The Moment of Bold Bet on AI Computing Power: Who's Paying the Price, and Who's Running Naked?
Hello everyone, I'm your financial journalist friend. Recently, something big has happened in the AI community: everyone has stopped talking about saving money and is now talking about borrowing and spending a lot of it.
JD.com has announced plans to build a supercluster with 100,000 graphics processing units (GPUs), ByteDance has secured a massive loan of $29.6 billion (about 210 billion RMB), and the budgets of Alibaba and Tencent have doubled. Across the ocean, OpenAI and Anthropic have signed contracts worth astronomical amounts.
Many people are confused: Isn't AI still not making much money? Why are companies buying equipment, building data centers, and taking out high-interest loans like crazy?
Today, I'll break down this news in simple terms, explaining the logic behind this “computing power arms race” and the potential risks involved.
---
Where's the Money Going? From “Telling Stories” to “Buying Shovels”
In the past, big companies talking about AI would focus on things like the size of their models, the number of users, and when they would start making money. But now, the focus has shifted.
The money is being spent on things you can't see: GPUs, data centers, electricity, and computing resources that need to be reserved years in advance.
It's like when people used to discuss where the gold mines were; now, they're just buying shovels, mining trucks, and building roads first.
- JD.com's approach is pragmatic:曹 Peng (JD.com's CEO) put it plainly: JD.com's AI is not for publishing papers or showing off; it's for use in warehouses, delivery stations, and robots. They plan to buy 3 million robots and 1 million autonomous vehicles. Building a cluster with 100,000 GPUs is to ensure that AI can control these devices. This is about “measuring your feet before making shoes” – buying as much computing power as the business needs.
- Alibaba's approach is long-term: Wu Yongming (Alibaba's CEO) said they will invest 380 billion RMB over the next three years, more than the total of the past decade. They want to completely rebuild their cloud and AI infrastructure, betting that all businesses will rely on AI in the future.
- ByteDance's approach is aggressive expansion: ByteDance plans to invest over 200 billion RMB this year, with half of that going towards buying chips. They got a loan of $29.6 billion at a low interest rate, indicating that banks believe ByteDance will be able to repay it. They're betting that with the strongest computing power, their AI applications (like DouBao and JianYing) will run the fastest and best.
- Tencent's approach is steady profitability: Tencent is investing less, but their cloud business is already profitable. They're using the profits from their existing businesses to slowly feed AI development, focusing on stability.
In summary: Before, it was about “using AI to make money”; now, it's about “first having the AI infrastructure, or else you're out.”
---
Different Approaches in China and the US: The US Prefees, China Does the Math Carefully
The ways China and the US are playing this game are very different, which determines who faces greater risks.
The US: Prepaying for “Certainty”
American companies (like Microsoft, Amazon, Meta) and AI firms (like OpenAI, Anthropic) are adopting a strategy of locking in the future.
- Anthropic’s example is extreme: In the past 11 months, it signed contracts worth $51.7 billion, eight times its annual revenue! Many of these contracts are “take-or-pay” (you have to pay regardless).
- Why do they do this? Because US capital is abundant, and they fear that if they don’t lock in computing power now, they won’t be able to buy chips and electricity when AI demand surges. So, they prefer to pay a lot now to ensure they have the resources needed in the future.
- Risks: If AI demand doesn’t materialize, these huge contracts become a huge liability, and the companies could go bankrupt.
China: Being Careful within Constraints
Chinese companies face two challenges:
1. Limited cash flow: Moody’s reports suggest that the top Chinese companies may see negative free cash flows in the next two years, and Chinese companies have relatively less financial flexibility.
2. Chip shortages: They can’t buy or only get a small amount of high-end NVIDIA chips.
Therefore, Chinese companies’ strategy is to link computing power to specific businesses. For example, JD.com’s computing power will be used in warehouses and delivery, Baidu’s in search and logistics, and Alibaba’s in cloud services.
- The difference between ByteDance’s loan and Anthropic’s contract: ByteDance’s money is borrowed, backed by its business credibility; Anthropic’s money is a prepayment, based on faith in the future.
In simple terms: The US is betting on a “big future explosion,” so they pay in advance; China is betting on “doing its current business well” while buying the necessary infrastructure.
---
Different Approaches in China and the US: The US Prefees, China Does the Math Carefully
The ways China and the US are playing this game are very different, which determines who faces greater risks.
The US: Prepaying for “Certainty”
American companies (such as Microsoft, Amazon, Meta) and AI firms (such as OpenAI, Anthropic) are adopting a strategy of locking in the future.
- Anthropic’s example is extreme: In the past 11 months, it signed contracts worth $51.7 billion, eight times its annual revenue! Many of these contracts are “take-or-pay” (you have to pay regardless).
- Why do they do this? Because US capital is abundant, and they fear that if they don’t lock in computing power now, they won’t be able to buy chips and electricity when AI demand surges. So, they prefer to pay a lot now to ensure they have the resources needed in the future.
- Risks: If AI demand doesn’t materialize, these huge contracts become a huge liability, and the companies could go bankrupt.
China: Being Careful within Constraints
Chinese companies face two challenges:
1. Limited cash flow: Moody’s reports suggest that the top Chinese companies may see negative free cash flows in the next two years, and Chinese companies have relatively less financial flexibility.
