SanDisk Rose 40x; the Next Underappreciated AI Hardware Re-Rating Now Runs Through Hybrid Bonding and the HBM Crossover
SanDisk is the reference point that started this. After spinning out of Western Digital in early 2025, the stock bottomed near forty dollars in April of that year and now trades close to nineteen hundred — a forty-five-fold move accomplished in roughly twelve months. It is the kind of chart that sends investors hunting for the next one. But the lesson of SanDisk is easy to misread. It did not climb because of a proprietary technology nobody else had. It climbed because NAND flash entered a brutal undersupply, pricing inflected, and a newly independent company captured the entire swing. That is a commodity supercycle, not a moat. Memory re-rated because the physics of supply and demand turned, and the same mechanism will eventually turn the other way.
That distinction matters because there are two entirely different archetypes hiding inside the phrase “underappreciated AI stock.” The first is the commodity supercycle that SanDisk represents. The second is the proprietary-moat story, of which Marvell’s optical networking franchise is the cleaner example: a structural advantage in a technology the rest of the market has not yet fully repriced. An investor asking for a unique advantage that has not caught its wave is asking for the second kind. The search below was an attempt to find it, and the most useful result was a negative one.
The screen for moats that have not repriced mostly comes up empty
Run a screen for AI-hardware names with a genuine structural advantage and the immediate problem is that the obvious candidates already ran. The optical and power-delivery picks-and-shovels that dominate the conversation — Vertiv, Astera Labs, Credo, Coherent, Lumentum — have substantially re-rated. They remain good businesses, but they are no longer SanDisk-before setups. The asymmetry, if it exists at all, has moved further down the stack and into smaller, stranger names.
Even there, the honest finding is sobering. The upstream indium-phosphide substrate supplier AXT looked, on paper, like the perfect chokepoint play: indium phosphide is the scarce input under every optical laser and transceiver, and AXT is one of the few suppliers. But the stock had already risen roughly fifty-fold over the prior year, from under two dollars to a high above one hundred forty, before pulling back violently. The chokepoint thesis is real, with an indium-phosphide backlog above one hundred million dollars and gross margins that swung from negative to near thirty percent, but the easy money was gone, and the position carries two live risks that have nothing to do with the thesis: a six-hundred-million-dollar capital raise that dilutes existing holders, and a revenue stream gated by Chinese export permits that arrive on Beijing’s schedule rather than the company’s. AXT is the cautionary tale of this entire cohort. A real moat is not the same thing as an entry point.
The same caution applies across the group. Fabrinet, the de-facto monopoly contract manufacturer for high-mix optical transceivers, has risen well over one hundred and fifty percent in a year to a roughly twenty-five-billion-dollar market capitalization and a price-to-earnings ratio in the forties to sixties; its FY2025 revenue of three and a half billion grew under twenty percent and its return on invested capital sits near thirty percent, but the multiple already prices the quality. SiTime, the near-monopoly in MEMS precision timing, rose from the mid-four-hundreds toward eight hundred dollars; first-quarter 2026 revenue of one hundred thirteen million grew eighty-eight percent with a sixty-four-percent non-GAAP gross margin, and its acquisition of Renesas’ timing business pushes pro-forma revenue past five hundred million toward a billion-dollar path — but again, the market has done the repricing. These are compounders to own through a pullback, not forty-baggers waiting to happen.
Where any asymmetry actually remains
Filtering for a moat that has genuinely not repriced narrows the field to two uncomfortable names, and the discomfort is the point.
POET Technologies is the closest structural analog to Marvell’s optical story at its earliest and riskiest stage. Its patented optical interposer and hybrid laser approach is genuinely differentiated, using lithographically defined wavelength tuning and roughly two-thirds less indium phosphide than conventional designs — a meaningful advantage given the substrate scarcity that drove AXT. The company signed an initial fifty-million-dollar purchase order inside a framework that could exceed five hundred million over five years. But it remains a sub-billion-dollar company generating only a few hundred thousand dollars in quarterly revenue, with heavy dilution and a binary outcome: a multi-bagger if the interposer wins designs at scale, near-zero if it does not.
Aehr Test Systems is the unloved-monopoly counterpart. It is the only firm offering both wafer-level and package-level burn-in across silicon photonics, gallium nitride, silicon carbide, and high-bandwidth memory. The stock sits near forty dollars at the trough of a silicon-carbide and electric-vehicle downcycle, with zero debt and a clean balance sheet, precisely as an AI and photonics order cycle begins to build, with second-half bookings guided to sixty to eighty million. It is the rare case in this group that has not caught the wave, because its legacy end market is in recession while its future one is just igniting.
