Anthropic-Google TPU Capital Stack: AI Infrastructure Becomes A Balance Sheet Trade

1. Executive Overview

The relevant events are not isolated funding announcements. They represent a coordinated capital-market, cloud-capacity, and custom-silicon inflection across the Anthropic-Google-Broadcom TPU ecosystem. The first transaction was Anthropic's confidential submission of a draft Form S-1 registration statement to the SEC for a proposed IPO of common stock. That filing came immediately after Anthropic disclosed a $65B Series H financing at a $965B post-money valuation, with run-rate revenue above $47B in May 2026, and after a period in which the company's annualized revenue reportedly moved from approximately $9B at year-end 2025 to more than $30B by April 2026. The second transaction was Alphabet's proposed $80B equity capital raise, structured through $30B of underwritten public offerings, $40B of at-the-market issuance expected to begin in Q3 2026, and a $10B Berkshire Hathaway private placement. The third transaction was the reported Apollo/Blackstone-led private-credit financing of roughly $36B to buy Google TPUs that Anthropic would lease. Together, the 3 events indicate that frontier AI compute has moved from ordinary hyperscaler procurement into a fully developed capital stack involving public equity, strategic equity, mandatory convertibles, private credit, supplier credit enhancement, and long-duration cloud commitments. The source hierarchy matters: the Anthropic, Alphabet, Google, and Broadcom facts are company-confirmed or filed; the Apollo/Blackstone tranche economics and Broadcom residual-value support are reported private-credit terms and should be read as subject to final documentation.

The report uses the following source-confidence ladder so confirmed company disclosures, reported financing terms, and investment inferences do not blend together.

The primary strategic conclusion is that Google's TPU ecosystem has crossed a threshold from internal proprietary infrastructure into a credible external compute platform for frontier AI labs. Historically, TPUs were best understood as an internal Google cost and performance advantage used to run Search, YouTube, Ads, DeepMind, and Google Cloud workloads. The Anthropic relationship changes that framing. Anthropic is not merely consuming ordinary Google Cloud compute; it has entered into a 5 GW Google/Broadcom next-generation TPU arrangement, while Broadcom's filed disclosure identifies approximately 3.5 GW of Anthropic access through Broadcom beginning in 2027, subject to Anthropic's continued commercial success. That distinction is important: 5 GW is the broader Anthropic Google/Broadcom commitment; 3.5 GW is the cleaner Broadcom-throughput figure. Broadcom has separately disclosed a long-term agreement to develop and supply future Google TPU generations and provide networking and other components for Google's next-generation AI racks through up to 2031. That elevates TPU from a Google-internal accelerator family into a bankable, externally financed AI infrastructure ecosystem.

Alphabet’s $80B equity raise is the most important signal for the broader hyperscaler capex cycle. Alphabet disclosed that demand for AI products and cloud services exceeds available supply, that 2026 capex is expected to be $180B-$190B, and that 2027 capex is expected to increase significantly from 2026. The company also disclosed $174B of operating cash flow over the 12 months ended March 31, 2026, more than $85B of debt raised in the prior year, and total debt outstanding above $100B. This implies that even one of the world’s largest and most cash-generative technology companies is moving from a self-funded AI capex model to a hybrid funding model using equity, debt, mandatory convertibles, capped calls, and strategic private placements. The transaction should not be interpreted as a liquidity stress event. It should be interpreted as a conscious decision to preserve balance-sheet flexibility, support extraordinary AI infrastructure demand, and avoid funding a 2026-2027 AI buildout exclusively through operating cash flow and incremental leverage.

Anthropic’s financing sequence is equally important. The confidential S-1 provides optionality for a public listing, but the more immediate investment implication is enhanced disclosure. A public S-1 would likely reveal gross margin structure, compute commitments, cloud vendor concentration, revenue durability, related-party strategic arrangements, customer mix, deferred revenue, lease obligations, model-training cost, inference cost, free cash flow, safety liabilities, and sensitivity to hyperscaler financing. The $65B Series H at a $965B post-money valuation, combined with the confidential IPO filing, indicates that Anthropic is attempting to convert extraordinary private-market demand into a public-market-quality capital base before compute commitments become even larger. This is rational. Frontier-model businesses cannot scale without pre-committing to accelerators, memory, networking, power, cooling, land, and cloud capacity years in advance. The capital needs precede the fully visible GAAP revenue and cash-flow realization.

The Apollo/Blackstone-Broadcom TPU financing is the most novel transaction in the group, but the terms should be treated as reported rather than company-confirmed. The reported structure uses an SPV that borrows money, receives an equity investment, purchases Google TPUs, and leases those chips to Anthropic. Debt service is expected to be supported by Anthropic lease payments and the residual value of the chips. Reported tranche sizing includes approximately $6B of A1 notes, $25B of A2 notes, and $4.5B of B notes, with the senior A1/A2 layers expected to benefit from Broadcom residual-value support. Broadcom's 8-K confirms discussions with financial and operational partners, but it does not file final guarantee mechanics. The structure is therefore not purely an Anthropic credit. It is a reported supplier-enhanced AI infrastructure financing that attempts to transform proprietary accelerators into an institutional private-credit asset class.

The investment conclusion is constructive for the Google TPU ecosystem and for several public-market suppliers, but the risk-adjusted implications vary materially by company. Alphabet receives strategic validation of Google Cloud and TPU as scaled AI infrastructure, but shareholders face a more capital-intensive model, possible dilution, higher depreciation, and greater scrutiny of Cloud backlog quality and incremental ROIC. Broadcom is the cleanest direct public-equity beneficiary because it controls critical custom silicon, networking, SerDes, and rack-level component exposure for the Google TPU roadmap; however, the reported residual-value support introduces a new contingent credit exposure that should not be ignored. Lumentum and Coherent benefit from optical intensity in multi-GW AI clusters, but direct customer allocation remains opaque. Celestica benefits from rack-scale integration, liquid cooling, hyperscaler platform programs, and co-packaged optics, but remains a lower-margin execution-sensitive beneficiary. Fabrinet, Credo, Marvell, Arista, memory suppliers, power-equipment vendors, cooling vendors, electrical contractors, and data-center infrastructure providers are second-order beneficiaries, with exposure determined by exact customer awards and architecture-specific content.

