Google Q2 Cloud Revenue Up 82%, Capital Spending Raised Further

On July 24, 2025, Google disclosed its second-quarter results, with cloud business revenue up 82% year-on-year and capital spending further raised. The next day, Google's share price fell, leading the hardware sector lower.

Enterprises are beginning to calculate the return on AI spending. A US mobility platform's technology head spent $1,200 in a two-hour internal demonstration, and the engineering team used up the full-year AI budget in four months. 84% of engineers use a programming assistant with autonomous execution capabilities, and 70% of submitted code comes from AI tools. The company subsequently set a monthly spending cap of $1,500 per employee per tool. The output price gap between frontier models and cheap models can be as high as 100 times. Open model routing platforms process 25 trillion tokens per week.

Less than a month after the release of leading overseas models, Zhipu's GLM-5.2 and Moonshot AI's Kimi K3 were launched, with public benchmark scores close to each other. There is debate in the US on whether to restrict Chinese open-source models. One side advocates restrictions from a national security perspective, while the other emphasizes the value of open weights for competition, cost, and safety testing. No conclusion has been reached.

As of July 23, 2025, the combined capital spending of the top five North American cloud providers in 2025 was $411.5 billion. By 2027, it is projected to be approximately $1.02 trillion, up 32% year-on-year. Google's second-quarter single-quarter capital spending was equivalent to 115% of its operating cash flow for the same period, and free cash flow turned negative for the first time since listing. The company is supporting its investment through equity and debt financing.

Starting in 2025, chip companies such as Nvidia and AMD have been providing financing to labs like OpenAI and Anthropic in exchange for chip orders. Nvidia's investment in one of them could reach as high as $100 billion. In July, AMD announced an investment of up to $5 billion in another lab in exchange for up to 2 GW of accelerator deployment. The labs that received the money then turned around to buy computing power from cloud providers; one signed a cloud services contract with Oracle worth approximately $300 billion. Cloud providers, in turn, purchase tens of billions of dollars worth of chips from chipmakers to deliver computing power. At its July 16 earnings call, TSMC was asked whether it would follow its customers in making similar financial arrangements; management replied, "Not at present."

In May 2025, three memory manufacturers simultaneously appeared in the financing list of a top AI lab, officially positioned as "strategic infrastructure partners," and simultaneously signed multi-year supply agreements. In June, one of them reached a package agreement with the lab covering joint design, multi-year supply, on-premise deployment, and participation in financing.

Aggregated from the bottom up across 31 major global manufacturing enterprises, capital spending in the semiconductor manufacturing segment is expected to grow from $168.1 billion in 2025 to $341.7 billion in 2028, doubling in three years, with a compound annual growth rate of 27%. The increment is concentrated in the most advanced logic foundry and three memory manufacturers. Corresponding wafer fab equipment investment in 2028 is approximately $241.4 billion, and total manufacturing investment for the three years from 2026 to 2028 is $861.2 billion.

Price levels are still hitting new highs, but the acceleration has turned negative. The shares of the five major overseas memory manufacturers peaked in mid-to-late June and retraced 27% to 50% within a month.

Overseas interest in Chinese technology assets is concentrated in two chains: first, Chinese companies in the global AI supply chain, represented by optical modules and printed circuit boards; second, localization, covering foundry, AI chip design, and equipment.

The US Remote Access Security Act (RASA) has passed the House of Representatives and still needs Senate approval and signature to become law. The bill aims to bring "leasing regulated computing power across the ocean" under authorized management. Another bill, the Hardware Technology Control Multilateral Coordination Act (MATCH), is at the proposal stage and requires the Netherlands and Japan to adopt export control rules equivalent to those of the US, covering at least all immersion lithography equipment.

On open model platforms, Chinese models already account for more than 30% of US enterprise call volume, with a peak close to 46%.