Micron's bubble confession: can the HBM wave carry the September 30 earnings?

A rare scene is playing out in memory: two close observers of the sector openly say they are in the middle of a bubble , then add in the same breath that they hold Micron shares. According to MarketBeat, the market expects $31.14 in earnings per share and $50.41 billion in revenue for the fourth quarter due September 30; Micron's investor notice schedules the earnings call for September 30 at 2:30 p.m. Mountain time. A bubble confession and a buy thesis in one sentence is what this article unpacks.

Micron's old identity is worth recalling because it is the inverse of today's pricing. The hosts describe a classic commodity cycle company whose gross margins giants like Apple could squeeze below zero. The same company now posts gross margins above 80 percent; that gap raises the question of whether the cycle or the company has changed.

One example captures the violence of the cycle: the hosts recall Micron earning about $31 billion in revenue and $9 billion in net income in fiscal 2022, then falling to $15 billion in revenue and a $6 billion loss a year later. Such swings explain why memory stocks trade at persistently single-digit earnings multiples . Old scars are the real weight on today's optimism.

HBM: the memory bottleneck of AI

At the heart of the structural-change thesis sits high-bandwidth memory (HBM): memory dies stacked vertically next to the processor so data flows multiply. The hosts say Micron's HBM3E ships across every front, from Nvidia chips to AMD's MI350, AWS Trainium and Google TPUs. According to the official statement carried by GlobeNewswire, the company began volume shipments of HBM4 36GB 12H in the first quarter of 2026, promising over 2.8 TB/s of bandwidth, pin speeds above 11 Gb/s and more than 20 percent better power efficiency than HBM3E, with 48GB 16H samples already with customers.

Nvidia's official Rubin page confirms the scale: 288 GB of HBM4 per Rubin GPU at 22 TB/s of bandwidth, 20.7 TB of HBM4 per NVL72 rack. The rumored 168 GB low-memory variant debated by the hosts should be read alongside the Groq 3 LPX companion rack on the same page (256 LPUs and 128 GB of SRAM per rack): Nvidia is shifting part of latency-sensitive inference to SRAM. The bulls' answer is the Jevons paradox : as memory per unit gets more efficient, total usage grows so much that the net effect still favors memory makers.

Contracts: the experiment to tame the cycle

Micron's answer to the cycle is 16 strategic customer agreements covering 2026 to 2030. According to the Trefis analysis carried by Yahoo Finance, these take-or-pay contracts carry floor prices that leave profitability above the peak margins of past cycles, while ceiling prices are fixed near second-quarter market levels. The hosts report locked-in minimum revenue around $100 billion with the deals extending toward half of revenue; customers buy insurance against price shocks, Micron sells the future.

The numbers back the claim: per Yahoo Finance and Trefis, the company posted a record 84.9 percent gross margin in the third quarter; DRAM prices rose by percentages in the low 60s in a single quarter while bit shipments grew single digits, meaning it sold dearer rather than more memory. The same analysis says DRAM contributed 76 percent of the $41.5 billion quarter and guides the fourth quarter to $50 billion in revenue at margins around 86 percent. The ceiling detail is critical: part of the record profitability is already capped by contract.

The litmus test of pricing power was Apple. The hosts recount how Apple first absorbed memory costs without passing them on, then raised the white flag about six months ago and increased prices. Failing to qualify alternative suppliers or trim premium memory volumes shows bargaining power has fully shifted to producers. Customers assigning engineering armies to escape memory dependence proves the same point.

Supply: no relief until 2028

Management's boldest claim is on supply: the imbalance stretches beyond 2027 and the 2028 relief will be gradual. Micron's Boise page says the company plans $50 billion in fab investment through 2030, with first Idaho fab wafer output in mid-2027 and the second fab online at the end of 2028. The TrendForce assessment carried by Chosun ties the three-way supply chain and qualification schedules to this capacity picture. So the capacity flood bears await arrives in 2028 at the earliest, and in controlled fashion.

The second demand wave is forming on the consumer side. Per Meta's Connect 2026 announcement, the Muse personal assistant is moving onto AI glasses; an always-on assistant processing contacts, emails and travel implies manifold the memory traffic of a classic chatbot. The hosts put Tesla's full self-driving software in the same basket. Data-center training was the first wave; continuous inference on edge devices could be the second.

Cracks: China, SRAM and the spending peak

The most concrete risk comes from China. Counterpoint data carried by China Daily shows CXMT reaching 10 percent global DRAM share in the second quarter; Samsung leads at 38 percent, SK Hynix at 25 percent and Micron at 24 percent. CXMT lifting wafer capacity from 320,000 toward 420,000 with plans to double by 2030 means price pressure in commodity DRAM. The knock-on effect the hosts stress: if China fills commodity capacity, Korean and American makers shift more lines to HBM and tightness is felt everywhere.

The second risk sits on HBM's throne. The SemiAnalysis assessment carried by Benzinga claims Micron missed Nvidia's 11 Gb/s bar and takes zero share of Vera Rubin HBM4, split 70 percent SK Hynix and 30 percent Samsung. The TrendForce view carried by Chosun corrects the picture: a three-way supply chain is forming, Micron completes qualification in the second quarter, SK Hynix takes about two-thirds of volume. The claim remains an unverified leak ; a single sentence on HBM4 share in the earnings call could move the stock hard.

The hosts also give a watch list: TSMC's cautious language, a hint from a cloud giant that capital-expenditure growth is peaking, and rhetoric shifts before any contract cancellations. There is also an accounting trap: Micron reports the fourth quarter on 14 weeks with the first quarter back to 13, so the sequential-growth slowdown could fool algorithms. Signal must be separated from noise .

On September 30 three questions matter more than headlines: will the agreement count and the half-of-2028-revenue language harden, will HBM4 share near 20 percent be evidenced, will the capital-expenditure calendar stay in 2028. MarketBeat's compiled $31.14 earnings expectation is beatable, but the 2027-2028 outlook is the real price setter. The hosts' valuation contrast is striking: Micron at five to six times earnings against Dell at 19, AMD at 60 and Intel at 100. The thesis stands despite the bubble confession: bubbles pay , until the exit door appears.