2. What Happened: The 3 Transactions

The first transaction was Anthropic’s confidential draft S-1 submission. The company announced on June 1, 2026 that it had confidentially submitted a draft Form S-1 registration statement to the SEC for a proposed IPO of common stock. Anthropic did not disclose the number of shares, price range, valuation, or timing, and stated that the IPO would depend on SEC review, market conditions, and other factors. The key point is that the filing creates an IPO option rather than an immediate public listing. It also creates a path toward public-company disclosure of the economics behind one of the largest and fastest-scaling AI labs globally.

The S-1 came directly after Anthropic announced a $65B Series H financing at a $965B post-money valuation. Anthropic disclosed that run-rate revenue had crossed $47B earlier in May 2026 and that proceeds would support safety and interpretability research, compute expansion, product scaling, and partnership growth. The financing included large public-market investors, alternative asset managers, strategic infrastructure partners, and hyperscalers, including $15B of previously committed hyperscaler investment. Anthropic also identified Micron, Samsung, and SK hynix as strategic infrastructure partners, which is significant because the company’s bottleneck is not limited to accelerators; it extends into HBM, DRAM, NAND, advanced packaging, storage, and supply assurance across the AI hardware stack.

The capacity disclosures need to be reconciled rather than collapsed into one headline number.

The second transaction was Alphabet’s proposed $80B equity capital raise. Alphabet announced $30B of underwritten public offerings, consisting of $15B of mandatory convertible preferred depositary shares and $15B of Class A/Class C common stock. Alphabet also announced a $40B at-the-market stock program expected to begin in Q3 2026 and a $10B Berkshire Hathaway private placement split evenly between Class A and Class C shares. The company stated that proceeds from the underwritten offerings and the Berkshire private placement would be used for general corporate purposes, including capital expenditures to scale AI infrastructure and global compute capacity. The ATM proceeds are more nuanced: approximately $30B is expected to be used to satisfy 2026 employee equity-award tax obligations under a change in tax administration, while any incremental proceeds would be available for general corporate purposes.

Alphabet’s capital raise should be interpreted in the context of 3 numbers: $180B-$190B of expected 2026 capex, $174B of trailing 12-month operating cash flow, and more than $100B of total debt outstanding after more than $85B of debt raised in the prior year. These numbers make clear that Alphabet is not raising equity because AI demand is weak. It is raising equity because AI demand is larger than its preferred internally funded balance-sheet model. Cloud backlog above $460B, with approximately 50% expected to convert into revenue over the following 24 months, provides strong demand visibility but also requires infrastructure to be installed before revenue can be recognized. The transaction therefore trades some dilution and financing cost for capacity certainty, strategic flexibility, and balance-sheet protection.

The practical funding bridge is not a simple $80B AI-capex bucket. The ATM component is partly tax-administration related, while the underwritten offerings and Berkshire private placement are cleaner general-purpose and AI-infrastructure funding sources.

The third transaction was the reported Apollo/Blackstone-led chip financing. According to Reuters-sourced reporting and Private Equity Wire, Apollo and Blackstone were working to bring additional investors into a roughly $36B debt financing that would help Anthropic build AI infrastructure by purchasing Google TPUs that Anthropic would lease. The article described the structure as potentially one of the largest-ever private-credit deals and the largest chip-financing debt transaction, with Broadcom expected to support the largest senior portions. This transaction differs from ordinary cloud procurement because it inserts institutional private credit and reported supplier residual-value support between the chip maker, the hyperscaler ecosystem, and the AI model customer. The terms were reported as subject to documentation and final allocation, so the report treats them as underwriting inputs rather than finalized company disclosure.

The reported structure is economically important because it converts TPUs into financeable collateral. The SPV borrows money and receives equity capital, uses that funding to buy TPUs, and leases the chips to Anthropic. The debt is backed primarily by lease payments and secondarily by chip residual value. Anthropic is reportedly expected to use leased TPUs at sites in New York, Texas, Louisiana, and Indiana. The financing is also reportedly delayed-draw, meaning capital is drawn over time as chips become available and Anthropic leases commence. This better matches funding to delivery schedules and reduces negative carry, but it also embeds execution risk around chip availability, site energization, and lease commencement timing.

The most consequential feature is Broadcom's reported residual-value support. If Anthropic fails to make lease payments for a defined period and the SPV sells the TPUs at a value insufficient to repay senior debt holders, Broadcom would reportedly support the senior A1 and A2 investors under the residual-value arrangement. This is why the A1/A2 notes reportedly trade conceptually closer to Broadcom-linked investment-grade exposure than Anthropic's standalone startup credit. The structure may be off-balance-sheet in legal form, but it is not economically irrelevant. It transfers a portion of proprietary-chip residual-value risk from lenders to Broadcom if final terms match reporting, in exchange for accelerating adoption of the Google TPU ecosystem and supporting a major future revenue stream.

The underwriting lens should follow the financing waterfall: asset ownership, repayment source, residual support, and unresolved documentation questions.

3. Why It Happened

The immediate driver is compute scarcity. Anthropic disclosed that run-rate revenue exceeded $30B in April 2026, up from approximately $9B at year-end 2025, and that business customers generating more than $1M of annualized revenue doubled from more than 500 in February to more than 1,000 in less than 2 months. Anthropic then disclosed run-rate revenue above $47B in May 2026. That growth rate creates a procurement problem: compute must be secured in multi-year blocks before the revenue base fully matures. Model companies that fail to secure capacity will face latency constraints, product-availability issues, slower model iteration, weaker enterprise reliability, and lower negotiating leverage with hyperscalers.

The deeper driver is that frontier AI has become an infrastructure business disguised as a software business. The market often values model providers on software revenue multiples, but the operating reality increasingly resembles a capital-intensive utility, with large fixed commitments to accelerators, memory, networking, power, cooling, data-center shells, cloud services, and supply-chain reservations. Gross margin and free cash flow depend on utilization, price per token, model-efficiency improvements, accelerator depreciation, power cost, vendor credits, and the pace of hardware refresh. Anthropic’s decision to pursue an IPO path while simultaneously securing equity and lease-financed compute reflects the need to reconcile these 2 business-model identities: public software-like valuation and infrastructure-like capital intensity.

