Citigroup Says Alibaba’s AI Compute Build Could Hit $160B in Cloud Revenue. The Crypto Market Is Already Pricing the Next Bottleneck.

Over the past week, the crypto market ignored a more volatile number than Bitcoin: 20 gigawatts. That is the number Alibaba’s CEO Wu Yongming reportedly used to define the company’s AI infrastructure ambition. By fiscal 2033, Alibaba wants 20 gigawatts of compute under management. Citigroup’s research team took that number, connected it to cloud revenue, and issued a clean verdict: Alibaba’s AI business potential may be further unleashed, external cloud revenue could reach $100 billion by fiscal 2031, and under the CEO’s more aggressive target, external cloud business could climb to $160 billion by fiscal 2033. Citi kept a Buy rating and a HK$189 target price. That is a very clean number. Too clean, maybe. The first thing I saw when I read the note was not the $160 billion. It was the word “external.” It was also the word “gigawatts.” It was the missing word “capital.” Nobody in blockchain should be slow to an AI-compute story anymore. AI agents are already spending money on-chain. Decentralized physical infrastructure networks are already renting GPUs to strangers. Stablecoin rails are already moving value between parties who would never use the same bank. And now one of the largest cloud empires in the world is telling the market that compute, not just consumer software, is the real AI business. This is not merely a stock story. This is a market structure story. The ledger remembers what the hype forgets. I remember the 2017 Ethereum time-lock mess. I rushed an article into the world before the audit was done, and I got the speed right even when my verification was wrong. That experience taught me to spot the difference between a narrative and a liability. Alibaba’s 20-gigawatt target is not just a narrative. It is also a liability. The question is who ends up paying for that liability: Alibaba shareholders, Chinese grid operators, or the decentralized compute networks that are quietly building the same outcome without a single corporate balance sheet. For the crypto crowd, the real breakthrough is not that Citigroup trusts Alibaba. It is that Citigroup has just translated raw electricity into a revenue forecast. That is exactly the kind of translation the decentralized compute market has been trying to do for years. Render, Akash, Bittensor, io.net, Gensyn — they are all attempts to make a watt of compute tradable, auditable, and liquid. Alibaba is about to make 20 gigawatts of compute seem so enormous that the market will start asking how any single company can be allowed to own that much of the future. Here is the context everyone is skipping. Alibaba is the largest cloud provider in China, and China has one of the most controlled, segmented, and policy-heavy digital asset markets in the world. Digital collectibles without secondary markets have already been debunked in China. Without a secondary market, an NFT is just a one-time sale that even speculators refuse to hold. The same logic applies to compute. If Alibaba builds 20 gigawatts of AI infrastructure and cannot resell idle capacity freely, cannot move those assets across borders, and cannot let global developers pay with the fastest settlement rails available, then a portion of that 20-gigawatt empire is a digital collectible with no secondary market. Capacity without liquidity is just a very expensive JPEG. That is the contrarian angle this report does not mention. It is also the angle the crypto market should care about most. Let me establish the baseline first. Alibaba Cloud, in its fiscal 2024, generated around $15 billion in revenue. Citi is saying external cloud revenue could be $100 billion by fiscal 2031. That is nearly a seven-fold increase in about seven years. The implied compound annual growth rate is roughly 27%. Then, by fiscal 2033, the company could double that again to $160 billion, implying a revenue increase of 60% in just two years. That is not a forecast. That is a mandate. It assumes that AI workloads grow faster than most software revolutions in history. It assumes that Alibaba holds enough market share to capture a giant chunk of China’s AI training and inference demand. It assumes that the chips arrive. It assumes that the power arrives. It assumes that customers keep paying. Now let’s do the math that Citi probably did in the background. 