Broadcom has signed landmark multi-year supply agreements with Google and Anthropic extending to 2031 for AI infrastructure components.
In a move that fundamentally redraws the global artificial intelligence supply chain, Broadcom Inc. (NASDAQ: AVGO) has secured a series of massive, long-term contracts with Google (Alphabet Inc., NASDAQ: GOOGL) and AI startup Anthropic. Disclosed in early April 2026, the agreements lock in Broadcom as the primary architect for custom Tensor Processing Units (TPUs) and high-speed networking components through 2031. The partnership includes a staggering commitment to provide 3.5 gigawatts (GW) of compute capacity for Anthropic, a scale that rivals the energy output of multiple nuclear power plants and signals a shift toward massive, vertically integrated AI "factories."
On April 6, 2026, Broadcom confirmed its selection as the lead design partner for Google's seventh-generation TPU, known internally as "Ironwood." Built on a cutting-edge 3-nanometer process, the chip is designed to offer a four-fold performance improvement over previous generations while significantly reducing power consumption per teraflop. The collaboration with Anthropic represents an even more ambitious component of the announcement, with Anthropic gaining access to 3.5 GW of compute capacity through 2027 and beyond, powered by Broadcom-designed silicon and hosted within Google Cloud's infrastructure. The estimated cost for this build-out is between $120 billion and $175 billion, marking one of the largest infrastructure investments in technological history, with leadership from Broadcom CEO Hock Tan and senior executives at Alphabet and Anthropic emphasizing that the 2031 timeline will provide the "compute certainty" necessary to develop next-generation artificial general intelligence (AGI) systems.
Market reactions proved overwhelmingly bullish. On April 7, 2026, Broadcom's stock jumped more than 6% to reach record highs near $324 per share, buoyed by the company's report of a $73 billion backlog in custom silicon orders. Analysts now project Broadcom's AI-related semiconductor revenue to eclipse $100 billion by 2027. By capturing 60% to 70% of the custom AI accelerator market, Broadcom has transformed from a diversified chipmaker into the "implementation layer" of the AI revolution, with its integration of custom silicon alongside high-end networking components—including Tomahawk and Jericho series Ethernet switches and industry-leading SerDes IP—creating a competitive "moat" difficult for others to breach.
The announcement also lifted Marvell Technology, Inc. (NASDAQ: MRVL), positioning itself as an "open" alternative and recently named as NVIDIA's preferred ASIC partner through the NVLink Fusion platform, with Marvell's stock rallying 6.9% following the news. NVIDIA's own stock rose 2.1% on the announcement of increased AI investment, though the long-term trend toward custom ASICs represents potential erosion of market share in the hyperscale data center segment. The significance of the 3.5 GW capacity figure underscores how the AI race has moved beyond software and silicon into heavy infrastructure and energy management: 3.5 GW is enough to power roughly 2.6 million homes. Broadcom's provision of optical circuit switching (OCS) technology proves critical here, as light-based switching can reduce power consumption of these massive clusters by up to 97% compared to traditional electrical methods, validating the broader industry trend toward verticalization where AI companies and cloud providers seek to control every layer of their technology stack.