Hook
On September 24, a filing lodged with the United States Bankruptcy Court for the District of New Jersey disclosed a number that should have stopped every miner mid-scroll. Hut 8 โ the Toronto-listed, Nasdaq-traded compute company โ had submitted a bid of $140 million for two West Texas data center campuses, Pyote and Tarbush, carved out of the estate of the collapsed mining firm Poolin.
The headline figure is not the interesting one. The interesting one is the $52 million that preceded it. Thor CALAP had entered the auction as the stalking horse, the floor bid that establishes where bidding begins and against which every later offer must be measured. Hut 8's number is roughly 2.7 times that floor. When a floor gets tripled inside a Chapter 11 proceeding, you are not watching a liquidation. You are watching a discovery โ the moment a market admits it has been mispricing an asset class for two years.
Silence in the code speaks louder than the hype. The loudest silence in this filing belongs to Poolin's wallet users: roughly $163.7 million of unsecured claims held by ordinary depositors, people who parked coins in a yield product and who are now, involuntarily, junior creditors against a power purchase agreement in the Permian Basin. They will not be made whole. The only open question is the size of the shortfall.
A sale hearing is calendared for September 29. Every unresolved variable โ the true recovery rate, the probability of a losing-bidder objection, whether the interconnection rights transfer cleanly to a new owner โ compresses into that single date.
So let's do the arithmetic the market has been avoiding.
Context: What Poolin Was Before It Was an Estate
Poolin did not start as a cautionary tale. It started as infrastructure. At its peak the pool commanded a double-digit share of global bitcoin hashrate, and for a stretch of 2019 and 2020 it was, by some measures, the largest mining pool on earth. Hashrate is a commodity business, though, and commodity businesses valued on growth are structurally tempted to grow into adjacent businesses. Poolin grew into hosting. Then it grew into something else: Poolin Wallet, a deposit product that paid users a yield for parking idle coins with the platform.
The architecture of that expansion is worth staring at, because it is not unusual and it is not subtle. Pool revenue funded hosting buildout. Hosting customer deposits and wallet liabilities funded the next round of buildout. The whole thing was one repayment schedule away from coherence, and repayment schedules are exactly what break in a drawdown. When withdrawals froze in 2022, the freeze was not a policy decision in the ordinary sense. It was an admission that the asset-liability match had never existed.
By July of this year, Poolin was in Chapter 11. The estate's liabilities total roughly $173.1 million. Of that, approximately $163.7 million โ about 94.6 percent โ is owed to Wallet users as unsecured notes. The remaining ~$9.4 million sits in other claims. Read that split again. This is not a creditor book dominated by hedge funds who priced the risk and bought the spread. It is overwhelmingly a consumer book. That composition is the single most important fact in the entire filing, and it is the one the auction narrative has been quietly stepping around.
Then there are the assets: Pyote and Tarbush. Two sites in West Texas. Not code. Not patents. Dirt, steel, transformers, switchgear, and โ the only genuinely scarce item in the entire stack โ interconnection rights.
Here it is worth being precise, because "data center" has become a phrase that hides more than it reveals. In ERCOT, the grid operator for roughly 90 percent of Texas load, a large load cannot simply announce that it intends to consume. It must secure an interconnection agreement, and it must survive a queue. That queue has congested as load growth โ data centers, industrial electrification, electrolysis, crypto โ collided with a transmission system that was never planned for this. An existing, energized, already-interconnected site is therefore not a building. It is a permit with a foundation.
Interconnection scarcity is not theoretical, and I have measured it. In 2024, following the spot ETF approvals, I spent two months building a dashboard that tracked institutional capital moving from brokerage accounts into self-custody addresses. The finding that stuck with me was not the volume โ it was the routing. Large inflows from specific entities were being moved almost immediately into cold storage, a pattern that said holding rather than trading. The same routing logic applies to watts. Capital does not buy power assets to trade them. It buys them to sit on them, because the queue behind you is the moat.
Poolin's Texas sites had already stopped mining and stopped hosting before the auction. That detail matters more than it appears. The assets are dormant but physically intact, which means the buyer inherits an engineering problem rather than a construction project. Restart risk is real; greenfield risk is larger. Hut 8 is buying the middle of that distribution.
Core: The Auction Was a Pricing Event, Not a Liquidation
Now the numbers, because the numbers are where this story stops being a bankruptcy and starts being a repricing.