20 gigawatts is 20,000 megawatts, or 20 million kilowatts. A modern AI server with eight H100-class GPUs can draw around 10 kilowatts under load. That means 20 gigawatts can support roughly two million AI servers. At eight GPUs per server, that is about sixteen million H100-class GPUs. At $30,000 per GPU, the GPU bill alone could approach half a trillion dollars. That is before the servers. Before the networking. Before the cooling systems. Before the buildings. Before the power substations. Before the failure rate. The ledger remembers what the hype forgets. Nobody builds sixteen million GPUs overnight. Even the largest hyperscalers on Earth, Microsoft, Amazon, Alphabet, and Meta, spent roughly $200 billion combined on capital expenditures in 2024. That includes data centers, real estate, networking, and custom silicon. Alibaba’s 20-gigawatt target, if fully realized, would require a multi-year, multi-billion-dollar procurement plan that dwarfs what any single cloud provider has ever done in AI infrastructure. The only way the math even comes close is if the target is stretched over years and generations of hardware, not just H100s. By 2033, AI accelerators will be dramatically more efficient. Power per chip might drop. Density will rise. But the social footprint of that spending will still be enormous. That is the real story for crypto. Compute is becoming the most important scarce resource of the next decade. The blockchain industry has spent years pretending that consensus algorithms and token flows are the only thing that matters. The truth is that every AI agent, every on-chain trading strategy, every decentralized autonomous organization that wants to run large-language-model inference, every stablecoin issuance platform that wants to fight fraud with AI, needs compute. The last few months have made that impossible to ignore. I spend most of my day now tracing the social footprints of autonomous agents on platforms like Farcaster, and the pattern is clear: AI agents are not just talking about tokens. They are coordinating resource usage. They are requesting compute. They are trying to pay for it in real time. They do not wait for a bank loan. They hold crypto. They want utility. The Citi report is useful precisely because it is so centralized. It shows the old world’s map of the future: a giant cloud platform, a giant CEO target, a giant revenue line. But the crypto-native map looks completely different. Decentralized compute networks are built around the idea that no single operator should hold the keys to all the GPUs. They tokenize compute. They create secondary markets for idle capacity. They allow anyone with a spare graphics card to become a cloud provider. They allow AI developers to hedge against the whims of hyperscaler pricing by renting from a global pool of suppliers. They allow stablecoins to settle those payments instantly, without waiting for a wire transfer from Shanghai to Oakland. That is not a toy use case. That is the next chapter. Let me be clear about one thing: I am not bearish on Alibaba. I am bearish on monopolies of the mind. The 20-gigawatt target is a breathtaking vote of confidence in AI. It tells the world that Alibaba believes artificial intelligence is not a feature or a product line but an industrial base. That belief is correct. I have been in this industry long enough, from the 2020 Uniswap V2 summer to the 2021 Bored Ape cycle, to understand that whenever an institution starts pouring serious money into a technology, the social narrative around that technology shifts. From code to culture, the way a protocol is perceived can become more powerful than the code itself. Alibaba is about to shift the culture of enterprise AI in Asia. That matters. But it does not mean the centralized cloud is the only path. Riding the peak of the ape mania wave taught me something about value. The Bored Ape Yacht Club was not valuable because the JPEG was scarce. It was valuable because ownership signaled identity, status, and community access. The same is true for Alibaba’s compute. Owning 20 gigawatts of AI infrastructure is a signal of national importance. It is a status symbol for China’s tech sector. It is a signal to regulators, to enterprise customers, and to global investors that Alibaba will not be left out of the AI supercycle. But signals are not dividends. Status is not utilization. Hype is not revenue. This is where the contrarian lens comes in. Most analysts will look at the Citi report and ask whether Alibaba can raise enough money, buy enough chips, and land enough customers. That is the wrong question. The right question is whether the world will let this much computing be controlled by a single corporate balance sheet. The answer may be no. Not because Alibaba is evil, but because AI is already becoming too big, too distributed, and too autonomous for one controller. Consider the post-2020 pattern in crypto. When DeFi Summer took off, centralized exchanges were still the gatekeepers of liquidity. Yet the smartest teams quickly realized that trustless liquidity pools could grow in parallel. Uniswap did not wait for permission from any bank. It created a new market structure and let liquidity follow. The same thing is happening with compute. While Alibaba is promising 20 gigawatts of centralized capacity, networks like Render are creating a permissionless market for GPU cycles. Akash