Author: Lin Wanwan No one initially expected that the real bottleneck for AI would not be capital or large models, but electricity. Large-scale training and continuous full-load operation of AI inference, running 24/7, has led to a problem: insufficient electricity, forcing chips to sit idle. The United States' power grid infrastructure has lagged behind in the past decade, with new large-load grid connections often taking 2-4 years, making "readily available electricity" a scarce commodity across the entire industry. Generative AI brings a primitive and brutal issue to the forefront: the problem isn't a lack of models, but a lack of electricity. The story took a turn from there, and crypto mining companies, the first people to treat electricity as a "means of production," began to move from the margins to the center of the capital stage. Iris Energy (IREN) is a case in point. This year, IREN's stock price surged nearly 600% at one point, rising from $5.12 to $75.73 over 52 weeks. Even when Bitcoin's gains were still attractive, it boldly diverted power to renovate its self-built AI data center. When giants like Microsoft put up long-term contracts worth a total of $9.7 billion, the market gained its first intuitive understanding of the real-world path "from mining to AI": electricity and land come first, followed by GPUs and customers. However, not all mining companies are like IREN, choosing to bet everything on AI. In this massive migration of computing power driven by electricity, there is another steady force worth our attention—BitDeer. Bitdeer Technologies Group (NASDAQ: BTDR), founded by crypto legend Jihan Wu and headquartered in Singapore, possesses nearly 3GW of power resources worldwide, avoiding the superficial trap of relying on others for power from the outset. When the AI wave arrived, Bitdeer did not choose the aggressive "all-in" approach like Iren, but instead retained its lucrative Bitcoin mining as its core business while steadily upgrading some of its mining farms into AI data centers. This "offensive or defensive" strategy makes it the best example for observing how global players think and strategize in this computing power race. To this end, we interviewed Wang Wenguang, Vice President of Global Data Center Business at Bitdeer, hoping to get insights into the current global AI power shortage and how they view mining companies' shift to AI data centers—whether they consider it capital speculation or a genuine need for AI. We had an in-depth conversation about these questions. Why is the power shortage in the United States so severe? Dongcha: Let me start with a basic question: Do you think electricity prices will continue to rise in the future? BitDeer: I think so, because this is a very important supply and demand relationship in the future. Dynamics: Regarding the power shortage in the United States, there is a saying in the market that it is very difficult to obtain a "power permit" in the United States? BitDeer: It's not that the so-called "electricity permits" can't be approved, but rather that the physical speed of grid expansion can't keep up. For many years after the relocation of heavy industry from the US, the US power grid did not expand systematically. After mining companies moved to the US in 2021, much electricity that was "already connected to the grid and had signed PPAs" was locked up by these companies. With the rise of ChatGPT and the entry of pure AI players, they discovered that a large amount of readily available electricity was stored in mining farms. This explains why large manufacturers are willing to cooperate with mining companies. Rather than waiting 2-4 years to build a 500MW park from scratch, they would rather spend 12 months transforming the existing park. When did the industry truly realize that "inference is also very power-intensive"? BitDeer: Around the time GPT-4 became more widespread, as companies embedded models into customer service, office work, search, risk control, and other applications, inference needs became more long-term and scenario-based, and power consumption did not decrease as initially expected. This leads to two types of changes. One aspect is the engineering upgrade: from stronger air cooling to liquid cooling/hybrid heat dissipation, with rack power, power distribution paths, fire protection and monitoring all being raised to the level of AI data centers. Another aspect is resource strategy: electricity has become the real bottleneck. People are no longer just talking about "buying cards," but are moving forward to obtaining electricity and grid connection, long-term power purchase agreements (PPAs), grid connection scheduling, cross-regional capacity backup, and, when necessary, obtaining electricity upstream like mining companies (self-generation/direct procurement). We've actually seen the same trend in the cryptocurrency mining industry: chip production can be expanded indefinitely (silicon comes from sand), but electricity cannot. We've used natural gas to generate electricity for mining farms in Canada, following the same logic. Today's AI is almost identical. Dynamic Observation: What are the differences in power consumption between AI data centers and traditional internet data centers? BitDeer: It's not a change in quantity, but a change in scale. In the past, 20-30 MW of traditional Internet data centers was already quite large, but now AI data centers often require 500MW or even 1GW. AI has transformed data centers from a "rack business" into an "electrical engineering" project, requiring