Anthropic’s multi-cloud strategy is also deliberate. The company has disclosed large compute commitments across AWS, Google/Broadcom TPUs, Nvidia GPUs, and other infrastructure partners. Anthropic stated that it trains and runs Claude across AWS Trainium, Google TPUs, and Nvidia GPUs, and that Claude is available through AWS Bedrock, Google Vertex AI, and Microsoft platforms. The company has also announced Amazon compute commitments involving up to 5 GW of capacity and more than $100B of AWS technology commitments over 10 years, while the Google/Broadcom arrangement provides multi-GW next-generation TPU capacity beginning in 2027.

That multi-vendor posture reduces single-supplier dependency and improves cost-performance optionality. Nvidia GPUs remain the most flexible and liquid accelerator platform, particularly for research, model experimentation, heterogeneous workloads, and broad developer compatibility. AWS Trainium and Google TPUs offer potential cost-per-token advantages for large-scale, more standardized workloads where hyperscaler-specific optimization can be justified. Anthropic therefore appears to be pursuing a portfolio approach to compute: GPUs for flexibility, Trainium for AWS-native scale, TPUs for Google-native scale and alternative custom-silicon economics, and public or quasi-public capital markets to fund the entire buildout.

Alphabet’s reason for raising equity is different but connected. Google Cloud is scaling rapidly, and Alphabet’s AI demand appears to exceed near-term infrastructure supply. Alphabet disclosed Q1 2026 revenue growth of 22% to $110B, Search & Other growth of 19%, Google Cloud growth of 63%, Cloud backlog above $460B, and approximately 50% of that backlog expected to convert into revenue within 24 months. That backlog is valuable only if Alphabet builds capacity in advance. A cloud backlog without available compute becomes deferred opportunity rather than recognized revenue. The $80B raise is therefore a way to fund the lag between signed demand and installed, energized, revenue-producing infrastructure.

Broadcom’s motivation is to convert custom-silicon incumbency into a multi-year platform annuity. Broadcom’s official 8-K states that it entered into a long-term agreement with Google to develop and supply custom TPUs for future Google TPU generations, plus a supply assurance agreement for networking and other components for Google’s next-generation AI racks through up to 2031. It also states that Anthropic is expected to access approximately 3.5 GW of TPU-based compute through Broadcom beginning in 2027. That puts Broadcom in the position of AI ASIC designer, networking supplier, rack-level component supplier, and financing enabler.

Private credit’s motivation is equally straightforward. AI infrastructure offers very large ticket sizes, long-duration contractual cash flows, identifiable collateral, and potential credit enhancement from investment-grade strategic counterparties. Traditional project finance has long funded power plants, aircraft, data centers, and telecommunications infrastructure. The new step is applying similar logic to accelerators. The challenge is that AI chips depreciate faster and have less stable residual values than power plants, aircraft, or fiber assets. The reported Broadcom support is therefore the linchpin that converts a potentially speculative Anthropic TPU financing into a senior institutional credit product.

4. Google And Alphabet: Strategic Implications

The positive Alphabet case is that Google now has the most credible non-Nvidia AI compute ecosystem at frontier scale. Google controls Gemini, DeepMind research, TPUs, Google Cloud, Vertex AI, Kubernetes/GKE, distributed training infrastructure, compiler tooling, and enterprise distribution. Anthropic adds an external frontier-model anchor tenant whose usage can improve TPU software maturity, capacity utilization, performance tuning, compiler optimization, and ecosystem credibility. This is a major strategic upgrade. A proprietary accelerator used mostly internally is a cost advantage. A proprietary accelerator adopted by a scaled third-party frontier lab becomes a platform.

The Anthropic TPU commitment also validates Google’s technical roadmap. Google introduced 8th-generation TPU architectures designed separately for training and inference, with TPU 8t oriented toward large-scale training and TPU 8i oriented toward latency-sensitive inference. Google described TPU 8t pods with 9,600 chips, 2 PB of shared HBM, 121 ExaFLOPs, double interchip bandwidth versus the prior generation, and the ability to scale to up to a 1M-chip logical cluster through its Virgo network. Google also emphasized 4th-generation liquid cooling and up to 2x better performance per watt versus Ironwood. These specifications matter because the industry bottleneck is shifting from raw accelerator availability to usable cluster-level throughput per watt, interconnect efficiency, memory bandwidth, cooling density, and uptime.

Google’s TPU positioning is especially relevant for inference. As AI usage shifts from training runs to agentic workflows, coding agents, enterprise copilots, search augmentation, long-context reasoning, multimodal generation, and real-time business process automation, inference volume can become the dominant cost driver. GPUs remain flexible, but custom ASICs can be structurally advantaged in predictable, high-volume inference if they deliver lower cost per token at acceptable latency and utilization. Google’s TPU 8i positioning around inference, combined with TPU 8t for large-scale training, suggests a roadmap designed to optimize the full model lifecycle rather than only peak training performance.

The negative Alphabet case is that strategic control comes with a materially more capital-intensive financial model and a more nuanced capital-raise story. Alphabet's disclosed 2026 capex of $180B-$190B is roughly equal to or above trailing 12-month operating cash flow of $174B before shareholder returns, taxes, working capital movements, debt service, and other corporate cash needs. The $80B raise is not entirely incremental AI capex cash because roughly $30B of ATM proceeds is expected to meet 2026 employee-equity tax obligations under a change in tax administration. Even so, the underwritten offerings, mandatory convertibles, Berkshire private placement, operating cash flow, and debt capacity together show that Alphabet is moving toward a hybrid AI infrastructure funding model. The old Alphabet model of extremely high organic cash generation, relatively asset-light incremental growth, and large recurring buybacks is being replaced by a model closer to a vertically integrated AI infrastructure utility.

This transition is not inherently negative for equity value. If Alphabet can monetize AI capacity through high-margin Cloud revenue, Search monetization defense, enterprise AI subscriptions, model APIs, TPU-based infrastructure, and improved internal productivity, the incremental ROIC could remain attractive. The issue is that the bar is high. A $180B-$190B capex year requires enormous future revenue and gross profit conversion to earn excess returns after depreciation and capital cost. The investment debate therefore shifts from “Does AI demand exist?” to “Will the marginal AI infrastructure dollar produce durable economic profit after pricing, utilization, depreciation, power, and obsolescence?”

The Cloud backlog is central to that debate. A backlog above $460B is strategically powerful, but its quality depends on customer concentration, pricing, duration, cancellation rights, cloud credits, strategic funding arrangements, GPU/TPU pass-through economics, and service mix. If backlog is dominated by a small number of frontier AI labs with capital-market-dependent funding models, the headline backlog multiple may overstate durability. If backlog is diversified across enterprises, public sector, developer platforms, and recurring AI-native workloads, Alphabet’s equity raise becomes a highly rational acceleration of a visible growth curve. Public disclosures over the next several quarters should be assessed through this lens.