is letting users bid on anonymous compute containers. Bittensor is turning machine intelligence into a tokenized commodity. These networks are not trying to replace Alibaba tomorrow. They are building the secondary market that Alibaba’s balance sheet cannot have. They are making compute liquid. They are turning the phrase “riding the real-time current of value” from a slogan into an ordering system. And here is the irony that the traditional finance world hasn’t yet noticed: some of the largest infrastructure bottlenecks in the 20-gigawatt build are also bottlenecks in crypto. Power constraints. Chip export controls. Cooling technology. Grid connection queues. Land availability. Labor constraints. These are all physical barriers. They will not be solved by a token alone. But the financial rails around them can be solved by crypto. Stablecoins are already being used in emerging markets to bypass local currency inflation and dollar shortages. The same logic applies to compute procurement. A data center operator in Latin America that wants to buy GPU time from a supplier in Southeast Asia does not need a correspondent bank. It needs a stablecoin and a smart contract. It needs a market that does not care about the buyer’s passport. That is the quiet truth behind the $160 billion forecast. The bank is modeling a world where Alibaba sells cloud services like groceries: supply, demand, and a margin. But the future is not only about selling cloud services. It is about the on-chain coordination of compute supply and demand. The real value will not just be in the 20 gigawatts Alibaba owns. It will be in the marketplace that can instantly route work to the cheapest, most available compute anywhere on Earth. Alibaba could be a major participant in that marketplace. Or it could be a giant, highly painted warehouse while the actual commerce happens somewhere else. I have a specific bias here, and I want to be honest about it. I became a crypto journalist because I love the speed of information. The 2017 time-lock incident taught me that a first-mover advantage is only worth something if you can survive the follow-up. The Terra/Luna crash in 2022 taught me that the human cost of financial complexity matters more than the charts. The AI-agent news loop in 2025 is teaching me something else: the next market makers will not all be human. AI agents are already trading NFTs, minting tokens, and communicating with each other through decentralized social protocols. They do not have to wait for a corporate cloud webinar. They will rent compute wherever it is cheapest. They will pay with crypto because settlement is instant and permissionless. They will leave a footprint in the ledger before a single analyst updates their spreadsheet. Decoding the pulse of the crypto zeitgeist now means decoding machine conversations. That is why I want to push back on the most conventional reading of the Citi note. The conventional reading is: Alibaba has a clever CEO, a realistic AI target, and a Buy rating. The better reading is: Alibaba is playing the centralized game, and the decentralized game is already being built under its feet. No matter how many gigawatts Alibaba builds, the true utility lies in the ability to make compute as easy to buy and sell as a stablecoin. Otherwise, the company is holding a future asset class without a secondary market. That is the same old digital collectible trap. When China’s digital collectibles market collapsed, it was not because the art was bad. It was because there was no liquid secondary market, no social layer that let ownership be carried forward. The same thing will happen to centralized AI compute if it cannot be exchanged, financed, or hedged through open markets. Let’s go back to the numbers, because the numbers matter. Citi says Alibaba’s external cloud revenue could be $100 billion by fiscal 2031. To get there, Alibaba will need to capture a growing share of China’s enterprise AI spending. That is possible. China is no longer waiting for Western models to do all the work. Domestic large-language models are being deployed across government, finance, education, and manufacturing. Alibaba’s own Tongyi Qianwen model family is a serious competitor in that market. The company also has a massive e-commerce ecosystem, and AI will drive more personalized shopping, logistics automation, and customer service. That creates internal demand for compute, but internal demand does not translate into external cloud revenue by itself. If Alibaba uses 10 gigawatts for its own businesses and only sells the remaining 10 gigawatts to outside customers, the revenue math changes. I cannot build a revenue model from a watt target alone. I need to know utilization. I need to know average price per compute-hour. I need to know the split between training, inference, storage, and value-added services. I need to know the capital cost of the electrical grid