Author | Lin Wanwan

Editor | Sleepy.txt

Initially, no one expected that the true bottleneck of AI is not capital, not large models, but electricity.

With large-scale training running at full capacity for extended periods and AI inference operating 24/7, a problem emerged: insufficient electricity, forcing chips to sit idle. Over the past decade, the United States' electrical grid infrastructure has relatively lagged behind, with new large loads connecting to the grid taking 2–4 years on average. This has made "readily available electricity" a scarce commodity across the industry.

Generative AI has brought a raw and harsh reality to the forefront: it's not the model that's lacking, it's the electricity.

As a result, the story took a turn, and cryptocurrency mining companies, the earliest to consider electricity as a "means of production," began transitioning from the periphery to the center stage of capital.

Iris Energy (IREN) is an exemplar of this trajectory. This year, IREN's stock price skyrocketed nearly 600% at one point, ranging from $5.12 to $75.73 within a 52-week period. Seizing the opportunity while Bitcoin's surge remained enticing, it decisively allocated power to transform its self-built AI data center.

When tech giants like Microsoft waved $9.7 billion worth of long-term orders, the market for the first time visually understood the real path from mining to AI: starting with electricity and land, then GPUs and customers.

However, not all mining companies, like IREN, chose to bet their entire fortune on AI. Amid this power-driven hash rate migration, there is a steady force worth our attention—Bitdeer.

Bitdeer Technologies Group (NASDAQ: BTDR), founded by the crypto legend Wu Jihan and headquartered in Singapore, strategically holds nearly 3GW of global electrical resources, steering clear of the shallow trap of relying on others for "power supply" from the start. As the wave of AI approaches, Bitdeer did not opt for IREN's aggressive "All-in" approach but retained profitable Bitcoin mining as its "cash cow," while prudently upgrading some mining farms into AI data centers.

This "advance while holding ground" strategy has made it the best example to observe how global players are contemplating and strategizing in this hash rate race.

To delve deeper into this, we interviewed Wang Wenguang, Vice President of Global Data Center Business at Bitdeer, hoping to shed light on the current global AI electricity shortage and understand their perspective on mining companies transitioning to AI data centers: do they view it as capital speculation or a genuine need for AI? In this in-depth conversation, we explored a series of questions on this topic.

Why is the Power Outage Situation So Severe in the United States?

Watcher: Let's start with a fundamental question, do you think electricity prices will continue to rise in the future?

BitDeer: I think they will, because this is a critical supply-demand relationship for the future.

Watcher: Regarding the power outage situation in the United States, there is a belief in the market that it is difficult to obtain an "electricity license" in the U.S. Is that true?

BitDeer: It's not that this so-called "electricity license" is hard to get approved, but rather that the physical expansion speed of the power grid cannot keep up. In the many years since heavy industry moved out of the U.S., the U.S. power grid construction has not been systematically expanded. When mining companies moved into the U.S. in 2021, a lot of "already grid-connected, already signed PPA" power was locked up by mining companies. With the influence of ChatGPT, pure AI players coming in, they only then discovered a large amount of immediately usable power in the mining field.

This explains why large factories are willing to partner with mining companies. Instead of waiting 2-4 years to build 500MW from scratch, it's better to renovate the existing site in 12 months.

Watcher: When did the industry truly realize that "inference is also power-intensive"?

BitDeer: Probably after the widespread adoption of GPT-4. As companies embed models into customer service, office work, search, risk control, etc., the long-term and scenario-based inference demand has not decreased as initially envisioned.

This has led to two types of changes.

One is engineering upgrades: from stronger air cooling to liquid cooling/hybrid heat dissipation, cabinet power, power distribution paths, fire protection, and monitoring have all been raised to the level of AI data centers.

The other is resource strategy: electricity has truly become the number one bottleneck. People are no longer just talking about "buying cards," but are instead focusing on obtaining electricity and grid connection upfront, long-term PPAs, grid connection schedules, cross-regional capacity backups, and, when necessary, upstream electricity acquisition like mining companies (self-generation/direct procurement).

In fact, in the mining industry, we have long seen the same trend. Chips can scale infinitely (silicon comes from sand), but electricity cannot. We have used natural gas self-generation in Canada to power mining farms, and that is the logic. Today's AI is almost the same.

Watcher: What are the differences in power consumption scale between AI data centers and traditional internet data centers?

