AI Financial Reporting Season: Giants Choose Different Directions in AI Development The most noteworthy aspect of this AI financial reporting season is not revenue, but the way the giants have demonstrated their financial performance. On the road of AI, no one wants to compete with anyone. Google is the most versatile. We will handle the full stack of cloud, chip, and model ourselves, and TPU will contribute tens of billions of dollars annually, with a capital expenditure of approximately 92 billion dollars by 2025. The entire chain needs to be held in its own hands. Amazon doesn't roll models, just focus on making Trainium chips. Trainium's annual revenue is billions of dollars, with over one million units in production. It gave Anthropic over 500000 pieces to build Project Rainier and sold its computing power to OpenAI. Not building the smartest brain, only selling machines that run the brain. NVIDIA pushed selling shovels to the extreme. It invests up to $100 billion in OpenAI to build a data center with at least 10 gigawatts and millions of GPUs. As both a supplier and a financial owner and shareholder, the balance sheet is directly attached. Microsoft's gaming platform is no longer tied to OpenAI. Azure AI Foundry has integrated OpenAI, Claude, Grok, Mistral, and has tens of thousands of models serving over 70000 customers. It turns AI capabilities and agents pushed into the enterprise into water and electricity, connecting anyone with a good model. Meta is the fiercest and most awkward. By 2025, capital expenditures will be between 66 billion and 72 billion US dollars, and will continue to rise in 2026. This confirms the Heart Roll model and drives the Hyperion cluster towards 5 gigawatts. However, advertising still dominates in terms of revenue, and AI has not directly earned any extra money. Reality Labs has also lost 4.5 billion yuan. All investment, returns float in the sky. Musk is the wildest. SpaceX traded $60 billion in stock for cursor, holding Memphis Colossus, equivalent to 1 million H100 cards, and rented computing power to Anthropic and Google for approximately $26 billion per year. Put computing power, models, and applications into one basket. Apple is alternative. The capital expenditure for fiscal year 2025 is only $12.7 billion, and by 2026 it will be approximately $14.3 billion. Don't build a large model, switch to ChatGPT, and later switch to Gemini. Use your own Private Cloud Compute to mix and rent computing power. But there is no such thing as a free lunch in the world. AI hardware and tariff costs are rising, and iPhone revenue is showing signs of fatigue. The rising prices are being eaten up by costs. In the same AI trend, some people are all inclusive, some only sell shovels, and some burn money to create models. The route is not right or wrong, and the account must be repaid in the end.