Bitcoin computing power pivots to AI

Arseniy Grusha on miners pivoting to AI, costs, energy, and the road to AGI.

The crypto hype has cooled and “smart” money is streaming into AI infrastructure. Large miners are already refitting data centres for neural networks, and construction costs have soared many times over. 

How the global economy is being retooled—and whether it is time to write off traditional bitcoin mining—was discussed on Podcast Society by Arseniy Grusha, founder of the Dataprana data centre in the United States.

From mining to AI

ForkLog (FL): Are we really witnessing the end of the crypto-mining era, or is this just another transformation? 

Arseniy Grusha (A. G.): Definitely not the end. Mining is not going anywhere, but the industry is transforming sharply. Over the past five years we have seen industrialisation and the arrival of large public companies, which triggered a rally in Bitcoin network difficulty.

Today AI giants have entered the market and need enormous capacity. Returns in AI look far better, so miners are massively converting their data centres to new tasks. This process will only accelerate.

FL: How realistic—and how hard—is it to convert mining data centres for AI needs?

A. G.: It is very hard and very expensive, because the designs are entirely different. Miners’ main asset is pre-allocated power and a legal grid connection. Everything else has to be built from scratch, with a different approach.

The cost gap is enormous: a bitcoin data centre averages about $400,000 per MW, whereas an AI facility runs around $10 million per MW. The 20–25-fold increase stems from stringent requirements for 100% redundancy. AI needs full backups: massive batteries and diesel generators ready to take over the load within seconds.

The design and liquid-cooling systems for GPU capacity also grow more complex. All these systems require professional engineering and strict certification to ensure uninterrupted operation.

FL: Are backups needed to ensure uninterrupted operation? 

A. G.: Exactly. The key is uptime. It must be 99.9999%, allowing just a couple of minutes of downtime a year.

FL: So miners’ only real advantage is access to power, and everything else is built anew?

A. G.: Yes. The crucial advantage is access to large volumes of electricity. In the United States utilities move slowly, and bringing in 100 MW from scratch takes a couple of years.

Big IT firms such as Microsoft, Google and Amazon pay vast sums for speed to win the AI race. It is therefore cheaper for them to use ready-made mining sites and refit them for their needs.

FL: In numbers, how much more attractive is it for a miner to switch to artificial intelligence?

A. G.: At Dataprana we do colocation hosting and invest in infrastructure rather than ASICs, because hardware depreciates quickly. As a business, traditional mining hosting can pay back in three to four years, which is very fast.

AI data-centre investments are roughly 25 times more expensive. Hence their payback can stretch to six or seven years.

FL: So the bet is more long term?

A. G.: Yes, but the difference is also in scale. Mining is a niche business with high volatility because it depends on Bitcoin’s price, and contracts are at most a couple of years.

AI is a global market with long contracts. A ten-year deal with, say, Microsoft provides stability that makes it easy to raise bank financing for such builds.

FL: What happens to the Bitcoin network if miners move en masse into AI?

A. G.: By my estimates, public miners control about 30% of the Bitcoin network, and they are actively moving into AI. This year we will definitely see a decline in network difficulty and a drop in hashrate as large players begin shutting off old hardware.

The result will be a glut of mining rigs on the market and, naturally, falling prices for them. That equipment will still need hosting somewhere, so demand for traditional bitcoin data centres will not disappear.

In the end the situation will balance out, but over the next year I expect a gradual decline in difficulty. Hardware makers will also be forced into self-mining to use their unsold inventory.

FL: Bitmain seems to have already started cutting prices.

A. G.: Of course: at the current price and historically high difficulty, mining is not very profitable. The market should balance out over the next three to six months.

The best time to invest in bitcoin mining for the next four years will come in the autumn. By then equipment will be cheaper, the entry barrier lower, and it will be possible to make 300–400% on one’s investment.

FL: Are chip quotas a big problem now? Is it better for providers to buy directly from Nvidia with long lead times, or from resellers?

A. G.: The main GPU buyers are the “Magnificent Seven”, who control 80% of the market. They are buying everything to win the technology race, or at least not fall behind competitors.

The remaining 20% is split between small startups and brokers who profit from reselling scarce cards. Chips are advancing incredibly quickly, accelerating the traditional Moore’s law.

Despite new models, even older cards are holding their value for now because of colossal AI demand. Sooner or later the market will saturate and there will be a pullback, but when is unclear. In AI today the situation resembles crypto’s golden age: absolutely everything works and makes money.

Energy constraints

FL: Is electricity really a problem in the US and globally?

A. G.: The US does not lack generation—Texas alone produces 85 GW, more than all of Germany consumes. The problem is ageing grids that were not designed to deliver gigantic volumes of power to a single point.

Building data centres requires large-scale upgrades to substations and transmission lines, which takes years. To balance the load, authorities require AI companies to invest in their own generation.

