--- title: "MU (Analyst briefing): Capex may decline after FY2027 ---" type: "Topics" locale: "zh-HK" url: "https://longbridge.com/zh-HK/topics/39358820.md" description: "Below is Dolphin Research's Trans from Micron's FY2026 Q2 analyst huddle. For the earnings read-through, cf. 'Micron: AI-fueled memory surge — can it break the cycle?'.Key takeaways from the analyst huddle: 1) Non-HBM carries higher GPM than HBM. The company will not materially rebalance HBM vs. non-HBM on short-term price swings.This is a strategic decision. Customers need a complete product stack to build AI systems, and you cannot sell an AI system that is short on HBM. The market requires a balanced portfolio to drive shipments at scale..." datetime: "2026-03-19T04:20:29.000Z" locales: - [en](https://longbridge.com/en/topics/39358820.md) - [zh-CN](https://longbridge.com/zh-CN/topics/39358820.md) - [zh-HK](https://longbridge.com/zh-HK/topics/39358820.md) author: "[Dolphin Research](https://longbridge.com/zh-HK/news/dolphin.md)" --- > 支持的語言: [English](https://longbridge.com/en/topics/39358820.md) | [简体中文](https://longbridge.com/zh-CN/topics/39358820.md) # MU (Analyst briefing): Capex may decline after FY2027 --- **Below is Dolphin Research's transcript from MU's FY2026 Q2 analyst huddle. For the earnings read-through, see '**[**Micron (MU): AI is driving a storage boom — can it break the cycle?**](https://longbridge.cn/zh-CN/topics/39356514?channel=SH000001&invite-code=294324&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=12482f65-395b-49b9-bb28-e4ef68a460a0)**'.** **Key takeaways from MU's analyst session:** **1) Non‑HBM GPM above HBM: MU will not shift mix aggressively between HBM and non‑HBM on short‑term price gaps.** This is strategic — customers need a full memory portfolio to build AI systems, and you cannot ship an AI system that lacks adequate HBM. The market requires a balanced kit to scale shipments. **2) CapEx: Still DRAM/HBM‑led. FY26 CapEx guided up to $25bn+, with added spend tied to announced capacity adds and U.S. expansion.** On construction, FY26 is guided to high‑single‑digit billions on a net CapEx basis. Looking to FY27, construction spend rises by Approx. $10bn on an incremental FY26→FY27 basis, and equipment outlays will also increase. **FY27 NAND CapEx begins to step up but remains a smaller slice vs. DRAM. After FY27, MU will stay highly disciplined and CapEx could decline, though no decision yet.** **3) DRAM capacity constraints: Persist through 2027, with meaningful cleanroom relief only in 2028.** Announced projects run into 2028–2030, and the catch‑up pace will hinge on demand; equipment orders and installs will stay flexible. **Management cannot yet judge when supply will meet demand.** **MU crushed estimates, benefiting from the current memory upcycle. Beyond near‑term beats, the market focuses on the durability of earnings,** including the depth of multi‑year agreements, demand assurance, the outlook for tight supply, and CapEx plans. **Across two recent interactions, management addressed key concerns: (i) LTAs extended to 5‑year terms but skew more strategic than take‑or‑pay; (ii) demand assurance remains at 50% to two‑thirds, unchanged QoQ; (iii) supply tightness seen through 2027, with cleanroom and capacity relief in 2028; (iv) CapEx rising into FY27, with potential declines thereafter, while supply/demand beyond that remains uncertain.** **Overall,** despite blowout results, management's tone trails market expectations, and investors will still worry about post‑2027 earnings. **Growth visibility for 2026–2027 is consensus already, and the cautious stance beyond FY27 keeps MU framed as a cyclical. Even with repeated beats, cyclicality may cap PE expansion.** **I.** $Micron Tech(MU.US) **Earnings highlights** 1\. **Big beat vs. consensus**: FQ2'26 revenue of $23.86bn (vs. $19.51bn, +22.3%), EPS of $12.20 (vs. $8.73, +39.8%). DRAM and NAND pricing both rose sharply, with larger gains in NAND. Both segments saw QoQ bit growth, larger for DRAM. 2\. **FQ3 guide**: Pricing remains the primary driver. DRAM and NAND shipments are expected to grow modestly. 3\. **CapEx raised materially**: FY26 CapEx lifted to $25bn+ from $20bn, mostly for DRAM and HBM adds. Construction spend guided to mid‑to‑high single‑digit billions. For FY27, construction up by Approx. $10bn and equipment also higher; NAND CapEx starts to rise but remains far below DRAM. 4\. **Start‑up and OpEx**: New line start‑up (e.g., ID1) at ~$100–200mn per quarter from next quarter through FY27, then tapering, with an impact of ~50bps or less at current revenue levels. FQ4 OpEx seen near $1.6bn, and FY27 at a run‑rate of ~$1.7bn on higher R&D. 5\. **HBM vs. non‑HBM margin**: Non‑HBM DRAM currently carries higher margins than HBM, and DRAM ex‑DC is also very strong. HBM3E yields continue to improve, while HBM4 is ramping earlier with a faster yield curve. **II. Details from the earnings call** **Q&A** **Q: What drives the decision to add NAND capacity, and how do you see NAND demand?