In late 2024, the capacity market that prices physical reliability across the largest power grid in the United States did something it had never done before. PJM's base residual auction cleared from roughly $29 per MW-day to $269, then to $329 per MW-day across successive delivery years โ a near-tenfold repricing of reliability in under three years. The finger-pointing arrived the same week. It landed, as it always does in an election-cycle energy debate, on data centers, with Bitcoin miners sharing the crosshairs. RWE, a European utility holding both generation and grid exposure, supplied the framing that the entire conversation then inherited: data center load growth raises one question โ who pays for the grid โ and the answer might be residential ratepayers.
That framing is not wrong. It is incomplete in a way that matters, and the missing variable is precisely the one my discipline has spent a decade mapping. The load that never blinks is not the load that mines.
Context: what a grid actually sells
Before the crypto layer, the accounting layer. A grid does not sell electrons. It sells three separable products: energy in kilowatt-hours, capacity (the promise to deliver at peak), and reliability services (frequency, voltage, spinning reserve). Energy is variable cost โ mostly fuel. Capacity and reliability are overwhelmingly fixed: depreciation on transformers, substations, and transmission corridors; the standing operations and maintenance that keep them alive; the regulated return on that rate base. Industry estimates consistently place 70 to 80 percent of total grid cost in this fixed bucket.
This distinction is not academic. It is the entire game. When a utility raises its revenue requirement to fund a new substation, the cost does not scale with how many kilowatt-hours flow through it. A transformer rated for 100 MW costs roughly the same whether it carries 40 MW or 90 MW. A transmission corridor costs what it costs whether it is congested or idle. So any tariff that recovers fixed cost through volumetric energy charges โ cents per kWh โ systematically misprices every customer whose load shape differs from the average residential profile. The residential customer pays for peak, for a diversified load, for a capacity factor near 40 percent. The large industrial customer negotiates a rate that reflects its own load shape and its own bargaining power. The two are not in the same market, even though they share the same wire.
Most retail tariffs are three-part: a fixed customer charge, a demand charge based on peak draw, and an energy charge per kWh. Large loads routinely negotiate all three. The reason is simple โ they can, and the utility would rather have them than not. That asymmetry of negotiating power is the seam along which cost allocation tears.
Capacity markets are where the mispricing becomes visible as a single number. PJM's auction does not price energy; it prices the promise to be available at peak. When that promise reprices tenfold, the market is telling you that the system's marginal cost of reliability has structurally changed. The market's instinct โ blame the newest large load โ is understandable and, in this case, analytically backwards.
In my 2020 MakerDAO stress-testing work, I built a thousand-scenario Python model to find where over-collateralization breaks under gas-fee shocks. The lesson generalized far beyond DeFi: a system's failure mode is never in its average, it is in its tails and in the mismatch between its pricing unit and its cost structure. The grid's pricing unit is the kWh. The grid's cost structure is fixed. That mismatch is the defect. Everything else โ miners, data centers, RWE's warning โ is downstream of it.
There is also a governance point buried in the sourcing. RWE holds generation and grid assets, and its stated preference โ that the beneficiary of grid upgrades should pay โ is not a neutral analytic claim. It is a stakeholder position. A utility that owns the wires benefits when a new large load is forced to fund the upgrades that connect it. That does not make the claim false; it makes it interested. Any analyst who accepts it without discounting for the source is not analyzing, they are transcribing.
Core: two load profiles, two opposite economics
Now the technical part, and the part the market keeps refusing to price.
Bitcoin mining is an interruptible load. This is not marketing language; it is architecture. A mining rig produces exactly one output โ hashes โ and hashes carry no delivery deadline. A miner can power down in seconds with zero customer-facing penalty, because there is no customer. The revenue it forgoes is the hashprice: a function of network difficulty, block subsidy, and transaction fees, divided by network hashrate. That forgone hashprice is the entire cost of curtailment. There is no service-level agreement to breach, no churn to suffer, no service credit to pay.