everything to be remeasured: lines, substations, cooling, fire protection, redundancy, PUE... The experience of traditional Internet data centers is still useful, but no longer sufficient. Dynamic Observation: Why has "electricity" become the scarcest factor in the upstream sector? Bitdeer: Chips can be scaled up because they come from silicon and capacity management; electricity is difficult to scale up because it comes from power generation and grid upgrades. In the past, the mining industry has tried "finding energy upstream," including doing self-generated power projects in Canada; the path for AI is similar—whoever controls electricity first gets the first deployment time. New Battlegrounds in AI: From "Gaining GPUs" to "Gaining Power Grids" Dynamic Observation: What specific changes are needed for mining companies to transform into AI data centers? Previously, it was said that "Bitcoin computing power can be used to run AI," but mining chips (ASICs) and the GPUs required for AI are incompatible. So why are mining companies now able to "provide AI computing power"? BitDeer: Global mining was once divided into two: Bitcoin relied on mining chips ASIC, which was highly efficient but had a single purpose; Ethereum relied on NVIDIA GPUs, which were versatile but have since exited the mining stage after switching to PoS. Therefore, the so-called "mining farms turning to AI" in the market today almost entirely refers to the transformation of Bitcoin mining farms. The core point is that mining farms are no longer "calculating hashes," but are upgrading themselves into AI data centers. This involves upgrading the infrastructure, removing ASIC racks and replacing them with GPU servers; transforming the "good enough" power system into a professional-grade power supply and distribution system with N+1/2N redundancy; upgrading traditional air cooling to a cooling system capable of supporting high-density GPUs; and standardizing and auditing the data center's sealing, dustproofing, and fire protection facilities. Once these four steps are completed, the crypto mining farm will transform from a "mining workshop" into an "AI server room". Why can mining companies build their own systems faster than major AI companies? The answer is electricity. AI is a business that relies on "electricity and heat," and the time cycle for building an AI data center is 3-4 years, making time the biggest hurdle. Mining companies happen to hold these "hard assets," thus giving them a head start in the transformation. Dynamic Observation: A few days ago, Microsoft and Amazon signed multi-year AI contracts with cryptocurrency mining companies. Iris Energy (IREN) signed a contract with Microsoft worth a total of 9.7 billion yuan for 5 years; another company, Cipher, signed a contract with Amazon Web Services worth 5.5 billion yuan for 15 years. This is seen as the first case of mining farms collaborating with major companies. What are your thoughts? BitDeer: Iris Energy is a forward-thinking Australian company that has been mining in the United States for a long time. Iris Energy's shift to AI served as a signal; while Bitcoin prices were high and its competitors were still expanding their mining operations, it diverted some of its electricity to build its own AI data center. As a result, AI companies began actively seeking its services. The real tipping point came from Hyperscaler's substantial financial investment—such as Microsoft's approximately $9.7 billion commitment—which allowed the market to see for the first time clearly that the relationship between mining companies and mega-scale manufacturers was not just "technology integration," but rather an "exchange of electricity and time." The surge in AI demand has amplified the need for infrastructure, opening up opportunities for collaboration. Dynamic Observation: Why are leading mining companies more likely to be chosen by major US AI companies at this stage? BitDeer: Because of "available electricity + fast engineering delivery". The site selection and grid connection of mining companies in the previous cycle have now become the upfront capital for AI data centers. Time is the biggest discount factor, which directly determines who can go online, acquire customers, and generate rolling cash flow within the window of opportunity. Dynamic Observation: So, how difficult are the land selection requirements for AI data centers? BitDeer: Not a big deal overall. In the US and most countries, what's truly scarce is electricity, not land. The reason is simple: most places with access to large-scale electricity are energy-rich areas (natural gas fields, coal mining areas, near hydropower stations, etc.), sparsely populated, and with cheap land. For example, Bitdeer's large data centers in Norway and Bhutan are located far from population centers, where power resources are concentrated and land costs are low. The same applies to the United States; these campuses are not located in the city center, but in more remote locations where land is easy to find and inexpensive. The "first principle" of site selection is power and grid connection; land usually follows the power supply and is not the main bottleneck. Dynamic Observation: AI is now being described as an upstream business of "steel, electricity, and land," even like another form of real estate. What's your take on this? BitDeer: After the large model was released, the power consumption of AI far exceeded most people's expectations. Initially, people thought that "training would consume more power, while inference