AI Picks and Shovels (Arc 2: Wiring the Mind): Micron (NASDAQ: MU) — The Memory

Picks and Shovels Series | Arc 2: Wiring the Mind | Part 2 of 3

Last week we featured Arista Networks, which gave the cluster a way to talk to itself — a switching fabric that turns a warehouse of GPUs into one training run instead of thousands of isolated boxes. Talking to each other is only half the problem. Every one of those GPUs also needs somewhere to hold what it is working on, and that somewhere is running critically short.

This week’s company does not build a single accelerator. It builds the memory that sits next to one.

Memory is the constraint nobody sized for

For most of the last three years, the AI infrastructure story has been told almost entirely through the lens of compute: how many GPUs, how much FLOPS, whose chip is faster. Memory was the boring, commoditised input sitting quietly in the background.

That story has flipped. High-bandwidth memory — HBM — is the stack of DRAM dies wired directly onto an AI accelerator, and it determines something compute alone cannot: how much context a model can hold in a single pass, and how fast it can move data in and out of the chip doing the work. A faster GPU sitting next to an undersized memory pool is a GPU waiting on data.

As model context windows and agentic workloads have grown, HBM has gone from a line item to a binding constraint on what an AI cluster can actually do. Only three companies on the planet manufacture it at commercial scale: SK Hynix, Samsung, and Micron Technology.

Micron — The Memory

Micron (NASDAQ: MU) is the smallest of the three by HBM share, and the only one of the three that is American. Founded in 1978 above a dentist’s office in Boise, Idaho, by four friends with no formal background in the industry they were about to spend five decades in, it survived boom-and-bust memory cycles that wiped out nearly every other Western DRAM maker. It is now, by a wide margin, the most consequential American company in a market that AI has turned from a commodity business into one of the tightest, highest-margin supply squeezes in semiconductors.

The numbers: a company being re-rated in real time

Micron’s fiscal year runs to late August. Its most recent reported quarter — fiscal Q3 2026, released 24 June — was a genuine record by any measure. Revenue came in at US$41.46 billion, more than triple the US$9.30 billion booked in the same quarter a year earlier and up from US$23.86 billion the prior quarter. GAAP net income was US$28.24 billion, or US$24.67 per diluted share; non-GAAP EPS was US$25.11. Operating cash flow was US$25.39 billion. Management guided fiscal Q4 to US$50 billion in revenue, plus or minus US$1 billion, with gross margin near 86% and non-GAAP EPS of US$31, plus or minus US$1.

Micron reports fiscal Q4 2026 results after the market close this Wednesday, 30 September. Street consensus sits around that US$31 guide. The company has beaten consensus EPS in each of the past four quarters, by an average margin a little above 20%.

The stock has already done a lot of pricing-in ahead of that print. MU closed Friday 25 September at US$1,082.28, up roughly sixfold over the past year, for a market capitalisation of about US$1.22 trillion. On trailing twelve-month earnings the multiple is in the mid-20s. Annualise the current-quarter guide and the forward multiple compresses into the high single digits — which is exactly how peak-cycle memory stocks always look cheap, right up until the cycle turns. Whatever the actual number on Wednesday, the more important disclosure will be management’s tone on fiscal 2027 supply commitments. The shortage driving this re-rating still has no near-term fix.

The shortage, in one number: 3-to-1

Micron’s own management has put a figure on the mechanism reshaping the whole memory industry: every unit of wafer capacity converted to HBM production costs roughly three units of standard DDR5 capacity. The company has since said the silicon penalty is not shrinking as generations advance. HBM and conventional DRAM are built on the same fabrication lines, so every wafer Micron, Samsung, or SK Hynix diverts to feed AI accelerator demand is wafer capacity that used to make the memory going into ordinary laptops, phones, and servers.

The result, as of September 2026, is what several industry analysts are now calling the most serious memory shortage in fifteen years. HBM is effectively sold out for all of calendar 2026 across all three suppliers, under multi-year Strategic Customer Agreements that lock in volume and, in Micron’s case, price bands whose floors the company itself has said support gross margins well above any previous cycle’s peak. Demand visibility has already extended into 2028. Meaningful volume from new greenfield fabs is not expected before then, because building and qualifying a leading-edge memory plant takes years, not quarters. Incremental conversions and yield gains can loosen the tape a little earlier than that headline date. They cannot refill a sold-out year.

