Retrofits, mining sites, and whether a new build is obsolete on day one

An apparently terrible trade

Here is a transaction that should not make sense. TeraWulf discloses a cost of $8 to $10 million per megawatt to convert one of its bitcoin mining buildings into a Tier III AI data center. That building cost well under $1 million per megawatt to build. A new AI facility, built for the purpose with nothing inherited, costs $8 to $15 million per megawatt of fit-out. Thus the conversion costs up to ten times the value of the original building, and it saves little or nothing against a new start.

Before you reject the trade, look at how the numbers agree with each other. Each one is a different scope of work:

• Riot’s retrofit of existing buildings for AMD has a budget of $89.8 million for 25 megawatts. That is about $3.6 million per megawatt.
• TeraWulf’s full Tier III job costs $8 to $10 million. It adds liquid cooling and redundancy.
• A purpose-built AI fit-out costs $8 to $15 million. It replaces everything.

Each figure is where the scope before it predicts. That consistency is the evidence: the building was never the expensive part.

On those numbers, the market for conversions should not exist. But it is one of the busiest parts of the industry. Reports put Core Scientific’s agreement with CoreWeave alone at $8.7 billion. Every former miner with a substation now gets calls from buyers. When sophisticated buyers pay a price that looks irrational again and again, do not assume that they cannot do arithmetic. Assume that they buy something that is not on your spreadsheet.

They do. It is not on a cost-per-megawatt line, because it is not a thing. It is a date.

What the mine actually contains

Remove everything from a mining site that does not survive the conversion, and the remaining inventory is short. The buyer usually removes or guts the shell. Mining rigs accepted heat, dust and outages that a training cluster never will. The cheap electrical equipment goes too. Three items remain:

• Land zoned for industrial power, in a place where power is cheap. Cheap power is the reason the miner chose it.
• A substation, built and commissioned.
• Above all, an energized interconnection agreement, the contract that lets the site draw power from the grid.

That last item is the full trade. A new grid connection takes four to seven years in most US markets. In April 2026, ERCOT, the Texas grid operator, tracked about 410 gigawatts of large loads that requested a connection. Every other part of a data center is a product with a lead time: turbines, transformers, chillers, even the building. A place in a queue that formed years ago is not a product. The miner did not especially plan it, but it spent the last decade buying the one input that nobody can expedite. Then the AI industry arrived and found the miners holding it.

TeraWulf’s own cost breakdown shows the same thing from the other direction. Of the $8 to $10 million per megawatt:

• Liquid cooling takes $2 to $2.5 million.
• Electrical upgrades take up to a million.
• Redundancy takes $1.5 to $2 million.

Redundancy alone costs more than the full original mining build. This shows how little of the mine, as a building, the buyer purchases. The conversion price is high because almost nothing transfers except the one thing that matters.

Pricing the date

The market puts an explicit price on the thing that matters. Reports say an already-energized site gets $3 to $4 million per megawatt for each year of lead time it saves.

With that number, the apparently terrible trade reverses, by a large margin. Take a 200 megawatt Tier III conversion at $9 million per megawatt. That is $1.8 billion of fit-out, energized in perhaps fifteen months. The greenfield alternative arrives in something like five years. The conversion saves about 45 months. At the market’s own price for time, 45 months across 200 megawatts is worth roughly $2.6 billion.

The time is worth more than the full conversion costs. On this accounting, the building is not only cheap. Its effective cost is negative, and it stays negative for every conversion path in the model below.

One check that the time price is realistic: a megawatt filled with GPUs and rented on Colossus-style terms produces revenue on the order of $50 million a year. Against that, a payment of $3.5 million to start one year sooner is not aggressive. If anything, it is conservative. And when accelerator supply becomes less tight, the value of the queue position increases further. Silicon then stops being the bottleneck, and the permission to connect is the bottleneck that remains.

Why most mines still fail the test

None of this saves the average mining site. The filters are physical, and you can state them plainly:

1. Power density per square foot. AI needs a large amount of power in a small, dense footprint. Many mining halls spread modest power across acres of cheap shell. A megawatt spread thin pays for unused floor.
2. Structure. Liquid cooling needs three things: floor loading for coolant distribution units, ceiling height for piping, and an envelope that holds a controlled environment. A ventilated shed provides none of them.
3. The grid relationship itself. Miners often bought power on interruptible terms. Their agreement to shut off within seconds is exactly what made their electricity cheap. A training run does not accept interruption. Thus the buyer must rebuild the site’s redundancy topology, and sometimes its full contractual identity. This is the reason for the $1.5 to $2 million per megawatt line item.

The honest sector summary is blunt: some bitcoin mines can become AI data centers, and most cannot. The ones that can are the ones whose megawatts arrive dense, structural and firm.

The obsolescence question, answered by layers

The same method answers a worry about every greenfield project. If construction takes four years and the technology changes every two, is a new build obsolete before it opens?

Divide the building by how fast each layer ages. Epoch AI’s model does this when it gives the facility a 14-year life and the IT equipment 5.

• Interconnect and substation age over decades. Arguably they increase in value, because the queue behind them gets longer.
• Shell and structure last fifteen or more years if the original specification is generous.
• Power distribution and cooling plant last ten to fifteen years and carry real technology risk. 800-volt distribution and liquid cooling are now changing this layer.
• Compute turns over in five years on the books, and in less time in economic reality.

Seen this way, the obsolescence question goes away. Sites do not become obsolete. Fit-outs do, on schedule, by design, and everyone’s depreciation table already includes that schedule.

The real risk for a new build is not that its compute ages. That is certain and acceptable. The risk is an incorrect specification of the slow layers. An example is a shell designed for 40 kW air-cooled racks in a 132 kW liquid-cooled market. It has no floor loading, riser space or plant room to adapt. A competent developer therefore follows three steps:

• Overbuild the layers that you cannot change later, the interconnect above all.
• Install the pipe routes and the plant space that you do not need yet.
• Expect to gut and replace the fast layers twice in the building’s life. Each replacement means the site outlived another generation, which is good news.

This is why the mining conversion works at all. The miner bought too much of the one layer that never becomes obsolete. It spent too little on all the layers that do. The conversion buys the permanent layer and rebuilds the short-lived ones. A well-run greenfield site does exactly the same thing to itself, by choice, ten years in. The retrofit market is not a workaround. It makes the layered nature of these buildings visible, and it puts a price on it.

What to take away

1. Conversion prices only look irrational on a cost map. The buyer purchases an energized interconnect, the one component of a data center that nobody can manufacture or expedite.
2. Time has a market price, $3 to $4 million per megawatt per year. At that price, every viable conversion returns more in saved time than its fit-out costs.
3. Most mines still fail. The megawatts must arrive dense, the structure must accept liquid cooling, and interruptible power contracts must become firm.
4. Buildings are layers that age at different speeds: decades for the interconnect, fifteen years for the shell, five for compute. Specify the slow layers for the market after next, and plan to replace the fast ones.
5. “Obsolete on day one” is a category error. Fit-outs become obsolete on schedule. Sites, chosen well, only increase in value as the queue behind them grows.