The top three cloud service companies: Amazon, Google, and Microsoft account for 65% of the global cloud service market and are regarded as “AI sellers”. The three companies greatly drive demand for computing power cards and have always been the focus of market attention.

Today, Google/Microsoft held quarterly results meetings one after another. Overall, driven by AI, Google's 24Q1 revenue and profit both exceeded expectations, while forecasting a year-on-year increase of at least 49% for Capex in '24; Microsoft's revenue and profit were also higher than expectations, and it is also guiding that AI investment will continue to expand in 2024, which is expected to increase significantly in the next quarter.

1. Overseas financial reports were gradually disclosed, and overall AI capital expenses were revised up

Overall, the demand for AI infrastructure is very strong:

• META: Annual capital expenditure increased from 300-370 to 350-400 to build data centers to support AI business;

• Google: The capital expenditure for each quarter of 2024 is expected to be more than 12 billion US dollars, an increase of at least 49% year-on-year for the whole year to maintain its leading position in AI infrastructure;

• Microsoft: said that demand for computing power cards exceeds supply, and decided to expand the scale of AI investment. The capital expenditure for the quarter was 14 billion yuan (expected to be 13.14 billion yuan), and capital expenditure is expected to increase sharply from month to month next quarter;

• Amazon: Awaiting the first quarter results meeting on April 30.

2. Strong demand, and the supply of computing power cards is in short supply

At the performance conference, Microsoft mentioned that the current market demand for AI is greater than the company's reserves. At the same time, during this earnings season, we have also seen increased capital expenditure by various manufacturers, all of which indicate the current boom in the AI industry.

On the other hand, from a supply chain perspective:

Training side ——

• As of the end of 2023, H100 shipped a total of 1.2 million copies (Omdia data)

• Train a GPT-4 with 1.8 trillion parameters: requires 8,000 H100

In other words, the H100 cards currently on the market are enough to train nearly 150 GPT-4 models. However, considering the later iterations of GPT-5 and GPT-6, as well as the demand for image models and future video models, demand on the training side is still very tight. Factorial Funds mentioned in a paper published in March of this year that after Sora was officially launched, it would take about 720,000 H100 copies to run.

Reasoning side ——

In 2023, with the optimization of the functions of various major models, many models have already been opened to the C-side, which has also catalyzed market demand for inference cards.

NVDA mentioned at a recent conference that demand for mystery cards in 2024 will increase significantly compared to 2023, and the shipment ratio will reach 70%.

The recent decline in delivery cycles that the market is concerned about is presumably more due to the release of CowOS and HBM production capacity.

But we still need to be wary of the cyclical nature of the industry. Referring to history, the semiconductor industry cycle is 4-5 years, and each cycle is catalyzed by new technology iterations. When the top of this AI cycle will arrive also requires continued attention.

3. AI empowers business improvement

After Meta applied AI to its products, it received good performance rewards. Google and Microsoft have also achieved good revenue and profits driven by AI.

Currently, Google has a clear AI commercialization path in ads/clouds/subscriptions:

• Advertisement: AI is connected to the push system to improve push accuracy

• Cloud service: computing power card rental

• Subscription: Launched a subscription service for big models

Microsoft has also launched corresponding products in Office/Windows/Cloud, etc.:

• Office: Connect to Copilot and launch a subscription service

• Cloud service: computing power card rental

At this performance conference, the cloud service business of both Microsoft and Google grew dramatically, and both companies also raised capital expenses to support the construction of AI infrastructure.

4. Earth-breaking dividends, Google ushered in a revaluation

Also an “AI seller,” Google's valuation is significantly lower than that of its peers.

On the one hand, the search business was challenged by GPT and Perplexity; but the other important reason is that Google decided not to pay dividends from the day the company was founded. Compared to giving back dividends to shareholders, Google is more inclined to invest profits in R&D work. This is very unfriendly to the company's market value management, so the company's valuation is also much lower than that of its peers.

However, unexpectedly, Google received its first dividend at this performance meeting. At the results meeting, Alphabet said that the board of directors has approved the payment of a cash dividend of 20 cents per share to shareholders registered as of June 10 on June 17, and “plans to continue to pay quarterly cash dividends in the future.” Additionally, the company authorized an additional $70 billion in share repurchases. After the results meeting, Google rose 12.18%, and its market capitalization increased by nearly US$236.2 billion (approximately RMB 1.7 trillion).

Overall, the continued boom in the AI industry has brought bright moments to “AI sellers.” As a key part of artificial intelligence infrastructure, demand for computing power cards will continue to grow. In the future, how will AI sellers further expand the market and tap new growth potential? It's worth looking forward to.

Here are the highlights of Google and Microsoft's performance meetings

Google's first quarter results:

Financial-related:

• Revenue: US$80.5 billion (yoy +15.4%), exceeding market expectations by 1.9%;

• Net profit: US$23.7 billion (yoy +57%), exceeding market expectations by 21%;

• Capital expenditure of US$12.012 billion in the first quarter (yoy +91%, QoQ +9.2%);

• Capex is expected to reach more than 12 billion US dollars in every quarter of 2024, with a year-on-year increase of at least 49%, exceeding expectations;

• “Unbreakable” first dividend+$70 billion repurchase plan

Business-related:

• Currently, there is a clear AI commercialization path in advertising/cloud/ subscription, etc.;

• In addition to TPU, it also provides NVDA computing power card services, including H100 and the newly released Blackwell;

• Reduced sales, general and administrative expenses, but increased research and development

Microsoft's first quarter results conference:

Financial-related:

• Revenue of US$61.9 billion, YoY +17%, exceeding Bloomberg's agreed estimate of US$60.9 billion;

• Net profit of US$21.9 billion, YoY +20%, exceeding Bloomberg's agreed estimate of US$21.1 billion;

• Capex +65% YoY and +13% Month-on-Month this season;

• Guide to continue expanding the scale of AI investment in 2024, which is expected to increase significantly over the next quarter

Business-related:

• Cloud services increased by 30% in Q1, leading to continued growth in cloud services in the next quarter;

• Indicate that the current supply of inference computing power is in short supply