Anthropic's TPU Pivot: Broadcom's Mystery Customer Revealed

Executive Summary: The Mystery Clarifies—But Not Favorably for Broadcom#

The Mizuho Hypothesis Under Fire#

Less than seventy-two hours after Broadcom's AVGO investment thesis received apparent validation through a Mizuho analyst report identifying Anthropic as the company's mystery ten-billion-dollar-plus customer, new developments have introduced material ambiguity into that narrative. Anthropic, the enterprise-focused artificial intelligence platform and one of the two most credible large-language-model providers alongside OpenAI, has announced a multibillion-dollar agreement with Google Cloud to access up to one million tensor processing units—a deployment pattern suggesting the company is pursuing a deliberately diversified, multi-vendor compute strategy rather than consolidating around AVGO's custom-silicon offerings. The timing and scale of the Anthropic-Google announcement, arriving just three days after the Mizuho thesis gained visibility in institutional investor circles, raises a critical question for AVGO shareholders: if Anthropic, the hypothetical fifth member of AVGO's supposed five-customer custom-silicon ecosystem, is deploying at scale on Google's TPUs rather than AVGO's custom accelerators, how many of the remaining four customers in the Mizuho narrative are actually committed to AVGO silicon, and what does this imply for the durability of the custom-silicon revenue projections that have underpinned the stock's recent valuation expansion?

The Anthropic-Google transaction itself is substantial and strategically significant. The agreement will grant Anthropic access to more than a gigawatt of compute capacity by 2026, a scale that industry experts estimate will translate to approximately fifty billion dollars in annual infrastructure spending at sustained deployment levels. Anthropic's Chief Financial Officer Krishna Rao characterized the deal as essential to the company's ability to "continue to grow the compute we need to define the frontier of AI." This framing is critical: it positions Google TPU capacity, not AVGO custom silicon, as Anthropic's primary infrastructure solution for the next phase of model training and inference. The company has emphasized that AWS, which has invested eight billion dollars in Anthropic and hosts the company's Project Rainier custom supercomputer on Trainium 2 chips, remains its "primary training partner and cloud provider." Yet the addition of Google TPU capacity and ongoing access to Nvidia's general-purpose GPUs suggests that Anthropic's compute strategy is deliberately distributed across three distinct hardware ecosystems: Amazon Trainium, Google TPUs, and Nvidia GPUs.

Direct vs. Indirect Relationships: The Critical Distinction#

For AVGO, which built its recent narrative momentum around the thesis that the company was capturing a diversified, multi-customer custom-silicon franchise, the emergence of Anthropic as a Google TPU customer rather than a AVGO custom-silicon customer represents a strategic pivot that the October 21 Mizuho analysis did not anticipate. This distinction between direct and indirect relationships is material to AVGO's valuation and growth trajectory. A direct custom-silicon engagement, where AVGO designs silicon optimized specifically for a customer's unique model architectures and workload patterns, generates higher gross margins, creates switching costs that enhance customer stickiness, and provides long-term revenue durability through proprietary design advantages. An indirect relationship, where AVGO manufactures silicon for Google that Google then deploys to customers like Anthropic, is capital-efficient and recurring but lacks the margin profile and customer lock-in characteristics of direct engagements.

The Mizuho hypothesis rested explicitly on AVGO capturing direct multi-gigawatt commitments from five major artificial intelligence platforms: OpenAI (confirmed), Anthropic (theorized), and three unnamed hyperscalers. If Anthropic's silicon strategy relies primarily on Google TPUs and AWS Trainium rather than AVGO, the hypothesis becomes substantially less credible. The company's entire narrative of customer diversification depends on the assumption that Anthropic represents one of five exclusive direct relationships. The revelation that Anthropic is using Google's commodity TPU offerings—manufactured through AVGO's partnership with Google, but not customized to Anthropic's specifications—introduces fundamental questions about whether the five-customer ecosystem truly exists or whether AVGO remains concentrated on the OpenAI engagement with three undefined and potentially non-exclusive relationships.

The Mizuho Thesis Under Revision: Analyst Methodology and Narrative Risk#

Gaps in Industry Check Methodology#

The Mizuho analysis, published on October 21, was grounded in two core methodologies: industry checks—confidential conversations with supply-chain participants and AVGO partners—and cross-referencing of publicly available financial disclosures regarding Anthropic's revenue trajectory and infrastructure investment plans. Mizuho analyst Vijay Rakesh argued that Anthropic's disclosed revenue targets for 2025 and 2026, combined with the timing and scale of AVGO's ten-billion-dollar-plus mystery customer commitment from September, made Anthropic the logical candidate for the mystery customer role. The logic was internally coherent: an artificial intelligence platform targeting nine billion dollars in annualized revenue by end-2025 and twenty billion dollars by end-2026 requires substantial infrastructure investments in compute capacity; such a company would logically engage AVGO for custom-silicon design to optimize its model architectures and inference patterns; the revenue scale and infrastructure footprint align with the magnitude of AVGO's disclosed commitments.

