Memory is having a moment. Revenues for SK Hynix and Micron have tripled in the last year, and Samsung’s has roughly doubled, according to the source. It is the obvious story: AI datacenters are guzzling high-bandwidth memory (HBM), DDR5, and NAND flash for GPU servers, so shortages turn into price hikes that ripple from AI infrastructure to consumer electronics.

Now zoom out one step, and the real twist shows up: this is not a quick fix cycle. Micron’s Thursday announcement that it would invest up to $3 billion to strengthen the US semiconductor supply chain lands in the middle of a longer timeline that typically cannot be compressed. The source also notes that the Idaho-based chipmaker is working to boost production across its Singapore, Taiwan, and Japan sites, but even with new fabs under way, anything SK, Samsung, or Micron starts today will take at least three years to bring online, and even longer to ramp production. That means the high-price environment does not just last through a quarter. It can last through an entire planning horizon.

Why the delay matters for more than memory investors is that semiconductors are slow by design, not slow by choice. Building a new DRAM or NAND wafer fab is described as complex and resource-intensive. Financing has to be secured, a location selected, permits won, and support facilities deployed. Those include power conditioning, air handling, and ultra-pure water filtration. Even after clean rooms are completed, hundreds of millions of dollars of specialized lithography, wafer transport, and test equipment must be installed and validated. Then the site has to be powered on and tuned to acceptable yields, a process that can take months, and often years even without delays.

This is where the AI boom changes the usual memory-market math. Historically, memory is a commodity with wild pricing swings driven by boom and bust cycles. Vendors rely on boom periods to finance fabs, even knowing that once additional capacity comes online, prices can crater. The source argues that AI has flipped the expectation. Instead of memory prices falling across 2025 and 2026, the opposite happened as AI infrastructure consumed every bit of DRAM and NAND it could get. The result is that startups building models, agents, and tools face a new reality: it is no longer whether the technology works, it is whether benefits justify continued investment at current or higher levels of infrastructure cost.

The cash-pressure point is straightforward in the source: sooner or later, AI startups have to turn a profit, and sky-high memory prices do not help margin in the cost per token. Those economics are especially relevant because OpenAI and others have already spent about the last four years developing increasingly capable systems, using hundreds of billions of VC capital. If memory prices remain elevated longer, the runway math gets tighter. The question becomes whether memory vendors can bring capacity online before AI “great houses” exhaust their VC-subsidized runway and the music stops.

The timeline for relief is the part that should make boards wince. A recent IDC report warns that relief from the RAMpocalypse may not arrive until at least 2028. Put differently: even if new capacity starts rolling into the market in the next few years, ramping yields and increasing output is not an on-off switch. It is a ramp, often a long one. For memory makers, that is great news because revenues stay inflated. For AI model developers, it is a cost headwind that has to be managed until the supply situation genuinely improves.

And there is still a bust-cycle cliff edge. If anticipated demand for AI falls short, everyone loses, and memory vendors could end up at the bottom of a bust cycle. In other words, the same mechanism that keeps prices high during a boom also sets up a painful reversal if growth expectations wobble. The source’s framing captures it: the AIpocalypse is the next test of whether demand stays strong enough to absorb the new capacity coming online.

So executives should read this as a supply-chain timing story with financial consequences. Micron’s $3 billion US-focused investment and SK Hynix and Samsung’s $576 billion plan are trying to extend supply resilience. But the manufacturing clock is unforgiving, and the market’s demand clock is tied to how long AI funding and scaling can sustain higher costs per token. Memory is still a commodity. The boom is driving the bus. The only question now is how long the ride can stay exciting before the next drop hits.