Four bidders converged on the same dirt. Thor CALAP anchored at $52 million. DigiPower X bid $36.5 million for Pyote alone. Pecos, working with the AI infrastructure firm Fluidstack, bid $100.5 million for Tarbush alone. Hut 8 bid $140 million for both.
Stop on the third bidder, because that is the anomaly. A single site โ Tarbush โ drew an AI-linked offer of $100.5 million, which is roughly 72 percent of what Hut 8 eventually paid for both campuses combined. There is no way to read that as a mining valuation. Bitcoin mining, post-halving, is a business whose gross margin is being compressed from two directions at once: block subsidies were cut in April, and network difficulty has kept climbing even as price has not. A rational miner valuing Tarbush on ASIC cash flows would land somewhere near the stalking horse, not near nine figures.
The bid only makes sense if the buyer is valuing watts, not hashes. Unraveling the thread that binds value to vision is the whole job here, and the thread runs straight through power density.
A bitcoin ASIC is a relatively forgiving load. It runs at moderate density, tolerates wide thermal envelopes, and can be curtailed aggressively without destroying the asset. A GPU cluster training a frontier model is the opposite: high density per rack, tight thermal tolerances, heavy dependence on liquid cooling in modern deployments, and brutal sensitivity to interruption. The same physical shell โ land, substation, transmission interconnect, water rights โ can host either. But the conversion is not free, and the resulting revenue per megawatt is not comparable.
That is the entire thesis of the miner-to-AI-datacenter trade in one paragraph. The market is not bidding up miners because mining is good. It is bidding up miners because a subset of them are accidentally holding the scarcest input in the AI buildout โ interconnected megawatts in jurisdictions that will actually let you build. Hut 8 is not buying a mine. It is buying the option to stop mining.
Put numbers on it. A post-halving ASIC fleet on a fixed power contract is fighting for a gross margin that a single difficulty adjustment can eat. At $0.03/kWh, a modern rig's contribution margin is thin enough that a 10 percent difficulty step and a 10 percent price drop in the same month turn a profitable site into a maintenance liability. That is the environment in which a $100.5 million bid for one Texas campus is not aggressive โ it is directional. The buyer is not underwriting hashes. The buyer is underwriting the assumption that hashes are about to stop being the highest-value use of that building.
There is corroborating evidence in Hut 8's own disclosure trail. The company has been expanding its development pipeline toward roughly 8.7 GW, a sequential increase of about 300 MW, and it has been deepening AI infrastructure investment alongside that expansion. An operator does not accumulate a multi-gigawatt development pipeline unless it believes the marginal buyer of compute is no longer a miner.
Which is precisely why the second set of numbers โ the ones about debt recovery โ deserves more attention than it is getting.
Hut 8's $140 million against the estate's $173.1 million in liabilities implies a naive recovery rate of about 80.9 percent. That number has been circulating as though it means something. It does not, and the way it fails is instructive.
Chapter 11 distributions are not a division problem. Before unsecured noteholders see a dollar, the estate pays professional fees โ debtor's counsel, the creditor committee's counsel, the financial advisor, the investment banker running the sale โ plus any administrative claims, plus any priority claims, plus the costs of winding the estate down. In a case with roughly $160 million of consumer claims and a contested asset sale, those frictions are not rounding errors. A more honest framing is that the Wallet users' realized recovery will land somewhere meaningfully below the headline ratio, with the gap determined by expenses nobody has itemized publicly. The 80.9 percent figure is a fantasy constructed by dividing two numbers that were never meant to be divided.
And a second-order effect hides behind the naive ratio entirely: the $140 million is not necessarily $140 million of cash. Hut 8's bid has been described as including non-cash consideration, which in distressed M&A conventionally means assumed liabilities, asset swaps, or earnout structures. If a portion of the consideration is assumption of obligations rather than wire transfer, the pool available for distribution shrinks further while the headline number stays frozen in place. I flagged the same pattern in 2017, when I spent six weeks dissecting the token distribution models of three ICOs and found that the vesting schedules โ the fine print, not the whitepaper โ revealed exactly who got paid first and who got paid last. The structure of a claim is more informative than the size of a claim.
We trace the ghost in the machine's memory. In a bankruptcy, the ghost is the repayment waterfall, and it always speaks last.
Contrarian: Asset Appreciation Is Not Creditor Recovery
Here is the counterintuitive angle, and it is the one the miners-are-getting-repriced crowd keeps skipping.
