Giants have significantly increased capital expenditures, and the "arms race" in AI computing power has escalated
Silicon Valley’s AI computing power “arms race” is escalating at an almost crazy speed. On July 22, local time, it was revealed that OpenAI had significantly raised its forecast for computing power expenditures before 2030 from approximately US$600 billion to US$750 billion. On the same day, Alphabet, the parent company of Google, handed over a financial report with revenue that exceeded expectations. However, due to negative net cash flow for the first time in history and another increase in capital expenditure guidance, the stock price fell after the market closed.
Silicon Valley’s AI computing power “arms race” is escalating at an almost crazy speed. On July 22, local time, it was revealed that OpenAI had significantly raised its forecast for computing power expenditures before 2030 from approximately US$600 billion to US$750 billion. On the same day, Alphabet, the parent company of Google, handed over a financial report with revenue that exceeded expectations. However, due to negative net cash flow for the first time in history and another increase in capital expenditure guidance, the stock price fell after the market closed. "China Business News" reporter noted that SpaceX is currently considering expanding its data center business in Texas; Microsoft is once again increasing its European AI footprint and signing a multi-billion-dollar data center cooperation agreement with Mistral; AMD announced that it will invest up to 5 billion US dollars in Anthropic, and Anthropic has committed to purchasing up to 2GW of AMD's latest generation AI chips. In this game where computing power determines life and death, global technology giants are taking an unprecedented radical stance, transforming from mere "computing power buyers" to "infrastructure maniacs". Behind this is not only the belief in Scaling Law, but also the desire for the right to speak in the future AI era. When the threshold for computing power infrastructure rises from tens of millions of dollars to tens of billions of dollars, it itself becomes an extremely wide moat. The continued higher-than-expected increase in the AI capital expenditures of giants will drive the further development of optical modules, fiber optic cables, servers, and storage sectors. There is no ceiling for computing power expenditures According to people familiar with the matter, the increase in OpenAI's computing power budget is mainly due to the company's new long-term agreements signed with multiple cloud computing providers, as well as its efforts to strengthen the construction of its own data centers after the "Stargate" data center project did not progress smoothly. As the latest move, OpenAI announced a $20 billion investment in the "Camellia Project" in Effingham County, Georgia, where it will independently design and develop data centers for the first time and contract with the local power company to obtain 3.2 GW of power supply from 2028 to 2032. "OpenAI is shifting from 'renting computing power' to 'building computing power.'" Li Yan, an investor who has long tracked AI infrastructure, told reporters that before, OpenAI mainly rented computing power from Oracle, Amazon AWS, etc., and now it has hired Brent Mayo, who built the Colossus data center for xAI, as the person in charge of data center construction. “This shows that model companies have realized that computing power cannot always be in the hands of others.” At the same time, AMD and Anthropic announced an "equity-for-order" cooperation on the same day: AMD will invest up to US$5 billion in Anthropic, and Anthropic promises to purchase up to 2GW of AMD's latest generation MI450 series GPUs starting in the first half of 2027. This deep binding model is becoming the new normal in the AI industry chain. Alphabet, Google’s parent company, is even bleeding. The financial report released after the market closed on July 22 showed that Google’s capital expenditures in the second quarter reached US$44.9 billion, resulting in negative free cash flow (-US$5.9 billion) for the first time in the company’s history. “Google is betting on AI with its wealth and life,” Li Yan pointed out. Despite strong revenue and cloud business growth, Google still raised its full-year capital expenditure forecast to $195 billion to $205 billion. Its CEO Pichai bluntly stated that the demand for AI still exceeds the amount of investment. To raise money, Google even raised $49.6 billion through a stock offering in June. Microsoft is not far behind. The company expects capital expenditures in natural year 2026 to be approximately US$190 billion, a year-on-year increase of approximately 48%, of which approximately US$25 billion will come from price increases for AI core components such as GPUs and CPUs. The company also plans to double its overall computing power within 2 years. Why are the giants still adding more money? The most direct reason is that demand far exceeds supply. Google CFO Anat Ashkenazy admitted that the company is "still in a supply-constrained environment" and that both external cloud customers and internal business demand are very strong. OpenAI’s flywheel logic is also very clear: “More computing power drives smarter models, smarter models drive better products, and better products drive faster adoption, more revenue, and more cash flow.”