Ranked by upside ceiling, the order runs POET first, Aehr second, then the relative laggard Vicor in vertical power delivery, then Camtek in high-bandwidth-memory metrology, with AXT, Fabrinet, and SiTime trailing because their re-rating is largely complete. Ranked by probability of success, the order inverts almost exactly: SiTime and Fabrinet first as entrenched profitable monopolies, Camtek third on its near-duopoly in advanced-packaging inspection, then Aehr and AXT in the middle on execution risk, Vicor lower, and POET last as the pre-revenue binary. Upside and probability trade off cleanly, which is the entire lesson. There is no name in this cohort that is simultaneously cheap, safe, and pre-wave.
BESI is the highest-quality expression of the one theme still compounding
If the optical and timing layers have already repriced, the layer that has not finished is advanced packaging — and specifically the transition to hybrid bonding. BE Semiconductor Industries, the Dutch company known as BESI, is the cleanest pure-play on that transition, and it deserves examination not as a hidden bargain but as the highest-quality vehicle for the one structural theme with years left to run.
BESI makes the bonding equipment that assembles stacked silicon. It trades on Euronext Amsterdam under the ticker BESI, with only a thinly traded over-the-counter American depositary receipt, BESIY, available to United States accounts. The shares sit near two hundred ninety euros after touching an all-time high above three hundred fourteen in mid-June, up from a fifty-two-week low around one hundred five euros the prior September — a move of roughly one hundred forty-seven percent. The market capitalization is about twenty-five billion euros against eighty-one million shares, and the valuation is demanding: a trailing price-to-earnings ratio around one hundred fifty, a gross margin near sixty-three percent, a net margin around twenty-four percent, and a return on equity above thirty percent. This is not a cheap stock. It is a quality monopoly priced as one.
The moat is hybrid bonding leadership. BESI is estimated to hold roughly two-thirds of the hybrid-bonding equipment market and better than thirty percent of a market expected to exceed five billion dollars this decade. The durability comes from qualification lock-in: once a fab designs a bonder into a certified production flow, switching vendors is slow and expensive. The fundamental picture underneath the rich multiple is improving sharply. Fiscal 2025 was a digestion year, with revenue of five hundred ninety-one million euros down nearly three percent and net income down about twenty-eight percent. But orders re-accelerated violently. First-quarter 2026 orders of two hundred seventy million euros rose one hundred four and a half percent year over year, revenue of one hundred eighty-five million grew twenty-eight percent, and net income of fifty-two million rose sixty-four percent, lifting the net margin to nearly twenty-eight percent. Hybrid-bonding adoption has expanded to eighteen customers and more than one hundred fifty installed systems. Management guided second-quarter revenue to grow thirty to forty percent sequentially at a sixty-four to sixty-six percent gross margin, implying first-half revenue near two hundred ninety-two million, up about forty-nine percent.
The market has noticed, which is why the contrarian entry that briefly existed has already closed. In early March 2026 the stock fell roughly fifteen percent on a report that AI chipmakers were considering slowing hybrid-bonding adoption. That dip reversed completely, and the shares went on to new highs within months. What remains is a business whose moat is intact and whose order book is inflecting, but whose multiple now prices a flawless ramp. The bear case is not that BESI is a poor company; it is that hybrid bonding for memory has been one year away for several years running, and at one hundred fifty times earnings any slip in that timeline hits the stock harder than the fundamentals alone would justify. Skeptics have modeled an installed base far below the company’s own long-term targets. The risk is multiple compression, not bankruptcy.
ASMPT is the incumbent BESI is trying to succeed
The competitive landscape clarifies the thesis rather than complicating it. BESI’s principal rival is ASMPT, the Singapore-headquartered, Hong Kong-listed firm majority-owned by ASM International of the Netherlands. The rivalry is asymmetric, and the asymmetry is the whole story.
The two companies lead different bonding methods. ASMPT is the global leader in thermocompression bonding, the technique that joins stacked dies through solder microbumps under heat and pressure, with its Firebird platform having shipped more than two hundred fifty units. Thermocompression bonding is what high-bandwidth memory uses today. BESI, by contrast, dominates the successor technology, true hybrid bonding. So the two firms are not fighting over the same socket so much as betting on different timelines: ASMPT owns the incumbent process, BESI owns its replacement. Across the broader thermocompression market the top five suppliers hold roughly eighty-eight percent combined, so even the incumbent layer is an oligopoly rather than a monopoly.
