At the end of 2024, Ilya Sutskever, co-founder and former chief scientist of OpenAI, delivered a statement that reverberated across the artificial intelligence community: all internet data had been consumed by large models. He suggested that pre-training—the core technology underpinning these models—might be nearing its conclusion. Though mature downstream applications remain scarce, Sutskever’s pronouncement marked an inflection point in AI’s development.
This statement, akin to a bombshell, delineated a significant boundary in the evolution of artificial intelligence. Coincidentally, just one day earlier, semiconductor giant Broadcom released its financial report. CEO Hock Tan projected that Broadcom’s custom XPU chips would generate between USD 60–90 billion in revenue for the fiscal year 2027.
Broadcom’s XPU is a custom application-specific integrated circuit (ASIC) developed in partnership with tech powerhouses like Google, Amazon, and Meta. Unlike Nvidia’s general-purpose GPUs, XPUs are tailored for specific tasks, highlighting the contrasting roles of these chips in large model applications: ASICs excel at inference, while GPUs are better suited for training.
Both Sutskever’s and Tan’s predictions hinted at the same underlying shift: the demand for AI chips is evolving, and GPUs could be ceding their dominance to ASICs.
The market responded swiftly. On December 13, Broadcom’s stock surged 24%, catapulting its valuation past USD 1 trillion. Meanwhile, Nvidia faced four consecutive days of stock declines. Even before this spike, Broadcom had already reached a valuation of USD 800 billion. This remarkable rise reflects a pivotal shift in investor sentiment: the belief that the era of intensive pre-training is winding down, opening the door for ASICs to erode Nvidia’s market share and drive Broadcom’s growth.
However, the new wave of AI chip competition may not be as intense as anticipated.
Broadcom’s ascent
Broadcom’s latest financial results exceeded market expectations, driven by robust growth across key segments. AI-related revenue for the fiscal year soared by 220%, propelling the semiconductor division’s revenue to a record USD 30.1 billion, a 7% year-on-year increase. Combined with CEO Hock Tan’s optimistic forecast for the 2027 market scale, Broadcom’s ASIC business is now viewed as entering a pivotal growth phase.
Broadcom’s business operates under two primary categories: semiconductor solutions and infrastructure hardware. The infrastructure hardware division, which experienced a 196% year-on-year revenue surge thanks to VMware’s integration, was the main contributor to Broadcom’s overall 44% revenue growth for the year. Meanwhile, the company’s original software business saw only marginal increases.
In the fourth quarter, Broadcom reported a net profit of USD 4.32 billion, bolstered by revenue growth and improved operational efficiencies. Business integration efforts further reduced expense ratios, accelerating profit realization.
Broadcom’s networking segment, part of its semiconductor solutions division, has become a cornerstone of its growth. ASICs now account for 32% of the company’s revenue, making them its largest revenue contributor. This segment also includes Ethernet switch and router chips, PHY chips, and optical transceivers, underscoring the breadth of Broadcom’s networking portfolio.
Broadcom’s foray into the ASIC business began with its 2016 acquisition of storage company LSI, and it has since become a major growth driver. Historically, Broadcom’s Ethernet switch products were the flagship of its data center networking equipment portfolio, where the company held a dominant position.
The rapid expansion of AI data centers has significantly increased demand for networking equipment. For instance, AI systems powered by Nvidia’s chips typically require around nine 1.6T optical modules per machine. This surge in demand has driven up the stock prices of optical module manufacturers like Marvell, whose valuation has doubled over the past year due to Nvidia’s strong market presence.
Broadcom initially benefited from this trend with its switch products. However, Nvidia’s acquisition of Mellanox in 2019 disrupted this dynamic, allowing Nvidia to enter the switch market and directly compete with Broadcom.
In response, Broadcom has leaned heavily into its ASIC business, which has emerged as a crucial cash cow within its networking division.
RELATED ARTICLE
End of Nvidia’s dominance?
“The deceleration of Nvidia’s growth is inevitable, but being surpassed or replaced is impossible,” one investor said.
While Broadcom’s ascent is clear, whether Nvidia can truly be replaced remains an open question. From a technical perspective, GPUs are also capable of performing inference tasks, despite ASICs being commonly associated with them. For instance, Nvidia’s GB200 chip delivers inference performance 30 times greater than its H100 predecessor.
That said, ASICs excel in scenarios requiring low power consumption and compact designs, making them better suited for edge applications. Meanwhile, Nvidia’s dominance and premium pricing have positioned it as a shared adversary for companies seeking cost-effective alternatives in the AI sector.
Broadcom, as the most viable competitor, has gained industry support from those looking to diversify away from Nvidia. Still, market analysts caution that Broadcom’s gains may be overhyped, while Nvidia’s stock could be undervalued. The slowdown in pre-training does not necessarily signal the decline of general-purpose models. Furthermore, widespread ASIC adoption could disrupt existing large language model frameworks, which remain central to AI’s current landscape.
KrASIA Connection features translated and adapted content that was originally published by 36Kr. This article was written by Song Wanxin for 36Kr.