would be lighter," but the opposite is true. Inference has also been a high-power-consuming process since it became more mainstream. As ChatGPT and DeepSeek become commonplace and more and more devices are connected, the noise floor of inference continues to rise. From an engineering perspective, AI is essentially a resource-intensive industry: On the chip side: During training, the accelerator card is running at almost 100% load, which naturally results in high power consumption; On the data center side: the heat density is much higher than that of traditional servers, the PUE is significantly increased, and the heat dissipation itself also consumes a lot of electricity; On the scale side: the power demand of AI data centers has jumped from 20-30MW of traditional Internet data centers to 500MW or even 1GW, which was almost unimaginable in the era of traditional Internet data centers. Therefore, comparing it to "real estate" is only half right. It does require land, factories, and a long construction period (often 3-4 years), but what determines its survival is electricity and heat—whether it can obtain large-capacity grid connection on time, achieve N+1/2N redundancy, and provide efficient heat dissipation. In this respect, it is very similar to the strong dependence on steel, electricity, and land. What are the characteristics of an AI data center? Dynamic Observation: What are the characteristics of the data center construction models in the United States? BitDeer: Due to power constraints and historical circumstances in the United States, Hyperscaler often has to personally intervene and cooperate with mining companies to obtain usable electricity. Dynamic Observation: Is it possible for foreign companies to build AI data centers in the United States? BitDeer: Simply put, AI data centers are a highly regional business. The deployment of massive systems with hundreds of megawatts and thousands of GPUs is still dominated by large US companies. We are only discussing AI data centers, not traditional internet data centers. Dynamic Observation: Could the AI Data Center become a tool for geopolitics? Will this affect your decision-making? BitDeer: I agree with this assessment. The foundation of AI is data, and data is naturally subject to sovereignty and security constraints. To prevent data leakage and security risks, various regions are tightening relevant policies: even though the United States allows foreign investment in data centers, as AI acquires more and more data, countries will likely move towards "local deployment, local compliance, and no data leaving the country". Simply put, AI is in the US, in the Middle East, and in Europe; regionalization will be a long-term trend. Industry landscape and potential Dynamic Observation: Besides IREN and Bitdeer, which mining companies have the most potential to transform into AI data centers? BitDeer: To see who has a chance, first check if they have large-scale power supply, then check if they can quickly convert their mining farms into GPU data centers. Those with grid connection + ground + substation, N+1/2N redundancy, liquid cooling/high density are the most likely to get AI orders. Another type is pure hosting/asset-light, where the company doesn't control the electricity and the park, making it passive when transitioning to an AI data center. In the US, companies like Riot, CleanSpark, Core Scientific, TeraWulf, and Cipher, which have their own resources and reliable expansion capabilities, are more likely to be targeted by major companies. So the conclusion is straightforward: electricity is the ticket, and transformative power is the speed; only when you have both can you get ahead. Overall, the key lies in who controls "high-quality, sustainable, high-load available electricity." For example, companies with more self-owned grid-connected resources have greater potential; those that mainly rely on managed services and lack their own energy resources and industrial parks are at a disadvantage in this round of structural transformation. What is BitDeer thinking? Dynamic Observation: What are Bitdeer's strategies and paths in "transforming from mining to AI"? Bitdeer: Wu Jihan's strategy has always been to build a complete industry chain. Bitdeer has about 3GW of power and industrial park resources, which is our biggest underlying advantage. When we first entered the AI field, we did not expect that "electricity" would become a core bottleneck, so we initially adopted a self-built and self-operated approach: we partnered with NVIDIA to become an NVIDIA PCSP, deployed a small-scale H100 cluster in Singapore, launched our own AI Cloud, and undertook training services for external clients. This project has been successfully implemented. Subsequently, we also established a second data center in Malaysia. As Hyperscaler entered this field and began collaborating with mining companies, we simultaneously upgraded our high-load campuses into AI data centers: we have announced the complete transformation of our approximately 180MW site in Norway into an AI DC, and the conversion of our approximately 13MW site in Washington State, USA. Ultimately, the essence of AI is very similar to that of cryptocurrency mining—both are businesses of "electricity + infrastructure." We have the full-chain capabilities from electricity and industrial parks to computing power operation, so the transition to AI is relatively smooth. Dynamic Observation: What are the core differences between BitDeer and other mining companies like IREN? BitDeer: Three