TSMC just reported a record $40.2 billion in Q2 2025 revenue. The market cheered. AI demand is exploding. Every hedge fund is chasing NVIDIA.
But look closer. That revenue surge is driven by HPC—High-Performance Computing for AI inference and training. The same advanced fabrication lines that produce NVIDIA’s H100 and B200 also produce the latest Bitcoin ASICs. And when capacity tightens, who gets priority? The client paying $30,000 per chip, not the one paying $3,000.
This is not a cyclical blip. It is a structural reallocation of the world’s most advanced manufacturing capacity. Proof-of-work miners are about to face their most uncomfortable reality: the era of cheap, abundant mining ASICs is ending.
Context: The Forgotten Supply Chain
For a decade, crypto mining has ridden the coattails of semiconductor innovation. Bitmain, MicroBT, Canaan—all depend on TSMC’s 7nm, 5nm, and 3nm nodes. The relationship was symbiotic: miners bought trailing-edge capacity that smartphone and PC makers no longer needed. But AI flipped that dynamic. Now, the leading edge is reserved for AI accelerators. TSMC’s own earnings breakdown shows that HPC now accounts for over 60% of revenue, up from 35% two years ago. The “crypto and other” category? Shrinking to near zero.
We didn’t account for AI’s appetite for fabrication capacity. The market assumed that Moore’s Law would continue to deliver cheaper, faster chips for everyone. That assumption is now broken. TSMC raised its capital expenditure forecast again, but that doesn’t change the fact that new fab lines take 3-5 years to come online. In the meantime, every wafer allocated to an AI chip is a wafer denied to a mining ASIC.
Core: The Mechanics of the Squeeze
Let’s trace the transmission mechanism. First, wafer cost. A 5nm wafer from TSMC costs roughly $17,000. An AI chip like NVIDIA’s H100 yields about 80 dies per wafer, each selling for $30,000. That’s a gross margin of ~80%. A Bitcoin ASIC yields maybe 200 dies per wafer, each die selling for $2,000 at the miner level. Total wafer revenue for the foundry: <$400,000 for AI, ~$400,000 for mining? Wait—no. The ASIC die is far smaller, so revenue per wafer is lower: roughly $200,000. But the catch is that TSMC prioritizes high-margin, high-volume clients. When AI orders surge, they get the first allocation. Mining ASICs get the leftovers.
Second, cost escalation. Even if mining ASIC manufacturers secure wafers, the foundry can raise prices. TSMC already announced a 10-15% price hike for 3nm and 5nm nodes in 2025. That flows directly into miner hardware cost. A next-generation Bitcoin miner (e.g., Bitmain S21 Pro) could see its price jump 20% in 2026. That pushes the payback period from 12 months to 18 months or more, assuming stable Bitcoin prices.
Third, delivery delays. In Q1 2025, MicroBT publicly warned of extended lead times for its M60 series due to TSMC capacity constraints. This is not anecdotal. It’s the new normal. Miners who pre-order machines may face 6-9 month delivery windows. That kills the advantage of deploying before the next halving.
The market doesn’t price in the structural shift from mining to inference. When you look at the hashrate growth curve, it’s still climbing. But the slope is flattening. In 2024, Bitcoin hashrate grew 45%. In 2025, it’s on track for 25%. The chip squeeze is the primary cause. New miners are harder to source; old miners are being kept online longer. That raises the floor for network security but also compresses miner margins.
Contrarian: The Adaptation Myth
“Miners always adapt.” That’s the standard narrative. They can move to cheaper energy, use immersion cooling, or even switch to AI compute themselves. But these adaptations have limits. Cheaper energy doesn’t solve the hardware shortage. Immersion cooling adds to upfront capex. And pivoting to AI compute requires buying NVIDIA GPUs, not ASICs—a completely different supply chain. The typical miner doesn’t have the expertise or capital to become an AI cloud provider.
The blind spot is the assumption that Moore’s Law applies equally to all ASICs. It doesn’t. Mining ASICs are designed for a single algorithm. They cannot be repurposed. When the chip supply tightens, the entire PoW ecosystem suffers because the hardware is non-fungible. Contrast this with Proof-of-Stake, where validators only need commodity hardware. The chip squeeze makes PoW structurally more vulnerable than its advocates acknowledge.
Some will argue that the market will clear: higher miner prices will reduce demand, and TSMC will reallocate capacity. That’s true in a textbook sense. But the adjustment is painful. Small miners with thin margins get wiped out. Centralization increases as only the largest players can secure forward capacity contracts. The argument that “Bitcoin mining is becoming a ‘real’ industry with mature supply chains” actually exposes a weakness: mature supply chains mean fixed costs and less flexibility.
Takeaway: The Next Narrative
The chip squeeze is not a one-time event. It’s a permanent re-pricing of computational resources. For the next 12-24 months, PoW miners will face rising hardware costs, longer lead times, and thinner margins. The market has not yet priced this into mining stocks or token prices. When it does, the re-rating could be severe.
Miners should rethink their capital allocation. Instead of chasing the latest 3nm machine, consider extending the life of existing fleets with efficiency upgrades. Or explore hybrid models that hedge with AI compute capacity. But don’t assume the supply chain will bend in your favor. AI ate the world—and now it’s eating the wafer supply.
We didn’t account for AI’s appetite for fabrication capacity. The next bear market in mining may not be caused by Bitcoin’s price. It may be caused by the simple fact that there aren’t enough chips to go around.