Decent Newsroom
Sign In

The Escalator Clause

By Roger November 14, 2026
Paraguay did not hand its hydropower to the highest bidder. It handed it to the better contract - a 15-year AI deal at $48.91/MWh with an inflation escalator, against miners paying a flat $44. At $67,116 a 30 J/TH fleet breaks even at $43.82. The AI bid did not kill the margin; the margin expired on its own schedule.
The Escalator Clause

The Escalator Clause

Paraguay did not hand its hydropower to the highest bidder. It handed it to the better contract. The distance between those two sentences is the entire bitcoin mining business.

On 22 October, Paraguay’s state utility signed away fifteen years of electricity, a deal CoinDesk detailed on 13 November. ANDE agreed to supply up to 286 megawatts to Kuntur, a Paraguayan company tied to Araico Corporation, which is building a $12 billion AI campus in Yguazú called Iguazú AI City. Crusoe, the American AI cloud company, will run the machines. Deliveries begin with 6 megawatts on 1 December 2026 and reach 146 megawatts in April 2028.

Fifty-four days earlier, the same utility’s own engineers had asked for something close to the opposite of a fifteen-year contract: an auction, a 700-megawatt cap on all mining, AI and data-centre load from 1 January 2028, and a floor of $56 per megawatt-hour. ANDE president Miguel Báez has called the miners’ deadline, 31 December 2027, improrrogable. Non-extendable. Four hundred and twelve days of runway, and no renewal path in current regulation.

The mining industry read this as an eviction. It reads better as a lesson in contract design that almost nobody bothered to parse. Because the AI campus did not outbid the miners. On the headline number, it very nearly undercut them.

What the AI campus actually pays

Here is the figure that got buried under the headline: Araico’s electricity averages $48.91 per megawatt-hour, escalating with U.S. inflation by at least 2.5 percent a year for fifteen years. The miners pay $44.

So the AI campus pays $4.91 more per megawatt-hour at the start (an 11.2 percent premium, by my own arithmetic) and then the number grows. The miners’ $44 has no annual adjustment at all. That, and not the $4.91, is the term sheet. A flat nominal tariff is not a fixed cost. It is a slowly shrinking number for the payer and a slowly shrinking number for the receiver, and only one of those two parties has to buy transformers, maintain a 5,000-megawatt system and meet national demand growing at nearly 8 percent a year through 2027.

Look at what ANDE collected under the flat rate: roughly $100 million in 2024, $295 million in 2025, and a projected $350 million in 2026. Revenue tripled in two years without a single cent of rate increase. It tripled because capacity grew to 944 megawatts contracted as of April, about 13 percent of Paraguay’s share of the 14,000-megawatt Itaipu dam, four companies (Cademot, Zuns, Muiden and W3X) holding 730 megawatts of it. Paraguay supplied about 4 percent of global mining power in the second quarter, the fourth-largest mining jurisdiction on earth.

Paraguay: 944 MW already sold to miners, a proposed 700 MW cap, and a 286 MW AI contract, with 6,300 MW of new requests waiting.

That growth curve just hit a wall the utility built itself. Under the engineers’ proposal, all mining, AI and data-centre load is capped at 700 megawatts combined. Kuntur’s 146 megawatts count against the cap. What is left for every miner in the country is under 560 megawatts, less than three-fifths of what they hold today, auctioned at a floor 27 percent above their current rate.

ANDE is not choosing AI over bitcoin out of fashion. It is choosing a revenue line that grows, that lasts, and that survives the currency it is denominated in. Everything else in this story follows from that.

The part the miners got right, and then stopped doing

For years the trade was genuinely beautiful, and it deserves to be stated properly rather than dismissed.

A miner earns revenue in bitcoin (an asset with a fixed supply) and pays its largest cost in a fiat currency at a nominally fixed rate. If the currency depreciates, the real cost of your power declines every year you hold the contract. You are short the dollar and long the hardest collateral in the world. On paper, that is the best trade in the industry.

The flaw is on the other side of the table. ANDE was holding the depreciating leg. It received a flat nominal $44 while its own input costs, its citizens’ demand and the alternative buyer all rose. A utility that signs a fifteen-year flat nominal contract is lending at a negative real rate to a foreign industrial customer. No treasurer survives that once a better offer exists. And the better offer did not even need to be much better. It needed an escalator and a signature.

The escalator against the flat rate: $48.91 rising at 2.5 percent a year, $44 frozen, and what that $44 is worth in 2026 money by year fifteen.

