Broadcom disclosed during its Q4 2025 earnings call that it received a combined $21 billion order from Anthropic for Google’s Tensor Processing Units (TPU), with $10 billion in the previous quarter and an additional $11 billion order for late 2026 delivery. The massive commitment underscores a strategic industry shift away from NVIDIA’s GPU dominance toward specialized, power-efficient custom AI accelerators. Broadcom now carries a $73 billion backlog of AI product orders expected to ship over 18 months, signalling a fundamental reshaping of AI infrastructure competition.
TPU Architecture and Market Position
TPUs are specialized processors designed by Google for AI workloads, now in their seventh generation. They are available to customers through Google Cloud and power Google’s internal systems, including training and deployment of the Gemini family of models. Broadcom’s role is manufacturing: Google designs the TPU architecture, while Broadcom converts those designs into manufacturable silicon and handles volume production.
This relationship mirrors Google’s long-standing strategy of controlling key AI hardware design while outsourcing fabrication to semiconductor partners with proven expertise. The arrangement allows Google to optimize chip design for its specific AI workloads while leveraging Broadcom’s manufacturing scale and semiconductor know-how.
Anthropic’s Massive Infrastructure Buildout
Anthropic’s $21 billion TPU order reflects an extraordinary commitment to AI compute capacity. The company intends to deploy one million TPUs, backed by more than one gigawatt of new compute capacity coming online in 2026—one of the largest dedicated AI compute buildouts in the industry. This infrastructure investment directly supports Anthropic’s goal of scaling its models, conducting advanced research, and serving enterprise customers at scale.
The size of Anthropic’s order—representing a single customer’s 18-month commitment—demonstrates the capital intensity of frontier AI development and the competitive pressure to secure compute capacity before rivals. For Anthropic, owning dedicated TPU infrastructure reduces dependency on cloud provider capacity constraints and enables proprietary model training at optimal cost and performance.
Competitive Dynamics: TPUs vs. NVIDIA GPUs
The rise of TPUs poses a direct challenge to NVIDIA’s historical GPU dominance in AI workloads. According to SemiAnalysis, TPU v7 delivers stronger performance-per-total-cost-of-ownership (TCO) than NVIDIA’s GB200 platform despite having roughly 10% lower peak floating-point operations per second (FLOPs) and memory bandwidth.
SemiAnalysis estimates that Google’s internal cost to deploy Ironwood (Google’s TPU-based infrastructure) is approximately 44% lower than deploying an equivalent NVIDIA system. Even when priced for external customers, TPUv7 offers an estimated 30% lower TCO than NVIDIA’s GB200, and roughly 41% lower TCO than the upcoming GB300. If Anthropic achieves realistic machine-fraction utilisation (MFU) rates of around 40%, the effective training cost per FLOP could be 50-60% lower than GB300-class GPU clusters.
Expanding TPU Customer Base
Broadcom disclosed that it now has five TPU/XPU (custom AI accelerator) customers, with Google and Anthropic named on the earnings call. Industry analysis and reports indicate that Meta and ByteDance are also among Broadcom’s custom AI chip customers, though Broadcom has not publicly confirmed the full roster. Meta is reportedly evaluating TPU deployment in its data centres beginning in 2027, signalling broader enterprise and technology sector adoption.
This customer diversification demonstrates that TPUs are transitioning from Google-internal infrastructure to a viable alternative for other frontier AI companies and technology firms. Each new customer validates the TPU architecture and creates competitive pressure on NVIDIA to innovate or reduce pricing.
Implications for AI Infrastructure Strategy
The TPU-Anthropic dynamic reshapes AI infrastructure competition. Rather than a single dominant vendor (NVIDIA), the market is fragmenting into platform-specific ecosystems: Google TPUs for customers aligned with Google Cloud and Anthropic’s ecosystem, NVIDIA GPUs for customers requiring maximum flexibility and vendor optionality, and potentially custom chips from other hyperscalers (Meta, Amazon) for their own workloads.
For enterprises and AI startups, this fragmentation introduces both opportunity and complexity. Access to cheaper, more efficient compute (TPUs) can lower barriers to building and training competitive AI models. However, platform lock-in risks increase, particularly for companies committing to single-vendor infrastructure at Anthropic’s scale.
Broader Market Impact
Broadcom’s $73 billion AI product backlog—expected to ship over 18 months—reflects the industry’s extraordinary capital deployment into AI infrastructure. This sustained, high-volume order flow validates Broadcom’s strategic positioning as a manufacturing partner to custom chip designers and suggests that custom AI chips will capture an increasing share of the total AI accelerator market.
For semiconductor supply chains, the shift toward custom chips increases demand for advanced packaging, interconnect, and memory technologies where Broadcom and other semiconductor suppliers excel. It also reduces NVIDIA’s near-term margin pressure from price competition, as NVIDIA’s customers will increasingly pursue custom designs rather than commodity GPUs.
