Excess Supply of Hashrate!?

👆Click the blue text above to follow us

Ken Griffin

Currently, there are two opposing voices in the AI computing power market. On one side, we have Elon Musk and Sundar Pichai, the CEO of Google, who believe that demand for AI computing power is greatly outstripping supply, and that there is still substantial room for further growth in demand. On the other side, Ken Griffin, founder of Citadel, is warning about the current enthusiasm for computing power investments, suggesting the market may be greatly underestimating the risks of computing power oversupply. So, who is correct—entrepreneurs concerned about tight supply, or investment experts worried about oversupply?

First, let's look at Musk's perspective. During yesterday's SpaceX earnings call, Musk stated that demand for AI computing power is currently extremely strong. Both major cloud providers and many small- to medium-sized companies are scrambling for access to computing resources. If SpaceX accepted every customer's request, its existing computing power could be rented out several times over.

With such strong demand, Musk says that investing in AI can pay back in less than a year, and by next year, this period could shrink to just six months. As AI models advance, the value created per unit of computing power is also rising. Based on renewal data, the price per watt in GPU leasing is still increasing.

Musk's assessment is supported by Sundar Pichai, CEO of Google. In an interview, he noted that it's very difficult for clients to reserve computing power for as far ahead as 2027, and that overall demand for computing resources has doubled compared to initial plans. This is mainly because it typically takes two to three years for data centers to go from planning to operation, but current AI demand growth is greatly outpacing previous expectations at the time of planning.

In addition, Google must prioritize its own AI model R&D, sacrificing some revenue from cloud services. Pichai did not deny that there might be some localized oversupply in the short term, but from a long-term perspective, AI is still in its early stages of development—the currently planned computing power will eventually be fully utilized (a rather obvious remark).

Now that we've heard from two Silicon Valley representatives, let's turn to the opposing voice from Wall Street. Ken Griffin, founder of Citadel, does not deny that demand for AI computing power is currently strong, but he warns that investors may be underestimating three major risks arising from a possible future oversupply.

The first risk is a collapse in GPU leasing prices. Griffin likens it to airline tickets: once a plane is set to depart, even selling an empty seat for $20 is better than letting it go unfilled. For GPUs, once a data center is built, idle equipment generates no revenue. Therefore, as long as the rental income covers additional operating costs, companies are incentivized to rent at lower prices. When GPU supply exceeds demand, the price drop for leasing is not linear—it could be a steep, cliff-like decline.

The second risk is GPU depreciation and valuation. Griffin points out that currently, GPUs that are five or six years old can still be rented out, but this situation may not last. If new generations of chips keep emerging, the rental rates and utility of older equipment will decline rapidly, so investment returns based on a lifespan of seven to nine years may be overly optimistic. This means that future actual returns could be much lower than forecasts suggest.

The third risk concerns the return on enterprise investment. Griffin says that savvy companies are starting to ask which operations really need to use the most advanced and costly AI models, and which tasks could be handled by less expensive open-source models. As companies begin to control AI spending more strictly, growth in demand for computing power may also be impacted.

Finally, Griffin does not deny the long-term value of AI technology, but he believes that major technological revolutions are often accompanied by localized bubbles. Especially when companies become more cautious in evaluating AI returns while AI firms continue to ramp up spending, the contradiction may lead the market to question the sustainability of the AI investment frenzy. Therefore, Griffin warns that the re-pricing of the AI industry may have just begun.

Jason believes that Musk is talking about the present—a short-term tightness in the computing power leasing market. Pichai speaks to the ultra-long term, similar to Internet infrastructure, meaning that current supply from infrastructure will eventually be absorbed by accumulated long-term demand; this point poses no issues, in his opinion.

Meanwhile, Griffin is referring to the medium term—current infrastructure supply rates are based on past demand projections. After demand growth starts to slow, the momentum of supply does not decrease in tandem but continues forward, creating the risk that supply may catch up with, or even surpass, demand. Jason also fully agrees with this view.

So, the viewpoints are very clear and not contradictory: in the short term, enjoy the pleasure of soaring leasing prices; in the medium term, brace for the pain of price collapses; but in the long run, current concerns about oversupply are nothing but clouds on the horizon.