Google Cloud’s Earnings Miss: Why It Needs to Invest Even More in GPUs

According to the Financial Times, Alphabet, which reported earnings on the 5th, plunged 6.94% because growth in its cloud division’s revenue fell short of expectations.

First, for Google Cloud at its current scale to grow roughly 30% year-on-year is an impressive achievement in itself. Nevertheless, the presumed reason for the stock’s sharp drop is that this performance was weaker than the previous quarter’s 35% growth, which in turn suggests that market expectations for results were extremely high.

In response, Alphabet announced that it would raise this year’s capex from the previous $60 billion level to $75 billion, a move backed by the CFO’s comment that the problem was “demand that exceeds the capacity we have”. The essence is that, even though they are already investing astronomical sums in buying GPUs, demand is outstripping supply by an even wider margin.

As investor interest in agentic AI and other AI-as-a-service offerings grows, and as the launch of Nvidia’s Blackwell servers is delayed, we are seeing liquidity increasingly flow into AI services. But investment in AI-as-a-service and in the underlying infrastructure needs to proceed in parallel.

As DeepSeek’s V3 model delivers performance close to OpenAI’s GPT-o1 at a much lower cost, concerns have emerged that Big Tech companies, which have poured hundreds of billions of dollars into buying GPUs, may have overinvested. Reflecting this, the market reacted extremely sensitively, with the share price of Nvidia, the world’s leading GPU company, plunging.

A useful framework for understanding this phenomenon is Jevons’ paradox. In economics, Jevons’ paradox describes situations where technological progress or government policy increases the efficiency of resource use, but lower costs lead to higher demand. As a result, resource consumption does not fall; instead, demand grows so much that usage actually increases. This mirrors how, over the past several decades, the unit cost of computing has steadily fallen, while demand for computing has risen inversely.

Looking at the chart of GPU computing rental price trends compiled by Semianalysis above, you can see that after the launch of DeepSeek’s V3 model, rental prices for Nvidia’s H100—which had been flat or slightly declining—have actually risen.

Given that what ultimately dragged down Google Cloud’s results was a shortage of GPU computing resources, this can be interpreted as a phase in which the unit cost of “intelligence” is falling and, as a result, demand for AI-based services is exploding.