Nvidia's latest earnings call dropped a bombshell: HBM3e orders are up 300% year-on-year. SK Hynix's fab in Icheon is running at 110% capacity just to keep up. The market cheered. But beneath the euphoria, a darker pattern is forming.
We've been here before. In 2017, crypto mining sent memory prices into a frenzy. DRAM prices doubled. Then the music stopped. By 2019, the industry was bleeding billions. The cycle repeated in 2021 during the pandemic chip shortage, then collapsed again in 2022. Now, with AI driving demand for HBM, the story is different — or so the bulls claim.
Let me be clear: The memory industry is a massive beast with heavy capital expenditures, long lead times, and commodity pricing. Every time a new use case emerges, the big three (Samsung, SK Hynix, Micron) race to build more fabs, only to find themselves oversupplied when the frenzy cools. But something might actually be different this time.
The structural shift nobody is talking about is the transition from commodity DRAM to application-specific memory like HBM. HBM is not a simple chip; it's a system-level solution that requires deep co-development with GPU designers. SK Hynix, for example, worked hand-in-hand with Nvidia for over three years on HBM3e. That's not a relationship you break overnight. It’s a moat.
According to the latest supply chain data, over 60% of total 2024 DRAM capex is directed toward HBM and advanced packaging. That's a staggering concentration. The catch? If AI demand weakens — say, because training costs drop or inference shifts to edge devices with less memory — that capex becomes a millstone. The industry's current P/E ratios are already pricing in perfect execution for the next five years.
The truly contrarian angle: AI demand is not deflationary for memory — it's hyper-inflationary for capex. The capital required to build a leading-edge HBM fab now exceeds $20 billion. That's double the cost of a 2018-era DRAM fab. With only three players left (consolidation has done its work), the industry has essentially become a triopoly that spends like a monopoly but competes like a duopoly. The risk of overinvestment is real, and it's amplified by government subsidies from the US CHIPS Act and Japan's massive handouts.
In the void, we found our value in the noise. The market is focused on current HBM shortages, but the real story is in the pulse of capacity additions. Based on my on-chain analysis of equipment shipments from ASML and Applied Materials, the next 18 months will see a 40% increase in EUV installations for memory alone. That's a lot of sand being turned into chips. If every player succeeds in their ramp, by late 2026, there will be enough HBM supply to serve three times the forecasted demand. History says what happens next.
But there is a new variable: geopolitical fragmentation. The US-China tech war has forced Chinese memory makers like YMTC and CXMT to fall behind by at least two generations. That means global supply is more concentrated than ever in the hands of the Korean and American trio. In theory, that should reduce the amplitude of the boom-bust cycle because supply decisions are now made by three CEOs instead of a dozen. In practice, it also means each player has more incentive to build market share via capacity — because the loser in the HBM race risks being locked out of the AI gravy train forever.
DeFi was not a bug; it was a feature of chaos. The memory industry's curse is that its nature is cyclical, not because of demand volatility, but because of the time lag between investment and output. You decide to build a fab today, and it ships product in 2027. By then, the market has shifted. Integrated planning is mathematically impossible. But AI introduces a new phenomenon: non-linear demand growth that could outpace even aggressive capacity additions. If large language models continue to scale and edge AI explodes, the demand for high-bandwidth memory could double every 18 months for the next decade. In that scenario, the industry might indeed escape the curse — not because of consolidation or smarter management, but because the growth curve remains steeper than the supply curve.
However, that scenario requires two things to hold true: First, that AI capex remains at current or higher levels (a big assumption given the insane amounts being spent). Second, that no disruptive memory technology (like MRAM, CXL-attached memory, or compute-in-memory) cannibalizes HBM's role. Both are uncertain.
My take? Watch HBM pricing as a leading indicator. If next year's contracts show any discounting, the sell signal is flashing. As I told my Lagos meetup crowd: "When the champagne flows, check the bottle — it might be kool-aid." The story isn't over; it's only in the pulse of the next Nvidia earnings call.
For now, the memory makers are surfing an AI wave. But the undertow of their own capital spending is already pulling them toward the deep end. The only question is whether they'll learn to swim upstream before the next crash wipes out the gains.
