SLMs on Devices: Klement Warns About Hyperscalers

The rise of local AI could change the game for cloud computing. Joachim Klement anticipates fewer data centers and pressured revenues if SLMs replace a large portion of LLM tasks. Competitive SLMs, including in reasoning, present significantly lower costs.

Fewer data centers anticipated and capex questioned

Joachim Klement believes that the industry may ultimately require fewer data centers than expected. He argues that if 70% to 80% of tasks currently performed by LLMs are handled by SLMs, the revenue growth of hyperscalers may not be sufficient to justify their capital investments. He specifies that significant investments in semiconductors will still be necessary, but that less expensive models suited for desktop computers could suffice, without the need to focus on the most advanced Nvidia chips.

SLMs winning in 81.2% of cases and 50 to 85% cheaper

Locally executed SLMs (QWEN 3, GEMMA 3, GPT-OSS, GRANITE 4.0) have been compared to cloud LLMs (ChatGPT 5, Claude Sonnet 4.5, Gemini 2.5 Pro). The top-performing SLMs achieve comparable or superior results in at least 90% of situations, with an average of 98.6%. When it comes to reasoning tasks, they outperform their competitors in 62.5% of cases and, overall, match or exceed cloud LLMs in 81.2% of cases. According to the same study, these performances are achieved with computing and energy costs that are 50% to 85% lower. In 2023, SLMs achieved a 99% success rate on the simplest reasoning tasks, falling short only on the more difficult ones.

An investor thesis: local, small, and cheap versus the cloud

Joachim Klement, who serves as an investment strategist in London, believes that local AI challenges the economic foundations on which centralized hyperscalers rely. He develops his reasoning around several axes: questioning the principle of "bigger is better," the balance between initial investment and on-device performance, shrinking margins and the difficulty of achieving profitability, as well as using hardware refresh as a safeguard. He asserts that the AI of tomorrow will be characterized by its small size, efficiency, and low cost, which would render the centralized cloud model structurally deficient. Klement also specifies a time horizon and market hypothesis: he anticipates a possible burst of the AI bubble between 2027 and 2028, with an S&P 500 that could be around 5,000 points, representing a decrease of 33% from a level of about 7,500 points. Among the frequently cited critical voices of AI are also Michael Burry, Yann LeCun, David Cahn, Ed Zitron, Jim Covello, Melanie Mitchell, Molly White, Abi Awomosu, David Rosenberg, Stacy Rasgon, Dean Baker, and Gary Marcus.