of previous cycles. A smaller group uses fixed prices or market-linked terms without formal bands.
For customers, the ceiling protects against an even more severe shortage. The floor is the price of supply assurance. For Micron, the floor protects returns if market pricing falls, while the ceiling sacrifices some upside if scarcity becomes more extreme. Both sides are exchanging part of a volatile future for planning certainty.
Fourteen signed contracts carry minimum revenue of about $100 billion over the remaining term. That figure is based on minimum committed volume and minimum pricing, making it a conservative representation of potential sales. Micron expects actual revenue under the agreements to exceed the recorded obligations.
The deposits are equally significant. Customers are expected to provide $22 billion of cash deposits and related financial commitments, with about $18 billion arriving as cash. The money will appear as financing-related cash flow and is intended to be returned or applied later in the contract period, rather than treated as operating free cash flow.
Economically, the deposits resemble customer participation in the capacity programme on which their own products depend. They strengthen Micron’s balance sheet while new fabs consume capital and give customers a financial reason to maintain the relationship. The arrangement begins to look less like component procurement and more like infrastructure finance.
This is a remarkable reversal of bargaining power. In previous downturns, customers could reduce orders, work down inventory and allow suppliers to absorb the price collapse. Today, large buyers are willing to commit cash and volume years ahead because unavailable memory could strand much more valuable investments in accelerators, data centres, vehicles and devices.
The durability of the model will be tested when spot prices fall below contract floors. Procurement teams that welcomed protection during a shortage may challenge agreements that appear expensive in a looser market. Take-or-pay terms are legally stronger than informal forecasts, but no supplier wants its largest customers to regard a five-year relationship as a burden.
Mehrotra must manage the contracts as partnerships, not financial instruments. Technology premiums, new product negotiations and allocation decisions need to remain credible on both sides. The agreement should give customers confidence to invest, not make them regret having insured against scarcity.
Contracts do not eliminate the incentive to overbuild
The paradox of long-term agreements is that they make expansion easier to finance. Stable demand commitments justify new fabs, support government incentives and reduce the risk attached to equipment orders. If every supplier responds similarly, contracted demand can still produce industry overcapacity.
Micron plans net capital expenditure of about $27 billion in fiscal 2026. Fourth-quarter spending alone is expected near $10 billion, and quarterly investment in fiscal 2027 will rise further. More than half of next year’s increase will relate to construction, pulling forward cleanroom capacity required later in the decade.
The company’s US ambition has expanded from approximately $200 billion to more than $250 billion of planned manufacturing and technology investment through 2035. It is building two fabs in Idaho, developing a vast site in Clay, New York, modernising production in Virginia and bringing advanced HBM packaging capability to the United States.
The projects have strategic logic. Micron is the only US-based producer of memory at scale. American customers and policymakers want a larger domestic supply of a component now recognised as essential to AI, defence, vehicles and critical infrastructure. The company’s customer agreements explicitly reflect demand for US production.
The timing remains difficult. Micron’s first new Idaho fab is expected to produce initial wafers in mid-2027, with the second following in late 2028. New York is moving from site preparation into vertical construction, but a multi-fab complex will unfold over years. By the time capacity arrives, the systems, products and competitive landscape used to justify it will have changed.
Government support reduces the cost but does not remove the economic requirement. Grants, tax credits and infrastructure assistance can improve returns on domestic manufacturing. They cannot create customers for output the market does not need. Nor can the strategic value of US production protect Micron from global price declines.
Mehrotra says the company will align supply plans with market conditions. That promise is familiar across memory cycles and difficult to maintain at the peak. Construction has long lead times, equipment contracts are inflexible and political commitments create momentum of their own. Once thousands of jobs and regional development plans depend on a fab, slowing it becomes harder than announcing it.
The strategic agreements provide better visibility than Micron has ever had. They do not make a ten-year forecast reliable. The executive discipline required during this boom is the willingness to distinguish contracted demand from extrapolated enthusiasm.
HBM creates differentiation by making manufacturing harder
For most of the memory industry’s history, process technology created value largely by reducing cost per bit. A smaller memory cell allowed more bits on each wafer, expanding supply and lowering unit economics. The company reaching a node first could enjoy a cost advantage, but the advantage eventually became more output for the market.
HBM changes that relationship. Performance depends on the design of the memory, the number of stacked die, interconnect, packaging, thermal management, power efficiency and qualification with the customer’s processor. Manufacturing more bits is not enough; those bits must be delivered in an integrated product that operates reliably beside the world’s most expensive compute.
