I still remember the first time I really sat down with the numbers behind crypto mining and realized how much of the story sits in the power bill. Then along comes a fresh comparison from Ethereum co-founder Vitalik Buterin that puts those figures next to the cost structure of AI inference, and suddenly the picture looks even more interesting. Electricity eats nearly half of Bitcoin mining expenses, while artificial intelligence workloads are almost entirely driven by the capital spent on computing hardware. That contrast feels worth unpacking properly.
Why This Cost Split Matters Right Now
In mid-October 2026 Vitalik shared a pair of tables that make the difference hard to ignore. Bitcoin mining, according to the breakdown he posted, lands at roughly a 50/50 split between computing capital expenditure and electricity. AI inference, by contrast, shows computing capital expenditure taking about 76 percent of the total. The rest of the line items, cooling, power equipment, financing, labor, rent and miscellaneous costs, fill in the gaps, but the headline numbers already tell a clear story.
What struck me most is how cleanly the two activities diverge once you look past the surface. Both need energy and both need silicon, yet the relative weight of those two inputs sits in completely different places. I’ve found that understanding this kind of structural difference helps anyone thinking about infrastructure investments or long-term technology bets see where the real pressure points lie.
Bitcoin Mining: Electricity Takes Almost Half the Bill
Let’s start with the Bitcoin side of the table. The figures are presented on a per-Bitcoin basis and total roughly 93,700 dollars. Electricity supplied directly to the computers comes in at 40,600 dollars, or 43.3 percent. Another 2,400 dollars, or 2.6 percent, covers other electricity costs. Add those two together and you get 43,000 dollars, which works out to 45.9 percent of the entire estimated cost of producing one Bitcoin under the assumptions shown.
Computing capital expenditure sits at about 37,600 dollars, roughly 40 percent. That category is broader than the sticker price of the mining machines themselves. It folds in hardware, cooling systems, power equipment and the financing costs attached to all of it. Labor lands at 3,200 dollars, facility rent around 2,500 dollars, and a mixed bag of remaining expenses, maintenance, spare parts, insurance, taxes, bandwidth and general operations, adds another 7,400 dollars or so.
These numbers do not claim to represent every miner on the planet. Local electricity rates, machine efficiency, climate and financing terms all move the needle. Still, the overall shape is useful. Roughly half the cost of creating a new Bitcoin, under this model, flows out through the power meter. That reality has shaped mining geography for years. Operators chase cheap power, and many have become increasingly sophisticated about securing long-term energy contracts or even co-locating with generation assets.
Bitcoin mining cost is split about 50/50 between compute capex and electricity. Meanwhile, AI inference is almost completely dominated by compute capex.
That short summary from Vitalik captures the essence. Once you dig into the supporting table, the electricity share settles a little under 46 percent, close enough to the round “half” description for most practical discussions.
AI Inference: Capital Expenditure Dominates the Picture
The AI table uses a different unit, dollars per GPU-year, and paints a markedly different cost landscape. Computing capital expenditure comes in at approximately 19,900 dollars, or 76 percent of the total. Again the category includes the hardware itself plus cooling, power equipment and financing. Electricity supplied to the computers is only 1,050 dollars, or 4 percent. Other electricity adds 210 dollars, bringing the combined power share to about 4.8 percent.
Labor contributes 1,400 dollars, facility rent roughly 900 dollars, and the remaining maintenance, insurance, tax and bandwidth items total around 2,800 dollars. The listed figures sum to about 26,260 dollars per GPU-year. Because one table measures cost per Bitcoin and the other per GPU-year, you cannot simply compare the absolute dollar amounts and declare one activity more expensive than the other. What you can compare is the relative weight of each cost category, and that is where the contrast becomes sharp.
AI inference, at least under these assumptions, is almost completely dominated by the capital locked up in computing infrastructure. Power is a real expense, of course, but it is a relatively small slice of the pie. The big money sits in the GPUs, the supporting gear, and the cost of capital needed to acquire and keep them running at scale.
How the Two Cost Structures Shape Business Decisions
This difference in cost structure is not just academic. It influences how companies finance projects, how they negotiate contracts, and even how they think about risk. Bitcoin miners have long treated electricity as the primary variable cost that must be managed tightly. Many have become experts at finding stranded or low-cost power. Some have even developed hybrid models that can switch between mining and other energy-intensive activities depending on market conditions.
AI infrastructure operators face a different set of priorities. Because capital expenditure is such a large share of the ongoing cost of running inference workloads, the ability to secure high-utilization rates on expensive GPUs becomes critical. Idle silicon is expensive silicon. That reality pushes operators toward long-term customer commitments, careful capacity planning, and financing structures that spread the heavy upfront investment over many years.
I have watched several former mining companies pivot toward AI hosting in recent years, and the cost comparison helps explain why the transition is not always straightforward. The skill set of hunting cheap power remains valuable, yet the capital intensity of GPU clusters and the need for denser, more sophisticated facilities change the financial equation. Some operators have responded by keeping the real estate and power infrastructure in-house while letting customers own or lease the actual servers and processors. That division of responsibility shows up clearly in several recent deals.