Micron has committed more than US$250 billion to US manufacturing investment through 2035, targeting 40% of its DRAM output onshore. Japan is now a genuine node in the same chain: Micron’s first HBM production wafer came off the Hiroshima line, and the company has broken ground on a roughly US$9 billion expansion of that site, with commercial shipments from the new capacity aimed around 2028.

That combination — sold-out capacity, multi-year contracts, and a multi-year lead time before large new supply arrives — is precisely the setup that hands a supplier real pricing power. Micron’s non-GAAP gross margin has moved from the high-40s a little over a year ago into the mid-to-high 80s on the current quarter’s guide. That is not a modest cyclical uptick. It is a structural re-rating of the business, for as long as the shortage holds.

The competitive order: a distant, growing third

Micron is not the market leader in HBM, and it is worth being honest about where it actually sits. SK Hynix has held the dominant position for years and still commands about half the market by most recent industry estimates, built on its early qualification as Nvidia’s primary HBM supplier. Samsung sits second: its share rose sharply through 2026 as HBM4 qualification progressed, reaching roughly a third in the second quarter. Micron is the clear number three, at about the high teens — a figure that itself is still moving as the HBM4 ramp gathers pace.

The gap matters less than the direction of travel. Micron’s product breadth — HBM, server DRAM, low-power DRAM, and NAND — is comparable to SK Hynix’s; its share within HBM specifically is simply smaller today. High-volume HBM4 shipments and progress on HBM4E for 2027 were both flagged in the most recent quarter. The bull case rests less on catching SK Hynix outright and more on the fact that, in a three-supplier market that is sold out regardless of who is winning share, being a well-capitalised, growing third player in an oligopoly is still an extremely good place to sit.

The genuine risk on the other side of that: Samsung’s HBM4 requalification at Nvidia, if and when it lands cleanly, would intensify competition for the same hyperscaler wallets Micron is currently locking in via long-dated agreements, and could eventually compress the extraordinary margins the shortage supports. Customer concentration is the same coin flipped over. HBM demand is still an Nvidia-and-hyperscaler story; a pause in accelerator deployments would hit Micron harder than a garden-variety DRAM downturn. None of the three suppliers has ever proven able to hold peak-cycle pricing power indefinitely. Memory has been a brutally cyclical industry for forty years, and Micron itself was posting losses as recently as fiscal 2023. The shortage is real and, on most supply forecasts, likely to persist through at least 2027. It is not a permanent state of the industry.

The read-through: margin pressure at the buyers

Every dollar of pricing power Micron is capturing is a dollar coming out of someone else’s cost base, and that read-through has started showing up well beyond the semiconductor sector — which is exactly the inflation-adjacent thread running through this week’s Market Moves piece.

Apple has flagged rising memory costs compressing iPhone margins, and has already lifted Australian MacBook and iPad prices by 20% or more. Microsoft has implemented price increases on some of its AI-adjacent software products, including Copilot for GitHub, partly to recoup infrastructure costs that include memory. Samsung’s own mobile division has warned of a challenging year for smartphone shipments as device makers pass memory-cost inflation through to consumers rather than absorb it. IDC is now projecting the average global smartphone price will hit a record high in 2026 as entry-level, sub-US$100 handsets effectively disappear from the market, and is forecasting a contraction in PC unit shipments of up to roughly 5% this year as vendors signal broad-based price increases. TrendForce measured conventional DRAM contract prices climbing close to 100% in a single quarter earlier this year — the fastest move on record for the category.

None of that is abstract for an Australian household. It shows up in the price tag on the next laptop or phone upgrade at JB Hi-Fi, at the exact moment central banks on both sides of the Pacific are trying to convince markets that goods inflation is under control. It is a small thread in a much larger inflation story, but it is a real one, and it is a direct consequence of the same shortage that is making Micron’s shareholders very happy.

The Aussie angle: two sides of the same shortage

Australia has no domestic memory manufacturer, so the direct way to hold this story is entirely indirect — through Micron itself as a US-listed name, or through a broad US index sleeve (IVV, VGS, NDQ, or an equivalent) or a superannuation growth option that already carries it as part of a market-cap-weighted Nasdaq or S&P 500 exposure, almost certainly without the average member ever having chosen it deliberately. Buying MU directly on Nasdaq is straightforward through most Australian brokers; it also means US dollar exposure, US brokerage, and US tax paperwork that an ETF sleeve quietly absorbs.