Micron’s 1-gamma DRAM and G9 NAND nodes are ramping towards what the company expects to be the highest-volume nodes in its history. Development of subsequent generations is intended to support volume production in the second half of 2027. The company has also signed a multi-year agreement with ASML to secure extreme-ultraviolet lithography systems for future DRAM nodes.
Each technology advance raises capital intensity and execution risk. HBM stacks can lose value if one layer fails. Packaging capacity becomes as important as wafer output. New nodes require more complex tools and longer learning cycles. A product delay can cause a customer to miss an accelerator platform with a limited qualification window.
This complexity is commercially attractive if Micron executes. It creates barriers to entry, makes customer collaboration deeper and supports pricing based on system value rather than raw bits. It also means the company can no longer treat packaging as a downstream operation whose economics are secondary to the fab.
Micron is expanding advanced packaging in Taiwan and building Singapore into another centre for HBM packaging, with meaningful capacity expected in the first half of 2027. The geographic network must function as one production system: wafers, stacks, packaging, testing and customer delivery cannot be optimised independently.
Mehrotra’s background in flash systems gives him an advantage here. SanDisk created value by combining memory technology, controllers, firmware and product design rather than selling undifferentiated NAND alone. Micron is applying a similar logic at a much larger scale, moving from memory components towards memory subsystems designed for specific computing architectures.
The transition will be proven by margins after supply normalises. If Micron retains a premium because customers value its products, technology and reliability, HBM will have altered the quality of the business. If margins collapse with conventional DRAM pricing, complexity will have raised capital requirements without ending commoditisation.
Asia remains the operating centre of an American champion
Micron’s US expansion has become part of national industrial policy, but the company remains inseparable from Asia’s semiconductor ecosystem. Its most advanced production, packaging, suppliers and customer relationships span Taiwan, Japan, Singapore, Malaysia, South Korea and the wider region.
In March 2026, Micron completed the acquisition of Powerchip’s Tongluo facility in Taiwan after agreeing to pay $1.8 billion. The existing 300,000-square-foot cleanroom is being equipped for DRAM, with meaningful output expected in mid-2027. A second cleanroom of similar size is under construction and will support extreme-ultraviolet equipment.
The acquisition is strategically different from a greenfield US fab. It adds existing cleanroom space near Micron’s Taichung operations, allowing the company to install tools and begin production more quickly. In a shortage, time has enormous economic value. A year gained in capacity can generate returns that justify a premium for the site.
Japan remains central to technology development and advanced production. Singapore supplies NAND and is being expanded for HBM packaging. Taiwan provides manufacturing depth and an experienced engineering base. The US projects add resilience and strategic capacity, but they do not replace the Asian network.
Mehrotra must manage that geography amid a semiconductor industry increasingly divided by policy. Micron has been restricted from selling certain products to critical information-infrastructure operators in China. Chinese memory companies are expanding with state support. Export controls affect equipment and customer access. Governments increasingly regard memory as a national-security input rather than a globally interchangeable component.
This fragmentation creates opportunities for Micron because customers value trusted supply outside China and are willing to support US capacity. It also raises cost. Duplicated production, local-content requirements and regional inventories reduce some of the efficiencies that made semiconductors global.
Micron’s principal competitors retain scale and manufacturing strength in South Korea, while Chinese producers continue to advance in DRAM and NAND. Current scarcity can make every supplier profitable; the competitive differences become clearer when new capacity arrives and pricing weakens.
Mehrotra’s task is to use geographic diversity as an operating advantage rather than a political collection of projects. Each site needs a defined role in technology, capacity or customer resilience. A global network can reduce concentration risk. An undisciplined one can multiply fixed costs.
The shortage is reaching consumers who did not order an AI boom
The memory cycle is now being shaped by data-centre economics, but most memory still enters a wider range of products. Smartphones, personal computers, vehicles, industrial equipment, medical devices and network infrastructure all compete for wafer capacity influenced by AI.
As HBM absorbs more DRAM resources, conventional memory prices rise. NAND supply is constrained by limited cleanroom space and the conversion of some capacity towards DRAM. Customers in slower-growth markets must accept higher prices even when their own unit demand is flat or declining.
Micron’s latest business-unit margins show how far pricing power has spread. Mobile and client gross margin reached 87 per cent. Automotive and embedded margin reached 79 per cent. These are not merely premium AI categories benefiting from unique product differentiation; they are broad markets experiencing scarcity.
The effect can become self-limiting. Device makers may reduce memory content, delay launches or raise prices. Consumers can postpone replacements. Automakers with long qualification cycles may accept higher costs for supply security, but industrial customers can redesign systems around lower specifications or alternative components.
