The AI chip race isn’t just a one-horse sprint led by Nvidia (NVDA).

As hyperscalers like Google (GOOG, GOOGL), Meta (META), and Microsoft (MSFT) race to lower the eye-watering costs of running massive AI models, a second front is opening in the custom silicon wars, with Broadcom (AVGO) as its primary architect.

"Broadcom is projected to retain its leadership as the premier AI Server Compute ASIC design partner with a 60% market share in 2027," according to a recent report from Counterpoint Research.

This dominance is underpinned by a symbiotic relationship with the world’s most advanced foundry, Taiwan Semiconductor Manufacturing Company (TSM), which remains the "dominant foundry choice ... with close to 99% wafer fabrication share for the top 10 players’ AI Server Compute and ASIC shipments."

This shift signals the industry is moving beyond Nvidia’s pricey, all-purpose GPUs. While Nvidia provides a powerful all-purpose AI tool, tech giants are increasingly designing their own Application-Specific Integrated Circuits (ASICS) tailored to their unique workloads.

Broadcom thrives here by acting as the bridge, turning these internal corporate blueprints into functional hardware. By hitching its wagon to the internal capital expenditures of the world’s wealthiest companies, Broadcom has seen its stock climb roughly 55% over the last year.

The cost-saving incentive for these giants is massive. Goldman Sachs analyst James Schneider noted that the Google-Broadcom TPU (Tensor Processing Unit) is rapidly closing the performance gap with Nvidia, estimating a staggering 70% reduction in "cost-per-token" as the technology evolves from the TPU v6 to the v7.

In a world where AI inference costs can severely impact a balance sheet, that efficiency is a powerful gravitational pull toward custom silicon. Google famously trained its Gemini 3 entirely on its TPUs.

However, the custom chip boom is not a rising tide that lifts all boats equally. Marvell Technology (MRVL), often cited as Broadcom’s primary challenger, is currently navigating "design win headwinds." Counterpoint’s analysis suggests Marvell’s design service share could slide to 8% by 2027, even as its total shipment volumes grow.

Goldman Sachs remains Neutral on Marvell with a $90 price target, noting that the company’s fortunes are heavily tied to Amazon Trainium program, which has faced its own performance hurdles and is oftentimes seen as playing catch-up to Nvidia’s chips.

While some on Wall Street, like Raymond James analyst Simon Leopold, remain bullish on Marvell as a long-term "share gainer," the immediate data favors Broadcom’s grip on high-volume contracts. The firm issued a Strong Buy rating on Marvell with a $121 price target, while giving Broadcom a $420 target.