Real-World Examples From Data Center Deals
One company that has drawn attention is TeraWulf. In the middle of 2026 it was exploring a potential 3.5 billion dollar financing package to support an AI campus in Kentucky. Plans reportedly included leveraged loans and high-yield bonds. The financing discussions followed a 20-year lease agreement with Anthropic at the Justified Data campus in Hawesville. The facility is expected to deliver roughly 401 megawatts of critical computing capacity and could generate about 19 billion dollars in revenue over the initial contract term.
Initial capacity is targeted for the second half of 2027, with full deployment aimed at early 2028. The revenue figure covers two decades and is not an upfront payment. The company’s chief financial officer has described an operating model in which customers remain responsible for servers, processors and technology upgrades, while TeraWulf focuses on power and physical infrastructure. Earlier, a subsidiary raised 3.2 billion dollars through senior secured notes carrying a 7.75 percent coupon and maturing in 2030.
Other firms have used crypto holdings to help fund AI data-center expansion. In late July one subsidiary sold 1,000 ETH for roughly 1.9 million dollars after fees to support Japanese AI data-center plans. Another company monetized about 100 BTC and set up a Bitcoin-backed credit facility for construction work in Michigan. Its initial customer deployment covered approximately 20 megawatts under a ten-year agreement with two optional five-year extensions, and the company estimated more than 1.2 billion dollars in revenue over the maximum term.
MARA announced a definitive agreement to acquire a 1,200-acre property in Texas intended to support high-performance computing, flexible computing services and Bitcoin mining. The site is expected to offer an initial gigawatt of grid capacity by October 2027 and up to two gigawatts by April 2028. Development is planned in phases with a partner, subject to regulatory approvals. The seller is set to retain a minority ownership interest once a high-performance computing lease is signed.
These deals illustrate a common pattern. Companies with experience in large-scale power and data-center operations are positioning themselves as landlords of compute capacity rather than pure owners of the silicon. That model aligns with the cost structure Vitalik highlighted. When capital expenditure on hardware dominates, there is clear value in specializing in the parts of the stack that are harder to move, power delivery, cooling, land and long-term energy contracts, while letting customers manage the rapidly evolving GPU layer.
What the Numbers Suggest About Future Infrastructure
Looking ahead, the divergence in cost structure may continue to influence how capital is allocated. Bitcoin mining remains sensitive to electricity prices and to the ongoing arms race in mining hardware efficiency. AI inference, by contrast, is more sensitive to the price and availability of advanced accelerators, to utilization rates, and to the cost of capital needed to keep those accelerators deployed at scale.
One practical implication is that improvements in energy efficiency matter for both activities, yet they deliver different relative benefits. A 10 percent reduction in power consumption helps a Bitcoin miner more than it helps an AI operator, simply because electricity is a larger share of the mining cost base. Conversely, longer hardware lifecycles, better financing terms, or higher utilization rates move the needle more dramatically for AI workloads.
I keep coming back to the idea that these two industries are becoming more intertwined even as their cost profiles remain distinct. Former mining sites offer ready access to power and land. AI demand is growing fast enough that operators are willing to adapt facilities originally built for ASICs. At the same time, pure-play AI companies continue to build greenfield campuses optimized from day one for dense GPU deployment. Both paths are visible in the market right now.
Digging Deeper Into the Categories Themselves
It is worth pausing on how the tables define computing capital expenditure. In both cases the category is broader than the purchase price of the chips. Cooling systems, power distribution equipment and financing costs are all rolled in. That choice makes sense from an economic perspective because those elements are tightly coupled to the decision to deploy the hardware. Yet it also means that improvements in cooling efficiency or cheaper financing can shift the reported capital share without any change in the price of the processors themselves.
Labor and facility rent appear in both tables at relatively modest percentages. For Bitcoin mining they total a bit over 6 percent. For AI inference they sit in a similar range. These are not trivial amounts, but they are secondary compared with the two dominant categories. The residual “other” bucket, covering maintenance, insurance, taxes, bandwidth and miscellaneous operations, is larger for mining than for AI in percentage terms, yet still secondary.
Perhaps the most interesting aspect is how sensitive the AI numbers would be to changes in hardware pricing or depreciation schedules. If next-generation accelerators deliver significantly more performance per dollar, or if operators can keep equipment productive for longer periods, the capital share could decline. Conversely, if demand for the latest silicon remains intense and prices stay elevated, the 76 percent figure could prove durable.
Putting the Comparison Into a Broader Context
Vitalik’s post is a snapshot, not a comprehensive industry survey. Different operators will show different numbers depending on their location, scale, power contracts and hardware mix. Still, the directional message is consistent with what many people in the industry already sense. Bitcoin mining is an electricity-heavy business. AI inference is a capital-heavy business. That distinction influences strategy, financing and even the kinds of partners each activity attracts.
In my experience, the most useful comparisons are those that force you to look past surface similarities. Both activities consume large amounts of electricity and silicon. Both have attracted significant investment. Both are sensitive to technological progress. Yet once you examine the cost structure carefully, the priorities diverge. Managing the power bill is existential for many miners. Maximizing the productive life and utilization of expensive accelerators is existential for AI operators.