points. First, it will not 100% transform into an AI company; based on calculations, at the current stage, Crypto Mining's profits are still better than AI data centers, and mining has stable cash flow and good returns. Our second advantage is our international engineering organization capabilities. Bitdeer's engineering organization and execution capabilities are unparalleled in the world. The same AI data center typically takes two years to build in the US, but we can usually do it in a year and a half. This is achieved through parallel development and supply chain collaboration, synchronizing key aspects such as civil engineering, electromechanical systems, power distribution, and cooling, compressing the usual 24-month cycle to about 18 months, and bringing usable capacity to market much faster. The third company maintains a steady strategy: the AI industry is young, even younger than Crypto, and it will not go "all in," but will pursue a longer-term development pace. Dynamic Observation: Where is Bitcoin's power infrastructure currently mainly distributed? Bitdeer: Bitdeer currently has a global deployment of approximately 3 GW of power and related infrastructure, covering five countries: the United States, Canada, Norway, Ethiopia, and Bhutan, to support the construction and operation of mining and AI data centers. Costs and Financing Dynamic Observation: I saw a Goldman Sachs report that mentioned that an AI data center could cost as much as $12 billion. Is it really that expensive? BitDeer: It's indeed much larger, on the order of magnitude "tens of times". Here's a more intuitive comparison in layman's terms: Bitcoin mining farm (USA): Building a 1 MW facility costs approximately $350,000–400,000. But building a 1 MW AI data center costs approximately $11 million. This is because AI data center investment is a complex undertaking involving heavy electromechanical and standard-setting processes. Add to that grid connection queues, environmental/energy assessments, and regional compliance, and the timeframe typically ranges from 18 to 36 months. You'll find that AI data centers are not essentially about "buying more cards," but about turning a piece of land into an "electricity city" that can handle 500MW-1GW of power. Connecting the power, dissipating heat, ensuring redundancy, and achieving compliance are all very expensive tasks. Observation: Where will the money come from? Is financing needed? BitDeer: To be honest, everyone needs to raise funds. Here are some common fundraising strategies used in the industry: 1. Project financing/infrastructure loans: Use the park and equipment as collateral, and rely on long-term leases or computing power offtake (customers promise to buy your computing power for many years) to reassure banks. 2. Equipment leasing/leaseback: Leasing GPUs and some electromechanical components over a longer period, avoiding the need to pay a large amount of cash upfront. 3. Long-term PPAs: Only by locking in electricity prices and available capacity can the debt side be willing to offer low interest rates. 4. Partner with major companies: Large clients/companies offer minimum spending requirements, prepayments, guarantees, or even joint ventures (JVs), allowing you to obtain cheaper funding. These terms can be seen in collaborations between IREN, CoreWeave, and Google/Microsoft. Observation: Will Bitdeer also need to raise funds? Will it announce its partnerships with major companies soon? BitDeer: I can't say more about this publicly right now. Conclusion Shortly after the interview, BitDeer revealed its next step in the capital market. On November 13, Bitdeer announced that it will raise $400 million through the issuance of convertible senior notes, and grant initial purchasers an option to subscribe for up to an additional $60 million of notes within 13 days, bringing the total fundraising to a maximum of $460 million. The new funds will be used for data center expansion, ASIC mining machine R&D, expansion of AI and HPC cloud businesses, and general corporate purposes. With electricity becoming the most scarce upstream resource in the AI industry, the final allocation of this $460 million to specific plots of land and the number of megawatts of new load it will largely determine BitDeer's ranking in the next round of computing power competition. For Bitdeer, this money is more like putting the judgments made in the interviews into its balance sheet: one end connects to the core cash flow of mining, and the other end connects to the long-term, high-growth business line of AI data centers. It may not be immediately reflected in the revenue and profit of the next quarter's financial report, but it will slowly rewrite the power structure of the computing power business in the coming years—who is qualified to sit at the negotiating table, and who can only wait in line for electricity on the grid connection list. Looking back from the outcome, the story of this round of AI infrastructure is not complicated: electricity has become the real upstream, time has become the new currency, and the industrial parks and grid connection quotas held by mining companies have become "old assets" that others cannot buy with money. Once the hype surrounding models and applications subsides, the market will likely have to reassess the situation: it will no longer matter who tells the loudest story; only companies that can connect every megawatt of electricity and keep it running steadily in a world of power shortages will be qualified to remain at the next stage of the