The more interesting local angle sits on the other side of the trade. NEXTDC (ASX: NXT), Australia’s own listed data-centre operator, posted record FY26 results in August: contracted utilisation nearly tripled to 740 megawatts on a pro forma basis, net revenue reached A$405 million, and actual capital expenditure came in at A$3.4 billion — already above the A$2.7–3.0 billion range it had guided. FY27 capex is now guided at A$5.25–5.75 billion. Since the start of FY26 the company has raised about A$9.8 billion of fresh capital to fund the build-out, and it is planning a multi-billion-dollar hyperscale AI campus at its S7 Sydney site with OpenAI under a memorandum of understanding.

Every new rack NEXTDC is contracted to build gets filled with memory-dense GPU servers, and the memory going into them is priced off exactly the same shortage inflating Micron’s margins. An Australian investor can be structurally exposed to both sides of this trade at once: a beneficiary through Micron sitting in an index sleeve, and a payer of the same input-cost inflation through the build costs embedded in every dollar NEXTDC spends on capacity — costs that eventually show up in what Australian businesses pay to rent a rack, or what Australian consumers pay for the device on the shelf this Christmas.

That is not a reason to avoid either side, and it is not a claim that NEXTDC is a victim. It is a growth compounder whose input costs are rising. It is a reason to understand that “the memory shortage” is not a single, cleanly one-directional bet. It is a cost that moves through the entire AI supply chain, landing differently on a chipmaker’s income statement than it does on a data-centre operator’s capex bill or a household’s upgrade budget.

The investment case — in one line per side

Bull case: Micron is one of only three companies on Earth that can supply HBM at commercial scale, sitting inside a sold-out, multi-year-contracted shortage with pricing structures set to support margins above any previous cycle’s peak, demand visibility already extending into 2028, and a growing share of a category that did not meaningfully exist in its current form two years ago.

Bear case: Micron is the clear number three in HBM behind two better-established Korean rivals, memory has a forty-year history of brutal cyclicality that has bankrupted or humbled every player in it at least once, new supply from all three suppliers eventually arrives regardless of today’s shortage, and a stock that has run about sixfold in twelve months is pricing in a lot of things going right in a row.

The honest read: The shortage is real, well-documented across independent industry sources, and likely to persist for at least another year or two on the current supply timeline. Whether US$1,082 is paying you fairly for that, given how cyclical this industry has always proven to be, is a question this Wednesday’s earnings print will move a great deal — and one every investor still has to answer for themselves.

Next Monday

Next Monday’s company doesn’t build a chip at all. It was started by two facilities engineers with an idea that had nothing to do with silicon: a neutral room where rival networks could plug into each other without any one of them owning the wires. Its first building went up in a Virginia town nobody outside the industry had heard of; that town is now the single densest patch of internet infrastructure on the planet. It survived the dot-com crash that took down most of its early peers, turned itself into a landlord two decades ago, and now owns — literally, by the square foot — a meaningful share of the rooms where the world’s cloud providers physically connect to one another.

A brain with memory still needs a place to live.

Data Sources:

• Micron Technology, Form 8-K, press release and prepared remarks, fiscal Q3 2026 results, 24 June 2026; fiscal Q4 2026 earnings date, 26 August 2026
• Micron Technology, US investment update, 9 July 2026; Hiroshima expansion remarks, July 2026
• Counterpoint Research and Reuters, HBM market-share estimates, 2025–2026
• IDC, memory-shortage impact on smartphone and PC markets, August 2026.
  TrendForce, DRAM and HBM contract pricing, 2026
• Reuters; The Guardian; ABC News, Apple, Samsung and consumer-device pricing, 2026
• NEXTDC Limited, ASX release and FY26 results presentation, 27 August 2026
• Market data as of 25 September 2026 close

Wall St. Down Under | Australia

Subscribe | wallstdownunder.com.au

Disclaimer: Wall St. Down Under is an independent financial newsletter for informational and educational purposes only. Nothing published here constitutes financial advice, a recommendation to buy or sell any security, or a solicitation of any investment decision. Always do your own research and consult a licensed financial adviser before making investment decisions. Australian investors should consider their own financial situation, objectives, and risk tolerance. Past performance is not indicative of future results.