Each side has a heavyweight ally, and those alliances move BESI’s stock more than ASMPT’s results do. Applied Materials took a nine-percent stake in BESI in April 2025, becoming its largest shareholder, and the two co-developed an integrated die-to-wafer hybrid-bonding line pairing BESI’s placement precision with Applied’s wafer-cleaning and metrology. ASMPT’s counter is a deep collaboration with IBM on both thermocompression and hybrid bonding for chiplet packages. The genuine wildcard sits above both: ASML has reportedly been studying hybrid-bonding equipment, and its ultra-precision lithography heritage could reshape the competitive map if it productizes a tool. That latent entrant, not ASMPT, is the real threat to BESI’s moat. One structural point favors BESI as the cleaner vehicle regardless: advanced packaging is only about a quarter of ASMPT’s revenue, the rest tied to more cyclical surface-mount businesses, whereas BESI is concentrated in exactly the high-value bonding an investor wants exposure to.
The physics: from microbumps and thermocompression to true hybrid bonding
The entire thesis rests on one transition in how stacked dies are electrically connected, and it is worth stating plainly. The incumbent method places tiny solder balls, called microbumps, between two chips, then joins them. Thermocompression bonding is the precise version of this: rather than reflowing a whole batch at once, a tool picks each die, aligns it, and presses it down with heat and force, bonding one die or a few at a time. This enables finer pitch and thinner stacks than mass reflow, which is why memory through the current generation, and twelve-high stacks, rely on it. But the method still depends on solder bumps and on underfill between the layers, and that dependence is the ceiling. Bumps add height to every layer, solder carries far more electrical resistance than copper, and as pitch falls below roughly ten microns the bumps begin to bridge and short, electromigration worsens, and the growing stack becomes too tall and too hot to manage.
Hybrid bonding removes the solder entirely. Both mating surfaces, a dielectric with embedded copper pads, are polished to near-atomic flatness, then brought together. The dielectric surfaces bond first at room temperature, and a low-temperature anneal causes the copper pads to expand and fuse directly, copper to copper, with no bump and no underfill. The payoff is large on every axis at once. Pitch falls below five microns and trends toward one, so the interconnect density rises by orders of magnitude. Each layer becomes far thinner, so more dies fit inside a fixed height budget. Direct copper contact slashes resistance and inductance, raising bandwidth and lowering power. And copper conducts heat far better than a sandwich of solder and underfill, easing the thermal problem that limits tall stacks.
The cost of all this is manufacturing difficulty, and that difficulty is precisely the moat. Surfaces must be defect-free and flat to sub-nanometer tolerances, because a single stray particle becomes a void and a dead connection, and alignment must be sub-micron, which is why BESI advertises placement-accuracy prototypes measured in tens of nanometers. Very few tools can do this reliably at production volume, and the firms that can are the ones worth owning.
The crux is the HBM crossover, and the signals are upstream
There are two adoption fronts, and they explain why the timeline is the whole debate. Logic-on-logic stacking, where compute dies are bonded directly atop one another, already uses hybrid bonding; that front is live. High-bandwidth memory is the larger volume prize and the one still transitioning. Through the current generation it remains on microbumps and thermocompression. The inflection arrives with the next memory generations, where the height budget for sixteen-layer stacks finally forces the change. One major memory maker has signaled it will apply hybrid bonding as early as its sixteen-high next-generation parts, with some intermediate product likely bridging on advanced thermocompression first.
That single decision, made not by the tool vendors but by the memory makers, sets BESI’s order trajectory. The signals worth watching therefore sit upstream of the equipment companies entirely. The high-bandwidth-memory roadmaps of the three memory makers determine the crossover, and the share leader’s volume commitment matters most. The foundry capacity commentary around advanced packaging and stacked logic is the readthrough for the front that is already moving. And ASMPT’s thermocompression bookings function as a thermometer rather than a trade: if its incumbent orders stay strong deep into the next memory generation, the crossover is slipping and BESI’s timeline lengthens; if that commentary softens while hybrid-bonding talk rises, the handoff has begun.
The synthesis is clean. SanDisk’s forty-fold run was a commodity supercycle that is mostly spent, and the proprietary-moat names that looked like its successors have largely repriced already. What has not finished repricing is the packaging layer, and within it the move from solder to direct copper bonding. BESI is the best business positioned at that hinge, expensive precisely because the market understands the prize. The position any investor takes there is not a bet on a hidden monopoly. It is a bet on a known monopoly’s timeline, and the timeline belongs to the memory makers. Watch what they commit to volume, and the rest of the thesis follows.