Yet the Anthropic-Google TPU deal announced just three days later revealed a critical gap in the Mizuho analysis: it did not account for the possibility that Anthropic, despite the capital available and the infrastructure needs, would choose to deploy on Google's commodity TPU offerings rather than pursue a custom-silicon partnership with AVGO. This gap raises broader questions about the reliability of industry checks as a methodology for identifying undisclosed customer relationships. If Mizuho's confidential conversations with supply-chain participants and AVGO ecosystem partners led to the conclusion that Anthropic was a AVGO custom-silicon customer, but Anthropic's actual compute strategy relies primarily on AWS Trainium and Google TPUs, then either the industry-check methodology failed to uncover material strategic information, or the AVGO partners and supply-chain participants who participated in the checks did not have access to Anthropic's compute roadmap. Either interpretation is concerning for investors who have allocated capital based on Mizuho's thesis.

Concentration Risk Reasserts Itself#

The October 13 OpenAI announcement, combined with the October 21 Mizuho analysis, created a narrative of AVGO as the winner in the artificial intelligence infrastructure arms race—a company capturing exclusive custom-silicon relationships with the two most credible large-language-model providers (OpenAI and Anthropic). The October 23-24 Anthropic-Google TPU announcements have undermined that narrative, introducing ambiguity into whether AVGO has truly captured the five-customer ecosystem that the Mizuho thesis envisioned or whether the company remains heavily concentrated on the OpenAI engagement. The Trefis analysis published on October 23 titled "Could AVGO Be In Trouble?" articulated, with notable prescience, the very concentration risk that the Anthropic-Google announcement has now validated. Trefis highlighted that AVGO's AI semiconductor revenue, which reached $5.2 billion in Q3 2025, is heavily dependent on a "limited number of hyperscale cloud clients," with forty to fifty percent of that revenue concentrated in a small set of customers.

The Anthropic case exemplifies precisely the risk that Trefis articulated: Anthropic, previously theorized to be a key AVGO customer, has instead chosen to diversify its suppliers, deploying on Google TPUs rather than consolidating around AVGO. The October 21 Mizuho thesis was intended to mitigate investor concerns about concentration by demonstrating that AVGO had diversified its customer base; the October 23-24 Anthropic-Google announcements have partially negated that mitigation by revealing that at least one of the purported five customers is not, in fact, a significant AVGO custom-silicon customer. AVGO's stock, which added more than $126 billion in market capitalization following the October 13 OpenAI announcement, is now trading at a price-to-earnings multiple of approximately 84.7 times trailing earnings—a valuation that presupposes near-perfect execution on multi-customer custom-silicon roadmaps and sustained hyperscaler capital expenditure at elevated levels. This elevated valuation leaves minimal room for disappointments or revisions to the custom-silicon pipeline.

Execution Imperatives: The OpenAI Relationship and Valuation Recalibration#

OpenAI as the Critical Test#

With the Anthropic customer identification now complicated by the Google TPU deployment pattern, the OpenAI relationship becomes the critical test of AVGO's custom-silicon strategy. The October 13 announcement of a partnership between OpenAI and AVGO was grounded in explicit statements from OpenAI regarding the company's commitment to a ten-gigawatt infrastructure deployment over multiple years. If OpenAI executes on its multi-year custom-silicon ramp as disclosed, AVGO will have a sustained revenue stream and a reference customer of immense prestige. Conversely, any delays in OpenAI's infrastructure buildout, any shifts in OpenAI's vendor strategy, or any disappointing performance characteristics of AVGO's custom accelerators could trigger a significant re-evaluation of the stock.

AVGO's management team, led by Chief Executive Hock Tan, will face intense scrutiny from institutional investors at the next earnings call regarding customer identification, revenue guidance, and the durability of the custom-silicon pipeline. The company has historically been reluctant to disclose specific customer names in its custom-silicon business, citing confidentiality agreements. However, the Mizuho hypothesis and the subsequent Anthropic-Google announcement have created investor demand for greater transparency regarding customer identities and roadmaps. The three unnamed hyperscalers referenced in the Mizuho thesis remain entirely opaque to institutional investors. The Anthropic experience—a company that appeared to be a AVGO customer based on supply-chain checks and financial logic, but that is actually deploying on Google TPUs—demonstrates the risk inherent in relying on inference to identify and assess customer relationships.