Two true statements are being fused into one false one. The true statements: Texas power assets were undervalued, and that undervaluation is now correcting. The false one: therefore the people who lost money in this ecosystem are being made whole.
They are not. The auction is a revelation about the value of watts. It is not a transfer to the people who funded the buildout of those watts. Poolin's wallet users financed an expansion whose residual value is now being captured by a well-capitalized public company purchasing assets out of a court-supervised sale. This is not a scandal. It is the ordinary mechanics of priority claims in a system where depositors were, structurally, last in line. But it should be named, because the industry's preferred narrative โ a rising tide โ is doing a great deal of unearned work here.
The ledger remembers what the market forgets: the $140 million and the $163.7 million are not the same $163.7 million, and the gap between them is not a rounding error but a distribution of loss.
There is a second blind spot, and it concerns the AI premium itself. The prevailing reading of the Pecos and Fluidstack bid is that AI demand is real and deep. It is real. But the bid tells us what one buyer was willing to pay for one site on one day, and it tells us nothing about the durability of AI capital expenditure through a full cycle. If hyperscaler capex decelerates โ and capex in concentrated buyer markets moves in waves rather than lines โ the marginal value of a converted site falls. A buyer who paid an AI multiple for an asset not yet converted to AI use has purchased a narrative option, not a revenue stream. The conversion capex is unspent. The cooling retrofit is unbuilt. The AI tenant is, as of the filing, unannounced.
A related assumption deserves scrutiny: that conversion is reversible. Once a site is retrofitted with high-density cooling, higher-capacity transformers, and a tenant holding an uptime SLA, the option to flip back to ASIC mining when hashprice spikes is not symmetric. The flexibility being priced into these assets is real in one direction and largely theoretical in the other.
Third blind spot: interconnection rights are not automatically assignable. The value of Pyote and Tarbush rests on the assumption that ERCOT and the relevant transmission counterparties will accept a change of control without material re-study. In practice, a load shift of this scale โ from curtailment-tolerant ASIC mining to a load a data center operator will not want curtailed โ can trigger re-evaluation. The filing does not disclose the transferability status of those agreements. That absence is itself a signal, and it belongs in the risk column next to the price.
Finally, the losing bidders are not inert. Pecos and Fluidstack bid $100.5 million for Tarbush and lost to a bundled offer that implicitly valued the site lower in isolation. DigiPower X bid on Pyote and lost. Bidders who have done the diligence and been beaten by a packaged bid have both standing and motive to object to the sale procedure. An objection does not kill a deal, but it delays it, and in distressed transactions delay has a cost. September 29 is not a formality.
Takeaway: Three Signals, Resolving Soon
Strip the narrative away and three signals remain, each of which resolves on a short clock.
The first is the hearing itself. Approval converts the auction into a repricing data point for every other North American operator holding an energized interconnect โ Riot, Marathon, Core Scientific, and the long tail of private miners sitting on 100 MW blocks across the Panhandle. Rejection, or a sustained objection, converts it into a lesson about procedural risk inside bundled stalking horse auctions.
The second is Hut 8's disclosure cadence over the next two quarters. A pipeline of 8.7 GW is a statement of intent. A signed AI or HPC lease is a statement of fact. Until a lease appears, the $140 million is best understood as a bet that conversion will be possible, priced at a premium that assumes it will be. Watch the capex line before you watch the megawatt line.
The third is the Poolin distribution plan, whenever it lands. That document โ not the auction price โ will tell the industry what unsecured consumer claimants actually recovered in a 2024-vintage crypto bankruptcy. My read, based on the debt composition and the non-cash structure of the winning bid, is that the realized rate comes in well below the 80 percent being quoted. If I'm wrong, I'll say so with the same data.
Finding the signal where others see only noise is not a slogan; it is a discipline that occasionally makes you unpopular. In the spring of 2022, I spent three weeks mapping the reserve degradation of an algorithmic stablecoin and published the trajectory weekly while the market treated the warnings as noise. The death spiral arrived within 48 hours of the final note. I do not raise that episode to claim foresight. I raise it because the same discipline applies here, and the lesson is consistent: when the asset side of a story gets loud, check the liability side before you celebrate.
Chaos is just data waiting for a lens. This filing is chaos to most readers and a lens to a few. The watts were always valuable. The tragedy is that the people who paid for them will read about the price in a headline, not in a distribution notice.
The auction closes a chapter. It does not close the ledger.
