Hey there 👋, so Google $GOOGL kicked off big tech earnings season and delivered roughly what I expected. Brutal profits, but at the same time increased capex, now $195 to $205 billion for this year, which spooked the market a bit after the results. The guys here have already broken down the results themselves and summed them up brilliantly, so I'll focus on one thing. Cloud grew 82% to $24.8 billion and the backlog of contracted orders swelled to $514 billion, compared to $106 billion a year ago. Who's actually driving that demand? Because the idea that it's just about computing power for training AI is already a bit off the mark.
1. The shift from training to inference, i.e. AI running 24/7
Training a model is like writing a textbook. A massive one-off computation that takes weeks to months. Inference is when millions of people and systems continuously browse that textbook and ask questions. Every company that has embedded AI into its products, from customer chatbots to searching internal documents, needs chips running 24 hours a day. And as voice, image and video generation are added to applications, the demands don't grow linearly, but exponentially. You can see it directly in Google's numbers. The Gemini API processes 22 billion tokens per minute, up from 16 billion last quarter, and nearly 500 cloud customers have burned through over a trillion tokens in the past year. Plus, the AI labs themselves are buying compute, for instance Anthropic has contracted up to a million TPU chips with Google.
2. Traditional corporations, i.e. banks, retail and carmakers
When you tap your card in a store, hundreds of parameters are evaluated in the cloud in milliseconds – where you are, what your shopping habits are and whether it's an attack – so the bank can approve the transaction. Carmakers, in turn, are developing autonomous driving, and their test fleets upload petabytes of video daily, which needs to be processed in the cloud and run through simulations. And e-commerce players have algorithms continuously recalculating the prices of millions of items based on inventory, competition and even the weather. Local servers simply aren't enough for these data volumes.
3. Pharma and research
Drug development used to take 10 to 15 years, because simulations of molecular bonds were done physically in labs. Today, so-called in silico research is underway, where computers in the cloud simulate billions of chemical combinations and protein folding within a few days. Just look at AlphaFold from Google DeepMind, which won a Nobel Prize. That's why big pharma and genomics startups are increasingly hungry cloud customers.
4. The SaaS multiplier as the hidden engine
This is, in my opinion, the most underestimated part. When a company buys an accounting system, CRM or graphic tool, that software almost never runs on the developer's own servers. Software companies themselves are huge buyers of compute from Google, AWS $AMZN or Azure $MSFT. As soon as thousands of companies start using a new tool, its developer must immediately buy massive additional capacity on the backend. Every successful SaaS product thus multiplies demand for the cloud.
5. Cybersecurity
Security has moved from local antivirus to the cloud. Modern systems analyze network traffic on millions of devices simultaneously, looking for anomalies in terabytes of logs per second. That's impossible without constant massive compute, and with the growing number of attacks, this item is only going to grow.
All in all, cloud demand isn't a single bet on AI training, but several layers that multiply each other. That's why the capex that spooked the market still makes sense to me. Google itself admits that capacity is still insufficient and companies are signing up for years in advance.
How do you see it? Is the record capex a justified investment into clear demand, or does it already look to you like an exaggerated bet that this growth won't slow down?
