Hook
Micron just dropped its Q3 2024 earnings. Revenue hit $8.1 billion—up 82% year-over-year. The beat sent shares soaring 4% after hours. But buried inside the fine print is a signal that should chill every Bitcoin miner and GPU farmer reading this: High-Bandwidth Memory (HBM) sales alone accounted for a double-digit percentage of total revenue, and the entire HBM backlog is sold out through 2025.
That’s not a flex. It’s a resource war declaration. Every Terabyte of HBM3E that goes into an NVIDIA H100 or a Blackwell B200 is a Terabyte that cannot go into a mining rig. And with HBM supply struggling to keep pace with AI demand, the spillover effect is already squeezing the lower-margin crypto mining hardware market.
Context
For those who don’t live inside silicon politics: HBM is the ultra-fast, vertically-stacked DRAM that makes AI accelerators work. Without it, a $30,000 H100 is a paperweight. Micron, Samsung, and SK Hynix control nearly 100% of HBM production. And right now, every fabs’ entire HBM capacity is locked into contracts by hyperscalers—Microsoft, Amazon, Google—for their internal AI clusters.
Meanwhile, cryptocurrency mining—especially for Bitcoin (ASICs) and Ethereum Classic (GPUs)—runs on standard GDDR6 memory and cheap DRAM. But as HBM consumes more wafer starts and advanced packaging lines, the capacity for producing GDDR6 shrinks. Prices rise. Lead times stretch. Miners who were already bleeding on thin margins get squeezed further.
This isn’t speculation. I’ve been auditing hardware supply chains since 2017, back when I caught that Mumbai DEX’s integer overflow vulnerability. The same math applies: supply is finite, demand is exponential, and whoever pays the highest unit economics wins the allocation.
Core: The Data Behind the Squeeze
Let’s break down the numbers feeding this narrative—because narrative without numbers is just hype.
First, Micron’s data center revenue—which includes HBM and server DRAM—grew over 400% year-over-year to $4.4 billion. That’s more than half their total revenue. Their “compute and networking” (which includes PC and mining-related memory) grew only 15%. Translation: AI is cannibalizing the rest.
Second, NVIDIA’s data center revenue hit $47.5 billion last fiscal year. Every one of those H100 and B200 GPUs needs at least 80GB to 192GB of HBM3E. In 2024, the total HBM market is estimated at ~20 billion GB-equivalent units. AI alone consumes 70-80% of that. Miners? They’re fighting for the leftover scraps of GDDR6X and older HBM2E.
Third, track the Bitcoin network hashrate. It’s still rising—hitting 600 EH/s recently—but the growth rate is decelerating. In 2023, we saw 60-70% year-over-year growth. In 2024? Closer to 20%. Miners are deploying more efficient S21 Antminers, but the rate of new machine deployment is slowing. Why? Chip shortages and higher prices. Bitmain’s S21 Pro costs $4,000+ per unit, up from $2,500 for the S19 a year ago. The margin squeeze is real.
I don’t just read these numbers; I rode them. In 2020, during the DeFi yield farming frenzy, I deployed $50k into Compound and watched the gas fees eat my profits. The lesson was visceral: capacity constraints don’t care about your farming strategy. The same dynamic applies here. The only yield that matters is the one that survives after hardware costs and depreciation. Right now, that’s negative for a lot of miners.
The Empirical Test
Go to eBay. Search for “RTX 3090” or “RTX 3080.” Six months ago, a used 3090 fetched $1,000-$1,200. Today? $700-$900. That’s a 30% drop. It’s not just Ethereum’s switch to Proof-of-Stake—that was two years ago. No, the real cause is the secondary market dump from miners who can’t make the economics work on altcoins. They’re selling their rigs, flooding supply.
Meanwhile, I’ve been tracking the price of H100 GPU cloud instances. On AWS, a single p4d.24xlarge instance costs over $32 per hour. That’s $230,000 a year per instance. Compare that to renting a mining rig: you’re lucky to get $0.10 per hour per GPU. The market is telling you exactly where capital should flow. AI is the 800-pound gorilla; crypto mining is the chihuahua.
Contrarian: The Pragmatist Test
But here’s where the story gets murky. The “AI vs Mining” narrative is too clean. It ignores a few messy realities:
- Miners adapt. Bit Digital, once a pure Bitcoin miner, now generates 20% of its revenue from AI cloud services, renting out its GPUs to AI startups. Hut 8 just signed a multi-year deal to provide colocation for AI compute. These aren’t victims—they’re pivoting.
- Old hardware trickles down. The H100 and B200 may go to hyperscalers, but the A100 and V100? Those get sold to smaller players—including altcoin miners. As NVIDIA releases new chips, older generation GPUs become cheap enough for mining and for AI inference at the edge. This could actually lower the barrier to entry for mining certain coins like Kaspa or Flux.
- Distributed GPU networks might win. Projects like Render Network and Akash are building marketplaces where anyone can rent out their GPU cycles—for AI or for rendering or for compute. If AI demand floods these networks, the token price pumps, attracting more suppliers. This creates a virtuous cycle that could partially offset mining revenue losses.
I’m not a trend predictor; I ride the volatility. But I’ve seen this movie before. In 2021, everyone thought NFTs were dead. In 2022, everyone said DeFi was finished. The infrastructure survived. Infrastructure is permanent. The specific use case—mining, AI, whatever—shifts, but the hardware stays. And the market always finds a new owner for those GPUs.
Takeaway
So where does this leave us? The Micron earnings are telling us one thing: AI is consuming the world’s advanced memory supply, and crypto miners are getting starved. If you’re a miner sitting on an S19 or a rig of 30-series cards, the clock is ticking. Your depreciation curve is steepening.
But for the decentralized compute networks—the ones that aggregate idle GPUs and sell them to AI startups—this is a golden opportunity. The narrative war between “AI vs Crypto” is a false binary. The real story is a reallocation of hardware capital, and those who pivot fastest will survive.
I don’t predict trends; I ride the volatility. The protocol is neutral; the user is the variable. Right now, the variable is telling you to bet on adaptable infrastructure, not rigid mining setups. Align your hashrate with the demand curve, or get left holding a bag of depreciating silicon.
— Matthew