There is a second asymmetry, and this one is arithmetic rather than monetary. I measured the network on the morning of 14 November 2026. Difficulty stood at 134.82 trillion at block height 975,770, which puts the global hashrate at roughly 965 exahashes per second. With a 3.125 BTC subsidy and fees at 0.8 percent of the block reward (0.0253 BTC out of 3.1503 in the template I pulled), the network grosses about $31,547 per exahash per day at a bitcoin price of $67,116.

Turn that into an electricity ceiling. A machine running at 20 joules per terahash needs 480 megawatt-hours per day per exahash. So the break-even power price at today’s difficulty and price is $65.72 per megawatt-hour. At 30 J/TH it is $43.82. At 40 J/TH, which is gear from 2021 still humming in cheap-power jurisdictions, it is $32.86.

Read those three numbers against the two contracts. At $44, a 30 J/TH fleet earns precisely nothing: minus $133 per exahash per day, before a single salary or spare part. At Araico’s $48.91, you need to be at or below about 26.9 J/TH just to reach break-even, and a 25 J/TH fleet clears it by $3.67 per megawatt-hour, a margin that one bad transformer, one curtailment event or one difficulty adjustment eats whole. At the engineers’ $56 floor, the threshold drops to 23.5 J/TH.

Break-even power price at $67,116 by fleet efficiency, against the miners’ $44 tariff, the AI campus’s $48.91 and the engineers’ $56 floor.

This is the finding that reframes the whole story. The AI bid did not kill the Paraguayan margin. The margin expired at the same time as the contract. At $82,000, roughly where a great deal of this capacity was financed and where Checkonchain notes the panic sellers of the last two days bought, the 30 J/TH break-even is $53.53 and $44 is comfortable. At $60,000 it is $39.17, and every 30 J/TH machine in the country is under water at any tariff whatsoever. The campus in Yguazú simply arrived in the same quarter the arithmetic turned. It is the beneficiary of a margin collapse it did not cause.

The steel-man: the escalator is a bet on someone else’s bubble

Now the honest part, because the case against my own reading is strong.

Araico has signed a fifteen-year contract with a 2.5 percent escalator underwritten by an equity market that is pricing an AI productivity boom. The Federal Reserve’s own vice chair is not yet sure that boom exists in the data. In a speech at UC San Diego on 13 November, Philip Jefferson put the numbers on the table: the Census Bureau’s survey of firms shows roughly 25 percent of all U.S. firms using AI in at least one business function as of the middle of 2026, against close to 40 percent in professional services and information, and around 10 percent in mining and in leisure and hospitality. Worker surveys put usage at 40 to 50 percent of working adults, but only a modest fraction use it daily. One study of senior managers estimated AI had lifted sales per employee by half a percent over three years.

Labour productivity growth has averaged 2.2 percent between 2020 and 2025, up from 1.5 percent over the previous business cycle, but Jefferson is careful to note the pickup began before the chatbots arrived in 2022, and that “the magnitude and timing of forthcoming productivity gains remain highly uncertain.” He also says the expectation of those gains is already raising aggregate demand: equity wealth, household spending, and a very large data-centre construction programme that is pushing up prices in memory and logic chips precisely because supply cannot keep up.

That is the risk embedded in every escalator clause signed this year. If AI capital spending is Cisco in 1999, which is the question Anthony Pompliano asked in his five charts on 13 November, then ANDE has traded a flat $44 from a solvent counterparty for a rising number from a counterparty whose willingness to pay was set by a share price. Fifteen years is a long time to hold the wrong side of that.

And the equity market is showing its nerves. Nvidia’s board announced the resignation of Jensen Huang after 33 years as chief executive on 12 November, with CFO Colette Kress stepping in as interim; bitcoin slid to $68,000 the same week as capital rotated toward Anthropic’s IPO and a broader risk-off move, with $1.6 billion of liquidations and a sixth consecutive weekly loss. The same market that underwrites the escalator is repricing the entities that must pay it.

Two clocks

Because while the price tape was doing that, the ownership tape was doing something else entirely.

JPMorgan’s Q3 2026 13F, filed on 12 November, discloses roughly $981 million in bitcoin-linked exposure as of 30 September. The three largest buckets are $549 million across spot bitcoin ETFs, $131.3 million in Strategy and $300.6 million in Coinbase, spread across 17 managed entities. The IBIT position alone went from 10.4 million to 11.4 million shares, up 9.6 percent in count and about 52 percent in dollar value, quarter over quarter. The Strategy share count rose 36 percent. Add Robinhood and listed miners and the total clears a billion. This is the bank whose chief executive called bitcoin worthless in 2021 and a pet rock in 2024, accumulating through the pipes while the rhetoric stayed fixed.