The TPU ecosystem also gives Alphabet a stronger bargaining position against Nvidia. The implication is not that Nvidia demand collapses. Aggregate AI demand remains large enough to support GPUs, TPUs, Trainium, and custom XPUs simultaneously. The more realistic medium-term implication is that Google reduces dependence on Nvidia for its own workloads and offers strategic customers a lower-cost alternative for certain training and inference workloads. Over time, this could reduce Nvidia’s share of incremental hyperscaler-controlled AI workloads, particularly where model architecture, software stack, and deployment patterns are stable enough to justify ASIC optimization.

5. TPU Ecosystem: Broader Structural Implications

The most important ecosystem implication is that TPUs are becoming bankable infrastructure. A chip architecture becomes an ecosystem when it has customers, software support, financing, supply-chain partners, manufacturing scale, memory allocation, networking integration, and data-center deployment pathways. The Anthropic-Google-Broadcom capital stack provides all of those elements. Google contributes the cloud platform and TPU architecture. Broadcom contributes custom silicon, networking, rack-level components, and supplier credibility. Anthropic contributes high-growth frontier-model demand. Apollo and Blackstone contribute private-credit distribution. Alphabet contributes balance-sheet scale and public-equity funding. The resulting structure makes TPU capacity more institutionalized and less dependent on Google’s internal capital budget alone.

The TPU stack also changes the supplier map. Scaling AI clusters is not simply a function of accelerator count. It requires high-bandwidth memory, optical interconnect, Ethernet or proprietary switching, retimers, DSPs, liquid cooling, power delivery, thermal management, facility integration, transformers, substations, and software orchestration. Google’s own TPU disclosures highlight memory scale, interchip bandwidth, storage speed, liquid cooling, and cluster goodput. This is why the beneficiaries extend well beyond Google and Broadcom. Optical suppliers, electronics manufacturing services providers, switch platforms, power vendors, cooling vendors, and data-center construction companies all become levered to TPU deployment.

The second-order implication is that AI infrastructure is moving toward specialized supply chains by accelerator ecosystem. Nvidia clusters use a different supplier map from Google TPU clusters, AWS Trainium clusters, and future custom XPU clusters. Some suppliers are common across all ecosystems, such as memory, optics, power, cooling, and high-speed connectivity. Others are architecture-specific, particularly custom ASIC design, board design, rack integration, interconnect topology, and cluster orchestration. This means equity investors should avoid treating “AI capex exposure” as homogeneous. The correct analysis is content per rack, share of architecture, margin structure, customer concentration, and whether the supplier is exposed to multiple accelerator ecosystems or concentrated in 1 stack.

The third implication is that financing capacity may become a competitive moat. Nvidia’s largest customers have historically financed GPU purchases through balance sheets, cloud revenue, private credit, data-center operators, and GPU-backed loans. Google and Broadcom are now showing that TPUs can also be financed through institutional credit when supplier residual support is available. That matters because frontier AI labs do not want to fund all compute with equity. A structure that converts capex into leases can materially accelerate model-company growth, preserve equity capital, and improve near-term reported scalability. However, it also increases fixed obligations and exposes the system to utilization and pricing downturns.

6. Broadcom: Direct Winner, New Credit-Adjacent Risk

Broadcom is the clearest direct public-market beneficiary of the TPU complex. The company is formally tied to Google’s future TPU generations and next-generation AI rack components through up to 2031, while Anthropic is expected to access approximately 3.5 GW of TPU-based AI compute through Broadcom beginning in 2027. This gives Broadcom multi-year visibility into custom AI accelerators, networking, SerDes, switching, DSP/retimer content, and rack-level components. Unlike more generic AI suppliers, Broadcom appears embedded in both the accelerator and cluster fabric of the Google TPU roadmap.

The Broadcom exposure should be separated into confirmed operating economics and reported contingent credit support.

The current financial trajectory supports that view. Broadcom reported Q1 FY2026 revenue of $19.3B, up 29% year over year, adjusted EBITDA of $13.1B, free cash flow of $8.0B, and AI revenue of $8.4B, up 106% year over year. Management guided Q2 FY2026 revenue to $22.0B and AI semiconductor revenue to $10.7B. AI is therefore not a marginal growth category inside Broadcom; it is becoming the dominant incremental semiconductor growth engine. The Google/Anthropic agreements extend that growth curve into 2027 and beyond.

Broadcom’s strategic position is attractive because hyperscalers increasingly need custom silicon to manage cost, supply, and differentiation. Nvidia’s platform remains dominant, but reliance on Nvidia alone creates cost pressure, supply risk, and limited architectural control. Broadcom offers custom accelerator design, high-speed SerDes, Ethernet switching, networking silicon, DSP/retimer capability, and deep hyperscaler operating relationships. That combination is rare. Google’s decision to extend TPU development and supply arrangements through up to 2031 reinforces Broadcom’s position as a critical supplier for one of the largest AI infrastructure roadmaps globally.

The underwriting issue is residual-value support. The reported Anthropic financing makes Broadcom more than a semiconductor vendor if final terms match reporting. It makes Broadcom a credit enhancer for senior AI chip-financing tranches. That support may not impair Broadcom's credit ratings or appear as debt in the ordinary sense, but economic exposure still exists. If Anthropic underperforms, stops paying leases, and TPU collateral is insufficient to repay senior debt, Broadcom could be required to make lenders whole on supported senior amounts. The probability of loss may be low in the base case, but the severity could be meaningful because senior tranches reportedly total roughly $31B before any final sizing changes.

The key valuation implication is that Broadcom's AI revenue should be capitalized differently across 3 buckets. The first bucket is high-value custom-silicon development and production revenue, which deserves a premium multiple due to scarcity, customer entrenchment, and multi-year roadmap visibility. The second bucket is AI networking and rack content, which also deserves premium treatment because bandwidth, scale-out networking, and SerDes are structural bottlenecks. The third bucket is financing-enabled revenue supported by reported residual-value guarantees or similar arrangements, which deserves a probability-weighted contingent-liability discount. Equity value creation remains likely if revenue and margin conversion are strong, but the risk profile is no longer pure semiconductor operating leverage.