2. Chip shortages: They can’t buy or only get a small amount of high-end NVIDIA chips.
Therefore, Chinese companies’ strategy is to link computing power to specific businesses. For example, JD.com’s computing power will be used in warehouses and delivery, Baidu’s in search and logistics, and Alibaba’s in cloud services.
- The difference between ByteDance’s loan and Anthropic’s contract: ByteDance’s money is borrowed, backed by its business credibility; Anthropic’s money is a prepayment, based on faith in the future.
In simple terms: The US is betting on a “big future explosion,” so they pay in advance; China is betting on “doing its current business well” while buying the necessary infrastructure.
---
The Rise of Domestic Chips: A Boom with Hidden Concerns
In a gold rush, the most stable ones aren’t the ones digging for gold but the ones selling the tools. One of the biggest beneficiaries of this AI boom are domestic GPU chip manufacturers.
Why have domestic chips become so popular?
1. Surging demand: In the first quarter of this year, domestic AI computing power demand increased by 417%, while supply only increased by 128%. High-end chips are in high demand.
2. Forced choice: Due to export restrictions, domestic chips have become a necessity rather than an option.
What’s happened in the capital market?
- Valuations soaring: Shuoyuan Technology’s stock price soared 188% on its debut day, with a market value of 170 billion RMB; other domestic GPU companies like Moore Threads, Muxi, and Beren also had huge gains.
- Huawei’s “super node” strategy: Although Huawei’s single-chip performance isn’t as good as NVIDIA’s, it combines thousands of chips into a supercomputer (super node). For example, the Atlas 950 has 6.7 times the performance of NVIDIA’s products at the same time. This is using “system advantages” to compensate for “single-point weaknesses.”
- From selling chips to selling systems: Previously, they sold individual GPUs; now, they sell complete solutions (chips + interconnectivity + software).
However, there are significant risks:
- Valuations far exceeding performance: Except for Cambricon, most domestic chip companies are still losing money. Shuoyuan’s market-to-sales ratio was as high as 171 on its debut, indicating high expectations that haven’t been met yet.
- Liquidity pressure: Companies like Moore Threads are about to face a large amount of restricted shares being released, which could put pressure on their stock prices.
- Circular trading risks: There’s a similar cycle in Silicon Valley, where Tencent is both a shareholder and a customer of Shuoyuan. While this is stable, it could lead to risks if the industry declines.
In summary: Domestic chip companies are doing well now, but whether they can turn their “market value” into “profit” depends on the actual success of their products.
---
The Biggest Paradox: The Most Aggressive Players Are Advising Caution
Here’s a very interesting phenomenon that’s also a key indicator of risk:
Sam Altman, CEO of OpenAI, has recently started advising the industry to be cautious in podcasts.
- He signed the most contracts, driving up global computing power investment.
- Now he says, “Many new cloud companies are emerging, claiming they’ll build massive computing power next year, but they have no revenue and no buyers.”
Anthropic is also facing huge contract pressures before its IPO.
- It signed contracts worth $51.7 billion, but much of the computing power won’t be available until 2027.
- If AI demand doesn’t meet expectations in 2027, these contracts will be a huge burden.
Why are the most aggressive players advising caution?
1. Risk transmission: They’ve realized that many of the new cloud companies (those that buy chips and rent them out) are shell companies, financing through hype. If these companies fail, chip manufacturers and data centers will be affected.
2. Physical limitations: Electricity and land are limited. When there’s so much money that it loses its meaning, people start talking in terms of “gigawatts” (units of electricity). This indicates that the bottleneck has shifted from “money” to “physical resources.”
What does this mean?
- A bubble may be forming: When even the most aggressive players are worried about no buyers, it suggests that market sentiment has shifted from “enthusiasm” to “anxiety.”
- Differentiation will intensify: Only the companies with real businesses and cash flows (like JD.com, Alibaba, Tencent, ByteDance) will survive. Those that rely on financing and have no actual business support may be the first to fail.
---
The Ultimate Question: Who Will Last in This Gamble?
Finally, let’s answer the core question: What exactly is this AI computing power gamble about?
It’s not about who has the most money, but about who can endure the longest.
1. For American companies: The bet is on “sustained high AI demand.” If they’re right, the computing power they lock in now will become a scarce asset, leading to huge profits; if they’re wrong, the long-term contracts and massive data center investments could be the final straw.
2. For Chinese companies: The bet is on “domestic substitution and business integration.” If domestic chips can keep up and AI can be effectively integrated into e-commerce, logistics, content, etc., the investments will translate into productivity. If the investments are too rapid and cash flows dry up, or if domestic chip technology falls short of expectations, they’ll face difficulties.
Implications for ordinary investors:
- Don’t blindly chase AI stocks with high valuations: Many chip companies’ valuations have already exceeded their expected performance for the next few years.
- Focus on companies with real businesses: Those like JD.com, Alibaba, Tencent, and ByteDance, where AI investments directly benefit specific businesses (such as logistics, cloud services, advertising), have stronger resilience.
- Be cautious of companies that rely solely on renting computing power: Those without core technology and that only buy chips to rent them out are at the highest risk.
In conclusion:
We’ve entered a moment of bold bet on AI computing power, and no one dares to stop first. Because if they do, they might be left behind by the times. But as the gamble continues, what’s being tested is no longer just courage, but endurance and efficiency.
Whoever can truly transform computing power into productivity will be the one who laughs last. Those that rely on hype and financing to stay afloat may disappear in the next cycle.
This game has just begun, but the elimination round has already started.