At the same time, Checkonchain counted over $680 million of realised loss locked in by investors in a single day: people who bought above $80,000 on the way up and sold at the price they once wished they had bought at. That is what a transfer of coins looks like from the outside: no coins are destroyed, they change hands at a discount to the seller’s cost basis.

Two clocks are running. The price clock is set at the margin by the most impatient holder, and this week that holder was frightened and leveraged. The ownership clock is set by the most patient one, and this week that one filed a 13F. Both are true. Only one of them is a forecast.

The layer that does not need a megawatt

There is a third story from the same forty-eight hours that got a fraction of the attention, and it is the one that will still matter in 2035.

On Delving Bitcoin, a proposal for a post-quantum address format called bc1z, built on ML-DSA-65, was posted as a draft BIP. Separately, a contributor published a STARK proof of bitcoin’s entire UTXO set at block 972,500 — with the circuits themselves proven in Lean, a formal verification language. That is not a marketing slide; it is a machine-checked claim that a proof system correctly describes a consensus-critical computation. Also this week, Bitcoin Optech’s 431st newsletter covered draft BIPs for exfiltration-resistant signing, the disclosure of three vulnerabilities affecting older Bitcoin Core versions, a plan for mining the testnet5 genesis block, and an investigation into degraded compact block reconstruction.

None of that required a single megawatt in Paraguay. None of it asked permission from a utility, an exchange, or an IPO calendar. The protocol’s roadmap is denominated in review cycles, not quarters, and it advances whether or not anybody is bidding for the electrons this month.

What to watch

Four markers will tell you whether the reallocation is structural or a squeeze.

First, any ANDE resolution before 31 December 2026 that softens the December 2027 date. A decree is a cheaper signal than a speech.

Second, the auction itself. If the 700-megawatt cap clears at the $56 floor with 3-to-5-year terms, it becomes a template that other state utilities (Bhutan, Quebec, Iceland, the Gulf) can copy. If miners win a meaningful share at $56 and still run profitably, then my break-even arithmetic says bitcoin has to be somewhere north of $85,000 for a 30 J/TH fleet to make that work. Watch the price-to-difficulty ratio, not the price alone.

Third, HIVE. It has roughly 77 percent of its 385 megawatts in Paraguay, at Yguazú and Valenzuela, both expiring the same day as everyone else’s. It has said publicly it wants ten more years, could go from 300 to 1,000 megawatts in eighteen months, and expects to hand ANDE $140 million this year; at the engineers’ floor that bill becomes about $178 million before a single new machine. Management calls the alternative “a hypothetical.” The next annual report will say whether that is still true.

Fourth, the counterparty. Q4 13Fs arrive in February 2027. If JPMorgan’s IBIT line keeps growing, the institutional pipe was real. If it unwinds, the $981 million was a hedge. And if Jefferson’s productivity data does not improve through 2027 while AI capex keeps compounding, then every escalator clause signed in 2026 becomes a claim on a boom that never showed up in the statistics, and the utility will have discovered that a growing nominal price from a shrinking customer is worse than a flat one from a solvent one.

The grid never asked what the electrons were for. It asked who signs, for how long, and whether the number in the contract grows. Paraguay answered all three questions in a single week in October, and the answer was not about bitcoin at all.

Bitcoin’s answer to the same question has never been to bid higher for the same electron. It has been to need fewer of them, and to keep producing blocks whether or not anybody is still bidding.

What I did not verify

I have not seen the Kuntur contract itself, only the terms reported by Paraguayan outlets and relayed by CoinDesk; the $48.91 average, the 2.5 percent floor on the escalator and the 286-megawatt ceiling should be treated as reported figures rather than audited ones. HIVE declined to confirm either its $140 million annual payment or what happens to its Paraguayan fleet after 2027, and I had no way to check its investor filings directly for a contingency plan. The 700-megawatt cap and the $56 auction floor remain a proposal by ANDE’s engineers’ association, not adopted policy. Araico, ANDE and the engineers’ group did not respond to CoinDesk’s requests for comment, and none of them answered mine either, because I did not ask.

Sources

Own measurements: bitcoin price $67,115.80 and network state (difficulty 134,823,837,323,272.7, block height 975,770, block template value 3.15028951 BTC) read on 14 November 2026 at 09:31 CET. Break-even power prices are a calculation, not a quotation: revenue per exahash per day divided by megawatt-hours per exahash per day, assuming 100 percent uptime, no cooling overhead, and no non-power costs. Real fleets do worse than these numbers, never better.

Write a comment