Broadcom Locks In Google and Anthropic for the AI Decade

Broadcom’s expanded silicon and networking pacts with Google and Anthropic signal a definitive shift toward purpose-built AI infrastructure designed to last through the decade.

Key Takeaways

• Broadcom has formalised a multi-year agreement with Google to design and supply successive TPU generations alongside a rack-level networking assurance pact extending to 2031, cementing the deepest custom-silicon partnership in the industry.
• Anthropic will access approximately 3.5 gigawatts of TPU-based compute capacity through Broadcom from 2027, reflecting the company’s accelerating commercial growth and its most significant infrastructure commitment to date.
• The deals reinforce a structural industry shift away from commodity GPUs toward co-designed accelerator stacks, positioning Broadcom as the essential integrator bridging hyperscaler ambition and manufacturable silicon reality.

The Filing That Rewrote the Map

On April 6, 2026, Broadcom submitted an 8-K to the Securities and Exchange Commission. Regulatory filings of this kind rarely command much attention beyond compliance desks, but this one carried the architecture of an industry realignment. The document formalised two separate agreements: a long-term technology development and supply framework with Google for successive generations of Tensor Processing Units, accompanied by a networking and component assurance pact covering Google’s next-generation AI racks through 2031; and, alongside Google and Anthropic, an expansion of an existing collaboration that will deliver approximately 3.5 gigawatts of TPU-based compute capacity to the AI laboratory beginning in 2027.

The language was measured. The implications were not. Together the agreements extend Broadcom’s position as the essential engineering intermediary of the AI infrastructure era, binding the company to two of its most consequential partners across timelines long enough to span multiple chip generations, data centre buildout cycles, and geopolitical seasons.

Google’s Deepening Dependency

The relationship between Broadcom and Google is not new. For more than a decade, the two companies have collaborated on the custom silicon programme that produced the TPU family, the accelerators that give Google its most durable edge in AI training and inference. What the April filing establishes is the formalisation of that partnership into a binding multi-year structure, converting what was previously an evolving engineering collaboration into a declared strategic commitment.

The distinction matters. Under the new framework, Google will continue to specify the architectural advances required for each successive TPU generation, while Broadcom assumes responsibility for ASIC implementation, physical verification, advanced packaging, and the supply-chain coordination required to reach hyperscale production volumes. The accompanying networking assurance pact extends Broadcom’s footprint into the optical and Ethernet layers that increasingly define system-level performance.

At the scale of modern AI training runs, interconnect quality is not peripheral. When a single model training job consumes tens of thousands of accelerators simultaneously, the speed and reliability of the fabric binding them can determine whether that run completes in weeks or in months. By extending the Google agreement into rack-level networking through 2031, Broadcom has secured a position not merely in the chips but in the connective tissue of the entire system.

Anthropic’s Compute Architecture

The Anthropic dimension of the filing reflects a different but equally consequential logic. Anthropic is not a hyperscaler. It does not operate its own silicon fabrication pipeline, and until recently it did not hold the kind of long-dated infrastructure commitments that characterise companies of Google’s or Amazon’s scale. The 3.5-gigawatt allocation, routed through Broadcom and drawing on Google’s TPU infrastructure, represents a deliberate architectural choice: access frontier compute at scale without building the vertical integration required to own it.

The numbers behind that choice are instructive. Anthropic’s annualised revenue run-rate has risen above $30 billion in 2026, up from approximately $9 billion at the close of 2025. More than 1,000 enterprise customers now each spend over $1 million annually, a figure that doubled in under two months following the company’s February fundraising round. Against that trajectory, CFO Krishna Rao’s description of the TPU allocation as the company’s “most significant compute commitment to date” reads as understatement rather than emphasis.

The filing notes, carefully, that actual consumption will depend on Anthropic’s continued commercial performance and that discussions with operational and financial partners to support deployment are ongoing. Such language reflects the capital intensity of contemporary AI infrastructure honestly. Even well-capitalised organisations are exploring co-investment structures, revenue-sharing arrangements, and in some cases government-backed financing to distribute the cost of multi-gigawatt commitments. The willingness to make the commitment publicly, regardless of those caveats, signals a degree of commercial confidence that is itself informative.

The Structural Shift in Silicon

The deeper significance of both agreements lies in what they confirm about the direction of AI infrastructure investment. The era of general-purpose GPU procurement is not over, but it is no longer the whole story. Hyperscalers and well-funded laboratories alike have concluded that owning silicon architecture while outsourcing fabrication and systems integration yields advantages that compound over time: performance tuned to specific workloads, cost structures that improve with scale, and supply visibility that cannot be purchased on the open market.

Broadcom’s role in this model is precise. It does not own the algorithms. It does not operate the data centres. It translates design intent into manufacturable silicon at TSMC and delivers the switching, optical, and interconnect components that turn individual chips into coherent exascale systems. The Tomahawk and Jericho switch families, once considered mature franchise assets, have found renewed commercial relevance as AI clusters demand higher port density, lower latency, and optical scaling beyond what copper interconnects can sustain. The 2031 horizon in the Google supply pact effectively embeds that integration into the forward planning of one of the world’s most consequential AI programmes.

What the Market Understood

Broadcom Inc. shares (NASDAQ: AVGO) rose roughly 2.5 percent in after-hours trading following the filing, moving from approximately $314 at the close of the regular session toward $322 in extended hours. The reaction was calibrated rather than dramatic, which may be the most accurate response available. The stock has already priced in a significant share of the custom-AI opportunity. What the pacts provide is duration and specificity: concrete evidence that the relationships underpinning the thesis are not transactional but structural, extending through product cycles that have not yet been designed.

For investors with long-dated positions, that distinction is meaningful. Design wins in custom silicon tend to be self-reinforcing. Engineering familiarity, supply-chain infrastructure, and the sheer institutional knowledge embedded in a multi-generational collaboration create switching costs that protect revenue visibility in ways that commodity component contracts do not. The Google-Anthropic agreements simply extend the runway on a dynamic that was already well established.

The Integration Advantage

Broadcom is not without competition. OpenAI is pursuing its own custom accelerator programme with Broadcom’s involvement; Meta, Amazon, and Microsoft are advancing internal silicon efforts with varying degrees of vertical integration. The competitive landscape for bespoke AI compute is becoming denser, not thinner.

What Broadcom holds, however, is the longest and most technically mature custom-silicon collaboration in the industry, combined with a networking portfolio that no semiconductor peer has assembled to equivalent depth. The company captures value across the stack, from ASIC design services through switching fabric to optical components, without bearing the full risk of owning the applications or the infrastructure they run on. It is a model that scales with its customers rather than against them.

The April 6 filings do not resolve every uncertainty. ASIC development cycles remain expensive and unforgiving; architecture choices made today will propagate through product generations for years. Power consumption at multi-gigawatt scale presents genuine engineering and regulatory complexity. And the software ecosystem surrounding Nvidia’s CUDA platform continues to offer a development velocity advantage that hyperscalers can reduce through internal tooling but cannot entirely neutralise.

None of that diminishes what the agreements establish. Broadcom has positioned itself as the connective layer of the AI infrastructure decade, and its most important partners have now said so on the record.