Bit Deer: It's not about incremental change, but about a magnitude change. In the past, a 20-30 MW traditional Internet data center was considered large, but today AI data centers have power demands of 500MW or even 1GW. AI has transformed data centers from a "rack business" to a "power engineering" business, requiring a reevaluation of everything: from power lines, substations, cooling systems, fire protection, redundancy, to PUE... The experience from traditional Internet data centers is still useful, but no longer sufficient.

Insight: Why has "electricity" become the scarcest upstream resource?

Bit Deer: Chips can be scaled because of silicon and capacity management; however, electricity is difficult to scale because it comes from power generation and grid upgrades. In the past, the mining industry has attempted to "source energy upstream," including self-power generation projects in Canada; AI is following a similar path—whoever secures electricity first gets the deployment advantage.

AI New Battlefield: From "GPU Grab" to "Grid Grab"

Insight: How do mining companies transform into AI data centers, and what specific changes are required? Previously, it was suggested that "Bitcoin mining power could be used for AI," but mining chips (ASIC) and the GPUs required for AI are not compatible. So why can mining companies now "provide AI computing power"?

Bit Deer: The global mining industry was once divided, with Bitcoin relying on mining ASIC chips, which are efficient but have a single purpose; Ethereum relied on NVIDIA GPUs, which are versatile but have exited the mining stage after transitioning to PoS.

Therefore, when the market talks about "mining farms transitioning to AI" today, it almost always refers to Bitcoin mining farms undergoing transformation. The key point is that mining farms are no longer "hashing" but upgrading themselves into AI data centers.

This involves infrastructure upgrades: removing ASIC racks and replacing them with GPU servers; upgrading the "just enough" power system to a professionally redundant power distribution system with N+1/2N redundancy; upgrading from traditional air-cooling to a cooling system capable of supporting high-density GPUs; and standardizing and making facilities such as room sealing, dust prevention, fire protection, and auditability.

After completing these four steps, a crypto mining farm transitions from a "mining workshop" to an "AI data center."

Why can mining companies self-build faster than AI giants? Electricity.

AI is a business of "electricity and heat," with an AI data center construction timeline of 3-4 years, making time costs the biggest barrier. Mining companies happen to hold these "hard assets," so their transition starting line is further ahead.

Insight: In the past few days, Microsoft and Amazon have successively signed multi-year AI contracts with crypto mining companies. Iris Energy (IREN) signed with Microsoft for a total value of 9.7 billion, spanning 5 years; another company, Cipher, signed with Amazon Web Services for 5.5 billion over 15 years. This is seen as the first batch of cases where mining farms collaborate with tech giants. What is your take on this?

BitDeer: Iris Energy is a forward-thinking Australian company that has been mining in the United States for a long time.

Iris Energy's pivot towards AI is like a signal flare. At a high point in the Bitcoin price, while peers are still expanding mining operations, it diverted some of its power to invest in a self-built AI data center. As a result, AI enterprises actively approached it.

The real tipping point came from the Hyperscalers' real money—such as Microsoft's approximately 9.7 billion commitment. For the first time, the market clearly saw that between mining companies and tech giants, it is not just "technology integration" but "an exchange of power and time."

The heat of AI amplifies infrastructure needs, opening up collaboration opportunities.

Insight: Why are top mining companies more easily chosen by American AI giants at this stage?

BitDeer: Because of "available power + engineering delivery speed." The site selection and grid connection in the previous cycle of mining companies have now become the upfront capital for AI data centers. Time is the greatest discounting factor, directly determining who can go live within the window, acquire customers, and generate rolling cash flow.

Insight: Is the land selection requirement for AI data centers challenging?

BitDeer: Overall, not very. In the United States and most countries, what is truly scarce is electricity, not land.

The reason is simple: places with ample electricity supply are mostly energy-rich areas (natural gas fields, coal mining belts, near hydroelectric power stations), sparsely populated, and with low land prices.

For example, BitDeer's large data centers in Norway and Bhutan are located far from population centers, with concentrated power resources and low land costs. Similarly in the United States, these industrial parks are not in urban cores but more remote areas where land is easily available and affordable. The "first principle" of site selection is electricity and grid connection, with land usually following the power, not the main bottleneck.

Observer: AI is now being referred to as the upstream business of "steel, electricity, and land," even being compared to another form of real estate. What is your take on this?

BitDeer: After the emergence of large models, the power consumption of AI far exceeds most people's expectations.

Initially, everyone thought "training consumes power, while inference is light," but the reality is the opposite. After inference became popularized, such as with ChatGPT and DeepSeek, with more and more end-user access, the underlying power consumption of inference has remained high.

