Texas's data-center debate has exposed a trade-off that the Bitcoin mining industry's AI pivot tends to hide. The same mining sites that look inefficient beside a long-term AI lease can be unusually useful to the power grid because they can stop consuming electricity within minutes.
That flexibility is not theoretical. Riot Platforms earned $24.7 million in power credits during the third quarter, mostly by selling contracted electricity back into the market and partly by participating in demand-response programs. Yet the company's new 910 megawatt lease with Anthropic restricts that same behavior at its Corsicana campus: Riot cannot voluntarily curtail the tenant's load or enroll it in voluntary demand response without Anthropic's consent.
The contract may still be an excellent commercial decision. It is a 15-year agreement with approximately $28 billion of total contract value. But it illustrates what Texas could lose as miners replace interruptible Bitcoin machines with higher-value computing that customers expect to remain online.
The Texas permit freeze did not begin with Democrats
Gina Hinojosa won the Texas governor's race on November 3 after campaigning for a data-center moratorium until lawmakers adopt stronger protections for households, water and the power grid. Her victory gives the industry a new political obstacle, but the scrutiny was already bipartisan.
On August 3, Governor Greg Abbott directed the Public Utility Commission of Texas and the Electric Reliability Council of Texas, or ERCOT, to audit data-center projects before allowing them to advance through the grid-connection process. In September, he also directed the state's environmental regulator to pause certain data-center permits while their effects were reviewed.
The headline queue explains the intervention. Abbott's directive cited more than 474 gigawatts of requested large-load connections, more than five times ERCOT's record peak demand. That is not a forecast of real consumption. Project queues contain duplicates, speculative requests and developments that will never be financed. ERCOT's November 5 update said 433 respondents to its data-center inquiry represented roughly 177 GW, including responses from all 245 identified projects of at least 75 MW.
Even the smaller number is enormous relative to the system. Electricity may be available somewhere in Texas, but a project still needs generation, transmission, transformers and permission to draw power at its specific location. In that environment, a load that can accept interruption is fundamentally different from one that must run continuously.
Bitcoin mining turns electricity into an option
Bitcoin miners do not accumulate half-finished hashes. Each calculation is an independent attempt to find a valid block. When a miner switches off, it gives up revenue during the interruption, but it does not lose a work-in-progress asset that must be recovered later.
That makes industrial mining unusually responsive to electricity prices. A miner can run when power is abundant, stop when the grid is tight and, if it has bought electricity under a fixed-price agreement, resell that power when market prices exceed expected mining revenue.
Riot's third-quarter filing separates the benefit. It recorded $22.9 million of manual curtailment credits and another $1.7 million from demand response. The company explicitly says this power strategy significantly reduces its cost to mine Bitcoin.
Texas law can still require large loads, including data centers, to reduce demand during a grid emergency. But mandatory emergency curtailment is not the same as an operator voluntarily switching off whenever prices signal scarcity. The former is a last-resort system protection; the latter can make flexibility part of the site's everyday economics.
AI revenue is higher, but the load is harder to interrupt
The attraction of AI is easy to understand. Hashrate Index estimated in October that an efficient Bitcoin ASIC fleet could generate about $135 of gross revenue per megawatt-hour, while Nvidia H100 GPUs rented at 70% utilization could generate about $1,175 per MWh.
Those figures are not profit comparisons. GPU clusters require much more capital, networking, cooling and data-center infrastructure. Hashrate Index estimated roughly $18 million to $23 million of GPU equipment per megawatt before the cost of a Tier III facility. Utilization and customer pricing can also move sharply.
Still, the revenue gap is large enough to explain why miners with scarce power connections are courting AI tenants. Signal21 has previously examined how AI compute is becoming an infrastructure asset class and why powered land can be more valuable than the machines currently installed on it.
The operational bargain is different. An AI provider has sold a service. Training can sometimes be checkpointed and resumed, while some inference can be shifted between locations or delayed. But those actions require spare capacity, software coordination and contracts that tolerate interruption. A Bitcoin miner can simply stop making new attempts.
