Memory makers are slaves to the boom-bust rollercoaster, and the AI boom is the wildest ride of all

The memory industry has always lived on a knife edge. DRAM and NAND makers — Samsung, SK Hynix, Micron, Kioxia, Western Digital — have spent decades riding cycles that turn boolean in months: glut turns to shortage, pricing collapses then triples, capacity comes online just as demand evaporates. The current AI supercycle is not a new ride. It is the same rollercoaster, just taller, faster, and with fewer safety bars.

The cycle never changes, only the amplitude

Every boom starts with a narrative. In 2017 it was smartphones and data centers. In 2020 it was work-from-home laptops and cloud build-out. Today it is large language model training and inference clusters. The narrative pulls in capital. Capital buys wafer starts. Wafer starts take 18 to 24 months to become bits. By the time bits hit the market, the narrative has often moved on. The result is inventory overhang that crushes margins for quarters.

This time the amplitude is historic. Nvidia's H100 and B200 GPUs consume 80 to 192 GB of HBM3E per board. Each HBM stack is eight to twelve DRAM die thinned, bumped, and stacked through-silicon vias. That is not commodity DRAM. That is a custom semiconductor product with yield risk, test complexity, and a single dominant customer. When one customer represents 40 percent of your high-margin output, you are not a supplier. You are a captive foundry.

HBM is the new choke point

High-bandwidth memory has replaced the CPU as the gating factor for AI training scale. TSMC's CoWoS packaging capacity is the real bottleneck, not wafer starts. Samsung and SK Hynix have both guided that HBM bits will grow 60 percent year-over-year in 2025, yet they cannot promise volume because advanced packaging slots are allocated quarters in advance. Micron entered HBM late and is still qualifying on Nvidia's test benches. The result: a seller's market where lead times stretch to six months and contract prices reset quarterly.

For CRM and RevOps platforms, this matters directly. Every AI feature — predictive forecasting, conversation intelligence, automated data enrichment — runs on GPU clusters that sit behind HBM supply. When HBM tightens, cloud instance prices rise. When cloud instance prices rise, the unit economics of AI features flip. A vendor that built its differentiation on "AI included" suddenly faces a cost structure that looks like 2018 GPU mining all over again.

The small-model pivot is a supply-chain signal

The industry chatter about small language models, distillation, and on-device inference is not purely architectural. It is a hedging maneuver. If you can serve 80 percent of the use case with a 7-billion-parameter model that runs on a CPU with commodity DDR5, you have just removed your dependency on the HBM allocation queue. OpenAI, Anthropic, and the hyperscalers will keep buying HBM by the container ship. The rest of the B2B software ecosystem — CRM, CPQ, sales engagement, marketing automation — will optimize for inference on general-purpose silicon because the economics of dedicated AI accelerators are becoming unknowable.

Microsoft's emissions jump of 25 percent in one year, driven by AI datacenter build-out, is the macro version of the same signal. The hyperscalers are capital-intensive in a way that makes memory makers' capex look like rounding error. But the memory makers are the ones who must commit silicon starts 24 months out based on hyperscaler forecasts that change quarterly. That is the definition of a captive supply chain.

What the next bust looks like

The bust will not arrive because AI demand disappears. It will arrive because the hyperscalers over-order HBM slots in 2024 for 2026 delivery, then realize in late 2025 that inference efficiency gains — quantization, mixture-of-experts routing, speculative decoding — have cut per-token memory bandwidth by 40 percent. The 2026 HBM capacity comes online. The hyperscalers cancel or defer. Spot prices collapse. Memory makers take impairment charges on advanced packaging equipment that has no alternative use.

We saw a preview in late 2022 and 2023 when PC and smartphone demand evaporated. DRAM spot prices fell 50 percent in six months. Samsung and SK Hynix cut wafer starts. Micron guided losses. The difference this cycle: the high-margin HBM layer did not exist last time. The next bust will strand specialized capital that cannot be repurposed for commodity DRAM or NAND.

Strategic posture for software buyers

If you are a RevOps leader evaluating AI-enabled CRM modules, ask three questions. First: what is the inference hardware dependency — dedicated GPU with HBM, shared GPU with commodity VRAM, or CPU with system DRAM? Second: what is the vendor's pricing model if cloud GPU instance rates rise 30 percent year-over-year? Third: does the vendor have a fallback path to smaller distilled models that run on general-purpose silicon?

The memory makers will survive. They always do. They have balance sheets built for cycles, long-term supply agreements with hyperscalers, and the discipline to cut capex when visibility evaporates. The AI boom is the wildest ride because the amplitude is largest and the choke point — advanced packaging — is hardest to expand. The software layer sits at the end of that whip. Understand the supply chain or accept the price volatility. There is no third option.