Anthropic expands its AI chip strategy with Google TPUs and Broadcom silicon, challenging Nvidia's dominance in model training infrastructure. Here is what it means for investors.

Key Highlights

  • Anthropic will deploy next-generation Google TPUs from 2027, expanding its partnership with Alphabet (NASDAQ:GOOGL) and Broadcom (NASDAQ:AVGO).
  • Broadcom's AI semiconductor revenue reached $8.4 billion in Q1 FY2026, up 106% year on year.
  • Anthropic deliberately trains Claude across three chip architectures to limit vendor pricing power.
  • Nvidia's (NASDAQ: NVDA) revenue growth projections remain intact at approximately 79% for its upcoming quarter.
  • Custom chip revenues from Broadcom's hyperscaler partnerships are projected to exceed $100 billion by end of 2027.

A shift in compute procurement strategy

The announcement that Anthropic will integrate next-generation Tensor Processing Units into its model training infrastructure from 2027 marks a notable evolution in how frontier AI laboratories manage compute procurement. The chips, co-developed by Alphabet (NASDAQ:GOOGL) and Broadcom (NASDAQ:AVGO), are purpose-built for large-scale AI workloads. Their adoption alongside existing infrastructure signals a structural shift from single-vendor dependency toward a more distributed compute architecture.

This development carries implications beyond Anthropic alone. It reflects a broader pattern among AI hyperscalers to reduce exposure to any single supplier while securing access to expanding compute capacity. The strategic calculus is straightforward: diversified sourcing limits the pricing leverage of any one hardware vendor and provides operational resilience as training workloads scale.

Broadcom's accelerating custom chip revenues

Broadcom's position in this landscape is increasingly significant. The company's AI semiconductor revenue for the first quarter of fiscal year 2026, which ended February 1, reached $8.4 billion, representing 106% growth compared with the same period a year earlier. Broadcom's chief executive, Hock Tan, has indicated that custom AI chip revenues alone could surpass $100 billion by the close of 2027.

This trajectory reflects the growing appetite among hyperscalers for chips designed to their own specifications rather than general-purpose GPU solutions. Several major cloud infrastructure operators are expected to launch Broadcom-designed custom silicon within the next few years, a trend that consolidates Broadcom's position as a pivotal player in AI compute supply chains.

The revenue impact on Alphabet is less transparent. It is not yet clear how TPU revenues will be accounted for within Alphabet's reporting segments. Google Cloud has already demonstrated strong momentum, with revenue growth accelerating from 34% year on year in Q3 2024 to 48% in Q4. If TPU-related revenues begin to flow through Google Cloud, further acceleration in that segment becomes plausible.

Risks to Nvidia's Dominance

The rise of custom silicon presents a credible long-term challenge to Nvidia's position. As AI laboratories scale, the economics of purpose-built chips improve: hyperscalers gain design sophistication, foundry access broadens, and alternatives like TPUs and Trainium mature. If custom architectures close the performance-per-dollar gap with GPUs, demand concentration around Nvidia could erode. The CUDA moat, while deep today, is not insurmountable. Sustained investment in competing software stacks by Google and Amazon may gradually lower switching costs. Supply constraints further expose a structural vulnerability: when allocation is tight, customers are incentivised to develop alternatives rather than wait.

Why Nvidia Retains Structural Importance

Despite the attention generated by Anthropic's TPU expansion, framing this as Nvidia's displacement requires qualification. Anthropic trains Claude across three chip types: Nvidia GPUs, Google TPUs, and Amazon's (NASDAQ:AMZN) Trainium. The 2027 announcement adds capacity; it does not represent an exit from Nvidia's ecosystem.

Nvidia's GPU production is broadly understood to be fully allocated through at least 2027. Anthropic's turn toward Alphabet and Broadcom reflects the difficulty of securing additional Nvidia compute rather than a preference-driven pivot away from it.

Anthropic is managing compute relationships with deliberate commercial intent. Remaining committed to multiple suppliers prevents any single vendor from exerting disproportionate pricing power. Locking into one architecture would expose the company to concentrated supply risk and reduce negotiating leverage. The multi-vendor model is not just operationally sensible; it is a leverage management strategy.

Wall Street's consensus reflects this reality. Revenue growth projections for Nvidia's upcoming quarter remain around 79%, with full-year estimates near 71%, consistent with sustained demand that absorbs whatever compute volume Anthropic routes toward alternative suppliers.

Capital allocation implications

For investors, the more relevant question is not whether Nvidia faces a threat from this development, but how capital is likely to flow across the broader AI semiconductor ecosystem over the next two to three years.

Broadcom's custom silicon strategy positions it as a beneficiary of hyperscaler investment in proprietary compute. Its revenue trajectory and the scale of projected custom chip demand suggest that Broadcom could attract institutional attention as a diversification play within the AI hardware theme. The concentration of custom chip design expertise and its relationships with multiple hyperscalers provide a structural moat that generic GPU makers cannot easily replicate.

Alphabet's exposure to TPU revenues, while not yet cleanly isolated in its reporting, adds a hardware dimension to its AI infrastructure story. If Google Cloud's growth rate continues to accelerate, TPU-related revenues may emerge as a contributor worth monitoring.

Nvidia, meanwhile, retains its position as the default training infrastructure provider at scale. Its software ecosystem, including the CUDA framework, represents a switching cost that no competing chip architecture has yet overcome in a commercially meaningful way. The installed base and developer tooling remain durable advantages.

Structural competition, not displacement

What this episode illustrates is the maturation of AI infrastructure procurement. Early in the generative AI cycle, Nvidia occupied the field almost uncontested. As training workloads grow and as AI laboratories gain financial scale and technical sophistication, procurement decisions become more deliberate. Custom silicon, multi-vendor architectures, and long-term supply agreements are the instruments of that sophistication.

The competitive dynamic between Nvidia, Broadcom, and Alphabet is not a zero-sum contest. Demand for AI compute is expanding rapidly enough to accommodate multiple winners. The more relevant structural question is whether any single vendor achieves the kind of concentration that would give it durable pricing power over AI laboratories. Anthropic's procurement approach suggests that leading AI developers are actively working to prevent precisely that outcome.macro conditions.