Late in August, Canaan sold 3,952 ETH and 54 BTC for about $13.9 million. It did not push the money into wafers, tape-outs, or a foundry prepayment. It bought back 16.4 million of its own American depositary shares.
Read that as a capital allocation decision. A Bitcoin mining-equipment maker, three months after reporting a $97.6 million net loss, converted part of its crypto treasury into claims on itself. The blended repurchase cost works out to roughly $0.451 per ADS โ $7.4 million divided by 16.4 million shares. There is no softer way to state what that number means.
Then came the September 8 disclosure that framed everything. Second-quarter 2026 revenue: $31.9 million, against a May forecast of $35 million to $45 million. Product revenue: $13.6 million, down from $42.9 million in the first quarter. Third-quarter guidance: $11 million to $15 million.
And the detail most readers will skim past. At June 30, more than half of Canaan's Bitcoin was pledged. Exactly 1,117 of 1,915.5 BTC sat as collateral for secured term loans.
That last line is the article. The rest is context.
Context
Canaan designs ASICs and sells them under the Avalon brand. It also self-mines and holds a treasury, which makes it three businesses wearing one ticker: a fabless chip designer carrying foundry commitments, a hashrate operator carrying power contracts, and a small crypto fund.
The chip business is structurally short the thing it sells. An ASIC's price is a function of the expected mining revenue it produces, which is a function of BTC price, network difficulty, and the hashrate price curve. When difficulty rises and BTC does not, machine prices fall faster than component costs โ because foundry commitments were signed eighteen months earlier against a volume assumption that no longer holds. The same compression logic that says post-Dencun blob space saturates within two years says ASIC margins converge on the cost of power. Both are markets that attract capital until margin approaches zero.
Canaan's first quarter of 2026 already showed the crack: a revenue collapse reported alongside a headline about a record BTC and ETH treasury approaching $148 million. May guidance of $35 million to $45 million assumed a demand recovery. September delivered $31.9 million. The third-quarter range steps down roughly 60% at the midpoint.
There is a quieter fact buried in the disclosure chain. Canaan counted paused Ethiopia mining as nearly 35% of its July operating hashrate total. A third of the operating fleet, idle, in the quarter the company is guiding to $11 million to $15 million.
Core
Start with the balance sheet, because the balance sheet is where the narrative and the ledger disagree.
At June 30, Canaan reported 1,915.5 BTC. The headline reads like a nine-figure position. The composition reads like something else:
1,117 BTC pledged for secured term loans โ collateral, not spendable. 100 BTC transferred to a fixed-term product โ locked, not spendable. 698.5 BTC in the cryptocurrency assets category โ unencumbered.
Roughly 36% of the headline Bitcoin balance was actually free. The other 64% was either a lender's margin or a term deposit. Code talks, but stories sell โ and the story here was a treasury table.
Now price the same asset three ways from the same filing, because the filing lets you.
Crypto assets: $47 million divided by 698.5 BTC equals $67,286 per BTC. Crypto receivables, meaning pledged plus fixed-term: $70.9 million divided by 1,217 BTC equals $58,258 per BTC. Mining revenue: $17.7 million divided by 243 BTC produced equals $72,840 per BTC.
Three implied marks for one asset, on one balance sheet, at one date. The spread between the highest and the lowest is about 25%.
Part of that is legitimate. Mining revenue is recognized when coins are received across the quarter, not at quarter-end. Receivables may be net of haircuts or measured on a different basis. A pledged coin is not the same instrument as a spot coin. But the range tells you something the headline does not: there is no single number for Canaan's Bitcoin, so there is no single number for Canaan's liquidity. Narrative is the new liquidity โ until the lender asks for margin.
I have run this exercise on dozens of treasury tables over the years. The tell is never the total. It is the gap between the total and the portion a lender would let you touch.
Which brings us to cash, the line that moved the right way.
Canaan reported $66 million in cash at June 30, up from $43.5 million at March 31. That is a $22.5 million rebuild in a single quarter, in a quarter carrying a $97.6 million net loss. Nothing in that P&L explains the movement.
The pledge does. In the same window cash rose by $22.5 million, 1,117 BTC went up as collateral for secured term loans. If those loans were written at a conventional 50% loan-to-value against the receivables mark, borrowing capacity is roughly $35 million. At 65%, closer to $46 million. Timing is not proof. It is, however, the only mechanism in the disclosure capable of producing that cash movement.