game.Author: Lin Wanwan No one initially expected that the real bottleneck for AI would not be capital or large models, but electricity. Large-scale training and continuous full-load operation of AI inference, running 24/7, has led to a problem: insufficient electricity, forcing chips to sit idle. The United States' power grid infrastructure has lagged behind in the past decade, with new large-load grid connections often taking 2-4 years, making "readily available electricity" a scarce commodity across the entire industry. Generative AI brings a primitive and brutal issue to the forefront: the problem isn't a lack of models, but a lack of electricity. The story took a turn from there, and crypto mining companies, the first people to treat electricity as a "means of production," began to move from the margins to the center of the capital stage. Iris Energy (IREN) is a case in point. This year, IREN's stock price surged nearly 600% at one point, rising from $5.12 to $75.73 over 52 weeks. Even when Bitcoin's gains were still attractive, it boldly diverted power to renovate its self-built AI data center. When giants like Microsoft put up long-term contracts worth a total of $9.7 billion, the market gained its first intuitive understanding of the real-world path "from mining to AI": electricity and land come first, followed by GPUs and customers. However, not all mining companies are like IREN, choosing to bet everything on AI. In this massive migration of computing power driven by electricity, there is another steady force worth our attention—BitDeer. Bitdeer Technologies Group (NASDAQ: BTDR), founded by crypto legend Jihan Wu and headquartered in Singapore, possesses nearly 3GW of power resources worldwide, avoiding the superficial trap of relying on others for power from the outset. When the AI wave arrived, Bitdeer did not choose the aggressive "all-in" approach like Iren, but instead retained its lucrative Bitcoin mining as its core business while steadily upgrading some of its mining farms into AI data centers. This "offensive or defensive" strategy makes it the best example for observing how global players think and strategize in this computing power race. To this end, we interviewed Wang Wenguang, Vice President of Global Data Center Business at Bitdeer, hoping to get insights into the current global AI power shortage and how they view mining companies' shift to AI data centers—whether they consider it capital speculation or a genuine need for AI. We had an in-depth conversation about these questions. Why is the power shortage in the United States so severe? Dongcha: Let me start with a basic question: Do you think electricity prices will continue to rise in the future? BitDeer: I think so, because this is a very important supply and demand relationship in the future. Dynamics: Regarding the power shortage in the United States, there is a saying in the market that it is very difficult to obtain a "power permit" in the United States? BitDeer: It's not that the so-called "electricity permits" can't be approved, but rather that the physical speed of grid expansion can't keep up. For many years after the relocation of heavy industry from the US, the US power grid did not expand systematically. After mining companies moved to the US in 2021, much electricity that was "already connected to the grid and had signed PPAs" was locked up by these companies. With the rise of ChatGPT and the entry of pure AI players, they discovered that a large amount of readily available electricity was stored in mining farms. This explains why large manufacturers are willing to cooperate with mining companies. Rather than waiting 2-4 years to build a 500MW park from scratch, they would rather spend 12 months transforming the existing park. When did the industry truly realize that "inference is also very power-intensive"? BitDeer: Around the time GPT-4 became more widespread, as companies embedded models into customer service, office work, search, risk control, and other applications, inference needs became more long-term and scenario-based, and power consumption did not decrease as initially expected. This leads to two types of changes. One aspect is the engineering upgrade: from stronger air cooling to liquid cooling/hybrid heat dissipation, with rack power, power distribution paths, fire protection and monitoring all being raised to the level of AI data centers. Another aspect is resource strategy: electricity has become the real bottleneck. People are no longer just talking about "buying cards," but are moving forward to obtaining electricity and grid connection, long-term power purchase agreements (PPAs), grid connection scheduling, cross-regional capacity backup, and, when necessary, obtaining electricity upstream like mining companies (self-generation/direct procurement). We've actually seen the same trend in the cryptocurrency mining industry: chip production can be expanded indefinitely (silicon comes from sand), but electricity cannot. We've used natural gas to generate electricity for mining farms in Canada, following the same logic. Today's AI is almost identical. Dynamic Observation: What are the differences in power consumption between AI data centers and traditional internet data centers? BitDeer: It's not a change in quantity, but a change in scale. In the past, 20-30 MW of traditional Internet data centers was already quite large, but now AI data centers often require 500MW or even 1GW. AI has transformed data centers from a "rack business" into an "electrical engineering" project, requiring