Recently, as a co-founder of Ethereum, Vitalik Buterin posted a power cost breakdown table on X, comparing Bitcoin mining and AI inference side by side. The table was quickly relayed by media such as Jinse Finance: in Bitcoin mining, the capital expenditure for the power equipment per Bitcoin is approximately $37,600, with electricity costs around $40,600. Together, both account for a significant majority of the total cost, with electricity costs representing about 43%, while hardware investment accounts for about 40%, and labor only takes up about 3.4%. For venues and other items, their proportions are single-digit leftovers. In contrast, Vitalik pointed out that the AI inference phase is almost entirely dominated by the procurement and depreciation of hardware like GPUs, with electricity costs constituting only a small fraction. The media even emphasized the stark difference in power proportions using “4% vs 43%.” This data and conclusion are still just Vitalik's personal analytical framework and a single media's estimated breakdown, rather than an officially verified study through cross-source validation, but it raises a sharp question: when we lay bare the true economic structure of power activities, does the simplistic narrative that all power usage is “extremely high energy-consuming” still hold?
Vitalik's Post: Cost Comparison Table of Bitcoin and AI
To bring the debate back to concrete numbers, Vitalik recently shared a cost breakdown table on X, placing Bitcoin mining and AI inference side by side on the same “power bill.” The most striking part of the table is the Bitcoin column: according to Jinse Finance, the capital expenditure per Bitcoin for power equipment is about $37,600, accounting for about 40% of the total cost; while the electricity costs are even higher, around $40,600, accounting for 43.3%. Labor costs occupy just a small corner—approximately $3,200, or 3.4%, while the proportion of venue rent is negligible, and not even fully captured in the media's version. Overall, the combined electricity and hardware capital expenditures for Bitcoin mining account for 80%–90% of the cost, with only these two items being able to create a real gap in the table.
On the other hand, the AI inference column presents almost a reverse story. Vitalik's core judgment is that in the AI inference scenario, costs are almost entirely dominated by capital expenditures on high-performance hardware like GPUs, while electricity is just a minor component. Jinse Finance directly pointed out in the title—electricity costs account for about 4% of the total AI inference costs, significantly lower than the 43% electricity cost proportion in Bitcoin mining. The specific dollar amounts of other cost items have not yet been disclosed, but the structural trends are already very clear. Through this “Bitcoin vs AI” cost comparison table, Vitalik did not target a specific industry, but rather a rough narrative that views all power activities as energy black holes, forcing people to re-examine the actual dimensions in which the cost structures of different power economies truly differ.
The Truth in Miners' Ledgers: Electricity and Hardware Share Equal Burden
In the world of proof of work, the miners' ledgers have long been locked into two primary cost lines: one for power equipment and one for electricity. Bitcoin uses a PoW mechanism; whoever wants to participate in the bookkeeping to receive block rewards must run a large number of machines 7×24 hours non-stop, directly locking the cost structure into two steps: “first invest in buying equipment, then continuously consume electricity.” Capital expenditures on power equipment include not only the mining machines themselves but also cooling systems, power supply equipment, and financing costs behind them. These become long-term burdens as soon as they are deployed at the mining site; while electricity is an operational expense that must be paid daily—if the machines stop, all the capital invested on the books immediately becomes meaningless.
The breakdown table provided by Vitalik lays this structure bare: according to Jinse Finance, the capital expenditure for power equipment per Bitcoin is about $37,600, accounting for about 40% of costs, and electricity costs about $40,600, making up approximately 43.3%, together constituting the majority of overall costs. Media further summarizes that hardware and electricity combined account for about 80%–90%, with labor and venues being single-digit proportions, almost pushed to the edges in this equation. The reason is not complex—PoW's competition is essentially a competition of power capacity and electricity prices. Miners will continuously optimize equipment efficiency and migrate to cheaper power sources over the long term, until the electricity costs and equipment depreciation create a balanced pressure on the books, allowing both cost lines to be roughly equal, while labor and venues are forced to be on the edges, becoming variables with very little influence in this battle for power.
AI Inference Almost Entirely Borne by Hardware
High computational demand does not equal high power consumption; Vitalik highlights this point regarding AI inference. Inference refers to putting a model into the real world after training is completed: receiving user text, images, or environmental signals, running the network based on given parameters, and generating answers, images, or decision outcomes. Unlike the training phase where parameters are repeatedly adjusted and gradients are backtracked, inference is more like continuously using an already assembled machine—there's no payment for “learning,” but rather for “continuous usage.”