In ERCOT, the Texas grid operator, miners have become the largest and fastest demand-response resource in the territory. They routinely shed gigawatts during scarcity events and buy back during surplus. The breakeven is trivial to compute and brutal in its clarity. Suppose a miner earns the equivalent of six cents per kWh in hashprice at current difficulty. It curtails whenever the grid price exceeds six cents. During a scarcity event, real-time prices can clear between $1.00 and $9.00 per kWh, with ERCOT's system-wide offer cap having sat at $5,000 per MWh in recent years. The miner sells the electron back for up to eighty times what mining with it would have earned. Curtailment is not a concession to the grid; it is the single most profitable action available. This is what a flexible load looks like when its incentives are aligned with the system's needs by construction rather than by mandate.
Demand-response programs formalize this. When a miner enrolls as a curtailable load, it receives a capacity payment for being available to drop load, on top of the energy it sells back during events. The grid buys optionality; the miner monetizes it. In liquidity terms, the miner is writing a put option on grid capacity and the grid is the buyer. That is a structurally sound trade for both sides, and it is precisely the trade a flat AI load cannot offer, because a flat load has no option to sell.
The consequence for fixed-cost allocation is counterintuitive and important. A flexible load consumes power when the grid is long and cheap โ frequently at negative or near-zero prices during West Texas renewable oversupply โ and releases capacity when the grid is short. From the standpoint of spreading fixed costs, a flexible load is nearly ideal. It uses capacity that would otherwise be stranded, it pays for energy during hours that would otherwise have no buyer, and it stops using the system precisely when the system is stressed. It increases usage hours without increasing peak stress. That is the opposite of a cost-shifting problem.
AI data centers are the structural inverse. Their load profile is flat and inelastic. Capacity factors commonly exceed 70 percent โ the source material's own figure โ and the workload cannot be interrupted without violating service-level agreements. A GPU cluster mid-training does not power down because the grid is short; it powers down because the operator decides, and the operator has signed contracts that forbid the decision. This is a high-capacity-factor, low-flexibility, high-reliability-demand load. Structurally, it is among the hardest loads a grid can absorb, because it adds to baseload and to peak simultaneously while contributing almost nothing to demand response.
Run the two profiles through the same fixed-cost arithmetic and the popular narrative inverts. The load the market blames โ mining โ is the one load class that helps amortize grid fixed costs. The load the market is racing to attract โ AI โ is the one that strains them. Logic is immutable; incentives are the variable, and the incentive structure has been aimed at the wrong target.
If I map the grid's load base the way I would map liquidity in a market, four strata appear. At the bottom, residential load: small, numerous, inelastic in aggregate, no exit option, no negotiating power โ the price-taker. Above it, small commercial: similar, marginally more elastic. Above that, flexible large load โ Bitcoin miners, some industrial process load, some electrolyzers โ interruptible, price-responsive, effectively the system's swing buyer and swing seller of capacity. At the top, inflexible large load โ AI data centers, hospitals, semiconductor fabs โ high-capacity-factor, contract-bound, and increasingly able to bypass the retail tariff through co-location and power purchase agreements. The pricing system treats all four as one customer class. They are not. They are four different instruments, and pricing them identically is like quoting the same spread on a Treasury bill and a distressed credit.
There is a second, less-discussed layer: large loads do not merely consume, they negotiate, and they exit. Three escape paths are now standard. First, direct physical connection โ co-locating generation behind the meter so the load never touches the transmission system, and therefore never contributes to the fixed-cost pool it still relies on for backup. Second, long-term power purchase agreements, most visibly nuclear PPAs, which lock in clean baseload at a negotiated price that bypasses the retail tariff entirely. Third, self-built microgrids โ gas, storage, and renewables serving a load islanded from the public grid.
Each is a rational response to a mispriced tariff, and each produces the same second-order effect. The load leaves the cost-sharing pool; the grid's fixed costs remain. The denominator shrinks; the numerator does not. The remaining customers โ residential and small commercial, with the least bargaining power โ absorb a larger share per unit. This is not a prediction of malice. It is arithmetic, and it is the same arithmetic that governs every shared-cost pool with heterogeneous exit capability.
