Summary of Key Points
The United States' embargo on high-end Chinese AI chips has been escalating, ranging from the ban on the A100/H100 to the A800/H800, and then to the H20/H200. Despite NVIDIA's efforts to create "specialized versions" of these chips with reduced performance, they still cannot escape the restrictions. In this context, the core issue for domestic AI companies has shifted from being unable to obtain chips to the inability to afford the infrastructure required to operate their computing centers—costs related to electricity, cooling, and data center facilities have become new major hurdles. Large companies are investing heavily in building these centers, but they face pressure on their profits. Meanwhile, green energy has become crucial for AI infrastructure, and the race for energy resources has become a central aspect of future AI competition.
I. The US Chip Embargo: Tightening Restrictions and the Rise of Domestic Alternatives
The US's control over Chinese AI chips is becoming increasingly stringent:
- October 2022: Ban on the A100/H100 (which exceed a certain computational power threshold); NVIDIA released the reduced-performance A800/H800 as a substitute.
- October 2023: The A800/H800 were also banned; NVIDIA then introduced the H20 with significantly lower performance, making them highly sought after.
- 2025: Sales of the H20 were suspended and later resumed under harsh conditions, resulting in virtually no actual purchases due to security concerns.
- 2026: NVIDIA's revenue from China accounts for less than 6%, and its market share has dropped to 37%, with domestic chips (such as Huawei's Ascend 910B and Cambricon's Synerge 590) surpassing it in performance at a quarter of the price.
With the import route for high-end chips essentially blocked, domestic alternatives have become the only viable option.
II. AI Computing Centers: Power-Hungry Monsters
AI computing centers are extremely energy-intensive:
- Energy Consumption: A single 700W AI chip consumes ten times the power of a traditional CPU; a cluster of tens of thousands of chips can use 200,000 kWh of electricity in a day (equivalent to the annual electricity consumption of a small county). A 1GW computing center uses nearly 10 billion kWh annually, costing 3.5 billion yuan at Xinjiang's current electricity rate of 0.35 yuan/kWh.
- Operational Costs: The daily operating cost for an AI center can range from 132 million to 240 million yuan, while its daily revenue may be less than 1 million yuan. According to Guolian Securities, ByteDance's operations could result in annual losses approaching 100 billion yuan.
- Rising Infrastructure Costs: The rental price of AI cabinets has increased by 10%-20%, with liquid-cooled cabinets experiencing even higher increases, pushing the competition for infrastructure expansion.
While chip costs are a one-time investment, electricity expenses represent a continuous financial burden that companies find increasingly difficult to bear.
III. Large Companies Spending Heavily on Infrastructure
To meet the demand for computing power, large companies are investing heavily:
- ByteDance: Invested 70 billion yuan in building a computing center in Ulanqab and another 120 billion yuan in Helingeer; planned capital expenditures for 2025 exceed 150 billion yuan, with 90 billion yuan dedicated to AI infrastructure.
- Alibaba: Plans to invest 380 billion yuan over the next three years in cloud and AI infrastructure, exceeding the total investment of the previous decade. Net profit in Q4 of 2026 plummeted by 84%.
- Tencent: Invested 847 billion yuan in the first half of 2026, more than the entire year of 2025, for its MixNeng model and cloud computing capabilities. These investments have affected their cash flow and profits.
These expenditures have significantly reduced the profit growth of these companies, with rumors of declining net profits.
IV. The Future of AI: Energy Competition and Green Energy
Energy is a critical foundation for AI, and green energy has become a key factor:
- Vision Technology: Its Ulanqab Star River Base (the world's largest single AI computing center) uses 67% green energy supplied by local wind power and plans to build 5GW of green computing centers in the Gobi Desert by 2030.
- National Policies: The national integrated computing network aims for green energy to account for more than 80% of total energy use by the end of 2025, with energy efficiency targets below 1.5%.
- The US Also Faces Energy Shortages: In 2025, AI will account for 4.3%-5.2% of the US's total electricity consumption. Tesla data centers in California have seen training efficiency reduced by 30% due to power rationing caused by high temperatures.
Companies that can access cheap and stable green energy will gain a competitive advantage in the AI race.
V. Shift in Cost Focus: From Chips to Infrastructure
The current situation is as follows:
- Chips Are No Longer the Biggest Problem: Domestic chips outperform the H20, with their market share exceeding 60%, making them easier to obtain.
- Infrastructure Costs Have Become a New Challenge: Operating costs (electricity, cooling, maintenance) are increasing significantly. Although these costs are not directly visible to users, they pose a substantial burden on companies.
- Shift in Capital Allocation: Companies are shifting their investments from chips to infrastructure. For example, ByteDance's 90 billion yuan in AI infrastructure purchases was largely related to infrastructure development.
The competition in the AI industry has evolved from a focus on chips to a battle for infrastructure and energy resources. Those who can control these costs will be able to survive and thrive.
Conclusion
The US chip embargo has forced China's AI industry to focus on improving its internal capabilities, particularly in infrastructure and energy management. Large companies are investing heavily in building computing centers while facing pressure on their profits. In the future, green energy, efficient cooling systems, and low-cost operations will become the core competitiveness for AI companies. This "infrastructure battle" is just beginning.