The two names at the top of our list are SanDisk (SNDK) and Marvell (MRVL), and the reason is the same in both cases even though the businesses look nothing alike. Each one has spent the past year converting a cyclical, spot-priced position into contracted or attached revenue. That conversion, not the AI headline, is what justifies the multiple. The ceiling on both is now set by how much compute the world can power, not by anything either company controls.
Start with SanDisk, because the case there changed on August 13. The stock closed recently at $1,657 for a market capitalisation of roughly $243 billion, up 541% in 2026 and up about 50% from its July low, yet still around 30% below the intraday record of $2,354.39 set on June 22. Fiscal Q4 delivered $8.97 billion of revenue, growth of 372% year over year, gross margin of 84.6% and better than $5 billion of free cash flow. The September quarter is guided to $10.3 billion to $10.8 billion. On trailing earnings the stock trades near 22 times. Those two facts sit uneasily together, and the market has spent two months trying to decide which one is the mistake.
The Investor Day answered it. Management laid out a financial model covering fiscal 2028 through fiscal 2030 that calls for mid to high teens annual revenue growth, adjusted gross margin around 80%, adjusted operating margin around 75%, operating expenses at roughly 5% of revenue, and adjusted free cash flow margin near 50% after tax, capex and working capital. The company also committed to returning 100% of excess cash after reinvestment. Shares rose 13.7% on the day to $1,528.11, extending a four-session gain to 25.8%, and dragged Micron, SK Hynix, Lam Research and Applied Materials up with them.
The number that matters is not the margin target. It is the coverage. SanDisk has signed New Business Model agreements with eight customers. Those agreements are expected to cover approximately 50% of bit shipments in fiscal 2027 and roughly two-thirds in fiscal 2028. They carry committed purchase volumes, binding contractual frameworks, minimum financial guarantees and structured pricing rather than quarterly spot negotiation, and they run three to five years. Bernstein estimates the floor pricing in the recent agreements at around $0.29 per gigabyte, broadly in line with the average selling price the company was already achieving, and raised its target to $3,000 from $1,700 on the view that these provide materially better downside protection than the take-or-pay contracts NAND vendors have signed in the past.
Understand what that does to the valuation framework. A NAND producer earns a low single-digit multiple at the top of a cycle because the market correctly assumes the pricing evaporates. A supplier with two-thirds of its output committed at guaranteed minimums across multiple years, at 80% gross margin, is not that business and should not be valued as though it were. The consensus target sits near $2,094 against a $1,657 price, and the Street’s own reluctance is instructive: even the equal-weight and sector-perform holdouts at Wells Fargo and RBC raised their targets after the event, to $1,550 and $1,600 respectively. Nobody left the room more bearish.
The moat is real and it is technological rather than contractual. CMOS directly Bonded to Array separates the logic wafer from the memory array, which lets SanDisk raise interface speed, plane count and performance tier without migrating the cell node. The ninth-generation 2Tb QLC part demonstrated exactly this: a 33% bandwidth improvement and six planes with bit density unchanged, meaning the company can move product up the price stack without adding a single bit of industry supply. BiCS9 QLC extends the same two-dimensional scaling strategy into custom derivatives built for individual customers in a capital-efficient way. The financial expression of this architecture is $674 million of net property, plant and equipment against more than $20 billion of annual revenue, because fabrication sits inside the Kioxia joint venture. There is no other NAND supplier that converts pricing into cash at that asset intensity.
Demand is being reshaped in SanDisk’s favour at the same time. AI inference has pushed key-value cache into a new tier of the memory hierarchy that sits between DRAM and conventional storage, and SanDisk models 1.2 zettabytes of bit consumption from AI data centres alone by 2030. The High Bandwidth Flash specification published jointly with SK Hynix through the Open Compute Project this month is the tell. Competitors do not co-author a standard for a category that is not going to be large.
Marvell is the same trade expressed through interconnect rather than storage. Shares sit near $222 for a market capitalisation of about $203 billion, inside a 52-week range of $61.44 to $329.88, roughly a third below the high. Fiscal Q1 revenue of $2.418 billion grew 27.6% year over year with data centre at $1.83 billion, or 76% of the mix. Management guided the July quarter to about $2.70 billion, roughly 35% growth, and raised its outlook for both fiscal 2027 and fiscal 2028, with the street now carrying something close to $16.5 billion for FY28. Custom silicon activity is at a record, with more than 50 opportunities across more than ten customers and custom revenue capable of exceeding $10 billion by fiscal 2029. Interconnect guidance was raised by more than 70%, which is where the largest single upgrade landed.
That last point is the thesis. The market prices Marvell as a bet on winning custom accelerator sockets against Broadcom, and treats hyperscaler in-sourcing as the existential risk. It has the polarity wrong. Marvell’s durable franchise is SerDes and optical DSP, and that content attaches to every accelerator in the rack regardless of who designed the compute die. If Google, Amazon or Microsoft pulls a processor design in-house, the interconnect, the retimers, the 800G and 1.6T optics and the 51.2 and 102.4 terabit switching still get bought, and Marvell still sells them. Attach revenue survives the decision that kills design-win revenue. Nvidia’s $2 billion convertible preferred stake, about 2.5% of the company, is best read as Nvidia buying exposure to Marvell wins that contain no Nvidia silicon at all. Jensen Huang has separately called Marvell a future trillion-dollar company, and the NVLink Fusion partnership puts Marvell inside the rack architecture of its supposed competitor.
The product cadence supports the multiple. Teralynx T100 delivers 102.4 terabits per second of switching. The Celestial AI acquisition, at $3.25 billion and up to $5.5 billion with earnout, plus XConn, bought the photonic fabric and CXL switching pieces outright rather than waiting on internal roadmaps. At Flash Memory Summit this month Marvell introduced the Bravera SC6 PCIe 6.0 SSD controller, sampling in the fourth quarter, alongside Structera X memory expansion and pod-level optical shared memory. Shares rose about 14% on that announcement alone. The company is also committing $250 million over three years to double engineering headcount in Bangalore and Hyderabad, making India its second-largest research hub. S&P 500 inclusion took effect on June 22 and brought permanent passive demand with it.
Owning both is not a doubled bet. It is coverage of the two physical constraints on AI inference, which are memory capacity and the bandwidth between chips. Neither company needs a particular model, a particular accelerator or a particular hyperscaler to win. SanDisk sells the bits regardless of which rack they land in. Marvell sells the connection regardless of what sits on either end of it. Compute vendors compete for share of a fixed power envelope. These two get paid on the envelope itself.
The Street has not caught up on either. Marvell carries 38 covering analysts, an average target near $257, a high of $400 and not a single sell rating, against a stock at $222. SanDisk carries a consensus near $2,094 against $1,657, with a high of $3,000. Both report into that gap shortly: Marvell on August 27, with the options market pricing a move of roughly 14%, followed by an Investor Day on October 6, and SanDisk on November 5.
The single number to watch is SanDisk’s fiscal 2028 New Business Model coverage. Management has guided it to roughly two-thirds of bit shipments. If the next update moves it higher, the argument that NAND remains a commodity cycle stops being available to anyone, and the multiple has a very long way to travel.