The strategic agreements give Micron committed volume, yet they also connect the company more directly to the health of customer roadmaps. A take-or-pay structure protects revenue contractually; it cannot ensure that the customer’s end market remains profitable. If expensive memory weakens demand downstream, both sides eventually need commercial flexibility.
Mehrotra must therefore manage allocation with a view beyond the highest immediate price. Customers remember how suppliers behave during shortages. A manufacturer that diverts too much capacity towards the richest segment may damage relationships in automotive or industrial markets where products remain in use for a decade.
Micron’s expansion in Virginia addresses that concern by producing long-lifecycle DRAM for automotive, defence, aerospace, industrial, networking and medical applications. Such products do not use the newest node, but their availability carries strategic value. A balanced portfolio can make the company more resilient when AI spending eventually moderates.
The memory boom is strongest because it has escaped the data centre. It will be most durable if Micron can serve the rest of the economy without making memory a constraint on the products that create demand.
The 2023 collapse should remain inside every capital decision
Nothing in Micron’s current performance resembles fiscal 2023. Revenue in the latest quarter alone was almost three times the full-year total then. Gross margin has moved from negative territory to the mid-80s. The balance sheet held $30.2 billion of cash and investments at the end of the third quarter, against $5.7 billion of debt.
The contrast is precisely why 2023 matters. It demonstrates how quickly memory economics can change when customers stop ordering and industry inventories rise. High fixed costs continue even when fabs are underutilised. Inventory loses value as prices fall. Technology spending cannot stop simply because current products are unprofitable.
Mehrotra responded to that downturn with a mix of supply cuts and continued innovation. The company reduced capital expenditure, slowed output and took inventory charges while protecting the nodes and HBM products required for the next cycle. That discipline positioned Micron to benefit when AI demand accelerated.
The strategic customer agreements are, in part, an institutional memory of that experience. Price floors, committed volumes and deposits are intended to stop customers from transferring all cyclical risk back to the supplier. The contracts may also give Micron the confidence to preserve research spending through future downturns.
But institutional memory fades fastest when profits are extraordinary. A company earning more than $18 billion of quarterly free cash flow can justify almost any expansion through current returns. Governments compete to host fabs. Customers demand capacity. Investors reward exposure to AI.
The danger is not that Mehrotra has forgotten the cycle. Few executives understand it better. The danger is that Micron’s commitments—to customers, governments, communities and its own technology roadmap—become collectively larger than management’s ability to slow investment when evidence changes.
Capital discipline in memory is not measured by spending less during a shortage. Some expansion is essential, and failing to build would cede strategic markets. It is measured by building at a pace that remains rational under less heroic pricing.
Micron’s price floors offer protection. They should not become an excuse to assume the floors will never be challenged.
Mehrotra is trying to convert a cycle into an institution
Sanjay Mehrotra has led Micron into the strongest financial position in its history. The company is shipping leading HBM products, gaining relevance inside AI architecture and generating enough cash to fund a global manufacturing programme while reducing debt. Customers are committing volume and capital on terms that would have been difficult to imagine during previous cycles.
His larger achievement may be recognising that supply assurance itself has become a product. Micron is not merely selling bits and bandwidth. It is selling the confidence that a customer can launch a data centre, accelerator platform, vehicle or device without discovering that memory is unavailable.
The strategic agreements monetise that confidence. They give Micron revenue visibility, protect margins and deepen technical collaboration. They give customers price boundaries and access to capacity. Properly managed, they can fund innovation and reduce the destructive oscillation between shortage and surplus.
They also create a new set of obligations. Micron must deliver the promised volume, qualify each product on time and allocate supply without compromising strategic relationships. It must expand enough to support customers while resisting the temptation to treat current prices as a permanent forecast. It must execute a $250 billion US vision while continuing to invest across Taiwan, Japan and Singapore.
The company’s extraordinary margins will inevitably moderate. The relevant question is where they settle. If product differentiation, HBM complexity and long-term contracts keep returns well above previous peaks, Mehrotra will have changed the quality of Micron’s earnings. If a wave of capacity restores conventional pricing, the latest boom will have been larger without being fundamentally different.
Mehrotra has been through enough cycles to know that declarations of structural change usually arrive near the top. This time, he has more than a declaration. He has binding contracts, customer deposits, leading products and supply constraints that cannot be resolved quickly.
But memory’s history is not overturned by one record quarter or even five years of agreements. It is overturned when the industry reaches the next downturn and Micron can continue investing, earning attractive returns and serving customers without destroying capital.
Mehrotra has persuaded customers to share the risk of scarcity. His legacy will depend on whether he can keep Micron from recreating abundance on uneconomic terms.