"The core strategy of technology giants to increase their computing power can be summarized as: build their own infrastructure to seize the initiative, and sign long-term agreements to lock in future production capacity." Li Yan believes that Google Cloud's revenue increased by 82% in the second quarter, and its backlog of cloud business orders rose to 514 billion U.S. dollars; Microsoft's AI annualized revenue has exceeded 37 billion U.S. dollars. Investments in computing power are already starting to yield some measurable business returns. Xu Bei, an analyst at Zhixun Think Tank, said that in the capital market in the first half of 2026, under the market shock, "AI infrastructure" sectors such as power, chips, and IDC liquid cooling bucked the trend and strengthened, becoming the absolute main line of capital pursuit. AI competition is accelerating into the "heavy industry era." The United States is "stacking scale" and China is "fighting for system" This computing power race takes on different faces on both sides of the ocean. The strategy of American giants is "big" - larger clusters, higher budgets, and longer contracts. But this "heap scale" model is encountering increasing financial pressure - Google's first negative cash flow in history worries the market, "the most reliable cash flow generator on the market now spends more than it earns." China’s computing power race presents another path. At the just-concluded 2026 World Artificial Intelligence Conference, “super nodes” became the protagonist. Huawei's Ascend 950 super node real machine debuted for the first time, achieving the industry's largest 1024 card scale; Alibaba Cloud demonstrated the Zhenwu 890 Panjiu super node; ZTE released the OEX super node. According to calculations by Huatai Securities, the domestic super node market space is expected to reach 341.4 billion yuan in 2028. "American giants are competing for capital scale and procurement volume, while Chinese manufacturers are competing for system integration and architectural innovation." A domestic chip company executive who participated in WAIC told reporters that the core of a super node is not to stack cards, but to allow thousands of cards to work like a computer - this tests the system engineering capabilities of interconnection, heat dissipation, and software stacks. In terms of the share of domestic chips, China will ship approximately 4 million AI accelerator cards in 2025, of which approximately 1.65 million domestic chips will be shipped, with the share exceeding 40% for the first time. Goldman Sachs predicts that the domestic AI chip market share is expected to exceed 50% in 2026. At the same time, the capital expenditure of domestic Internet giants is also accelerating. Alibaba’s capital expenditures in the next three years may exceed the 380 billion yuan previously promised; Tencent’s capital expenditures in the first quarter of this year were 31.9 billion yuan, a new single-quarter high. Alibaba CEO Wu Yongming said that "almost no card in the server is empty" and the return on investment in the AI data center is very certain. But the challenges cannot be ignored either. Some analysts pointed out that although the capital cost per unit of computing power of domestic chips has been reduced by 40% to 50%, the performance per watt is only 10% to 30% of that of foreign chips. How to move from “affordable construction” to “well-used” is still the threshold that domestic computing power must cross. From a global perspective, competition in AI computing power has escalated from a “card-buying race” to a “system war.” OpenAI is competing for power, land, and talent; Huawei is developing full-stack capabilities in chips, interconnection, and heat dissipation; Google is seeking a balance between self-developed TPU and external procurement. Tokens are becoming the “steel” of the digital era, and whoever masters the lowest-cost Token production capabilities will master the pricing power in the AI era. In Xu Bei's view, giants are no longer just buying clouds and fighting for parameters, but are frantically building computer rooms and grabbing energy. "Stable, low-cost, large-scale token production" has become a new production capacity indicator, and data centers are changing from back-end IT cost centers to value centers that can directly contribute to profits. "We used to think that buying graphics cards was an arms race, but now we find that buying graphics cards is just the ticket. Now the logic has changed. Whoever can connect the electricity and build the cabinet is the real winner." Li Yan said that as for who can survive, it depends not only on who can run fast, but also on whose chassis is stable and whose energy consumption is low. Fang Lei, founder of Jiuzhang Yunji, believes that for large model companies, the industry generally shows that 1 US dollar of revenue corresponds to a computing power cost of about 0.8 US dollars. But currently, 1 US dollar of income also corresponds to several times the market value or valuation, and the scale of computing power is still the most important means of growth for large model companies.
Industry insiders pointed out that when computing power changes from "R&D budget" to "valuation narrative" and from "cost center" to "profit engine", the giants are no longer just calculating a financial account, but a ticket to the right to speak about the future industry.