The cash line rose because the Bitcoin line got encumbered. That is the trade. It is not fraud and it is not even unusual. It is a different story than the treasury table tells, and the two do not reconcile to the same number.
The mining segment did contribute. Canaan produced 243 BTC and booked $17.7 million from mining in the quarter. Management described those operations as cash-positive before depreciation โ a narrower measure than profitability, and the qualifier matters. Depreciation on ASICs is not a rounding error; it is the entire economic life of the machine. Cash-positive before depreciation is what a hardware company says when its assets are still on the book.
As for the $97.6 million net loss, it is mostly noncash. It included $25.3 million in inventory and prepayment write-downs and purchase-commitment provisions, plus $9.2 million in property and equipment impairment. Write-downs do not spend cash. But purchase commitments are the tail risk in this business model: obligations to buy wafers and components sized for a demand curve that no longer exists. When a fabless vendor writes down inventory in the same breath as it provisions for purchase commitments, it is admitting its order book was priced off a forecast, and the forecast was wrong.
Now the September disposition. 3,952 ETH and 54 BTC for approximately $13.9 million, realized after quarter-end, with $5.4 million of proceeds going into buybacks in late August. Back-solve the blend. If BTC traded near the $67,286 implied by the June crypto assets line, 54 BTC is about $3.63 million, leaving $10.27 million for the ETH โ roughly $2,600 per ETH. Using the higher $72,840 mark from mining revenue, the ETH comes out nearer $2,520. Using the $58,258 receivables mark, it climbs past $2,700.
Either way, the ETH side of the record-treasury narrative was sold at a relative price that looks like a cycle low against BTC. Canaan liquidated its diversification and kept its collateral. The narrative was the liquid asset all along.
Consider the alternative uses of $5.4 million inside a hardware cycle. A tape-out on a competitive node runs into eight figures. A foundry prepayment buys priority. A buyback buys nothing the company can sell. As capital deployment it is the inverse of Optimism's RetroPGF, the one retroactive funding mechanism that has measurably moved builders toward shipping โ capital released after the work, allocated by peer review rather than by a board. Canaan also pays after the work. The difference is that the work is done, the machines are depreciating, and the recipients hold a residual claim on a fleet that is worth less each month.
One final reconciliation. At June 30, the two crypto lines sum to $117.9 million. The first-quarter headline referenced a record treasury nearing $148 million. A gap of roughly $30 million opened inside two quarters โ part disposals, part marks, part reclassification. That is what record treasury means in a bull market: a mark-to-market assertion with a one-quarter half-life. Hype decays; utility endures. The utility here is the loan.
Contrarian
The consensus read is that Canaan is a distressed miner dumping coins to survive. I think the consensus is examining the wrong asset.
If the equity trades at a steep discount to unencumbered crypto per share โ and a blended buyback cost of $0.451 per ADS strongly suggests it does โ then buying stock with coins is accretion, not panic. It is the one capital allocation in that building a rational CFO would defend.
Here is the blind spot. Every analysis I have read this week argues about the Bitcoin count. Nobody is asking about the loan covenants on the 1,117 pledged coins. Those coins are marked by someone else's price feed, on someone else's schedule, against someone else's margin threshold. Oracle latency is the Achilles' heel of every collateralized crypto market โ and at least in DeFi that feed is auditable, however imperfectly permissioned its node set. In a secured term facility with a single counterparty, the mark on $70.9 million of recorded receivables is a private number that surfaces in a filing four months after it was set.
Here is the second blind spot. The buyback is funded by the liquid half of the treasury. The liquidation risk is carried by the encumbered half. Those are not the same shareholder. A treasury program that returns capital to equity while pledging the reserve that backstops the balance sheet is a duration mismatch wearing a shareholder-friendly label.
Takeaway
Watch three numbers into the fourth quarter, none of them the Bitcoin total. The first is unencumbered BTC, which stood at 698.5 at June 30 and remains the only coin the company can spend or sell without asking permission. The second is whether Q3 revenue of $11 million to $15 million is mining-only. If a third of the fleet sits paused in Ethiopia and mining runs near $12 million a quarter, then hardware revenue is being guided toward zero and the guidance is a mining report wearing a revenue label. The third is the blended ADS repurchase price against crypto net asset value per share โ that ratio is the entire investment case now, and it moves every time the collateral mark does.
When a treasury carries three prices, which one does the lender use?