connections. Based on my audit experience, I can tell you this much: a watt is not a dollar until someone else pays for it. There is also the chip problem. The United States has restricted exports of advanced AI chips to China. Alibaba has had to design its own processors and work with domestic alternatives. Domestic chip supply is improving, but the performance gap with NVIDIA remains large. If 20 gigawatts must be built with domestic chips, Alibaba may need significantly more accelerators to reach the same useful compute. That increases power consumption, heat, and maintenance costs. It makes the 20-gigawatt target a bigger physical footprint than the same raw number would imply in a world where H100-class GPUs are freely available. This is the kind of nuance that Citi’s revenue model probably smooths over. The bank wants to see the upside. I want to see the physical plan. Crypto can actually help fix that information gap. If Alibaba announced a tokenized compute bond, or a proof-of-reserves dashboard for its AI capacity, or a stablecoin settlement rail for enterprise cloud customers, the market could monitor utilization in real time. That would be a genuine leap forward for transparency. But that is unlikely. Alibaba is not built to open-source its financial infrastructure. So the only way the market can price Alibaba’s AI potential with more precision is through publicly disclosed capital expenditure, quarterly cloud revenue, and AI workload metrics. Until then, the $160 billion number is a monument, not a map. What about the blockchain industry itself? The Citi report should be a wake-up call. DePIN projects have struggled to prove that decentralized compute can reach hyperscale reliability. But Alibaba’s target shows the scale of demand. If AI training truly consumes 20 gigawatts, then there will be enormous residual demand for verifiable inference, backup compute, and ultra-cheap spot capacity. That is the niche DePIN can win. It does not need to win 20 gigawatts. It needs to win the long tail. It needs to win latency-tolerant jobs. It needs to win privacy-sensitive use cases where data cannot be sent to Alibaba Cloud. It needs to win the idle GPU cycles sitting inside small data centers, crypto mining farms, and individual gaming rigs. The market is not either-or. It is both-and. But decentralized networks need to be ready to actually serve that demand. I have seen this story before. In 2020, when Uniswap V2 made automated market making accessible to everyone, the traditional financial world did not disappear. But the enthusiasm for open protocols forced centralized exchanges to improve their user experience, reduce fees, and become more transparent. The same dynamic will happen with compute. Alibaba’s 20-gigawatt build will legitimize the AI infrastructure conversation. That legitimization will spill over into decentralized compute. It will make enterprise users more comfortable buying compute services through new marketplaces. It will make hardware suppliers more interested in working with crypto-native platforms. It will make investors more willing to fund DePIN projects. The rising tide lifts all high-performance racks. But there is a darker side, too. The Citi report is also about geopolitical positioning. China wants AI sovereignty. The United States wants to maintain a lead in silicon. Every giant AI infrastructure project is now a strategic asset. In that environment, decentralized networks that are borderless and permissionless will attract suspicion. Blockchains allow value to flow across boundaries. That is the point. But it also means regulators will try to control the rails. Stablecoins are already under scrutiny. DePIN platforms that allow buyers and sellers to transact without jurisdiction control will be next. I do not think the Ape-mania wave prepared the industry for this. A JPEG of an ape does not threaten a nation-state. A global marketplace for AI compute does. If Alibaba is building 20 gigawatts, the regulators of every crypto-friendly jurisdiction will ask who is building the other 20 gigawatts. The answer is: a lot of people. Radiant local grids, industrial parks with surplus power, mining farm operators that pivoted from SHA-256 to AI, and independent GPU owners. They are all potential DePIN suppliers. Their computational assets may not be as polished as a hyperscale cloud, but they are real. They are already connected to the internet. They are already runnable. The only missing piece is the coordination layer. Tokens are a natural coordination layer because they can reward suppliers based on uptime, latency, and reputation. They can create a global spot market for compute. They can settle transactions instantly. They can make the secondary market for idle capacity as liquid as the secondary market for an NFT. That is the real convergence between Alibaba’s $160 billion dream and the blockchain’s actual