The most important Broadcom questions are now: how much of 2027-2031 AI revenue is already locked by Google TPU programs; what gross margin and working-capital profile should be assigned to TPU-related revenue; how Broadcom accounts for residual-value support; whether additional customers require similar credit enhancement; how much AI revenue is concentrated in Google/Anthropic; and whether Broadcom’s balance sheet is being used strategically to accelerate custom-silicon adoption. The answers determine whether Broadcom should be valued as a superior AI infrastructure semiconductor compounder or as a compounder with underappreciated supplier-finance tail risk.

7. Lumentum: High-Beta Optical Beneficiary, Direct TPU Share Not Proven

Lumentum is a highly relevant derivative beneficiary because multi-GW TPU clusters require enormous optical bandwidth. Large-scale AI clusters generate traffic across accelerator pods, racks, data-center halls, and regional data-center fabrics. Training requires high-bandwidth synchronization and scale-out networking. Inference increasingly requires low-latency, high-throughput connectivity for agentic workflows, retrieval, mixture-of-experts routing, distributed serving, and long-context applications. Google’s own TPU architecture disclosures emphasize pod scale, interchip bandwidth, liquid cooling, and large logical clusters. That creates demand for lasers, optical engines, transceivers, optical circuit switching, co-packaged optics, and related photonic components.

Lumentum’s recent reported results show that AI/cloud optical demand is already flowing through the P&L. The company reported Q3 FY2026 revenue of $808.4M, up 90.1% year over year, with GAAP gross margin of 44.2%, non-GAAP gross margin of 47.9%, GAAP operating margin of 21.6%, and non-GAAP operating margin of 32.2%. Management attributed the result to AI/cloud data-center demand, favorable mix, pricing discipline, laser chips, pump lasers, narrow-linewidth laser assemblies, scale-across components, and future expected contribution from optical circuit switches and co-packaged optics. Q4 FY2026 guidance of $960M-$1.01B implies another significant sequential step-up.

The Google/Anthropic TPU buildout is positive for Lumentum because it expands the number of scaled AI accelerator ecosystems that need optical bandwidth. Lumentum does not need Nvidia-only clusters to grow if TPUs, Trainium, and custom XPUs all require higher optical intensity. In fact, a multi-accelerator world may be better for high-quality optical suppliers because every architecture must solve the same underlying problem: moving more data at lower energy per bit, with lower latency, higher reliability, and tighter thermal constraints. Lumentum’s product exposure to lasers, optical components, optical circuit switches, and co-packaged optics aligns well with that requirement.

The principal caveat is that no public source establishes a specific direct incremental Lumentum award tied to Anthropic’s Google TPU financing. Hyperscalers use complex sourcing strategies and often dual-source or multi-source optics. Competitors and adjacent suppliers include Coherent, Innolight, Eoptolink, Fabrinet-assembled products, Accelink, Source Photonics, and internal or semi-custom hyperscaler reference designs. Co-packaged optics can also shift value away from traditional pluggable module suppliers toward optical engines, external light sources, silicon-photonics platforms, and switch-adjacent integration. Lumentum benefits most if its laser and photonic-component content remains central as architectures shift; it benefits less if value migrates primarily to module assemblers, switch vendors, or hyperscaler-specified captive designs.

The investment debate for Lumentum is therefore not whether AI optical demand is strong. It clearly is. The debate is share, margin durability, customer concentration, and whether current valuation already discounts a multi-year AI optical supercycle. The Google/Anthropic/Broadcom transactions increase confidence in the size and duration of the optical demand curve, but they do not eliminate allocation and pricing risk. For Lumentum, the key indicators are backlog duration, long-term supply agreements, hyperscaler customer concentration, capacity additions, pricing on high-speed lasers, optical circuit switch adoption, and co-packaged optics qualification.

8. Celestica: Rack-Scale Systems, CPO, Liquid Cooling, And Hyperscaler Execution Leverage

Celestica is a major derivative beneficiary because the TPU ecosystem will require complex rack-scale systems, high-speed switching, liquid cooling, optical integration, supply-chain orchestration, and manufacturing execution. The company reported Q1 2026 revenue of $4.05B, up 53% year over year, GAAP operating margin of 6.7%, adjusted operating margin of 8.0%, GAAP EPS of $1.83, and adjusted EPS of $2.16. Celestica raised 2026 guidance to $19.0B of revenue and $10.15 of adjusted EPS, and stated that 2027 visibility had strengthened. Its Communications and Cloud Solutions segment revenue rose 76% year over year to $3.24B, with hyperscaler revenue of approximately $1.7B, up 63%.

The most relevant disclosure is Celestica’s hyperscaler program win for the design and manufacturing of a co-packaged-optics Ethernet switch optimized for AI scale-out networks. The program uses 1.6Tb switch silicon, co-packaged optical interconnects, and liquid cooling, with production expected to ramp in 2027. The timing is notable because 2027 is also when Anthropic’s next-generation TPU capacity through Google/Broadcom begins to come online. The disclosure does not prove that Celestica is directly supplying the Anthropic-Google TPU buildout, but it does prove that Celestica is qualified for exactly the class of AI infrastructure hardware that hyperscalers need for next-generation clusters.

Celestica’s DS6000 switch platform further supports the strategic relevance. The platform is powered by Broadcom Tomahawk 6 and provides 102.4 Tbps of non-blocking bandwidth, 64 ports of 1.6TbE, and AI-oriented scale-up and scale-out support. This places Celestica in the intersection of Broadcom switching silicon, high-density Ethernet, AI rack design, cooling, and hyperscaler infrastructure deployment. That is precisely where value is migrating as AI clusters move from accelerator scarcity to networked system performance.

Celestica’s economics are structurally different from Broadcom’s. Broadcom owns scarce semiconductor IP and should capture high incremental gross profit on custom silicon and networking content. Celestica captures value through engineering, integration, manufacturing, supply-chain execution, test, quality, and ramp management. The margin ceiling is lower, but operating leverage can still be substantial if hyperscaler programs scale rapidly and production complexity supports higher value-added pricing. Celestica’s risk is not strategic irrelevance; it is execution, concentration, and price-down pressure. Hyperscaler programs can be large, but they are demanding, margin-sensitive, and dependent on flawless supply-chain coordination.