"The core strategy of technology giants to increase their computing power can be summarized as: build their own infrastructure to seize the initiative, and sign long-term agreements to lock in future production capacity." Li Yan believes that Google Cloud's revenue increased by 82% in the second quarter, and its backlog of cloud business orders rose to 514 billion U.S. dollars; Microsoft's AI annualized revenue has exceeded 37 billion U.S. dollars. Investments in computing power are already starting to yield some measurable business returns. Xu Bei, an analyst at Zhixun Think Tank, said that in the capital market in the first half of 2026, under the market shock, "AI infrastructure" sectors such as power, chips, and IDC liquid cooling bucked the trend and strengthened, becoming the absolute main line of capital pursuit. AI competition is accelerating into the "heavy industry era." The United States is "stacking scale" and China is "fighting for system" This computing power race takes on different faces on both sides of the ocean. The strategy of American giants is "big" - larger clusters, higher budgets, and longer contracts. But this "heap scale" model is encountering increasing financial pressure - Google's first negative cash flow in history worries the market, "the most reliable cash flow generator on the market now spends more than it earns." China’s computing power race presents another path. At the just-concluded 2026 World Artificial Intelligence Conference, “super nodes” became the protagonist. Huawei's Ascend 950 super node real machine debuted for the first time, achieving the industry's largest 1024 card scale; Alibaba Cloud demonstrated the Zhenwu 890 Panjiu super node; ZTE released the OEX super node. According to calculations by Huatai Securities, the domestic super node market space is expected to reach 341.4 billion yuan in 2028. "American giants are competing for capital scale and procurement volume, while Chinese manufacturers are competing for system integration and architectural innovation." A domestic chip company executive who participated in WAIC told reporters that the core of a super node is not to stack cards, but to allow thousands of cards to work like a computer - this tests the system engineering capabilities of interconnection, heat dissipation, and software stacks. In terms of the share of domestic chips, China will ship approximately 4 million AI accelerator cards in 2025, of which approximately 1.65 million domestic chips will be shipped, with the share exceeding 40% for the first time. Goldman Sachs predicts that the domestic AI chip market share is expected to exceed 50% in 2026. At the same time, the capital expenditure of domestic Internet giants is also accelerating. Alibaba’s capital expenditures in the next three years may exceed the 380 billion yuan previously promised; Tencent’s capital expenditures in the first quarter of this year were 31.9 billion yuan, a new single-quarter high. Alibaba CEO Wu Yongming said that "almost no card in the server is empty" and the return on investment in the AI data center is very certain. But the challenges cannot be ignored either. Some analysts pointed out that although the capital cost per unit of computing power of domestic chips has been reduced by 40% to 50%, the performance per watt is only 10% to 30% of that of foreign chips. How to move from “affordable construction” to “well-used” is still the threshold that domestic computing power must cross. From a global perspective, competition in AI computing power has escalated from a “card-buying race” to a “system war.” OpenAI is competing for power, land, and talent; Huawei is developing full-stack capabilities in chips, interconnection, and heat dissipation; Google is seeking a balance between self-developed TPU and external procurement. Tokens are becoming the “steel” of the digital era, and whoever masters the lowest-cost Token production capabilities will master the pricing power in the AI era. In Xu Bei's view, giants are no longer just buying clouds and fighting for parameters, but are frantically building computer rooms and grabbing energy. "Stable, low-cost, large-scale token production" has become a new production capacity indicator, and data centers are changing from back-end IT cost centers to value centers that can directly contribute to profits. "We used to think that buying graphics cards was an arms race, but now we find that buying graphics cards is just the ticket. Now the logic has changed. Whoever can connect the electricity and build the cabinet is the real winner." Li Yan said that as for who can survive, it depends not only on who can run fast, but also on whose chassis is stable and whose energy consumption is low. Fang Lei, founder of Jiuzhang Yunji, believes that for large model companies, the industry generally shows that 1 US dollar of revenue corresponds to a computing power cost of about 0.8 US dollars. But currently, 1 US dollar of income also corresponds to several times the market value or valuation, and the scale of computing power is still the most important means of growth for large model companies.
Industry insiders pointed out that when computing power changes from "R&D budget" to "valuation narrative" and from "cost center" to "profit engine", the giants are no longer just calculating a financial account, but a ticket to the right to speak about the future industry.