Why we remain negative on AI names
We have remained highly cautious of market conditions. Indeed, the recent AI sell-off has justified our calls for shorting names like SoftBank and Kioxia, among a few others in the supply chain that have been notably frothy in valuations. Our primary apprehension concerns AI houses like OpenAI whose spending ultimately cascades down the supply chain to feed all otherAI-related plays. We will focus on addressing our concerns here.
Put simply, we see a killer combination of rising costs coupled with falling prices for AI compute power that we think leave these LLM business models looking highly vulnerable, with lofty valuations in the private markets that appear disconnected from the reality we are seeing unfold on the ground.
While competition between OpenAI, Anthropic, Google, and xAI has intensified at the high end, Chinese LLMs—which charge a fraction of their token fees—are capturing market share in less demanding segments and starting to challenge their American rivals for more complex applications. Moonshot AI’s release of Kimi K3 last week, as the world’s largest open-weight AI model, serves as a prime example of this growing threat from the East.
We remain convinced that the Trump administration is poised to ban the usage of Chinese models by US corporates to protect homegrown AI firms from this market share assault. While this may not stop Chinese players from securing a substantial portion of the global market, US government meddling could isolate their own firms from international arenas.
Moreover, data centre expansion in the US faces mounting challenges as regulatory and local community opposition to such monoliths grows exponentially, creating obstacles regarding access to local power grids and water. Furthermore, for projects employing onsite generators, growing awareness regarding the low-frequency noise they generate—which has rendered neighbouring homes effectively worthless—has become a significant headache.
Given that these issues are unlikely to prove significant obstacles to data centre construction in China—which is aggressively adding nuclear capacity, constructing large dams, and designating strategic locations—it is difficult to see how US compute costs can remain competitive. We contend that their primary advantage of early access to advanced silicon is being eroded rapidly as other critical resources become the bottleneck.
Finally, secular inflation across the AI supply chain is accelerating, raising the costs of building data centres and straining the massive AI data centre budgets estimated at $450bn this year. This is a subject we have tackled a few times in this publication arguing that AI-related build-outs could prove inflationary for the world.
From foundry quotes and memory chips to multilayer PCBs, fibre optics, gas turbines and onsite power generators, construction machinery leasing, and high-grade steel, nearly everything we have seen from the ground up has experienced price hikes of 20% or more.
