Goldman Sachs Says ASICs Will Match GPU Demand by 2027. 2 AI Chip Stocks to Load Up on Right Now.

The article argues that ASICs are becoming the preferred hardware for AI data centers, with Goldman Sachs seeing demand for custom-built processors matching GPUs by 2027 and Bloomberg Intelligence projecting 27% annual growth through 2033. Broadcom and Marvell are identified as the two main beneficiaries, with Broadcom reporting $8.4 billion in Q1 AI revenue, up 106% year over year, and Marvell’s custom chip business doubling last fiscal year. Several analysts recently raised price targets on Marvell, citing continued demand for AI computing hardware.

Analysis

The market is still underestimating how quickly ASICs can cannibalize incremental AI capex from the GPU stack. This is not just a component mix shift; it is a bargaining-power shift toward the hyperscalers, who will increasingly force vendors to fund more engineering upfront in exchange for multi-year volume commitments. That favors the two scaled design houses with deep customer intimacy and punishes anyone relying on generic accelerator attach rates.

Second-order winners are likely in adjacent, less obvious layers: advanced packaging, HBM-adjacent interconnect, and networking/optics that sit around custom silicon deployments. The losers are not just legacy CPU/GPU suppliers on the margin, but also smaller merchant ASIC designers that lack anchor customers and cannot amortize NRE across enough sockets. Nvidia is not impaired in aggregate, but the implied mix shift could cap upside in AI compute ASPs and force more value capture into the platform/control plane rather than the chip itself.

The key risk is timing. Consensus is projecting a multi-year adoption curve, but custom silicon ramps can be lumpy and customer-specific; one delayed tape-out or yield issue can defer revenue recognition by 2-3 quarters. The other risk is that valuation already discounts durable growth, so the near-term trade is more vulnerable to multiple compression than to fundamental deterioration.

Contrarianly, the move may be underowned in the hyperscalers themselves. If internal ASIC adoption lowers training/inference cost per token by even mid-teens percentages, the real upside is not only vendor revenue growth but improved AI gross margins at GOOGL, AMZN, and MSFT over the next 12-24 months. That creates a more attractive asymmetry than chasing the most crowded beneficiaries after a strong run.

Key Decisions for Investors

• Long AVGO on 6-12 month horizon; use any post-earnings drawdown as entry. Risk/reward is favorable because custom silicon is still early in the revenue mix, but position sizing should reflect elevated multiple risk.
• Long MRVL versus short INTC as a pair trade over 3-6 months. MRVL has direct custom-design torque while INTC remains the cleaner short if investor attention rotates from turnaround narratives to AI capex quality.
• Add to GOOGL and AMZN on 12-month horizon as hidden beneficiaries of lower AI compute cost. The thesis is margin expansion rather than headline revenue, which the market is likely underappreciating.
• Buy AVGO/MRVL call spreads rather than outright calls for a 6-9 month view. This captures continuation in the ASIC re-rating while defining risk against a valuation reset.
• Avoid chasing NVDA here unless it de-risks materially; the stock still benefits, but the market may be overestimating its share of incremental AI infrastructure spend as custom chips take a larger slice.