Valuation and Investment Implications#

If the Mizuho hypothesis regarding the three unnamed hyperscalers is subject to the same methodological limitations that the Anthropic case has revealed, then investor confidence in the overall five-customer ecosystem thesis should be correspondingly reduced. AVGO's strategic challenge is to demonstrate that the three unnamed customers are, in fact, committed to exclusive or near-exclusive custom-silicon relationships that will generate sustained, high-margin revenue over multiple years. This demonstration must occur either through management commentary at earnings calls, through customer disclosure and confirmation, or through other explicit channels rather than through supply-chain inference. The absence of such disclosure leaves institutional investors dependent on analyst hypotheses, which the Anthropic case has shown can be incomplete or incorrect.

The October 21 to October 24 narrative arc has traced a reversal in investor sentiment regarding AVGO's customer diversification. On October 21, the Mizuho thesis appeared to validate the narrative that AVGO had successfully diversified away from a single-customer dependency to a five-platform ecosystem anchored by OpenAI and Anthropic. By October 24, the revelation of Anthropic's multi-vendor compute strategy has introduced substantial doubt regarding whether AVGO has truly achieved meaningful customer diversification or whether the company remains heavily concentrated on OpenAI and an undefined set of hyperscaler commitments. This recalibration is material and carries implications for AVGO's long-term valuation multiple and investor positioning. AVGO's current valuation at 84.7 times trailing earnings embeds aggressive assumptions regarding execution on the custom-silicon roadmap, durability of customer relationships, and sustenance of hyperscaler capital expenditure at elevated levels.

Outlook: Catalysts, Competitive Threats, and Investor Positioning#

Monitoring Execution and Competitive Dynamics#

Institutional investors should monitor several catalysts that will determine whether the October 24 sentiment shift proves to be a temporary valuation dip or the beginning of a more sustained re-evaluation of AVGO's custom-silicon strategy. First, AVGO's next earnings call (expected in early December for fiscal Q4 results) will be critical for management commentary regarding the mystery customer disclosure, the durability of the OpenAI relationship, and any articulation of roadmaps with additional hyperscaler customers. Second, watch for any additional announcements from major hyperscalers regarding their custom-silicon strategies and partnerships with other chip vendors. Third, monitor technical benchmarks and performance comparisons between AVGO custom accelerators and Google TPUs, Amazon Trainium, and other alternative silicon offerings to assess whether AVGO's custom silicon commands a premium or faces competitive pricing pressure.

Fourth, pay attention to any industry commentary or supply-chain reporting regarding AVGO's market share in custom ASIC AI accelerators, which currently stands at an estimated seventy-five percent but depends on continued customer selection of AVGO for custom-silicon design. Fifth, assess the implications of any shifts in hyperscaler capital expenditure plans, as the AVGO thesis is contingent on sustained or increasing levels of infrastructure investment across the industry. Beyond the customer concentration concern, the Anthropic-Google announcement reveals a competitive dynamic that poses a longer-term threat to AVGO's custom-silicon business. Google, through its TPU program and partnerships with companies like Anthropic, is demonstrating that custom silicon designed by Google (with manufacturing partnerships from AVGO and others) can compete effectively against specialized custom accelerators designed in exclusive partnerships between chip vendors and AI companies.

Conservative Investment Approach Warranted#

Anthropic's choice to deploy on Google TPUs rather than to engage AVGO for custom silicon suggests that the economics, performance characteristics, and operational convenience of Google's TPU ecosystem may be superior to alternative custom-silicon offerings. If this pattern—hyperscalers or independent AI companies choosing Google TPUs over vendor-specific custom silicon—becomes more common, the total addressable market for AVGO's direct custom-silicon relationships could be compressed, and the company's long-term growth assumptions would require downward revision. For conservative institutional investors, the prudent approach is to treat the AVGO custom-silicon narrative as materially more uncertain than the October 13-21 discourse suggested. AVGO's success or failure in the custom-silicon market will be determined not by analyst hypotheses or supply-chain inference, but by the company's ability to execute on disclosed customer commitments (OpenAI) and to articulate credible roadmaps with additional customers.

Until AVGO provides explicit disclosure regarding additional customer relationships or until other customers independently confirm their custom-silicon partnerships with AVGO, the stock should be valued with a discount to account for the customer concentration risk that the Anthropic case has now validated. For investors holding AVGO at post-October-13 prices, the recommended approach is to maintain exposure while monitoring the catalysts outlined above. For those considering new positions, entry at current prices should be tempered by awareness that the October 21-24 developments have introduced meaningful execution risk that the October 13-21 discourse had not fully captured. The custom-silicon opportunity for AVGO remains substantial, but the path to achieving the Mizuho thesis's ambitious revenue projections appears significantly more uncertain than October 21 commentary suggested.