Samsung's FMS 2026 Keynote Is The Easy Part - Execution Is Not

Samsung took the stage at the Flash Memory Summit in Santa Clara today and talked about "breakthrough 3D memory and storage innovations." The keynote was designed to project momentum. The problem is that momentum is not the same thing as market share, and market share is exactly what Samsung has been losing.

The HBM4 spec is competitive. The cost structure is not.

Samsung's HBM4 ships at 11.7 Gbps per pin, upgradeable to 13 Gbps, with a total bandwidth of 3.3 TB/s per 12-layer stack. That's roughly 2.7x the bandwidth of HBM3E and exceeds JEDEC's HBM4 baseline of 8 Gbps by about 46%. On paper, Samsung's 1c DRAM core paired with an SF4 logic base die delivers pin speeds that can meet NVIDIA's Rubin requirements while staying below 1 volt.

SemiAnalysis confirmed this at ISSCC 2026. Samsung was the only top-three memory vendor to present a technical paper on HBM4, and the independent engineering analysis found that Samsung has closed the gap on raw performance. The spec is not the problem.

The problem is that Samsung chose SF4 - a near-leading-edge logic node - for the base die. SK Hynix is using TSMC's N12 logic process for the same purpose, and Micron is using its own lower-cost CMOS base-die technology. Even with vertical integration discounts from its own foundry, SF4 is structurally more expensive. Samsung's HBM4 may match or exceed SK Hynix on bandwidth, but it costs more to build. In a market where hyperscalers negotiate brutal per-unit pricing, that's a margin killer.

And Samsung has its own yield history to contend with. The 1c DRAM node - the front-end process Samsung skipped a generation to reach - was sitting at roughly 50% front-end yield throughout 2025. Samsung originally planned HBM4 mass production for the second half of 2025. That got pushed to 2026. The company finally shipped commercial HBM4 in February, six months behind its internal target.

One more missed timeline does not constitute a turnaround.

SK Hynix didn't just win. It eclipsed Samsung.

Here's the context that the FMS keynote doesn't mention. SK Hynix overtook Samsung in total operating profit for the full year 2025 for the first time in the two companies' histories. SK Hynix posted 47.2 trillion won in operating profit. Samsung's entire group posted 43.6 trillion won - and Samsung's memory division alone generated only 24.9 trillion won. SK Hynix's total operating profit was nearly double Samsung's memory segment profit.

In Q1 2025, SK Hynix surpassed Samsung in overall DRAM market share for the first time, taking 36% to Samsung's 34%, according to Counterpoint Research. In the HBM segment specifically - the part of the memory business that actually matters for AI infrastructure - Counterpoint estimated SK Hynix held a 57% revenue share in Q3 2025 versus Samsung's 22%.

By December 2025, SK Hynix had reportedly secured over two-thirds of HBM supply orders for NVIDIA's next-generation Rubin products. Samsung's HBM4 may be technically competitive, but SK Hynix already has the design wins, the customer relationships, and the scale economics.

Samsung didn't lose HBM to a better product. It lost to its own yield failures while SK Hynix executed.

The HBF pivot is where Samsung actually has leverage.

This is the part of the FMS narrative that's worth paying attention to. Samsung has begun conceptual design and early development of high-bandwidth flash - HBF - a technology that stacks NAND flash, not DRAM, to deliver roughly 80–90% of HBM's speed with 8–16x the capacity at roughly 40% less power consumption.

SK Hynix and SanDisk are already collaborating on HBF standardization, targeting samples in 2026 and mass production in 2027. Kioxia has shown prototypes. But Samsung holds 32.9% of the global NAND flash market - the largest share of any vendor - and its foundry division brings logic design and process expertise that could differentiate its HBF control architecture.

The Korea Herald reported in February that Samsung is approaching HBF as part of a "broader restructuring of AI memory and storage architecture" rather than treating it as a complementary product. That's a fundamentally different strategic posture from SK Hynix, which is positioning HBF as an HBM adjunct.

HBF is still years from meaningful revenue. HBM itself took seven to eight years from development in 2015 to mainstream adoption, and a KAIST professor who pioneered HBM's basic structure expects HBF demand to surpass HBM only around 2038. But Samsung's NAND dominance is a real foundational advantage in this next phase. It's the one area where Samsung enters as the incumbent rather than the challenger.

The cross-currents

Samsung's FMS 2026 keynote was designed to signal that the company has arrived in AI memory. The engineering evidence supports a narrower conclusion:

• HBM4 performance: Samsung has closed the spec gap. ISSCC data confirms competitive pin speeds and bandwidth. The question now is volume, yield stability at scale, and whether the SF4 cost penalty prevents Samsung from pricing aggressively enough to win share.
• Market position: SK Hynix's lead is structural, not temporary. It has the design wins, the yield history, the lower-cost base die, and the customer trust. Samsung's HBM market share of roughly 22% reflects years of yield failures, not a temporary gap. Closing it would require either sustained margin sacrifice or a performance breakthrough that SemiAnalysis has not yet observed.
• HBF trajectory: This is the asymmetric opportunity. Samsung enters with the largest NAND share and foundry expertise. But HBF is a 2027+ story with no established revenue baseline, and the standard itself is still being defined by SK Hynix and SanDisk.

Directionally, Samsung's AI memory story is a two-phase problem: survive on HBM4 margins despite a cost disadvantage, then compete on HBF before the standard and its supply chain lock in around SK Hynix's ecosystem. Both are harder than the keynote implies.

The investor implication is straightforward. Samsung's memory division generated roughly half of SK Hynix's operating profit last year. For the stock to re-rate on an AI memory thesis, Samsung needs to demonstrate not that it can ship a competitive chip - which today's keynote did - but that it can do so profitably at scale against a competitor that already controls the customers, the supply contracts, and the cost curve. Until that gap narrows in the financials, not just the spec sheets, the FMS keynote remains exactly what it was designed to be: a signal of intent, not proof of execution.