Back the growth rates out and the forecast, as reported by cryptobriefing citing Gartner, starts arguing with itself. A 246.6% rise in DRAM revenue is a multiple of 3.47, and 371.9% for NAND is 4.72 [3][4], so a 2026 memory line of $837.3 billion [1] implies a 2025 base somewhere between $177 billion and $242 billion [6]. The bracket is approximate, since memory as a reporting category is wider than DRAM plus NAND. If that base was 27% of the 2025 market [2], the 2025 market was between roughly $657 billion and $895 billion [7], which means the $1.555 trillion 2026 total requires the whole semiconductor industry to grow between 74% and 137% in twelve months [8]. That is the figure to interrogate before anyone argues about memory's share of it.
The second year of the forecast is where the caution shows up in arithmetic rather than prose. Memory passes $1 trillion in 2027 against a $1.94 trillion total [6], which is 51.5% rather than 53.8% [5], and it gets there on 19.4% growth [4] while everything that is not memory runs from $717.7 billion to $940 billion, or 31% [2][3]. The total market grows 24.8% across that step [10]. Analysts quoted in the same report warn that this cycle is unlikely to be different in kind even if the demand wave is different in scale [10], and the 2027 line is that warning with a number attached.
Nvidia's procurement increment is the term worth studying. The quarter-on-quarter jump alone is $160 billion, which is 19% of the entire 2026 memory forecast, and the full commitment runs to a third of it, scheduled out to 2029 [9][9]. What that buys Samsung, SK hynix and Micron [8] is a visible order book; what it costs them is optionality, because a wafer start committed to high-bandwidth memory is a wafer start unavailable to the commodity DRAM and NAND that goes into phones and enterprise SSDs. SK hynix, on 56.4% of HBM [7], carries the most revenue and the most concentration at once. Cryptobriefing's own framing is that a modest slowdown in Nvidia accelerator demand would be amplified downstream by the sheer scale of those commitments [13].
This is probably wrong, but the position I would rather hold is the one with the least capex committed rather than the most HBM share, because fabs take years and cost billions [10] and concrete poured in 2026 is priced against a 2027 line that already decelerates [12]. Three ways it goes otherwise. Pricing holds because AI data centres genuinely reach 36.5% of semiconductor revenue in 2026 and 53% by 2030 [5], in which case today's capex looks cheap in hindsight. Or the 246.6% is mostly price rather than volume, a split the forecast does not publish and one that matters enormously to a buyer signing a 2027 server contract. Or Nvidia discloses another $160 billion next quarter and capacity ownership stops being an interesting question for two more years. What would falsify the caution is checkable: 2027 memory revenue arriving above $1 trillion on units rather than on average selling prices.