** A: Historically we noted DRAM needs new wafer capacity over the LT, while NAND can lean more on node transitions. We still believe NAND technology migration supports strong bit growth. This decision reflects demand confidence and the fact that migrations consume cleanroom space, as well as our move to consolidate more NAND R&D in Singapore near manufacturing. To clarify, this is not a greenfield site but an added cleanroom within an existing campus. On demand, DC‑led NAND needs are very strong. AI servers are deploying high‑capacity, high‑performance SSDs, which plays to our strengths — we were first to market with Gen‑6 SSDs, which pair well with NVIDIA systems, and demand far exceeds our supply. We have set market‑share highs in DC SSD for four consecutive years. We are excited about the portfolio, and supply will be far short of demand for the foreseeable future. Even if we break ground now, **the new cleanroom would not add capacity until 2H28**. More broadly, **the industry faces mid‑term cleanroom constraints, especially as some NAND space has been converted to DRAM**. **Q: How will added DRAM capacity affect pricing?** A: The **Idaho facility and capacity from the Powerchip acquisition will not contribute to output and revenue until FY28**. Large cleanroom projects across the industry share similar timelines — meaningful supply only starts to show in late 2027 and more so in 2028. This underpins why we keep saying tightness extends beyond 2026. Customer forecasts for 2026–2027 keep moving up. We are adding some supply in 2026, but it is far from enough to close the gap. AI‑driven demand is broad‑based beyond DCs. We expect tightness to persist at least beyond 2026, and we can always modulate equipment installs as needed. **Q: Any split assumptions behind the FQ3 DRAM/NAND guide?** A: We are not providing a split. Clues from FQ2: DRAM and NAND pricing rose sharply, with NAND up more; bits grew QoQ in both, with DRAM larger. **In FQ3, pricing remains the biggest driver**. We have given 2026 industry bit growth views, and industry shipments are supply‑constrained. Our bit growth should track the industry. So you can assume **modest bit growth in both DRAM and NAND for FQ3**. **Q: How do CXL and memory pooling affect DRAM demand?** A: CXL will be trialed at some customers, and companies are at different stages — some evaluating, some considering adoption. We will offer memory products configured for CXL. Given the supply/demand gap, customers will try any solution that can scale, provided the architecture, software, and systems line up. There is still significant work and technical limits before broad deployment, but trials and limited rollouts will proceed. **Q: With non‑HBM margins above HBM, how do you allocate capacity? Are you leaving money on the table by not tilting to non‑HBM?** A: Relative margins move by quarter. **Most HBM volumes in calendar 2026 were priced in late 2025**, which provides stability and visibility with strong ROI and profitability. When operations create incremental HBM supply, those bits sell at higher prices. **But we will not swing mix heavily between HBM and non‑HBM on short‑term price deltas.** Strategically, **customers need a balanced memory kit to build AI systems; you cannot ship an AI system without sufficient HBM**. HBM margins are healthy, non‑HBM margins even higher, and DRAM ex‑DC is very strong. Our goal is to meet customer needs holistically, not constantly re‑slice allocations across customers and segments. **Q: If LPU chips move to foundries, will SRAM stay on‑die or be separately packaged?** A: There are on‑die SRAM solutions in use today, and some discussion of bonded SRAM. We will not comment on customer choices ahead. Our focus is balanced system evolution. SRAM‑based systems in smaller deployments complement large‑scale systems such as Vera Rubin‑type ASIC accelerators. DRAM content in these systems keeps rising, with strong demand for high‑performance DRAM and AI DRAM — all tailwinds for our DRAM biz. **Q: How large is KV Cache within DC NAND demand?** A: Our DC growth outlook already included KV Cache as a key driver. As **customers evolve SSD deployment models, KV Cache has increased** our TAM expectations. We also see **HDD shortages boosting DC SSD demand significantly**, and this looks set to persist. Net‑net, DC NAND is in severe undersupply with demand rising — partially KV Cache‑driven and also because AI servers have insatiable fast‑storage needs. Our portfolio is well positioned here, including KV Cache use cases. **Q: DRAM vs. NAND split within FY26 and FY27 CapEx?