Riot's Anthropic agreement makes the distinction concrete. The lease permits curtailment when ordered by a utility or grid operator, but Riot cannot voluntarily reduce Anthropic's load or place the premises into a voluntary demand-response program without the tenant's approval. The AI load is not physically incapable of stopping; control over that decision has moved from the site owner to the customer contract.
Riot traded flexibility for contracted revenue
Riot sold 20,264 BTC for approximately $1.48 billion during the first nine months of 2026. Its filing says the proceeds were used primarily for data-center development and to repay a $200 million credit facility. The company is therefore converting part of its Bitcoin balance sheet and Texas power position into a capital-intensive AI campus.
At Corsicana, Anthropic has leased 910 MW of critical IT capacity for 15 years. Riot describes average annual rent of approximately $1.8 billion and estimates that it must fund about $5.5 billion of development capital. The scale makes surrendering some power-market optionality economically rational: predictable lease income can be worth much more than intermittent mining and curtailment revenue.
For ERCOT, however, the accounting is different. A fleet of mining machines can reduce demand when air-conditioning load and wholesale prices surge. A contracted AI campus may keep drawing power until the grid formally orders a reduction. Replacing one with the other can increase the amount of generation, storage or transmission Texas needs for its most difficult hours even if annual electricity consumption is unchanged.
An immediate exit from mining creates transition risk
Keel Infrastructure, formerly Bitfarms, shows the other side of the pivot. The company stopped Bitcoin mining at four US sites during the second quarter as it prepared them for high-performance computing. At the end of September, it still had no revenue-generating HPC capacity in service.
Keel reported a $273.1 million net loss for the first nine months and used $159.5 million of cash in operations. Its remaining revenue came from Quebec Bitcoin mining and energy sales at Pennsylvania power sites. The figures do not prove the conversions will fail. They show the financing gap that opens when mining revenue stops before AI infrastructure, permits and customers are ready.
That gap matters more when approvals are delayed. A powered site can look extremely valuable in a presentation, but the owner must continue paying for development while waiting for environmental permits, interconnection studies, construction and a tenant that can support project financing.
The hybrid model may be more valuable than it looked
MARA is pursuing a less binary strategy. In its latest filing, the company says it intends to retain Bitcoin mining as a flexible, revenue-generating load while expanding AI and critical-IT infrastructure. At Matagorda, Texas, MARA is evaluating a campus where existing mining can coexist with more than 1 GW of potential data-center capacity.
That structure has three possible advantages:
- ASICs can monetize power before an AI building and tenant are ready.
- Mining can absorb energy that would otherwise be curtailed or stranded.
- The site can cut mining first during scarcity without interrupting contracted AI services.
The idea is not free of constraints. Electrical architecture, interconnection rights and customer agreements must allow the two loads to coexist. AI tenants may also demand priority access to generation and backup systems. But it preserves an option that a full conversion destroys.
MARA has not yet proved that its largest proposed AI leases will close; our analysis of its third-quarter results found that much of the AI pipeline remains uncontracted. The strategic logic is nevertheless clear: Bitcoin mining can serve as a bridge load while higher-value infrastructure is built, then remain the first load switched off when the grid tightens.
The real competition is over firm power, not computing labels
Texas is not choosing between Bitcoin and AI in the abstract. It is deciding which projects may connect, who pays for the required grid upgrades and how much demand can disappear during critical hours.
AI operators could eventually reproduce much of mining's flexibility by shifting workloads, building storage and contracting across multiple regions. Google and other hyperscalers are already pursuing more flexible power agreements. If those systems become routine, AI's much higher revenue per megawatt-hour will make pure mining difficult to justify at premium data-center sites.
Until then, Bitcoin miners retain a capability that is becoming more valuable as the connection queue grows: they already know how to monetize electricity without requiring it every minute. The winners may not be companies that abandon mining fastest. They may be those that secure long-duration AI revenue while keeping enough interruptible compute to make their sites useful to the grid.
This publication is market analysis, not a recommendation to buy or sell.