From an engineering perspective, AI is fundamentally a resource-intensive industry:

· On the chip side: during training, the accelerator card is running at almost 100% load, naturally high power consumption;

· On the data center side: the heat density is much higher than traditional servers, PUE is significantly pushed up, and cooling itself also consumes a large amount of power;

· On the scale side: the power demand of AI data centers has leaped from the 20–30MW of traditional internet data centers to the level of 500MW or even 1GW, which was almost unimaginable in the era of traditional internet data centers.

So, comparing it to "real estate" is only half correct. Indeed, it requires land, factory buildings, and a long cycle (construction periods often take 3–4 years), but its life and death are determined by electricity and heat, whether it can timely access a large capacity power grid, implement N+1/2N redundancy, and efficient cooling. In this respect, its strong dependence is very similar to steel, electricity, and land.

What are the characteristics of AI data centers?

Observer: What is the characteristic of the model for constructing data centers in the United States?

BitDeer: Due to power constraints and historical paths, Hyperscalers in the United States often need to personally intervene and cooperate with mining companies to obtain available power.

Observer: Is it possible for foreign companies to build AI data centers in the United States?

BitDeer: Simply put, AI data centers are a strongly regional business. For truly massive deployments involving hundreds of megawatts and thousands of cards, the leading players are still domestic giants in the United States. We are only discussing AI data centers here, not traditional internet data centers.

Insight: Will AI Data Centers evolve into a geopolitical tool? Will this affect your decision-making?

BitDeer: I agree with this assessment.

At the core of AI lies data, which is inherently subject to sovereignty and security constraints. To prevent data leaks and security risks, regions around the world are tightening relevant policies: even though the U.S. allows foreign investment to build data centers, as AI gains more and more data, most countries will likely move towards "on-site deployment, local compliance, and data localization."

In simple terms, American AI stays in America, Middle Eastern AI stays in the Middle East, European AI stays in Europe, and regionalization will be a long-term trend.

Industry Landscape and Potential

Insight: Besides IREN and BitDeer, who in the mining industry has more potential to transition to AI data centers?

BitDeer: To see who has a chance, first look at whether they have access to significant electricity supply, and then see if they can quickly convert their mining farms into GPU data centers. Those with on-site power supply, land, substation, capable of N+1/2N redundancy, liquid cooling/high density, are the most likely to attract AI contracts.

On the other hand, those with pure hosting/light asset models, without control over electricity and industrial parks, will have a passive transition to AI data centers.

In the U.S., companies like Riot, CleanSpark, Core Scientific, TeraWulf, Cipher, with self-owned resources and reliable expansion capabilities, are more likely to be targeted by major players.

So the conclusion is straightforward: electricity is the ticket, and transformation capability is the speed; only when both are in place can you stay ahead.

Overall, it is crucial to see who controls "high-quality, sustainable large-load available electricity." For example, companies with more self-owned on-site power resources have more potential; in this round of structural transformation, models primarily based on hosting, lacking self-owned energy and industrial park, are at a disadvantage.

What Is BitDeer Thinking?

Insight: What is BitDeer's strategy and path for "Mining to AI"?

BitDeer: Mr. Wu Jihan's strategy has always been to cover the entire industry chain. BitDeer holds approximately 3GW of power and industrial park resources, which is our biggest underlying advantage.

When we initially entered the AI field, we did not anticipate that "electricity" would become a core bottleneck. Therefore, we started with a self-built approach: we partnered with NVIDIA and became an NVIDIA PCSP. We deployed a small-scale H100 cluster in Singapore, launched our proprietary AI Cloud, and offered external training services. This project has been successfully implemented.

Subsequently, we also set up a second data center in Malaysia. As Hyperscalers entered this field and began collaborating with mining companies, we concurrently upgraded high-load campuses to AI data centers. It has been announced that we will transform the entire 180MW site in Norway into an AI DC and convert the approximately 13MW site in Washington State, USA.

Ultimately, the essence of AI is very similar to Crypto mining—both are businesses involving "electricity + infrastructure." We have end-to-end capabilities in power supply, campuses, and computing power operations, making the transition to AI relatively smooth.

Insight: What is the core difference between BitDeer and other mining companies like IREN?

BitDeer: Three points. Firstly, we will not completely transform into an AI enterprise. Based on calculations, the current stage of Crypto Mining is still more profitable than AI data centers, and the mining industry has stable cash flow and better returns.