There will be no global collapse; grid adaptation is gradual. But uninterrupted server operation is critical: if AI data centres go down in future, global payments and millions of workflows will stall.

FL: If there is no global problem and earthly resources suffice, why are Elon Musk and Sam Altman interested in putting data centres in space?

A. G.: Space offers two main advantages: automatic cooling and endless solar energy without atmospheric losses. On Earth, about 30% of a data centre’s electricity goes just to cooling.

The main drawback today is complex logistics and the high cost of lifting equipment into orbit. To build gigawatt-scale clusters there, rockets would have to launch as often as planes fly today.

I think we will get there in 20–30 years. For now humanity must learn to erect such unprecedented facilities efficiently here on Earth.

FL: We used to hear scare stories that an AI query consumes tons of water. How true is that?

A. G.: Cards are getting more efficient, but six months ago one standard ChatGPT query consumed about 3 W. So 300 queries cost a data centre roughly 10–30 cents, accounting for electricity and infrastructure amortisation.

It is not cheap, but unit costs will fall as we scale. Meanwhile demand for compute will grow by thousands of times, as today we are using perhaps one percent of AI’s potential.

As for water, data centres use closed-loop cooling. Water circulates continuously through pipes to cool chips rather than simply evaporating, so actual consumption is small.

FL: Has AI already hit a ceiling?

A. G.: Definitely not; we are at the very beginning. Everyone is now waiting for AGI—artificial general intelligence—that will surpass humans in absolutely all tasks.

Elon Musk predicts AGI as soon as this year; other experts say 2027–2028. Such a superintelligence will radically change the world and take over a huge share of human functions.

Economics and markets

FL: Is the US a safe harbour for infrastructure builders, or a battleground of tough regulation?

A. G.: The main problem in the US is bureaucracy: permits take a long time, and power connections can take years. Neighbours in remote areas are often against industrial sites, and their plots have to be bought out.

But regulatory policy is highly favourable. The state incentivises construction with tax breaks, because leadership in AI is foundational to US national security.

In the global technology race with China, America is currently winning by a wide margin. I hope that continues.

FL: When raising capital, do you see investors really pivoting towards artificial intelligence?

A. G.: Let’s be frank: today bitcoin mining is of little interest to investors. The industry has come to resemble traditional oil extraction with high barriers to entry; you can no longer make a quick x100 on your money.

All capital is flowing into AI, as this is a new industrial revolution. In 20 years routine tasks in every household will be done by robots, and we will fully delegate domestic functions to cloud agents.

FL: In your view, is AI already a bubble?

A. G.: There is definitely an economic bubble, and it will burst—that is a standard market cycle. The situation resembles the dotcom boom: Amazon’s shares fell from hundreds of dollars to $6, yet the internet still changed the world forever.

About 80% of companies will be washed out because they make blatantly useless products. The main players will survive the crisis and continue aggressive growth.

I expect the downturn to last about a year, but certainly not decades. AI and robotics are advancing so fast that they will quickly pull the economy out of recession.

FL: In a recent report by Citrini Research, they predicted an economic crash caused by AI. What do you make of that?

A. G.: That is a perfectly realistic scenario. Consulting firms will lose relevance, and routine white-collar functions will be automated. But AI itself will not replace people—it will replace those who cannot use it.

To soften the social shock, with 99.9% probability we will see a basic income introduced within the next five years. To fund it, governments will have to impose a hefty robot tax on corporations.

The economy will undergo a painful 10–20-year restructuring, accompanied by depressions. In the end the system will stabilise—hopefully in a way that benefits humanity.

The future

FL: How do you view the concept of sovereign AI? Will states build their own closed data centres, and will there be room for private business?

A. G.: Sovereign AI will emerge with 99.9% probability, as happened with a sovereign internet in China. Today user data are the prime currency and a powerful instrument of political influence.

Every country will want to store citizens’ information strictly within its borders. Governments will develop their own chips, AI tools and search engines.

Building such protected systems will take governments five to 20 years. The public sector always moves far more slowly than private business.

FL: Your top three technologies that will change the world in the next couple of years.

A. G.: First, AGI. It will replace many employees; its impact will be huge. Second, longevity. Medicine is progressing such that in the coming decades life expectancy may reach 120 years. Third, space. In 5–10 years we will travel there far more often, which will profoundly change civilisation’s mindset.

FL: Advice for entrepreneurs who want to build infrastructure rather than just write code.

A. G.: Infrastructure is harder than online businesses because it requires enormous investment in hardware that cannot simply be moved if things go wrong.

The main advice: build infrastructure only within a reliable legal framework. Do not go to unstable countries chasing cheap electricity if there is no guarantee your assets are safe.

Build where a business can scale legally, where the rules are predictable and where investors are not afraid to commit capital.

This interview has been edited for length. More insights in the full episode: