Hook
Last week, JPMorgan released a report that should have sent shivers through every crypto miner's spine. The bank's semiconductor analysts predict that by 2028, server CPU shipments will soar from 26 million to 68 million units annually — with over 80% of those chips powering AI inference workloads. That's a staggering 160% increase in server demand, all driven by what they call 'Agentic AI.' But here's the ghost in the code: the same report also forecasts that memory prices — especially DRAM and HBM — will continue their relentless climb, suppressing PC demand by 8% year-over-year in 2026. For the crypto mining world, this isn't just a tech cycle shift. It's a structural chokehold on the very hardware that keeps the network hashing.
I've been tracing this narrative since the first whispers of AI inference eating GPU supply back in 2023. Now, with JPMorgan's hard numbers, the story is undeniable. The AI gold rush is starving crypto miners, and memory prices are delivering the final blow.
Context
To understand the magnitude, we need to rewind to the 2021 bull run. Crypto miners were the darlings of the semiconductor supply chain. NVIDIA's CMP (Cryptocurrency Mining Processor) cards were a thing, and gaming GPUs were sold out everywhere due to mining demand. Fast forward to 2025 — the narrative has flipped. AI training and inference now command the lion's share of advanced GPU production. NVIDIA's H100, B100, and the upcoming G-series are designed for data centers, not mining rigs. The latest JPMorgan analysis dissects this shift with surgical precision: the server market is entering a 'super cycle' while PC demand — and by extension, consumer-grade GPUs — faces a headwind from memory price hikes.
The report, authored by JPMorgan's semiconductor team, highlights two parallel forces: the AI inference explosion (powered by large language models and autonomous agents) and the memory price inflation (driven by HBM3E demand for AI and limited DRAM capacity expansion). For crypto miners, this means two things: 1) fewer GPUs allocated to the consumer/enthusiast market, and 2) higher costs for the memory components used in mining rigs. It's a double squeeze that the market hasn't fully priced in.
Core
Let's dig into the numbers. JPMorgan's forecast for server CPU shipments is not just a linear growth line — it's a hockey stick.
- 2025 baseline: 26 million server CPUs shipped.
- 2028 target: 68 million server CPUs shipped.
- Of those 68 million, 53 million will be used for AI inference (Agentic AI workloads).
That implies AI inference alone will consume more than 80% of all server silicon by 2028. Compare that to today, where AI inference consumes maybe 30-40% of server chips. The ramp is absurd.
Now, overlay the memory picture. JPMorgan notes that memory prices have already risen 30-50% year-over-year, driven by HBM3E shortages and DDR5 demand from AI servers. They expect this trend to continue through 2026, with DRAM prices potentially doubling from their 2024 lows. For crypto miners, this is a direct hit. Mining rigs — especially those using GPUs — rely on GDDR6 or GDDR7 memory. As HBM eats up fab capacity and pushes up DDR5 prices, GDDR follows suit. The cost of building a new mining rig today is roughly 20-30% higher than it was 18 months ago, purely due to memory components.
But the more insidious effect is on GPU availability. NVIDIA and AMD are allocating almost all of their advanced 4nm and 3nm wafer starts to AI inference chips. The result? Consumer-grade GPUs (like the RTX 5090) are not only scarce but also priced at a premium. Crypto miners used to be able to snap up last-generation cards at a discount when new gens launched. That strategy is dead. The same advanced packaging (CoWoS) that makes H100 possible also limits the supply of gaming dies.
Let me put a forensic lens on the supply chain. JPMorgan identifies 'supply bottlenecks' in CPUs, motherboards, memory, PCBs, and power components. What does that mean for a mining farm? A typical mining rig uses a motherboard with multiple PCIe slots, a CPU (usually low-power), and a power supply. If these components are in short supply because AI servers are hoarding them, then the cost and lead time for mining rigs explode. I've seen reports of lead times for high-end server power supplies stretching to 16-20 weeks. Mining farms can't afford to wait that long.
Contrarian Angle
The mainstream narrative is that AI inference is a magic bullet for all tech — NVIDIA stock is up, AMD is finally a viable second source, and memory makers are printing money. But the contrarian view is that this cycle is creating a massive hidden tax on non-AI hardware consumers. The PC market is shrinking by 8% in 2026, according to JPMorgan, because memory inflation makes new laptops too expensive. That's a canary in the coal mine for crypto miners: if casual PC buyers are priced out, mining rig builders will be even more squeezed.
More importantly, the AI inference narrative itself might be overblown. 'Agentic AI' — autonomous agents that act on behalf of users — is still a concept in search of a killer app. The JPMorgan report assumes a smooth adoption curve, but history shows that enterprise AI deployment often hits cultural and integration barriers. If the AI inference boom falters, the server chip forecast will collapse, and memory prices could crater. That would flood the market with cheap GPUs — good for miners but bad for the narrative that AI is unstoppable.
But here's the real blind spot the report doesn't address: the electrification bottleneck. AI inference servers are power-hungry beasts. Each inference node can draw 2-3 kW, and a rack of them can consume 40-50 kW. The US grid is already strained by data center buildouts. If power constraints cap AI server deployment, then the memory price hike is just a phantom — demand won't materialize. JPMorgan's model assumes infinite power, which is naive.
Takeaway
For the crypto mining community, the message is clear: the days of cheap, abundant GPUs are over — at least for the next two years. The AI inference supercycle is sucking up the supply chain, and memory prices are making it worse. Miners need to rethink their hardware strategy: look at ASICs for specific algorithms, consider undervolting existing rigs to extend life, or even pivot to staking and DeFi plays that don't require physical hardware. The next bull run in crypto may not have the hash rate growth we saw in 2021.
I hunt the story that the chart hides. And this chart hides a quiet war over silicon — a war that AI inference is winning, and crypto mining is losing. The question isn't if memory prices will suppress PC demand; it's whether crypto miners can survive the collateral damage.