Our second advantage is our international engineering team organization capabilities. BitDeer's engineering team's organization and execution capabilities are unmatched worldwide. For the same AI data center, a typical pace in the US takes around two years, but we usually achieve it in a year and a half. This is achieved through parallel advancement and supply chain coordination, aligning key aspects such as construction, electromechanical work, power distribution, and cooling simultaneously. We compress the usual 24-month cycle to about 18 months, establishing usable capacity more quickly.

Thirdly, the company's strategy remains conservative: the AI industry is very young, even younger than Crypto, so we do not go "all-in" and instead pursue a more sustainable development pace.

Insight: Where is BitDeer's current power infrastructure primarily located?

BitDeer: BitDeer is currently mainly positioned around the world with approximately 3 GW of power and related infrastructure, covering five countries: the United States, Canada, Norway, Ethiopia, and Bhutan. This infrastructure supports the construction and operation of mining and AI data centers.

Cost and Financing

Insight: I saw a Goldman Sachs report mentioning that building an AI data center could cost around $12 billion. Is it really that expensive?

BitDeer: Indeed, it is a significant amount, on the scale of "tens of times." Here's an easy-to-understand comparison in "layman's terms":

Bitcoin Mining Farm (USA): Building 1 MW costs approximately $350,000–$400,000, but building 1 MW of an AI data center costs around $11 million. This is because the investment in an AI data center is a complex combination of "heavy mechanical and electrical work" plus high standards, along with grid connection waiting, environmental/energy assessments, regional compliance, with a typical timeline of 18–36 months.

You will find that the essence of an AI data center is not just "buying more cards," but rather connecting a piece of land to form a "city of electricity" that can consume 500MW–1GW, ensuring proper electrical connections, heat dissipation, redundancy, compliance, all of which are very costly.

Insight: Where does the money come from? Is financing needed?

BitDeer: Honestly speaking, financing is required in all cases.

Let me share some common financing methods used in the industry:

1. Project Financing/Infrastructure Loans: Using the industrial park and equipment as collateral, relying on long-term leases or hashrate offtake (customer commitment to buy your hashrate for several years) to reassure the banks.

2. Equipment Leasing/Sale-Leaseback: Leasing out GPUs and some mechanical and electrical components, spreading the cost over a long period, avoiding the need to pay a large sum of cash upfront.

3. Long-term PPA (Power Purchase Agreement): Locking in electricity prices and available capacity first to make bondholders agree to lower interest rates.

4. Partnering with Major Companies: Large clients/companies provide minimum consumption guarantees, prepayments, guarantees, or even joint ventures (JV), allowing you to obtain cheaper funding.

In partnerships such as IREN, CoreWeave, and Google/Microsoft, you can see shadows of these terms.

Insight: Does BitDeer also need financing? Will there be any announcements soon regarding partnerships with major companies?

BitDeer: I can't reveal much about this publicly at the moment.

Conclusion

Shortly after the interview, BitDeer revealed its next step in the capital market.

On November 13, BitDeer announced that it would raise $400 million through the issuance of convertible preferred notes, and grant the initial purchaser the option to subscribe for an additional $60 million of notes within 13 days, with a maximum fundraising size of up to $460 million. The new funds will be used for data center expansion, ASIC miner development, AI and HPC cloud business expansion, and general corporate purposes.

In an era where electricity has become the most scarce upstream resource in the AI industry, how and where this $460 million will ultimately be invested, and how many megawatts of new load will be connected, will largely determine BitDeer's position in the next round of hashing power competition.

For BitDeer, this money is more like writing the assessment from an interview into the balance sheet: one end is connected to the cash flow bedrock of the mining industry, the other end is connected to the long slope and thick snow of the AI data center business line. It may not immediately reflect in the revenue and profit of the next quarter, but it will slowly reshape the power structure of the hashing power business in the coming years—deciding who is qualified to sit at the negotiation table and who can only wait in line on the grid connection list for electricity.

Looking ahead from the results, the story of this round of AI infrastructure is not complicated: electricity has truly become the upstream, time has become the new currency, and the industrial parks and grid connection indicators in the hands of mining companies have become "old assets" that others cannot buy even with money.

As the noise about models and applications gradually fades, the market is highly likely to go through the ledger once again: who tells a compelling narrative is no longer important, but rather the company that can connect every megawatt of electricity and run steadily in a power-hungry world is qualified to stay at the next stage of the table.

Welcome to join the official BlockBeats community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Official Twitter Account: https://twitter.com/BlockBeatsAsia