The key investment question for Celestica is whether the 2027 earnings power implied by hyperscaler AI program ramps remains under-modeled. Alphabet’s capex is expected to rise significantly in 2027, Google/Broadcom/Anthropic TPU capacity begins in 2027, and Celestica’s CPO Ethernet switch program also ramps in 2027. That timing alignment is suggestive but not dispositive. The upside case is that Celestica becomes a preferred high-complexity AI systems integrator across multiple hyperscaler architectures. The downside case is that growth is already capitalized at elevated multiples while program concentration and working-capital requirements become more visible.

9. Coherent, Fabrinet, Credo, Marvell, Arista, Memory, Power, And Other Ecosystem Companies

Supplier exposure should be underwritten by proof level. The Google/Broadcom TPU buildout is directionally positive for optical, EMS, connectivity, networking, memory, power, and cooling demand, but only Broadcom has direct public filing support for the Google TPU/rack role.

Coherent is a direct photonics peer and likely beneficiary of the same bandwidth bottleneck that supports Lumentum. The company reported Q3 FY2026 revenue of $1.81B, up 21% year over year, with GAAP gross margin of 37.7% and non-GAAP gross margin of 39.6%. Management cited strong demand across data-center and communications businesses and ongoing capacity expansion to meet AI data-center infrastructure demand. Coherent’s opportunity in the TPU ecosystem depends on its share of lasers, transceivers, optical components, and photonic subsystems used in Google, Broadcom, or supplier-qualified AI networks.

Fabrinet is a high-quality derivative beneficiary through precision optical manufacturing. The company reported Q3 FY2026 revenue of $1.214B versus $871.8M a year earlier, with GAAP EPS of $3.45 and non-GAAP EPS of $3.72. Management cited ongoing and ramping programs and new datacom customer agreements as demand drivers, and guided Q4 FY2026 revenue to $1.25B-$1.29B. Fabrinet’s value capture is manufacturing and packaging rather than ownership of all optical IP, but the company can benefit materially if AI optical suppliers outsource complex production at higher volumes.

Credo is exposed through active electrical cables, retimers, DSPs, and high-speed connectivity used to improve cluster reliability, power efficiency, and time to stability. Credo reported Q4 FY2026 revenue of $437.0M, up 157.0% year over year, with GAAP gross margin of 68.2%, non-GAAP gross margin of 68.3%, GAAP net income of $169.1M, non-GAAP net income of $226.7M, and $1.4B of cash and short-term investments. Management described its products as improving cluster time-to-stability, GPU utilization, network reliability, and infrastructure power efficiency. The same attributes are relevant to TPU clusters, but the degree of direct TPU exposure depends on whether Google/Broadcom architecture decisions include merchant connectivity content or rely more heavily on internally specified alternatives.

Marvell is both a beneficiary and competitive comparator to Broadcom. The Google TPU arrangement validates the broader trend toward hyperscaler custom silicon, which is positive for Marvell’s custom XPU, optical interconnect, and Ethernet strategy. However, Google TPU appears to be a Broadcom-led program, so Marvell is not the clean direct beneficiary in this specific ecosystem. Marvell reported Q1 FY2027 revenue of $2.418B, up 28% year over year, and management cited exceptional AI bookings across 800G and 1.6T scale-out optics, 51.2T Ethernet switches, NPO/CPO optical solutions, scale-across data-center interconnect, custom XPU, and XPU-attach solutions. The implication is that TPU adoption is not negative for Marvell’s thematic setup, but Broadcom has the stronger direct Google TPU read-through.

Arista’s exposure is more nuanced. AI clusters increasingly use Ethernet-based scale-out networks, and Arista has a strong position in cloud networking, high-radix switching, and AI networking architectures. However, Google’s TPU ecosystem is vertically integrated, and Broadcom’s rack-level agreement may limit the addressable share available to independent switch-system vendors in Google-specific TPU deployments. Arista benefits more clearly from broader AI Ethernet adoption across non-Google hyperscalers, cloud customers, and enterprise AI fabrics than from a direct read-through to the Anthropic-Google TPU financing. The broader sector implication is still constructive because TPU scale reinforces the need for high-performance networking across all AI infrastructure.

Memory suppliers are central beneficiaries because Anthropic explicitly named Micron, Samsung, and SK hynix as strategic infrastructure partners in its Series H announcement. The TPU ecosystem cannot scale without HBM, DRAM, NAND, storage, and advanced memory packaging. Google’s TPU architecture disclosures also underscore the importance of shared HBM and bandwidth at pod scale. The investment implication is that Anthropic’s capital raise and Google TPU commitments should be modeled not only as accelerator demand but also as memory, storage, and advanced packaging demand. Memory suppliers benefit most if AI demand remains undersupplied and HBM mix continues improving; they benefit less if industry capacity additions compress pricing before utilization ramps.

Power, cooling, electrical infrastructure, and data-center construction suppliers are structural beneficiaries. Multi-GW compute commitments require substations, transformers, switchgear, backup power, thermal systems, liquid cooling, pumps, power distribution units, electrical contracting, grid interconnections, and real estate development. Alphabet’s $180B-$190B 2026 capex guidance and expectation of significantly higher 2027 capex place power availability at the center of AI infrastructure execution. Likely public beneficiaries include Vertiv, Eaton, Schneider Electric, Siemens, ABB, GE Vernova, Quanta Services, Comfort Systems, Modine, selected data-center REITs, and private data-center developers, although direct award visibility varies materially by company and region.

10. Implications For Nvidia And The Broader Accelerator Market

The Google TPU ecosystem should be treated as an incremental alternative to Nvidia, not an immediate replacement. Nvidia’s position remains strongest in flexible training, research workloads, large developer ecosystems, model experimentation, heterogeneous deployments, and workloads requiring the broadest software compatibility. TPUs are more specialized and can be more attractive where workload scale, model architecture, and deployment environment justify optimization. Anthropic’s multi-accelerator strategy supports this interpretation: the company is not abandoning Nvidia GPUs; it is adding TPUs and Trainium capacity to diversify supply, cost, and performance.

The more important medium-term risk to Nvidia is not lost 2026 demand. It is the gradual migration of incremental hyperscaler-owned inference workloads toward custom ASICs where cost per token, power efficiency, and supply certainty matter more than general-purpose flexibility. If Google TPUs and AWS Trainium demonstrate strong economics at scale, hyperscalers and frontier labs will increasingly reserve GPUs for the workloads where Nvidia has the highest relative advantage and move repeatable high-volume inference to custom silicon. This could lower Nvidia’s share of incremental tokens served even if absolute Nvidia revenue remains high.