everything to be remeasured: lines, substations, cooling, fire protection, redundancy, PUE... The experience of traditional Internet data centers is still useful, but no longer sufficient. Dynamic Observation: Why has "electricity" become the scarcest factor in the upstream sector? Bitdeer: Chips can be scaled up because they come from silicon and capacity management; electricity is difficult to scale up because it comes from power generation and grid upgrades. In the past, the mining industry has tried "finding energy upstream," including doing self-generated power projects in Canada; the path for AI is similar—whoever controls electricity first gets the first deployment time. New Battlegrounds in AI: From "Gaining GPUs" to "Gaining Power Grids" Dynamic Observation: What specific changes are needed for mining companies to transform into AI data centers? Previously, it was said that "Bitcoin computing power can be used to run AI," but mining chips (ASICs) and the GPUs required for AI are incompatible. So why are mining companies now able to "provide AI computing power"? BitDeer: Global mining was once divided into two: Bitcoin relied on mining chips ASIC, which was highly efficient but had a single purpose; Ethereum relied on NVIDIA GPUs, which were versatile but have since exited the mining stage after switching to PoS. Therefore, the so-called "mining farms turning to AI" in the market today almost entirely refers to the transformation of Bitcoin mining farms. The core point is that mining farms are no longer "calculating hashes," but are upgrading themselves into AI data centers. This involves upgrading the infrastructure, removing ASIC racks and replacing them with GPU servers; transforming the "good enough" power system into a professional-grade power supply and distribution system with N+1/2N redundancy; upgrading traditional air cooling to a cooling system capable of supporting high-density GPUs; and standardizing and auditing the data center's sealing, dustproofing, and fire protection facilities. Once these four steps are completed, the crypto mining farm will transform from a "mining workshop" into an "AI server room". Why can mining companies build their own systems faster than major AI companies? The answer is electricity. AI is a business that relies on "electricity and heat," and the time cycle for building an AI data center is 3-4 years, making time the biggest hurdle. Mining companies happen to hold these "hard assets," thus giving them a head start in the transformation. Dynamic Observation: A few days ago, Microsoft and Amazon signed multi-year AI contracts with cryptocurrency mining companies. Iris Energy (IREN) signed a contract with Microsoft worth a total of 9.7 billion yuan for 5 years; another company, Cipher, signed a contract with Amazon Web Services worth 5.5 billion yuan for 15 years. This is seen as the first case of mining farms collaborating with major companies. What are your thoughts? BitDeer: Iris Energy is a forward-thinking Australian company that has been mining in the United States for a long time. Iris Energy's shift to AI served as a signal; while Bitcoin prices were high and its competitors were still expanding their mining operations, it diverted some of its electricity to build its own AI data center. As a result, AI companies began actively seeking its services. The real tipping point came from Hyperscaler's substantial financial investment—such as Microsoft's approximately $9.7 billion commitment—which allowed the market to see for the first time clearly that the relationship between mining companies and mega-scale manufacturers was not just "technology integration," but rather an "exchange of electricity and time." The surge in AI demand has amplified the need for infrastructure, opening up opportunities for collaboration. Dynamic Observation: Why are leading mining companies more likely to be chosen by major US AI companies at this stage? BitDeer: Because of "available electricity + fast engineering delivery". The site selection and grid connection of mining companies in the previous cycle have now become the upfront capital for AI data centers. Time is the biggest discount factor, which directly determines who can go online, acquire customers, and generate rolling cash flow within the window of opportunity. Dynamic Observation: So, how difficult are the land selection requirements for AI data centers? BitDeer: Not a big deal overall. In the US and most countries, what's truly scarce is electricity, not land. The reason is simple: most places with access to large-scale electricity are energy-rich areas (natural gas fields, coal mining areas, near hydropower stations, etc.), sparsely populated, and with cheap land. For example, Bitdeer's large data centers in Norway and Bhutan are located far from population centers, where power resources are concentrated and land costs are low. The same applies to the United States; these campuses are not located in the city center, but in more remote locations where land is easy to find and inexpensive. The "first principle" of site selection is power and grid connection; land usually follows the power supply and is not the main bottleneck. Dynamic Observation: AI is now being described as an upstream business of "steel, electricity, and land," even like another form of real estate. What's your take on this? BitDeer: After the large model was released, the power consumption of AI far exceeded most people's expectations. Initially, people thought that "training would consume more power, while inference