Strategist: Structural Misjudgment in AI Investment Direction — Apple and Dell May Emerge as Surprise Winners
Nashnova编辑部
Panmure Liberum's Joachim Klement argues AI's future belongs to small, local models — not cloud-scale giants. If he's right, Apple and Dell, widely seen as AI laggards, are actually the biggest winners.
Thousands of billions — bet on the wrong path?
Klement's core thesis: AI's future lies in small language models running on PCs and phones, not cloud-dependent large models like ChatGPT and Claude.
His key evidence comes from a Stanford study published in May 2025: small models on consumer hardware already handle over 80% of everyday tasks correctly.
This means → if eight out of ten jobs can run locally, hyperscalers are building "hundreds of billions of dollars" of infrastructure for the remaining two.
How much cheaper is running AI locally?
Klement's cost comparison: measured by initial capex per GB of memory, local AI runs roughly 80% cheaper than data-center equivalents.
Even at retail electricity rates, local power costs are 70–80% lower than large data centers.
In plain terms = for the same AI task, spending $1 on your own machine costs about $5 in a data center — same compute, several times the bill.
Are big companies already shifting to "hybrid deployment"?
AT&T says it allocates tasks across models based on cost and performance; JPMorgan expects enterprises to mix large and small, open-source and proprietary models.
Thomson Reuters this week launched an in-house system built on Alibaba's open-source Qwen model, saying it runs at a fraction of the cost of comparable frontier systems.
This reflects a shift in enterprise AI logic — from "all-in on the cloud" to "pick the model per task." Cost sensitivity is overtaking technology worship.
Who are the underestimated "surprise winners"?
Klement named Apple and Dell, calling device makers "the best trade I can think of." These companies are often labeled AI losers, but under a local-AI thesis they flip to winners.
On the chip side, edge-computing designers Arm and Qualcomm will "absolutely" benefit as compute migrates from cloud to device.
Memory makers also hold up: local AI still needs memory, but demand shifts from data-center HBM (high-bandwidth memory) to conventional DRAM installed in PCs.
Has Nvidia already sensed the wind shifting?
Klement reads Nvidia's desktop AI product DGX Spark as a "plan B" — a pivot ready if the data-center boom cools.
He says he will watch edge-business revenue growth in Nvidia's earnings; that segment currently accounts for less than 10% of total revenue.
This means → Nvidia is hedging its own bet — if data-center growth slows, edge is the escape route it has kept for itself.
Will data centers disappear?
Klement is explicit: data centers will not vanish, and cloud software will keep running.
But he believes the market has passed an inflection point and data centers are heading toward overbuilding.
In plain terms = data centers are not useless — they are just being built too fast with too much money. Whether hyperscaler capex converts into real returns is the single variable that will prove or disprove this call.
市场有风险，内容仅供研究参考，不构成投资建议。