For the supplier ecosystem, this creates a broadening rather than a simple rotation. Nvidia-specific suppliers remain important, but ASIC and TPU ecosystems expand demand for Broadcom, Marvell, optical suppliers, memory vendors, EMS partners, cooling vendors, and power infrastructure. The best-positioned suppliers are those with content across multiple accelerator architectures. The riskiest suppliers are those whose AI exposure is highly dependent on a single hyperscaler, a single form factor, a single module standard, or a single customer’s procurement cycle.

11. Financing Structure And Capital-Market Implications

The combined transactions show that AI infrastructure is becoming a capital-stack trade. The first AI cycle was dominated by GPU scarcity and Nvidia earnings revisions. The next phase is broader: public equity raises, mandatory convertibles, strategic equity placements, private-credit SPVs, vendor residual-value support, hyperscaler commitments, cloud backlog conversion, and power procurement. This creates more investable opportunities, but it also makes the analysis more credit-sensitive. Equity investors now need to underwrite off-balance-sheet commitments, lease structures, collateral values, counterparty risk, and residual-value guarantees.

The reported Anthropic TPU financing is especially important because it treats proprietary AI accelerators as collateral. That is financially powerful but structurally fragile. Nvidia GPUs have deeper resale markets because they are widely deployable across customers and software environments. TPUs are more specialized and more dependent on Google’s hardware/software ecosystem. This means TPU residual value is likely less liquid and more path-dependent than GPU residual value. Broadcom’s reported support solves the lender confidence problem but does not eliminate the underlying economic risk; it reallocates that risk to a strategic supplier.

The structure also increases circularity risk. Alphabet is raising equity to fund AI infrastructure and global compute. Anthropic is raising private equity and preparing an IPO while securing large cloud and chip commitments. Private-credit investors are financing the chips, supported by supplier residual-value guarantees. Cloud backlog may include commitments from AI labs whose funding depends partly on the same capital-market enthusiasm that supports hyperscaler equity issuance and private credit. This is not necessarily a bubble, but it is a system with feedback loops. If end-user AI revenue and utilization continue compounding, the capital stack works. If AI monetization disappoints, the leverage and fixed commitments become more visible quickly.

The critical distinction is between productive leverage and reflexive leverage. Productive leverage finances assets that generate durable, high-utilization cash flows with strong margins. Reflexive leverage finances capacity whose demand depends on continuously rising external capital availability, customer prepayments, or speculative future usage. The Anthropic-Google-Broadcom ecosystem can fall into the first category if Anthropic’s revenue quality, enterprise retention, and utilization remain strong. It moves toward the second category if run-rate revenue is concentrated, promotional, contractually fragile, or insufficiently profitable after compute costs.

12. Key Company-Specific Implications

For Alphabet, the transaction set is strategically positive and financially dilutive. The positive case is stronger Cloud growth, TPU validation, customer backlog conversion, AI infrastructure leadership, and reduced dependence on Nvidia. The negative case is lower near-term free cash flow, higher depreciation, more debt, share issuance, and a more infrastructure-like multiple framework. Alphabet’s valuation debate should increasingly focus on AI capex ROIC, Cloud gross margin after depreciation, backlog quality, and whether AI products defend or expand Search economics.

For Broadcom, the transaction set is directly positive for revenue visibility and strategic positioning. The company becomes the most important public-market derivative of the Google TPU roadmap. However, the reported residual-value support means Broadcom should be analyzed as a semiconductor supplier with credit-enhanced ecosystem exposure. This is not necessarily negative, but it changes the risk profile. Upside comes from custom accelerator ramps, networking content, and multi-year Google/Anthropic volume. Downside comes from concentration, pricing, execution, and potential residual-value obligations if Anthropic underperforms.

For Lumentum, the transaction set is positive but indirect. The optical bandwidth requirement in multi-GW AI clusters supports lasers, photonic components, optical circuit switching, and co-packaged optics. Lumentum's recent revenue acceleration and margin expansion show that AI/cloud demand is already material. The main caveat is direct share opacity. TPU scale increases the optical TAM, but it does not automatically assign that TAM to Lumentum, and the report should not imply direct Google TPU awards without new public evidence.

For Celestica, the transaction set is positive because TPU and custom AI clusters require complex rack-scale integration, high-speed switching, liquid cooling, and optical integration. Celestica’s 2027 CPO switch ramp and hyperscaler growth align with the timing of the Google/Broadcom/Anthropic TPU capacity ramp. The caveat is that Celestica’s direct customer allocation is not disclosed, and EMS/system-integration economics are lower-margin and more execution-sensitive than semiconductor IP.

For Coherent and Fabrinet, the transactions reinforce AI optical demand and high-speed datacom manufacturing growth. Coherent has photonics and datacenter component exposure; Fabrinet has outsourced precision manufacturing exposure. Both can benefit from a multi-accelerator AI capex cycle, but both remain dependent on customer-specific allocations, pricing, and architecture-specific share. The correct framing is derivative beneficiary, not confirmed TPU supplier.

For Credo, the events reinforce the need for power-efficient, reliable high-speed connectivity. The company’s very high growth and gross margin profile make it a leveraged AI connectivity asset, but direct TPU exposure remains architecture-dependent. Content certainty is lower than Broadcom’s but thematic relevance is high.

For Marvell, the events validate the custom XPU and AI networking thesis but also highlight Broadcom’s stronger Google-specific position. Marvell should benefit from hyperscaler custom silicon broadly, particularly outside Google TPU, but the Anthropic-Google transaction is not a clean direct Marvell catalyst.

For Arista, the read-through is positive for AI Ethernet and high-performance cloud networking but ambiguous for Google TPU specifically. Google vertical integration and Broadcom’s rack-level role may reduce direct Arista participation in this stack. The broader acceleration of AI networking remains constructive.

13. Key Risks And Disconfirming Evidence

The first risk is overbuild. AI usage is currently capacity-constrained, but the industry is building multi-year infrastructure on aggressive assumptions for enterprise adoption, inference volume, pricing, and model performance. If inference optimization, open-source model improvement, smaller specialized models, price competition, or enterprise budget discipline reduce paid token growth, the sector could face excess capacity. The risk would appear first in lower utilization, lower pricing, weaker Cloud margins, renegotiated commitments, and rising depreciation burden.

