Are investors backing the ‘wrong future’ in AI? Analyst sees unexpected winners — and bad news for hyperscalers

Panmure Liberum’s Joachim Klement warns hyperscalers could waste “hundreds of billions of dollars” in AI data-center capex if AI shifts toward small language models on PCs and mobile. He cites Stanford research suggesting local models can solve 80%+ of common tasks, with local inference running ~80% cheaper on RAM capex and ~70–80% cheaper on electricity. The market focus remains on Nvidia’s Q2 results, with an eye on potential growth in edge revenues (currently <10% of total) as a “fallback solution” if the data-center boom slows.

Analysis

The market is likely underestimating how much AI spend can migrate from centralized infrastructure to the endpoint stack. If inference economics keep compressing, the first-order winners are not just the obvious handset/PC names but the picks-and-shovels around local compute: silicon content, memory attach, and enterprise device refresh cycles. That makes AAPL, DELL, ARM and QCOM more interesting than they screen, while the weakest link is the concentrated hyperscaler capex trade that has supported NVDA and the broader AI infrastructure complex.

The key second-order effect is mix shift, not zero growth: local deployment does not eliminate AI demand, it reallocates dollars from data center GPUs, networking, and power to lower-ASP CPUs/NPUs, DRAM, and endpoint hardware. Over 1-3 months, the catalyst path is NVDA commentary on edge revenue and any softening in forward hyperscaler spending assumptions; over 6-18 months, the bigger question is whether enterprise AI becomes a device replacement cycle rather than a cloud buildout cycle. That would be structurally bullish for AAPL and DELL, but it also caps the multiple expansion for the AI infrastructure basket.

Contrarian risk: the consensus may be swinging too far from “all cloud” to “all edge.” Most high-value workflows still need cloud orchestration, model training, and shared data access, so the near-term overbuild risk is real but not a collapse in spend. The trade is therefore a relative-value expression, not a macro call: short the parts most levered to hyperscaler capex intensity, while owning the endpoint monetization beneficiaries that can prove content gains without needing perfect adoption assumptions.

Key Decisions for Investors

• Initiate a relative-value pair: long AAPL / short NVDA into the next 2-6 weeks of earnings follow-through risk. Thesis: the market is paying for persistent data-center scarcity, but endpoint AI monetization can be captured with less execution risk; stop if NVDA guides hyperscaler demand materially higher or edge revenue inflects above expectations.
• Buy DELL on pullbacks and pair it against a weaker AI-infra proxy if available. The setup is 3-6 months: if enterprise procurement shifts toward local inference, DELL benefits from refresh cycles and higher attach, while the downside is smaller than for compute-only names.
• Add ARM and QCOM as a 6-18 month basket on any post-earnings weakness in NVDA. These are the cleaner beneficiaries if AI content migrates into devices; invalidation would be a slowdown in OEM AI adoption or no uplift in handset/PC bill of materials.
• Avoid chasing the AI infrastructure complex on strength until hyperscaler capex visibility improves. If upcoming capex guidance or order commentary rolls over, use that as confirmation to rotate out of the most crowded AI buildout names and into endpoint-exposed semis and OEMs.