That does not mean the two worlds will remain separate. Quite the opposite. The same companies often appear in both conversations. The same land parcels and power contracts can serve either purpose, or both. What changes is the internal allocation of capital and the metrics that management teams watch most closely.
A Closer Look at the Electricity Component
Electricity is the most visible cost for Bitcoin mining, and the table breaks it into two pieces. Direct power to the computers accounts for the bulk, 43.3 percent. The smaller “other electricity” line, 2.6 percent, likely covers facility lighting, cooling overhead that is not already captured in the capital category, and similar items. Together they approach half the total.
For AI the same two lines add up to less than 5 percent. That does not mean AI data centers are power-sipping operations. Modern GPU clusters draw substantial electricity. The difference is simply that the capital cost of the equipment is so high that power becomes a smaller relative share. In absolute terms both industries remain major electricity consumers, and both are increasingly focused on securing reliable, preferably low-carbon, supply.
Some operators have begun exploring co-location with renewable generation or with facilities that can provide both power and cooling. Others are experimenting with demand-response programs that allow them to curtail load during peak grid stress. These strategies appear in both mining and AI contexts, yet the economic incentive is stronger for the activity that has a higher electricity share of total cost.
Capital Intensity and the Financing Angle
Because AI inference is so capital intensive, financing terms matter a great deal. Longer depreciation schedules, lower interest rates, or creative structures that shift ownership risk can materially improve the economics. The tables include financing within the capital expenditure category, which is a useful reminder that the cost of money is part of the cost of compute.
Bitcoin mining has its own financing dynamics. Machine prices, secondary market liquidity and the ability to collateralize equipment all influence the true cost of capital. Yet because electricity is such a large ongoing expense, miners often focus first on securing favorable power rates and only secondarily on optimizing the financing of the hardware itself.
The recent wave of large bond and loan packages for AI campuses reflects this reality. Billions of dollars are being raised to build the physical and electrical infrastructure that will host customer-owned or customer-leased accelerators. The operators taking on that debt are betting that long-term leases will cover the fixed costs and leave room for profit. The customers, for their part, are betting that owning or controlling the silicon gives them the flexibility and performance they need.
Where the Two Industries Overlap and Where They Diverge
There is clear overlap in the talent pool and in the physical assets. Engineers who understand high-density power distribution and advanced cooling are valuable in both settings. Land with access to substantial grid capacity is scarce and therefore valuable to both. Yet the day-to-day operating metrics differ. A mining operation watches hash rate, power cost per terahash, and uptime. An AI inference operation watches GPU utilization, tokens per second, latency and the cost per unit of useful work.
Those different metrics flow directly from the different cost structures. When power is nearly half your cost, every incremental improvement in energy efficiency shows up clearly on the bottom line. When capital is more than three-quarters of your cost, every incremental improvement in utilization or hardware longevity matters more.
I’ve found that the most successful operators in either domain tend to be those who stay disciplined about their primary cost driver while still managing the secondary ones competently. Ignoring electricity in mining is a fast way to lose money. Ignoring capital productivity in AI is equally dangerous.
What Comes Next for Cost Structures
Technology does not stand still. Newer mining machines continue to improve joules per terahash. Newer accelerators continue to improve performance per watt and performance per dollar. Cooling techniques, from advanced liquid systems to immersion, keep evolving. Power markets themselves are changing as more renewable capacity comes online and as grids become more flexible.
Any of these developments could shift the relative shares shown in the tables. A major leap in mining efficiency could reduce the electricity percentage. A sharp drop in GPU prices or a lengthening of useful life could reduce the capital percentage for AI. Financing conditions, interest rates and investor appetite for long-duration infrastructure assets will also play a role.
For now, though, the snapshot Vitalik shared remains a useful reference point. Bitcoin mining sits near a 50/50 split between compute capital and electricity. AI inference sits closer to a 76/24 split, with capital taking the larger share. Those ratios help explain why the two industries, despite sharing many of the same physical requirements, often pursue different strategies and attract different kinds of capital.
The comparison also serves as a reminder that cost structure is strategy. The numbers are not just accounting details. They shape where companies choose to locate, how they finance growth, which risks they emphasize, and ultimately which business models prove sustainable over time. Watching how those structures evolve over the next few years will tell us a great deal about the future of both digital asset production and large-scale artificial intelligence.
In the end, the most valuable insight may be the simplest one. Two activities that look similar from a distance, heavy electricity use, dense compute, specialized facilities, turn out to have quite different economic centers of gravity. Recognizing that difference early is half the battle for anyone trying to navigate the intersection of crypto infrastructure and AI infrastructure in the years ahead.
The tables will be updated, the hardware will improve, and the financing terms will shift. Yet the core observation is likely to remain relevant for some time: Bitcoin mining lives and dies by the power bill, while AI inference lives and dies by the capital deployed in silicon. Understanding that contrast is a practical starting point for anyone trying to make sense of the capital flows currently reshaping both industries.