would be lighter," but the opposite is true. Inference has also been a high-power-consuming process since it became more mainstream. As ChatGPT and DeepSeek become commonplace and more and more devices are connected, the noise floor of inference continues to rise. From an engineering perspective, AI is essentially a resource-intensive industry: On the chip side: During training, the accelerator card is running at almost 100% load, which naturally results in high power consumption; On the data center side: the heat density is much higher than that of traditional servers, the PUE is significantly increased, and the heat dissipation itself also consumes a lot of electricity; On the scale side: the power demand of AI data centers has jumped from 20-30MW of traditional Internet data centers to 500MW or even 1GW, which was almost unimaginable in the era of traditional Internet data centers. Therefore, comparing it to "real estate" is only half right. It does require land, factories, and a long construction period (often 3-4 years), but what determines its survival is electricity and heat—whether it can obtain large-capacity grid connection on time, achieve N+1/2N redundancy, and provide efficient heat dissipation. In this respect, it is very similar to the strong dependence on steel, electricity, and land. What are the characteristics of an AI data center? Dynamic Observation: What are the characteristics of the data center construction models in the United States? BitDeer: Due to power constraints and historical circumstances in the United States, Hyperscaler often has to personally intervene and cooperate with mining companies to obtain usable electricity. Dynamic Observation: Is it possible for foreign companies to build AI data centers in the United States? BitDeer: Simply put, AI data centers are a highly regional business. The deployment of massive systems with hundreds of megawatts and thousands of GPUs is still dominated by large US companies. We are only discussing AI data centers, not traditional internet data centers. Dynamic Observation: Could the AI Data Center become a tool for geopolitics? Will this affect your decision-making? BitDeer: I agree with this assessment. The foundation of AI is data, and data is naturally subject to sovereignty and security constraints. To prevent data leakage and security risks, various regions are tightening relevant policies: even though the United States allows foreign investment in data centers, as AI acquires more and more data, countries will likely move towards "local deployment, local compliance, and no data leaving the country". Simply put, AI is in the US, in the Middle East, and in Europe; regionalization will be a long-term trend. Industry landscape and potential Dynamic Observation: Besides IREN and Bitdeer, which mining companies have the most potential to transform into AI data centers? BitDeer: To see who has a chance, first check if they have large-scale power supply, then check if they can quickly convert their mining farms into GPU data centers. Those with grid connection + ground + substation, N+1/2N redundancy, liquid cooling/high density are the most likely to get AI orders. Another type is pure hosting/asset-light, where the company doesn't control the electricity and the park, making it passive when transitioning to an AI data center. In the US, companies like Riot, CleanSpark, Core Scientific, TeraWulf, and Cipher, which have their own resources and reliable expansion capabilities, are more likely to be targeted by major companies. So the conclusion is straightforward: electricity is the ticket, and transformative power is the speed; only when you have both can you get ahead. Overall, the key lies in who controls "high-quality, sustainable, high-load available electricity." For example, companies with more self-owned grid-connected resources have greater potential; those that mainly rely on managed services and lack their own energy resources and industrial parks are at a disadvantage in this round of structural transformation. What is BitDeer thinking? Dynamic Observation: What are Bitdeer's strategies and paths in "transforming from mining to AI"? Bitdeer: Wu Jihan's strategy has always been to build a complete industry chain. Bitdeer has about 3GW of power and industrial park resources, which is our biggest underlying advantage. When we first entered the AI field, we did not expect that "electricity" would become a core bottleneck, so we initially adopted a self-built and self-operated approach: we partnered with NVIDIA to become an NVIDIA PCSP, deployed a small-scale H100 cluster in Singapore, launched our own AI Cloud, and undertook training services for external clients. This project has been successfully implemented. Subsequently, we also established a second data center in Malaysia. As Hyperscaler entered this field and began collaborating with mining companies, we simultaneously upgraded our high-load campuses into AI data centers: we have announced the complete transformation of our approximately 180MW site in Norway into an AI DC, and the conversion of our approximately 13MW site in Washington State, USA. Ultimately, the essence of AI is very similar to that of cryptocurrency mining—both are businesses of "electricity + infrastructure." We have the full-chain capabilities from electricity and industrial parks to computing power operation, so the transition to AI is relatively smooth. Dynamic Observation: What are the core differences between BitDeer and other mining companies like IREN? BitDeer: Three