Bitcoin's 22% rally paid the median miner 1.8%

The two operators furthest into high-performance computing came out of bitcoin's run below where they went into it: lag a 22 percent coin move by 27 points and you are down roughly 5 percent, lag it by 24 and you are down about 2 ^[1][5][1]. The returns sit close to flat in both directions, not leveraged either way.

What makes it a puzzle rather than a verdict on mining is that the mine kept up. Luxor's August figures have revenue per petahash rising 24.4 percent against a 24.5 percent move in the coin ^[8], a pass-through of almost exactly one ^[2]. Unit economics followed the price while the equity did not, and the gap sits in the capital structure rather than in the margin per hash: Compass Point's Michael Donovan frames the AI build-out as a financing problem, one that needs money which neither dilutes heavily nor costs too much ^[7]. The market looks like it is marking the financing.

Underneath sits the question of what the subsidy is worth. With 99.3 percent of the block reward coming from issuance rather than transaction fees ^[4], a rising coin lifts revenue per hash and does nothing else; the levers that would widen margin are network difficulty and the power bill ^[13]. Luxor's own forward curve is not offering either, averaging about 6 percent below where hashprice closed August ^[3].

Coinbase's own disclosure leaves 12 percent of second-quarter net revenue coming from bitcoin spot trading ^[5], alongside a record 10.3 percent of trading volume and $20 billion of average USDC held across its products ^[15], and Brian Armstrong's summary is that "Coinbase is no longer a bet just on the price of Bitcoin" ^[16]. Circle ended the quarter with $73.3 billion of USDC outstanding, up 19 percent on the year ^[17], against $701 million of revenue and reserve income ^[18]. Annualise that quarter against the float and it is about 3.8 percent ^[6], a figure that says as much about short-term rates as about adoption, and the disclosure as reported does not split reserve income out from the rest. The evidence shows the fee and float businesses matched the coin ^[3] rather than capital bidding them above it.

The HPC revenue evidence points another way: if HPC is around 70 percent of 2026 revenue at IREN, Core Scientific and TeraWulf ^[6], benchmarking them against bitcoin is the wrong test, and CoinShares' tally of more than $70 billion of AI and HPC contracts across public miners ^[19] is a re-rating waiting on delivery. MARA's $1.5 billion Long Ridge acquisition and IREN's $3.4 billion Nvidia cloud contract together are about 7 percent of that book ^[20][7], so most of it is still promise. July's tape favours the financing read: Cipher fell 8 percent, Riot 5 and MARA 3 while Coinbase, BitGo and Figure rose ^[9].

The read breaks if the fee share turns out to be a cycle low rather than a level, or if one of these operators funds a build-out without touching equity, either of which would put the discount back on the mine where the old model expects it. Until one of those happens, an allocator buying bitcoin exposure through a miner is paying for 8.2 cents of every dollar the coin moved ^[4], plus a construction schedule.