AI Capex Grew over 80% Annually in Q1 to $90 billion
Led by hyperscalers, AI Capex reached $90 billion in Q1, an increase of over 80% annually.. As you can see from the chart below, the large cloud providers plus TSMC accounted for over 90% of the industry’s capital expenditures. If you’d like to a little reverse engineering, a couple of data points to help:
NVIDIA reported $39 billion of data center revenue in the quarter, the vast majority of which is booked by customers as “capital expenditures”.
NVIDIA’s data center revenue grew 73% year-over-year to $39 billion, close to the 81% growth in capex for the top 4 hyperscalers
AI data centers on average need about $15,000 per kW of capex to construct the facility
Capex is more or less trending to NVIDIA’s GPU revenue. The large cloud providers need to buy them, and TSMC is building out the capacity to manufacture them. Both NVIDIA and AMD are using TSMC’s “Chip on Wafer on Substrate” or CoWoS technology for fabrication for their latest designs, for which TSMC has been expanding production capacity at its new U.S. fab in Arizona.
The major cloud providers are expected to continue the 80% growth rate based on comments in their earnings calls. If you start compounding this growth you can get to $1 trillion pretty quickly, and $3-5 trillion not long after that, and I’m sure someone somewhere has a forecast that says exactly that , but spend is ultimately grounded by its share of revenue and AWS, which provides over half of Amazon’s operating profit, has hit a capex level of 70% of revenue, and it’s not going to stay there for more than a couple of years. That said, even without low interest rates, smaller GPU rental providers like CoreWeave and Lambda have had no problems raising capital, which in turn is putting pressure on the large clouds to act now.