purpose. Let’s talk about the phrase “where liquidity meets the human story.” When I attended meetups in Bali and Jakarta during the NFT cycle, I saw how people used ownership to define themselves without waiting for a bank or a gallery to validate them. The same energy is now flowing into compute. Small operators in cities that the traditional financial system ignored want to own a piece of the AI economy. They can do that by staking tokens in a compute network. They can do that by hosting GPUs in their garage. They can do that by earning stablecoins for providing decentralized storage. This is not a charity experiment. It is a survival strategy for people whose local currency is inflating, whose access to dollar-based cloud services is limited, and whose trust in centralized institutions is low. Crypto payments are not some ideological preference. They are the only realistic way for many small compute providers to reach a global market. The same inflation pressure that drove crypto adoption in developing countries is now driving crypto compute adoption. It all points to the same conclusion: Alibaba’s centralized AI build and blockchains are not separate universes. They are opposite ends of the same power grid. The takeaway is not that Citigroup is dumb. It is not that Alibaba is doomed. The takeaway is that the market is underpricing the alternative. Alibaba’s $160 billion cloud revenue forecast is a big number, but it will be created in a world where compute is increasingly tradable, borderless, and tokenized. Alibaba will have to face that world. Maybe it will partner with blockchain networks. Maybe it will issue its own stablecoin or enterprise blockchain service. Maybe it will build a regulated secondary market for idle AI capacity. Or maybe it will ignore crypto entirely and simply try to sell its 20 gigawatts the old way. The old way still works for the first few years. Yet over time, the cost of operating a global compute empire without a permissionless settlement layer will become impossible to ignore. As for the immediate market signals, here is what I am watching. First, Alibaba’s capital expenditure guidance. If the company is serious about 20 gigawatts, capex has to go parabolic. Second, the quality of Alibaba Cloud’s AI revenue disclosures. If management starts breaking out AI-related external cloud revenue, the forecast becomes more credible. Third, the state of domestic chip supply. Every new export-control regulation will shape the physical plan. Fourth, the growth of decentralized compute clusters. Every additional thousand GPUs on DePIN networks is a tiny raid on the centralized cloud’s future revenue. Fifth, the behavior of AI agents. The ledgers are already showing that autonomous agents prefer fast, low-friction payment rails. They will not always wait for a centralized credit department. There is a line I keep coming back to: the ledger remembers what the hype forgets. Hype sees 20 gigawatts. The ledger sees the missing audits, the undeclared utilization, the idle racks, the stranded assets, the canceled grid connections, and the long trail of collateralized debt. Blockchain is not magic. It cannot build a GPU. It cannot lay a power cable. But it can make the terms of compute ownership transparent. It can create a liquid market for capacity. It can let an AI agent on the other side of the world pay for compute in under ten seconds. It can turn idle hardware from a liability into a yield-bearing asset. Those are not small features. Those are the building blocks of the next major market. I want to end with a forward-looking thought, not a summary. For the next five years, the valley will be won by whoever can control the most efficient compute network, not necessarily the company that controls the most physical data centers. Alibaba has a head start on physical scale. Crypto networks have a head start on digital liquidity. The winner of the AI compute era might be the hybrid: a protocol that can plug into hyperscale clouds, DePIN clusters, and the open internet at the same time, while settling payments in stablecoins and providing proof through public ledgers. Alibaba will make billions from its cloud business either way. But whether the $160 billion forecast becomes real depends on whether it can sell compute the way the internet sells attention: fast, open, and with as little friction as possible. The cheetah in me loves the speed of the Citi target. The auditor in me wants to see the receipts. And the crypto native in me knows that the receipts will eventually be written on a ledger, not a spreadsheet. Chasing the ghost of Ethereum used to mean finding the next smart-contract platform. Now it means finding the next way to make compute liquid. Citigroup just built a ghost the size of 20 gigawatts. The market should chase it fast — but also watch for all the smaller, harder, decentralized shadows moving underneath it.