Broadcom reveals Anthropic as mystery customer behind $21 billion Google TPU order

Broadcom Unveils Anthropic as $21 Billion Customer

Broadcom CEO Hock Tan ended months of speculation during the company's fourth-quarter earnings call on Thursday, revealing that Anthropic is the mystery customer behind massive chip orders totaling $21 billion. The AI lab initially placed a $10 billion order in September for Google TPUs, specifically the latest TPU Ironwood racks, followed by an additional $11 billion order in the fourth quarter for delivery in late 20261

1

2

2

. The disclosure came after Broadcom's September earnings call sparked significant investor attention amid the AI infrastructure boom, with the company initially declining to name the customer.

Custom AI Chips Challenge Nvidia's Market Dominance

Broadcom manufactures custom AI chips called ASICs and XPUs, which experts believe offer superior efficiency for certain artificial intelligence algorithms compared to Nvidia's market-dominating GPUs. The company converts Google's TPU architecture designs into manufacturable silicon and handles volume production, while Google maintains control over the hardware design2

2

. Tan emphasized that Broadcom is delivering entire server racks to Anthropic, not just chips, making the AI lab Broadcom's fourth XPU customer1

1

. The chipmaker also secured a fifth customer during the fourth quarter with a $1 billion order, though the company has not disclosed this customer's identity.

Anthropic Plans Massive AI Compute Buildout

Anthropicintends to deploy one million TPUs backed by more than one gigawatt of new compute capacity coming online in 2026, representing one of the largest dedicated AI compute buildouts in the industry2

2

. This ambitious expansion reflects the growing demand for specialized AI accelerators as companies scale their machine learning operations. Tensor Processing Units are now in their seventh generation and are available to customers through Google Cloud, powering many of Google's internal systems including the training and deployment of Gemini models. Google recently highlighted that it trained its state-of-the-art Gemini 3 model entirely on its TPUs1

1

.

TPU v7 Offers Superior Cost Efficiency Over NVIDIA GB200

Analysis from SemiAnalysis reveals compelling economics behind Anthropic's TPU investment. Although TPU v7 has roughly 10% lower peak floating-point operations per second and memory bandwidth than NVIDIA GB200, it delivers a stronger performance-per-total-cost-of-ownership profile2

2

. Google's internal deployment cost for Ironwood is estimated at about 44% lower than an equivalent Nvidia system. Even when priced for external customers, TPU v7 offers approximately 30% lower total-cost-of-ownership than NVIDIA GB200 and roughly 41% lower than the upcoming GB300. SemiAnalysis notes that if Anthropic achieves around 40% machine-fraction utilization on TPUs, the effective training cost per FLOP could be 50-60% lower than what GB300-class GPU clusters are expected to deliver2

2

.

Growing TPU Adoption Across Tech Giants

Broadcom disclosed a $73 billion AI product backlog expected to ship over the next six quarters, signaling robust demand for custom AI chips2

2

. Several major companies have confirmed their use of Google TPUs, including Meta, Cohere, Apple, and Ilya Sutskever's Super Safe Intelligence. A report from The Information indicates that Meta is evaluating TPU deployment in its data centers starting in 20272

2

. The growing adoption stems from TPUs' power efficiency and tight optimization for AI workloads, creating increasing competitive pressure on Nvidia's GPU dominance. Industry analysis suggests that Meta and ByteDance are also among Broadcom's custom AI chip customers, though the company has not publicly confirmed the full roster.