Converting a bitcoin mining site to AI or HPC

A mining facility was built to one specification: convert electricity into computational output at the lowest achievable cost per megawatt-hour, in a market where interruption is an accepted operating condition rather than a failure. Every decision above the meter followed from that — the building, the airflow, the density, the redundancy, the standard the electrical system was designed to.

That specification produced sites which are, in one narrow respect, extremely valuable. They hold an energized, operating electrical position at scale, established at a time when such positions were easier to obtain than they are now. In a market where the binding constraint is the interconnection queue, an operating connection is the scarce object and everything else is replaceable.

Which sets up the framing this page runs on. In most transactions of this kind the buyer is not acquiring a data center that needs improving. The buyer is acquiring a power position with a building on it, and the building is closer to a cost line than to an asset. Sellers price the facility. Buyers who understand the trade price the position and then subtract the cost of everything they will have to remove.

That is a different exercise from the [brownfield case](brownfield-conversion), where the electrical position is historic and may not have survived the site's retirement, and where the dominant risk is environmental. Here the position is live and current. The risk sits elsewhere.

The division is unusually clean, which is why the diligence can be run as a sequence rather than as a general investigation.

Generally transfers, subject to the documents:

• The interconnection and the queue standing behind it — the asset the transaction exists for, and the one with the most conditions attached. Whether it survives a change in the character of the load is a separate question, worked through in [interconnection under a change of load](interconnection-under-load-change).
• The utility agreement and the tariff position — including whatever rate treatment the site currently receives, which was frequently designed around an interruptible load and may not be available to a firm one.
• The substation, transformers and medium-voltage distribution — real, expensive, long-lead, and the single largest item that does not have to be rebuilt from nothing.
• The land, the site works, and the access, water and fiber that reached it.
• Permits, entitlements and the operating history — a site with an established industrial use and an existing relationship with its jurisdiction starts ahead of a greenfield application.

Generally does not transfer, in the sense that it cannot serve the new use:

• The buildings. Sheds and lightly enclosed structures built for high-volume airflow are not a shell an AI tenant will accept, and the honest treatment of many of them is demolition rather than retrofit.
• The rack and containment layout. Specified for a density and a thermal approach that an accelerator hall does not share.
• The mining hardware itself. It has whatever value it has as equipment on its own terms; it is not part of the data-center asset, and a buyer who does not want it is buying a decommissioning obligation.
• The availability standard. This is the item most often missed and it is the subject of the next two sections.

The useful discipline is to run this as two columns before pricing anything, and to be explicit that the second column is a cost line rather than an asset that merely needs work.

This is the central point of the page, and it is where most conversion arithmetic goes wrong.

A mining site's headline capacity describes what its connection can draw. That figure is quoted in marketing, it is the figure the seller has in mind, and it is not the capacity the site can sell to an AI or HPC tenant. Three deductions sit between them, and they compound rather than overlap.

Redundancy. A mining load can be dropped. A training or inference tenant contracts for availability, and availability is bought with redundant electrical paths, redundant cooling capacity and equipment held in reserve rather than in service. Capacity committed to redundancy is capacity that cannot be sold. A facility built with essentially none has to find that headroom inside its existing connection, or extend the connection.

Cooling and facility overhead. The share of the connection consumed by the facility itself rather than by the tenant's equipment is not the same under the two regimes. A building whose cooling strategy suited one load and has to be re-engineered for another does not carry the same overhead, and the difference comes off the sellable total.

The tier the tenant will actually sign to. Availability standards are contractual, not aspirational. The level a counterparty requires determines the electrical and mechanical configuration, and therefore how much of the connection is left over. A site can be perfectly capable of serving a lower standard and simply not be able to serve the one the tenant's own commitments require.

The result is that a site marketed on its nameplate is being marketed on a number that no tenant will ever contract against. Ask for the megawatts of critical IT load the site can deliver at a stated availability standard, and ask what the site's own consultant assumed to get there. If the answer is the same as the nameplate, the analysis has not been done.

Once the two columns are drawn and the capacity has been derated, what remains is a build — and it should be planned and financed as a build, not as an improvement to an operating asset.

The practical consequences are three.

The retrofit competes with new construction on its own merits. The comparison is not conversion against greenfield in the abstract. It is conversion against greenfield *for the portion of the work that is actually shared*, which is the electrical position and the land. Everything above that is being built either way. Where a conversion wins, it wins on schedule and on the queue position, not on construction cost.

The existing revenue does not survive the work. A site being converted is a site whose current operation is being wound down, and the period between the two uses generates nothing while consuming capital, carrying the land and holding the connection. That gap has to be funded by someone, and it is a distinct financing question from the acquisition — the point taken up in [financing a mining-site conversion](/structures/financing-a-mining-site-conversion).

Phasing is usually the answer and it has to be designed, not discovered. Large sites rarely convert in one movement. Which portion of the connection is released to the new use, in what order, and what that does to the site's obligations under its existing arrangements, is a question with contractual answers rather than operational ones.

What none of this is, is a hardware exercise. How a hall is cooled and at what density it can be populated is the physical layer's question. What matters here is that the answer changes the sellable capacity, the schedule and the capital requirement — and that all three should be established before a price is agreed rather than after.

One feature distinguishes this fact pattern from every other site acquisition, and it is a counterparty fact rather than a site fact.

An operator converting its own site, or selling it, typically has a balance sheet whose value moves with a single volatile input. That is not a comment on the merits of the underlying asset; it is an observation about credit concentration, and it has direct consequences for a buyer:

• Ability to complete. Where the seller is retaining an interest, funding part of the retrofit, or performing obligations after closing, its capacity to do so is correlated with something neither party controls. A post-closing obligation from a counterparty whose resources move with an external price is worth less than the same obligation from a diversified one.
• The durability of reps, warranties and indemnities. These are only as good as the entity behind them. Where that entity's capacity is volatile, the market answer is escrow, holdback or third-party support rather than covenant language.
• Timing pressure that is not about the site. A seller under balance-sheet pressure is a motivated seller, which is commercially useful and diligence-relevant in equal measure. Pressure produces speed, and speed is where conditions go unread.
• Existing encumbrances over the estate. Operating equipment, the facility and sometimes the electrical infrastructure may already secure existing obligations. What has to be released, by whom, and on what terms is a closing-mechanics question that should be asked early because the answer is frequently slow.

None of this makes the transaction a poor one. It makes the counterparty a variable to be underwritten alongside the asset — which is exactly what a buyer would do in any other acquisition and is skipped here surprisingly often, because the site presents as an operating industrial facility rather than as a credit.