** A: **CapEx remains DRAM/HBM‑led. For FY26 we raised CapEx to $25bn+** from $20bn, reflecting adds we announced in Feb and U.S. expansion. On construction, FY26 is mid‑to‑high single‑digit billions on a net basis. **For FY27, we expect construction up by around $10bn and higher equipment spend. NAND CapEx starts to increase in FY27**, but remains a smaller share vs. DRAM. **Q: Is the per‑generation HBM capacity increase sustainable on each XPU?** A: **We do not make long‑range capacity calls; it depends on customer architectures. But AI trends — stronger inference, longer context windows, agents and multi‑agent flows — all demand more DRAM capacity and bandwidth**. Whether GPUs or ASIC accelerators, they will need more HBM and DDR5/LP5 capacity. The trend is clear so far, and disclosed roadmaps corroborate it. Importantly, how end customers extract value from AI aligns with needing more DRAM capacity and bandwidth at efficient power. HBM delivers that, which is why we say 'memory is a strategic asset in the AI era' — high‑performance AI architectures and infrastructure are impossible without DRAM and HBM. **Q: Any update on LT bit growth for DRAM and NAND?** A: No new LT numbers yet. Historically we cited mid‑to‑high‑teens for DRAM, but last year and this year are tracking higher. We see a sustained strong demand cycle, with HBM trade ratios putting heavy pressure on industry capacity and ours. For the foreseeable future, these are supply‑constrained figures rather than true demand. We expect strong industry growth again next year. **Q: When will MU's cleanroom capacity catch up with customer demand?** A: **For all major DRAM makers, constraints persist through this year and next, with meaningful cleanroom relief only in 2028.** Our projects run through 2028–2030, and timing will depend on demand; we will stay flexible on equipment POs and installs. On demand, we are discussing 5‑year multi‑year SCA agreements with several customers. We also see exciting new demand vectors emerging and scaling, including robotics, which we expect to become a major driver. Putting it together, **we do not yet have high‑confidence visibility on when supply will catch demand**, given the momentum across these vectors. **Q: Is FY27 the peak year for construction CapEx? What is a normal construction‑to‑equipment split?** A: No further split beyond what we disclosed. Spend will fluctuate. To clarify, **$10bn is the FY26→FY27 increment. We will modulate spend to maintain stable bit share, and CapEx could decline after FY27**, though we are not making that call now. We are investing to secure the supply our customers need over the next several years. **Q: Key swing factors for GPM over the next few quarters, and how to think about depreciation and start‑up costs?** A: On costs, we continue to execute well on reductions. FY26 benefits from DRAM 1‑alpha and NAND G9 transitions with significant efficiency and cost gains — for example, 1‑gamma displacing 1‑alpha capacity. Expense control is also strong, including management of geopolitical risks, with no impact on operations today. Start‑up costs tied to ID1 are **~$100–200mn per quarter from next quarter through FY27, then roll off**. At current revenue and margin levels, the **impact is ~50bps or lower**. Depreciation depends on when wafer volume ramps and asset lives; as a greenfield program, depreciation lives are very long. On HBM, HBM3E yields have kept improving after recent high‑volume ramps. **HBM4, though early, is tracking a faster yield curve than HBM3E**, with a favorable cost structure. On geopolitics, despite media reports around the Middle East and inputs shortages, we do not see supply risks and the cost impact is minimal. On OpEx, FQ4 is expected near $1.6bn (partly an extra week, partly higher R&D). FY27 should run around $1.7bn, then stabilize. **Risk disclosure and statement:**[**Dolphin Research Disclaimer and General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### 相關股票 - [Micron Tech (MU.US)](https://longbridge.com/zh-HK/quote/MU.US.md) - [NVIDIA (NVDA.US)](https://longbridge.com/zh-HK/quote/NVDA.US.md) - [GraniteShares 2x Long NVDA Daily ETF (NVDL.US)](https://longbridge.com/zh-HK/quote/NVDL.US.md) - [XL2CSOPNVDA (07788.HK)](https://longbridge.com/zh-HK/quote/07788.HK.md) - [XI2CSOPNVDA (07388.HK)](https://longbridge.com/zh-HK/quote/07388.HK.md) - [YieldMax NVDA Option Income Strategy ETF (NVDY.US)](https://longbridge.com/zh-HK/quote/NVDY.US.md) - [Direxion Daily NVDA Bear 1X ETF (NVDD.US)](https://longbridge.com/zh-HK/quote/NVDD.US.md) - [T-Rex 2X Long NVIDIA Daily Target ETF (NVDX.US)](https://longbridge.com/zh-HK/quote/NVDX.US.md) - [T-Rex 2X Inverse NVIDIA Daily Target ETF (NVDQ.US)](https://longbridge.com/zh-HK/quote/NVDQ.US.md)