points. First, it will not 100% transform into an AI company; based on calculations, at the current stage, Crypto Mining's profits are still better than AI data centers, and mining has stable cash flow and good returns. Our second advantage is our international engineering organization capabilities. Bitdeer's engineering organization and execution capabilities are unparalleled in the world. The same AI data center typically takes two years to build in the US, but we can usually do it in a year and a half. This is achieved through parallel development and supply chain collaboration, synchronizing key aspects such as civil engineering, electromechanical systems, power distribution, and cooling, compressing the usual 24-month cycle to about 18 months, and bringing usable capacity to market much faster. The third company maintains a steady strategy: the AI industry is young, even younger than Crypto, and it will not go "all in," but will pursue a longer-term development pace. Dynamic Observation: Where is Bitcoin's power infrastructure currently mainly distributed? Bitdeer: Bitdeer currently has a global deployment of approximately 3 GW of power and related infrastructure, covering five countries: the United States, Canada, Norway, Ethiopia, and Bhutan, to support the construction and operation of mining and AI data centers. Costs and Financing Dynamic Observation: I saw a Goldman Sachs report that mentioned that an AI data center could cost as much as $12 billion. Is it really that expensive? BitDeer: It's indeed much larger, on the order of magnitude "tens of times". Here's a more intuitive comparison in layman's terms: Bitcoin mining farm (USA): Building a 1 MW facility costs approximately $350,000–400,000. But building a 1 MW AI data center costs approximately $11 million. This is because AI data center investment is a complex undertaking involving heavy electromechanical and standard-setting processes. Add to that grid connection queues, environmental/energy assessments, and regional compliance, and the timeframe typically ranges from 18 to 36 months. You'll find that AI data centers are not essentially about "buying more cards," but about turning a piece of land into an "electricity city" that can handle 500MW-1GW of power. Connecting the power, dissipating heat, ensuring redundancy, and achieving compliance are all very expensive tasks. Observation: Where will the money come from? Is financing needed? BitDeer: To be honest, everyone needs to raise funds. Here are some common fundraising strategies used in the industry: 1. Project financing/infrastructure loans: Use the park and equipment as collateral, and rely on long-term leases or computing power offtake (customers promise to buy your computing power for many years) to reassure banks. 2. Equipment leasing/leaseback: Leasing GPUs and some electromechanical components over a longer period, avoiding the need to pay a large amount of cash upfront. 3. Long-term PPAs: Only by locking in electricity prices and available capacity can the debt side be willing to offer low interest rates. 4. Partner with major companies: Large clients/companies offer minimum spending requirements, prepayments, guarantees, or even joint ventures (JVs), allowing you to obtain cheaper funding. These terms can be seen in collaborations between IREN, CoreWeave, and Google/Microsoft. Observation: Will Bitdeer also need to raise funds? Will it announce its partnerships with major companies soon? BitDeer: I can't say more about this publicly right now. Conclusion Shortly after the interview, BitDeer revealed its next step in the capital market. On November 13, Bitdeer announced that it will raise $400 million through the issuance of convertible senior notes, and grant initial purchasers an option to subscribe for up to an additional $60 million of notes within 13 days, bringing the total fundraising to a maximum of $460 million. The new funds will be used for data center expansion, ASIC mining machine R&D, expansion of AI and HPC cloud businesses, and general corporate purposes. With electricity becoming the most scarce upstream resource in the AI industry, the final allocation of this $460 million to specific plots of land and the number of megawatts of new load it will largely determine BitDeer's ranking in the next round of computing power competition. For Bitdeer, this money is more like putting the judgments made in the interviews into its balance sheet: one end connects to the core cash flow of mining, and the other end connects to the long-term, high-growth business line of AI data centers. It may not be immediately reflected in the revenue and profit of the next quarter's financial report, but it will slowly rewrite the power structure of the computing power business in the coming years—who is qualified to sit at the negotiating table, and who can only wait in line for electricity on the grid connection list. Looking back from the outcome, the story of this round of AI infrastructure is not complicated: electricity has become the real upstream, time has become the new currency, and the industrial parks and grid connection quotas held by mining companies have become "old assets" that others cannot buy with money. Once the hype surrounding models and applications subsides, the market will likely have to reassess the situation: it will no longer matter who tells the loudest story; only companies that can connect every megawatt of electricity and keep it running steadily in a world of power shortages will be qualified to remain at the next stage of the game.