Cantor Fitzgerald: Memory Supply Will Be Tighter in 2027 Than 2026, Upcycle May Extend Beyond 2028
N.R. Finch
Cantor Fitzgerald analyst CJ Muse argues memory-chip supply will be tighter in 2027 than 2026, with the upcycle potentially lasting past 2028; Phison Electronics' CEO echoes the call from the supply-chain side, suggesting memory earnings expectations are systematically underpriced.
Why does the Street underestimate memory-chip earnings?
CJ Muse's core call: 2027 supply will be tighter than 2026, and the upcycle could extend to 2028 or beyond.
This means → this is not a short-cycle bounce but a structural upswing that may span three to five years.
On bull-case math, the companies in question could earn close to $200 per share next year, yet the stock trades at roughly 5× P/E — well below historical peaks.
In plain terms = if earnings reach that level, today's price is set for "the cycle is about to end," but the supply-demand data say something very different.
What is the supply-chain signal confirming?
Phison Electronics (群聯電子) CEO K.S. Pua says the NAND flash market faces a shortage "with no end in sight", with order visibility stretching into Q1–Q2 2027.
He characterizes the situation as "prolonged and irreversible", expecting the Q4 2026 shortage to be worse than Q2.
This means → the investment-bank macro view and the front-line supply-chain read point the same way — the supply gap is widening, not narrowing.
Why will AI inference become a lasting driver of memory demand?
AI model training leans on compute (GPUs), but once a model enters inference — the phase where it actually answers user queries — demand for flash storage and SSDs jumps sharply.
As the AI user base grows and content shifts from text to images and video, data volumes over the coming years will far exceed today's levels.
In plain terms = training is "teaching the model"; inference is "putting the model to work" — the working phase stores far more data, and flash capacity cannot expand fast enough to keep up.
This reflects a market overly focused on the compute race; the real bottleneck is shifting from GPUs to storage capacity.
Where do Phison's own inventory and orders stand?
At end-Q1 2026, Phison's inventory exceeded NT$70 billion (roughly US$2.2 billion) and is still growing.
Customers are asking to increase purchase volumes nearly every month; the company can only prioritize strategic accounts.
This means → even with aggressive stockpiling, supply still cannot keep up with demand growth — securing supply now depends not just on capital but on maintaining NAND-vendor relationships.
What is new in Phison's AI business and internal efficiency?
The aiDAPTIV+ business generates roughly NT$200–300 million per quarter, still in proof-of-concept and early market adoption; more visible growth is expected in Q3–Q4.
Since rolling out AI-assisted development in 2025, Phison had saved the equivalent of 49 engineers' workload by April 2026, rising to 75 engineer-equivalents in May.
The company says it will not cut headcount; freed capacity is redeployed to customer service, test validation, and new-product development.
What assumption is baked into the current valuation?
A ~5× P/E implies the market believes this upcycle is about to peak.
In plain terms = the market is betting "the good times are almost over," but supply-demand data and supply-chain signals both say "the shortage is just getting started."
Whether supply meaningfully improves in 2027 is the pivotal test of which side is right.
Content is for reference only, not financial advice.