Micron Technology (MU): Q2 FY2026

Micron delivered what CEO Sanjay Mehrotra called an "exceptional" quarter, with $23.9 billion in total revenue up 196% year-over-year and 75% sequentially — the company's fourth consecutive quarterly revenue record.¹ The more important story is not the magnitude of the beat, but the structural reasons Mehrotra laid out for why conditions remain tight, and what Micron is doing to entrench its position as the only U.S.-based volume producer of advanced HBM.

“AI has not just increased demand for memory; it has fundamentally recast memory as a defining strategic asset in the AI era.”
— Sanjay Mehrotra, Chairman, President & CEO, Micron Technology, Q2 FY2026 Earnings Call¹

The HBM4 announcement was the headline of Micron's quarter. At NVIDIA's GTC conference, Micron confirmed it had begun volume shipment of its HBM4 36GB 12-Hi product in 2026, designed specifically for NVIDIA's Vera Rubin platform.¹ The company also sampled its HBM4 16-Hi product, providing 48GB of HBM capacity per package — a 33% increase over the 12-Hi — and confirmed that HBM4E development is well underway with volume ramp targeted for calendar 2027. Mehrotra added that Micron expects to "reach mature yields faster than HBM3E," a direct reference to the yield ramp challenges that plagued the industry's early HBM3E production runs.¹

“With our HBM4 production ramp and volume shipments underway, we expect to reach mature yields faster than HBM3E. We have also sampled our HBM4 16-Hi product, which provides 48GB of HBM capacity in each HBM — a 33% increase in HBM capacity compared to HBM4 12-Hi.”
— Sanjay Mehrotra, Chairman, President & CEO, Micron Technology, Q2 FY2026 Earnings Call¹

The demand-supply imbalance is not an abstraction at Micron — it has a specific fulfillment number. Mehrotra confirmed that Micron can fulfill only 50% to two-thirds of demand from some of its key customers in the medium term, a figure unchanged from the prior quarter.¹ Against that backdrop, the announcement of Micron's first five-year strategic customer agreement takes on particular significance. Unlike prior one-year LTAs, these SCAs carry "specific commitments over a multiyear time horizon" designed to provide supply certainty for customers and visibility for Micron's business model across the full memory cycle.¹ CFO Mark Murphy noted that DRAM prices increased "in the mid-sixties percentage range" in fiscal Q2, driven by tight conditions and favorable mix — with Q3 gross margin guided to approximately 81%, up 600 basis points sequentially.²

“We project our fiscal 2027 CapEx to step up meaningfully to support HBM and DRAM-related investments. We expect construction-related CapEx to increase by over $10 billion year over year in fiscal 2027.”
— Mark Murphy, CFO, Micron Technology, Q2 FY2026 Earnings Call²

NVIDIA (NVDA): Q4 FY2026

NVIDIA reported $68.1 billion in quarterly revenue — up 73% year-over-year — with data center contributing $62 billion.³ For full fiscal year 2026, data center generated $194 billion in revenue, up 68% year-over-year, representing roughly a 13x increase since the launch of ChatGPT in fiscal 2023.³ While NVIDIA did not dissect HBM economics the way Micron does, memory supply constraints surfaced explicitly in two key areas: gaming and the Vera Rubin ramp.

“While end demand for our products remain strong and channel inventory levels are healthy, we expect supply constraints to be the headwind to Gaming in Q1 and beyond.”
— Colette Kress, EVP & CFO, NVIDIA, Q4 FY2026 Earnings Call³

Kress confirmed that NVIDIA has "strategically secured inventory and capacity to meet demand beyond the next several quarters," noting this time horizon is "further out in time than usual" — reflecting longer demand visibility than the company typically holds.³ This language maps directly onto Micron's SCA announcement: NVIDIA is almost certainly among the customers locking in long-term supply commitments. On Vera Rubin, Kress confirmed NVIDIA shipped first samples to customers in late February 2026 and remains on track for production shipments in H2 2026, expecting that "every cloud model builder" will deploy the platform.³ Each Rubin GPU requires up to 288GB of HBM4 memory, and the NVL72 rack accumulates over 20TB of HBM4 per rack — making the Vera Rubin ramp the single largest discrete driver of HBM4 demand in 2026.¹³

“We believe we have inventory and supply commitments in place to address future demand, including shipments extending into calendar 2027.”
— Colette Kress, EVP & CFO, NVIDIA, Q4 FY2026 Earnings Call³

CEO Jensen Huang framed the broader demand logic: "In this new world of AI, compute is revenues. Without compute, there's no way to generate tokens. Without tokens, there's no way to grow revenues."⁴ Huang also flagged that cloud CapEx across the top five providers is "approaching $700 billion" for 2026 — the total addressable demand signal that memory suppliers are scrambling to serve.⁴ NVIDIA's free cash flow of $35 billion in Q4 alone and $97 billion for fiscal year 2026 gives it the balance sheet firepower to lock up supply commitments, placing it in a fundamentally stronger position than smaller or less capitalized AI customers competing for the same HBM allocation.¹¹

AMD: Q4 FY2025

AMD reported $10.3 billion in revenue for Q4 FY2025 (+34% year-over-year), with its data center segment reaching a record $5.4 billion (+39% YoY).⁵ While the quarter reflected solid momentum in MI350 series deployments and record EPYC server CPU share, the HBM story at AMD is more about roadmap positioning than current-quarter dynamics. AMD's explicit HBM4 moment is still ahead: its current Instinct MI400 family (Helios/MI455X) ships with HBM3E, and the forthcoming MI500 — powered by CDNA6 on TSMC's 2nm process with HBM4E memory — represents AMD's next HBM generation, targeted for 2027.⁵

“Development of our next-generation MI500 series is well underway. MI500 is powered by our cDNA6 architecture, built on advanced two-nanometer process technology, and features high-speed HBM4e memory. We are on track to launch MI500 in 2027 and expect MI500 to deliver another major leap in AI performance.”
— Lisa Su, Chair & CEO, Advanced Micro Devices, Q4 FY2025 Earnings Call⁵

Su confirmed active discussions with customers on multi-year, multi-gigawatt deployments beginning with Helios and MI450, and reiterated AMD's target to grow data center revenue at over 60% annually over the next three to five years, scaling AI to "tens of billions in annual revenue in 2027."⁵ As a fabless designer sourcing HBM from the constrained market, AMD's Q4 transcript did not feature explicit commentary on HBM supply tightness for its own products, but an inventory build of approximately $70 million sequentially suggests AMD is carefully managing its HBM3E allocation into the Helios ramp.⁶

“Based on the strength of our EPYC and Instinct roadmaps, we are well-positioned to grow data center segment revenue by more than 60% annually over the next three to five years and scale our AI business to tens of billions in annual revenue in 2027.”
— Lisa Su, Chair & CEO, Advanced Micro Devices, Q4 FY2025 Earnings Call⁵

CFO Jean Hu highlighted that while data center GPU revenue — including $390 million from MI308 sales to China — drove Q4 outperformance, adjusted gross margin excluding the China inventory reserve release was approximately 55%.⁶ This is a meaningful gap versus Micron's 75% and NVIDIA's 75.2%, underscoring AMD's earlier-stage position in the AI premium earnings curve. AMD's CapEx of just $222 million in Q4 FY2025 against $10.3 billion in revenue reflects its fabless model — the capital intensity of HBM manufacturing falls entirely on suppliers, creating an asymmetric relationship between AMD's financial profile and its dependency on memory supply availability.¹²

HBM4 & the AI Memory Chokepoint: What MU, NVDA, and AMD Execs Are Saying

The Power Delivery Gap