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

AI data center spending is shifting from GPUs to custom ASICs, with Goldman Sachs arguing ASIC demand will match GPUs by 2027 and Bloomberg Intelligence forecasting 27% annual growth through 2033. Broadcom is cited as the dominant player with an estimated 60% share of the custom-built processor market through at least 2027, while Marvell is also benefiting from rising custom chip demand and analyst price-target increases. The article is bullish on both stocks for long-term exposure, though it emphasizes that neither is cheap.

Analysis

The market is moving from a broad “AI picks and shovels” trade into a more economically rational, supply-constrained custom-silicon cycle. That favors the few vendors that can monetize design wins, software co-optimization, and long customer integration cycles, while compressing the narrative premium on generic AI compute exposure. The second-order winner is the networking/interconnect layer attached to these custom chips, because ASIC adoption only works when data-center architects can improve watts-per-token and reduce cluster bottlenecks.

Broadcom and Marvell are not just selling chips; they are embedding themselves into customers’ architecture roadmaps, which raises switching costs and extends revenue visibility beyond a single generation of silicon. The more custom the workload, the more value migrates from raw compute toward system-level integration, validation, and packaging — a structural advantage for incumbents with process depth and hyperscaler relationships. That also implies less upside for legacy GPU vendors in marginal AI inference deployments where cost discipline matters most.

The key risk is timing: the trade is strong over 6-24 months, but consensus may be underestimating how lumpy custom silicon revenue can be if a hyperscaler delays ramps or re-specs a design. If capex budgets tighten, ASIC programs can slip one cycle even if the long-term thesis remains intact. Another risk is that NVIDIA’s platform strategy narrows the performance gap enough to keep some workloads on GPUs longer than expected, capping near-term share transfer.

Contrarianly, the market may be overpaying for certainty in AVGO and MRVL while underappreciating the beneficiaries one layer down the stack: advanced packaging, HBM supply chain, and networking optics tied to denser AI fabrics. The better expression may be to own the custom-silicon enablers while fading overextended cyclical expectations in the broader semiconductor basket. The article’s message is not that all AI chips win, but that economic gravity is shifting toward the most application-specific and power-efficient architectures.

Key Decisions for Investors

• Long AVGO vs. short a basket of broad AI compute beta (NVDA + semi ETF proxy) on a 3-6 month horizon: thesis is continued share shift toward custom ASICs, with better revenue durability and fewer multiple compression risks if AI capex normalizes.
• Initiate a tactical long MRVL position with a 6-12 month target; size smaller than AVGO because execution is more customer-concentrated, but upside is higher if additional hyperscaler wins convert into a faster re-rating.
• Pair trade: long AVGO / short INTC for 6-12 months. AVGO benefits from custom-silicon monetization; INTC remains a relative laggard if the market stays focused on AI inference economics and embedded design wins.
• Consider a basket long of AI networking/interconnect names rather than pure compute exposure. The ASIC buildout increases demand for adjacent infrastructure, and this is a cleaner way to express the theme with lower single-name design-win risk.
• Avoid chasing all AI semiconductor names equally after the recent run; use pullbacks to add only where there is visible hyperscaler commitment and multiyear revenue visibility, not just thematic exposure.