Talk of an AI chip peak in 2028 is ‘utterly wrong': CLSA

CLSA estimates demand for AI GPUs and ASICs exceeds supply by roughly 73%, with the imbalance worsening as AI infrastructure investment accelerates. Analyst Bhavtosh Vajpayee expects supply-demand equilibrium only by 2030, extending the expected AI semiconductor upcycle beyond market expectations for a potential 2028 peak.

Analysis

The investable implication is not simply higher GPU revenue: persistent allocation tightness shifts value toward companies controlling the scarce layers around compute—advanced packaging, HBM memory, high-speed networking and power delivery. NVDA and AVGO retain pricing power if system-level supply remains constrained, while TSM and ASML benefit from prolonged capital-intensity without needing to win the accelerator architecture battle. MU is the higher-beta second-order beneficiary because HBM mix expansion can lift gross margin disproportionately versus conventional DRAM pricing.

Near-term, the risk is that the market already capitalizes a multi-year scarcity premium in the obvious AI leaders. The more actionable 1-3 month catalyst path is quarterly evidence on hyperscaler capex, TSM advanced-packaging utilization, HBM contract pricing, and lead-time commentary from VRT, ETN and ANET; these data points determine whether infrastructure spend is broadening beyond chips. A deceleration in Microsoft, Alphabet, Amazon or Meta capex guidance would compress the entire stack before physical supply materially loosens.

The contrarian view is that an extended shortage can eventually hurt accelerator vendors by forcing customers to diversify architectures and improve utilization rather than buy incremental hardware. ASIC adoption is therefore both a demand tailwind for AVGO and a medium-term bargaining-power risk for NVDA. Over 6-18 months, grid interconnection and data-center power availability may become the binding constraint, favoring power and thermal infrastructure over another leg higher in pure compute multiples.

This remains an analyst forecast rather than independently verified industry capacity data. The thesis is falsified by falling HBM/advanced-packaging lead times, sub-scale AI revenue at hyperscalers despite maintained capex, or a meaningful reduction in 2027 foundry and equipment spending plans.

Key Decisions for Investors

• Maintain a 6-12 month basket long NVDA, AVGO, TSM and MU, tilted toward AVGO/MU rather than adding concentrated NVDA exposure. The basket captures compute, custom silicon, foundry and memory bottlenecks; reduce if HBM pricing or TSM advanced-packaging utilization weakens for two consecutive monthly industry checks.
• Initiate a 3-9 month relative-value trade: long VRT and ETN versus short SMH in equal beta-adjusted notional. Power and cooling revenue can continue to compound even if accelerator ASP expectations normalize; exit if hyperscaler capex guidance falls by more than 10% year-on-year or data-center order growth materially decelerates.
• Use ANET as a 1-3 quarter confirmation position, adding only after evidence that AI back-end network orders are converting to revenue rather than remaining supply-constrained backlog. Upside comes from cluster scaling; the key risk is a shift toward vertically integrated networking designs by hyperscalers.
• Avoid broad semiconductor shorts solely on a future supply-normalization thesis. Instead, set alerts around HBM lead times, TSM CoWoS capacity updates and cloud-provider capex guidance; a simultaneous easing in two of the three would support rotating from the AI hardware complex into less capex-sensitive software and internet exposure.