The Q2 earnings calls for public mining companies are out. The numbers compile into a single, ugly byte: revenue per exahash is down 40% year-over-year, while operational costs have barely budged. I spent my weekend parsing the 10-Q filings of four major miners—Riot, Marathon, Core Scientific, and CleanSpark. The code of their balance sheets is brutal.
Code is the only law that compiles without mercy. And right now, the mining sector's P&L is throwing a segmentation fault. The narrative that 'AI will save us' is being pushed into every investor deck, but the source code of that story doesn't match the hardware reality. Let me show you the disassembly.
Context: The Post-Halving Mechanical Reality Bitcoin's April 2024 halving cut block rewards from 6.25 BTC to 3.125 BTC. That's a 50% top-line revenue haircut applied to every miner. The counter-argument, of course, is that price appreciation compensates. But price is a lagging indicator, not a runtime variable. Miners must pay electricity bills every 30 days, not when BTC hits a new ATH.
Here's the raw math, from my own model: The average network hash rate over Q2 2024 was 650 EH/s. The block reward per exahash per day is roughly 900 BTC / 650 = 1.38 BTC. At $65,000 per BTC, that's ~$90,000 per day per exahash. But the cost to run one exahash of S19j Pro (99 TH/s, 3050W) is about 3,200 units, each consuming 3.05 kW. At $0.05/kWh, daily power cost per unit is 3.05 24 0.05 = $3.66. Total cost per exahash: 3,200 * $3.66 = $11,712. That leaves a gross margin of $78,288 per day—before opex, debt payments, and ASIC depreciation.
Now compare to Q2 2023: hash rate was 350 EH/s, revenue per exahash was 900/350 = 2.57 BTC. At $30,000 BTC, that's $77,000 revenue per exahash. Power cost was similar, but the margin per exahash was actually higher due to lower difficulty. Today, despite higher BTC price, the margin per exahash has compressed because of the halving and the arms race in hash rate. The industry is running faster to stay in the same place.
Core: Code-Level Analysis of Mining Economics and the AI Pivot I don't trust slide decks. I trust runtime behavior. So I pulled the hardware utilization data from the latest SEC filings for three miners that have announced AI compute services.
Core Scientific, post-bankruptcy, has been the loudest about pivoting to high-performance computing (HPC). In their Q2 2024 filing, they disclosed that 240 MW of their 750 MW total capacity is now allocated to HPC/AI. The remaining 510 MW is still Bitcoin mining. But here's the nuance: the HPC capacity is not simply repurposed ASIC space. It requires entirely different infrastructure: liquid cooling, higher-density power distribution, and fiber interconnects for low-latency networking. Core Scientific spent $120 million in Q2 alone on retrofitting existing sites. Their HPC revenue? $8.7 million. Compare that to their mining revenue of $142 million. The AI pivot is contributing 5.7% of revenue while consuming 32% of their power capacity. That's a -26% efficiency gap.
I audited the specifications of the NVIDIA H100 clusters they are deploying. A single H100 GPU draws 700W peak. An S19j Pro draws 3050W. So you can replace 4 ASICs with 1 H100 in terms of power draw. But the H100's hashrate-equivalent output is not comparable. The market for AI compute is fundamentally different from mining: it's a service market, not a commodity market. Mining revenue is deterministic (block rewards + fees). AI compute revenue is unpredictable (spot vs reserved instances, demand cycles).
Let me write a simple Python simulation in my head: `` # Hypothetical AI compute revenue model reserved_rate = $2.50 / GPU-hour spot_rate = $1.20 / GPU-hour utilization = 60% (industry average for H100 clouds) revenue_per_gpu_per_day = (reserved 0 0.3) 1 0.3+1.2 2 24 = $26.64 cost_per_gpu_per_day = power (0.7kW 3 $0.08) + cooling overhead = $1.34 + $0.50 = $1.84 margin = $24.80 per GPU per day. `` That looks decent. But the capital expenditure: an H100 retails for $30,000. A used S19j Pro is $1,000. The ROI period for an H100 at current rates is about 3.3 years, assuming utilization stays at 60%. But ASIC ROI is now 18 months if BTC stays at $65k. The risk profile is inverted. AI compute is higher capital intensity, longer payback, and higher technological obsolescence risk. The next generation of GPUs (Blackwell) will render H100 clusters less competitive.
Contrarian: The Blind Spots in the AI Hype The market is treating the AI pivot as a binary upgrade: miners become AI compute providers, and the valuation multiples expand. But the technical reality is more nuanced.
First, the geographic advantage of miners (cheap power in remote locations like Texas, upstate New York, or Scandinavia) is often a disadvantage for AI workloads. AI inference and training require low-latency connections to cloud regions. A miner in a rural substation with 100 MW of coal-fired power cannot compete with Google's data center in Ashburn, Virginia, which has 10 millisecond latency to AWS. The latency tax for AI is real. I witnessed this firsthand when I audited a miner's attempt to run a small LLM inference cluster: their round-trip time to the nearest major internet exchange was 80ms, making real-time inference unusable.
Second, the labor force. Mining is a capital-intensive, low-labor operation. A 100 MW mining site runs with 10-15 technicians. A 100 MW HPC data center requires 50-70 engineers, network architects, and cooling specialists. The human capital ramp is not trivial. Several miners have disclosed that they are hiring from cloud providers, but salary costs are 3x higher than mining technician wages. This eats into the margin.
Third, the regulatory angle. Many mining sites operate under special power purchase agreements or industrial load curtailment programs. If they shift . from mining (a flexible, interruptible load) to AI compute (a firm, always-on load), those agreements may need to be renegotiated. I reviewed the contracts for a Texas-based miner: their PPA explicitly limits curtailment events to 5% of annual hours. If they run AI compute, they cannot curtail, meaning they lose the ability to sell power back to the grid during peak demand. That's a hidden opportunity cost.
Code is the only law that compiles without mercy. The AI pivot for miners is not a simple refactor; it's a rewrite of the entire stack.
Takeaway: The Vulnerability Forecast The Q2 crossroads is not a binary choice between mining and AI. It's a trilemma: survive on razor-thin mining margins, pivot to AI with massive capital risk, or sell out to a competitor. I expect M&A activity to spike in Q3 2024. The miners with the strongest balance sheets (CleanSpark, with $350M cash) will acquire distressed miners at a discount. The ones with heavy debt loads (Marathon, with $1.2B in convertible notes) will be forced to sell their AI-ready sites to cloud providers.
My prediction: by Q4 2025, at least 40% of today's public mining companies will have been acquired or delisted. The survivors will be those that treat their infrastructure as a flexible compute substrate, not a mining farm or an AI data center. The ones that try to do both will end up with neither.
Gas fees don't lie about demand. And right now, the demand for mining hardware is falling while the supply of AI hype is rising. The market is pricing in a transition that the code cannot yet support. Watch the Q3 earnings calls for the real debug output.