The second risk is residual-value fragility. TPU-backed financing assumes that proprietary accelerators retain meaningful resale value in a stress case. That assumption is materially less robust for TPUs than for Nvidia GPUs because TPU secondary markets are narrower and more dependent on Google’s software and infrastructure stack. Broadcom’s reported residual-value support protects senior lenders but creates tail risk for Broadcom. The correct analytical approach is not to assume a loss, but to assign probability-weighted value to the contingent exposure and monitor disclosure.

The third risk is customer concentration. Alphabet Cloud backlog, Broadcom AI revenue, Lumentum AI optical growth, Celestica hyperscaler programs, Credo connectivity growth, and Fabrinet optical manufacturing can all be influenced by a small number of hyperscalers and frontier-model customers. Concentration can create strong growth when customers scale, but it also increases earnings volatility if architectures shift, procurement pauses, funding conditions tighten, or large customers demand price concessions.

The fourth risk is power and facility timing. Multi-GW commitments require grid interconnections, substations, transformers, switchgear, land, cooling, water or alternative thermal systems, construction labor, and permitting. Chips may be available before facilities are energized. Facilities may be ready before networking components arrive. Power may become the gating factor even when capital and silicon are available. This creates timing risk for revenue recognition across Broadcom, optical suppliers, EMS providers, and data-center infrastructure vendors.

The fifth risk is margin compression. AI infrastructure revenue can grow rapidly while value capture differs sharply across the stack. Broadcom can capture high-margin custom silicon and networking economics. Lumentum and Coherent can capture high-value photonics economics but face capacity and pricing dynamics. Celestica and Fabrinet can grow revenue rapidly but generally capture lower structural margins as manufacturing and integration partners. Alphabet can grow Cloud revenue quickly while depreciation and energy costs pressure incremental margins. High AI exposure is not equivalent to high economic profit.

The sixth risk is circular funding. Anthropic’s private valuation, Alphabet’s AI capex financing, private-credit TPU leases, supplier residual-value support, and cloud backlog all reinforce each other. This works if end-market AI demand is durable and profitable. It becomes risky if capital availability itself is supporting demand that has not yet proven sustainable unit economics. The eventual Anthropic S-1 will be one of the most important documents for testing this risk because it should disclose revenue quality, gross margins, compute commitments, cash burn, and customer concentration.

14. Catalysts And Data Points To Monitor

The most important catalyst is Anthropic’s public S-1, if filed publicly. The essential fields are revenue recognition, gross margin, compute expense, cloud commitments, lease obligations, related-party transactions, customer concentration, enterprise retention, deferred revenue, free cash flow, capex equivalents, and sensitivity to model-training versus inference workloads. The most important question is whether Anthropic’s $47B run-rate revenue is translating into durable gross profit after compute costs or whether revenue growth is being accompanied by escalating fixed obligations.

The public S-1 should be read through the following diligence grid.

For Alphabet, the key catalysts are final pricing and demand for the $80B capital raise, fully diluted share-count impact, Cloud backlog conversion, AI infrastructure capex updates, Cloud operating margin after depreciation, and disclosures around direct TPU hardware revenue or selected customer infrastructure sales. The central metric is not capex itself; it is incremental gross profit and operating cash flow generated per dollar of AI infrastructure invested.

For Broadcom, the key catalysts are AI revenue guidance, Google/Anthropic concentration disclosure, gross margin on custom TPU programs, working-capital intensity, accounting treatment for residual-value support, and any repetition of similar supplier-supported financing structures. Broadcom’s AI revenue trajectory appears strong, but the quality of that revenue depends on margin, cash conversion, and risk-sharing terms.

For Lumentum, Coherent, and Fabrinet, the key indicators are hyperscaler optical backlog, long-term supply agreements, customer concentration, high-speed laser capacity, OCS adoption, CPO qualification, 1.6T and 3.2T roadmap timing, and margin sustainability. For Celestica, the key indicators are hyperscaler program ramps, 2027 visibility, working capital, margin trajectory, CPO switch production, liquid-cooling content, and customer concentration. For Credo, Marvell, and Arista, the key indicators are AI connectivity content per rack, hyperscaler customer wins, Ethernet scale-out adoption, and whether TPU/ASIC ecosystems use merchant components or vertically integrated alternatives.

15. Investment Conclusion

The 3 transactions mark a structural transition in AI infrastructure. The market is moving from a simple GPU scarcity cycle to a broader AI capital-stack cycle in which public equity, private credit, strategic capital, supplier guarantees, custom ASICs, optical networking, power, cooling, and data-center execution are all central to value creation. Google’s TPU ecosystem is the largest visible beneficiary of this transition outside Nvidia. Anthropic becomes the anchor external customer, Alphabet provides the public-equity-funded infrastructure balance sheet, Broadcom provides custom silicon and ecosystem credit support, and private-credit investors provide leverage against future compute utilization.

The highest-conviction direct public-equity beneficiary is Broadcom, subject to careful underwriting of contingent residual-value exposure. The strongest derivative infrastructure beneficiaries are Lumentum, Coherent, Celestica, Fabrinet, Credo, selected memory suppliers, and power/cooling vendors, with company-specific outcomes determined by share, margin, and customer awards. Alphabet’s strategic position improves, but its equity story becomes more capital intensive and more dependent on AI infrastructure ROIC. Anthropic’s eventual public disclosures will be critical for determining whether the capital stack is financing a highly profitable software-scale business or a rapidly growing but structurally heavy infrastructure consumer.

The objective base case is constructive but not unqualified. The Anthropic-Google-Broadcom ecosystem has meaningful strategic substance: real revenue growth, multi-GW compute commitments, a validated TPU roadmap, large Cloud backlog, and a growing supplier base. The principal risk is that the same capital flows validating the ecosystem also increase leverage, fixed commitments, dilution, residual-value exposure, and overbuild risk. The revised report's core underwriting stance is therefore source-disciplined: treat company-confirmed compute, capex, and Broadcom roadmap facts as high-confidence evidence; treat Apollo/Blackstone tranche mechanics and residual-value support as reported and subject to final terms; and treat non-Broadcom supplier exposure as derivative until direct award evidence appears. The correct investment posture is to favor companies with demonstrable content, multi-year visibility, strong margin capture, and balance-sheet resilience, while discounting names where AI exposure is indirect, customer concentration is extreme, or revenue growth depends on opaque financing loops rather than verified end-user demand.