I have watched this pattern in a different domain. When I analyzed MakerDAO's collateral model during the 2020 gas spike, the fragility was never in the average collateral ratio; it was in the tail, where the most agile participants exited first and left the least agile holding the liquidation risk. A DeFi protocol and a power grid share a structural signature: a shared pool of cost or risk, populated by participants with wildly unequal exit options. Those who can exit, do. Those who cannot, pay. History repeats not in price, but in pattern.
This is the death-spiral risk that the RWE-style framing gestures at but never names. It is not that data centers raise consumer bills. It is that a fixed-cost pool with heterogeneous exit capability concentrates cost on the least mobile participants. Crypto miners are not the cause; as flexible load, they are a partial mitigation. The cause is a rate design that recovers fixed cost volumetrically from a customer base whose load shapes and exit options differ by orders of magnitude.
The regulatory sensitivity is acute, and it is where the crypto-energy thesis becomes investable. FERC's rulings on co-location โ whether a data center directly connected to a nuclear plant must pay the transmission charges it avoids by not using the grid โ are the current frontier. RWE's warning, stripped of its corporate interest, is essentially a European utility watching this same dynamic arrive on its own network, where grid fees are a larger share of the retail bill than in most US markets. The warning is a signal about the direction of rules, not a data point about cost.
Now the convergence, priced honestly, because this is where it turns actionable. The fusion of AI and crypto has blurred the load profiles that were previously separable. The same operator now runs mining rigs, which are flexible, and GPU clusters, which are not, on the same site, behind the same interconnect. The flexible half has been quietly subsidizing the economics of the whole operation: curtailment revenue, negative-price power, demand-response payments. As the capital mix shifts toward AI, the flexible half shrinks relative to the inflexible half, and the site's net contribution to grid flexibility declines. The load that looked like a grid asset begins behaving like a grid liability. The audit passed โ the site still balances its books โ but the economics failed, because the very flexibility that justified the interconnection is being retired by the workloads the site is pivoting toward.
The on-chain corollary is a structural information asymmetry, and it favors the disciplined analyst. Mining economics are transparent in a way AI data center economics are not. Hashprice, difficulty, and network hashrate are public and continuous; a miner's curtailment behavior is observable in grid demand-response data in near real time. An AI data center's marginal economics are opaque โ capital-intensive, contract-bound, disclosed quarterly at best, often buried in a hyperscaler's segment reporting. The flexible, transparent load is being scapegoated. The inflexible, opaque load is being courted. That is not a market failure in the pricing of crypto. It is a market failure in the pricing of load.
Contrarian: the decoupling thesis
The consensus view treats crypto mining and AI data centers as the same problem: big loads that strain the grid and shift cost onto consumers. The contrarian read is that they are opposite problems, and that treating them as one guarantees a bad policy outcome.
If regulators respond to the RWE-style warning with a uniform large-load tariff โ a flat capacity charge per MW, or a single non-differentiated rate class โ they will penalize the flexible load that improves grid economics while under-pricing the inflexible load that strains them. The flexible load will curtail more or relocate to a friendlier jurisdiction; the inflexible load will stay, because it has nowhere else to go and no incentive to flex. The grid ends up with less flexibility, the same peak stress, and a smaller cost-sharing pool. A uniform tariff applied to non-uniform loads is the same category error as pricing grid fixed costs by the kWh โ the same mismatch between pricing unit and cost structure, now committed at the policy layer.
The correct intervention is load-shape-differentiated. Pay flexible load for the flexibility it provides โ demand-response payments, capacity credits, curtailment compensation. Charge inflexible, high-capacity-factor load for the fixed capacity it monopolizes โ a coincident-peak demand charge that reflects its true contribution to system cost. That is neither a subsidy to miners nor a penalty on AI. It is the alignment of pricing unit with cost structure, the only durable fix. Structural integrity precedes market sentiment, and here the market sentiment is pointed in exactly the wrong direction.
Takeaway
The next signal is not a price. It is a ruling. Watch the FERC co-location dockets and the state public utility commission proceedings on large-load tariffs. If the emerging rule differentiates by load flexibility, the flexible crypto-energy layer gets repriced upward as a grid asset. If it does not, expect flexible load to exit first โ and ask who is left holding the fixed cost when it does. The grid will still need the transformer. The only open question is who signs for it.