At this point, the cost structure changes entirely. Vitalik points out that the core expenditure of inference is the prior acquisition and deployment of a large number of GPUs and other high-performance computing hardware, with subsequent depreciation and capital occupation almost becoming the main focus of the entire table. According to Jinse Finance, the electricity costs in that comparison table for AI inference only account for about 4%, while electricity costs for Bitcoin mining are as high as about 43%, presenting almost a reverse proportion structure in similarly computationally intensive scenarios. The reason lies in that inference typically processes requests in bulk at high utilization rates, and the electricity consumed during a single inference is just a decimal figure before the expensive hardware; miners must keep their equipment operating at full capacity to convert electricity into a continuous stream of hashes. Therefore, in the Bitcoin world, electricity prices are almost as vital as machine prices, while in the AI inference ledger, the real heavy burden is the entire row of GPU cabinets themselves.
High Power Consumption Narrative Challenged: How Will Regulation and Public Opinion Change?
For a long time, Bitcoin mining has become synonymous with “high power consumption,” facing regulatory pressure and public criticism in multiple countries and regions. The common logic in public discussions is that any computationally intensive activity must inevitably be accompanied by massive electricity consumption. The cost breakdown table Vitalik proposed directly contrasts sharply with this point — according to a single source estimate, Bitcoin mining electricity costs about $40,600, accounting for about 43.3% of total costs, while capital expenditure for power equipment is about $37,600, accounting for about 40%; whereas within the same framework, AI inference costs are almost entirely dominated by hardware capital expenditures, with electricity costs constituting only about 4%. When the electricity cost proportions for these two activities shift from “nearly half” to “single digits,” the instinct to bundle all computing power activities together as energy monsters is forced to step back.
If these numbers continue to ferment in policy circles and the media, the immediate effect will be that energy regulation and industry compliance discussions will find it difficult to continue using a template copied and pasted from Bitcoin mining. In mining, decision-makers focus on electricity prices, load, and grid pressure; while in AI inference, the real focus may be more on the scale of equipment investments and capital turnover, rather than consumption per kilowatt-hour. On the public opinion front, criticisms of computing power may also shift from “energy consumption is original sin” to questioning cost structures and profit distributions: who is paying the electricity bill, who controls the hardware, and who bears the resulting social costs. However, so far, there has been no systematic policy response reported regarding this table, and Vitalik's discussion remains more of a provocation in terms of framework and perspective. It is also necessary to be cautious that this comparison is based on a single-source estimate, with samples and methods not disclosed, especially since the dollar figures for venue rents and other costs of AI inference are not complete. Under these premises, any energy policy and compliance inference can only be regarded as a signal rather than a directly executable conclusion.
From Bitcoin to AI: The Imaginary Space of Next-Generation Power Economics
When Bitcoin mining sites and AI data centers are placed in the same cost table, the focus of the story subtly shifts: in the proof of work world, “half electricity costs, half mining machines” has almost become common knowledge, with the electricity expenditures and capital outlays for power equipment in Bitcoin mining together accounting for a large portion of the overall costs; while on the AI inference side, according to Vitalik's personal analysis and media reports, the narrative flips to “almost entirely mining machines, very little electricity,” with electricity cost proportions in single digits, and the actual burden being the upfront purchase and depreciation of GPU clusters. This structural difference implies that the layout of the infrastructure for the next generation of power economy is no longer merely about who can get cheaper electricity, but rather a long-term competition among hardware investment scale, capital costs, equipment utilization, and energy efficiency: miners tend to bet on continuous power supply and power density, while AI service providers must extract every bit of effective load from each GPU in their capital-intensive data centers. For this reason, how hardware investment and energy use are rebalanced is directly related to the sustainability of the industry—whether to continue amplifying “buying mining machines” or to optimize “using mining machines” will decide the next reshaping of the power landscape. However, all of these deductions have a clear premise: Vitalik’s table comes from a single estimate, with samples and methods not disclosed, and specifics regarding venue rents and other costs related to AI inference being missing. On such a data foundation, any grand narrative can only be viewed as an assumption rather than a conclusion. Therefore, what is genuinely needed is more independent and transparent cost breakdown research to ensure that public discourse about Bitcoin and AI stands on a credible ledger before discussing the long-term trajectory of power economics.
Join our community, discuss together, and become stronger!
AiCoin Exclusive Hyperliquid Benefits: https://app.hyperliquid.xyz/join/AICOIN88
AiCoin Exclusive Aster Benefits: https://www.asterdex.com/en-US/referral/9C50e2
On-chain Telegram Community: https://t.me/AiCoinWhaleData
On-chain Community: https://www.aicoin.com/link/chat?cid=N6OVMor5g
AiCoin On-chain Twitter: https://x.com/aicoinwhaledata
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到[email protected]，本平台相关工作人员将会进行核查。
