Are investors backing the wrong future in AI? Analyst sees unexpected winners — and bad news for hyperscalers
Investors pouring money into giant artificial intelligence data centers may be betting on the wrong technological future — and hyperscalers risk wasting “hundreds of billions of dollars” on the same false premise. That’s the view of Joachim Klement, managing director at Panmure Liberum, who believes the future of AI lies in small language models run on desktop computers and mobile devices — not frontier AI systems powering large language models such as ChatGPT and Claude, which rely on data center compute. “I’ve become very bearish on the AI capex boom over the last nine months because I have basically come to the conclusion that, from a technology perspective, investors, in my view, are investing in the wrong future,” Klement told CNBC’s “Europe Early Edition” on Wednesday. Klement pointed to Stanford University research published in May that found some smaller AI language models running on consumer hardware could correctly answer more than 80% tasks and queries that people perform, or could perform, bolstering the case for routing more AI workloads to local devices. ‘The best trade I can think of’ A decisive AI activity shift onto PCs and mobile devices would leave hyperscalers with excess capacity as value migrates toward device makers and edge-chip suppliers, which investors often consider AI laggards. Klement said device makers are “the best trade I can think of” to capitalize on this development, specifically naming Apple and Dell . AAPL YTD mountain Apple. “These [companies] are sometimes considered AI losers, but in my view are actually the AI winners,” he added. “80% of the tasks we can already move to cheaper models on local PCs.” The economics behind that shift are already influencing how some large companies deploy AI, with businesses becoming increasingly selective in matching individual tasks to models based on cost and required performance. AT & T says it routes tasks among models based on factors including cost and performance, while JPMorgan has said it expects businesses to use a mix of large and small, open and closed models. Thomson Reuters this week launched its own proprietary in-house system, built around Alibaba ‘s open-source Qwen model, which it says can run at a fraction of the cost of comparable frontier systems. ‘Fallback solution’ Markets are awaiting Nvidia ‘s second-quarter earnings, due after Wednesday’s closing bell, as the chipmaker’s dependence on spending by a small group of hyperscalers attracts growing scrutiny. Underlining his thesis, Klement noted how Nvidia’s launch of DGX Spark — its desktop AI solution — could be seen as its “fallback solution” if the data center boom slows down. NVDA 1M mountain Nvidia. “Nvidia, I think, has already kind of smelled where the wind is coming from,” Klement said. “It is going to be interesting within the Nvidia numbers to look at the growth in the edge revenues of Nvidia, which arguably are less than 10% of their total revenues, but it is kind of the fallback solution, if the data center boom slows down.” Meanwhile, chip designers active in edge computing, including Arm and Qualcomm , will “absolutely” benefit from such a shift, Klement said. Memory makers should also be relatively insulated, he argued, because local AI systems still require memory. The mix would shift from the high-bandwidth memory used in large data centers toward conventional DRAM installed in PCs. DELL YTD mountain Dell. “We’re not going to not use any data centers whatsoever because cloud software is still going to run,” Klement added. “But I think we’re past that tipping point where we are starting to overbuild data centers. Plus, you have to be aware that running an AI model on a local computer is massively cheaper. If you compare the initial capex per gigabyte of RAM, which by now unfortunately is the cost factor, it is about 80% cheaper than running it in a data center, and in terms of electricity bill, even if you do retail electricity prices, it still is 70-80% cheaper to run it locally than in a massive data center.”
