The reported 36 billion United States dollar structured debt facility that Apollo Global Management and Blackstone are arranging to fund Google-designed Tensor Processing Units for Anthropic marks a structural break in the way artificial intelligence infrastructure is financed.
Instead of relying on highly dilutive venture equity to pay for compute as an operating expense, Anthropic is pivoting to an off balance sheet, lease-backed special purpose vehicle that treats high performance silicon as bankable collateral in a private credit format.
At the core of the structure is an SPV that borrows capital from private lenders, purchases custom Google TPUs, and leases that hardware to Anthropic on long term contracts, with lease receivables servicing the debt. The senior tranches of this capital stack, totalling approximately 31 billion, are protected by a 100 per cent residual value guarantee from Broadcom, which effectively transforms rapid chip obsolescence into corporate credit exposure to an investment-grade semiconductor designer.
The transaction sits at the intersection of AI infrastructure, private credit, custom silicon, and power-constrained data centre build out, and forms part of a broader wave of hyperscaler and AI lab capital expenditure that is projected to reach roughly 725 billion in 2026.
For ultra high net worth investors and family offices, this represents a migration channel from volatile equity exposure into senior secured private credit and infrastructure-linked yield, but with a distinctive risk matrix incorporating collateral illiquidity, technological depreciation, and concentrated counterparty exposure.
Bancara views this development as a market structure story rather than a single company event, illustrating how the funding stack of artificial intelligence is being rewired around asset-backed lending and sovereign scale capex commitments.
Executive summary
- A 36 billion private credit facility reframes AI compute from speculative operating expense into a bankable, collateralised infrastructure asset class.
- A ring fenced SPV purchases Google TPUs, leases them to Anthropic, and services multi tranche debt from contracted lease cash flows.
- Broadcom’s 100 per cent residual value guarantee on senior notes transforms rapid chip obsolescence into structured corporate credit exposure.
- Multi gigawatt deployments across United States data centres expose AI scale to grid, power, and siting constraints.
- Hyperscaler capex and private credit flows reshape global rates, portfolio construction, and family office allocation, but concentrate technological, liquidity, and geopolitical risks.
The genesis of the compute cartel
The reported Apollo and Blackstone 36 billion TPU financing package for Anthropic is not simply the largest private credit and chip financing facility on record, it is the clearest signal yet that AI compute has completed its transition from speculative operating cost to a formal infrastructure asset class. Historically, frontier AI laboratories raised enormous venture equity rounds only to recycle that capital into cloud providers as prepaid compute, locking in dilution without securing resilient balance sheet assets.
In this new architecture, compute is financed in the same manner as aircraft, pipelines, or energy transmission assets, through SPVs that own the hardware and lend against contracted cash flows. Lenders fund a ring fenced vehicle which purchases Google TPUs, leases them to Anthropic, and is repaid from lease receivables rather than from unconstrained corporate cash flow. The deal size places the transaction at sovereign scale, comparable to major energy or telecom infrastructure financings, and underscores the role of non bank credit platforms in funding digital industrialisation.
For generative AI builders, the appeal is straightforward. Using private credit instead of incremental equity rounds allows Anthropic to preserve ownership while still accessing gigawatt scale compute capacity.
The 36 billion structure therefore represents a pivot away from a software era mindset, where capex requirements were modest, and into a capital intensive regime in which the cost of silicon, power, and physical plant dominate the economics of the technology stack.
Venture capital is not disappearing, but its role is being redefined. Equity now funds research, talent, and product, while private credit is being drafted in to fund the physical layer of AI. The Apollo–Blackstone transaction sits at the apex of this shift, providing a template that other labs and hyperscalers can replicate as AI infrastructure demand continues to expand.
Bancara views this as a market structure story, not a single company event, signalling a persistent reallocation of capital from speculative equity into secured, asset backed lending anchored on AI compute.
Inside the SPV: the anatomy of a compute lease
The core financing architecture is a ring fenced Special Purpose Vehicle that isolates the hardware and lease cash flows from Anthropics corporate balance sheet. Rather than lending directly to the operating company, Apollo, Blackstone, and co investors lend to the SPV, which then acquires custom TPUs directly from Google Cloud on a staged basis using a delayed draw facility.
The flow of capital operates as follows. First, institutional lenders commit approximately 36 billion to the SPV under a multi year delayed draw structure that aligns capital deployment with chip manufacturing and data centre readiness milestones. Second, as Google delivers physical TPU clusters, the SPV calls down portions of the facility and purchases the hardware, retaining legal title to the assets. Third, the SPV leases the TPUs to Anthropic on long term contracts, with Anthropic’s lease payments pledged as the primary source of debt service.
To match different risk appetites, the SPV issues three classes of notes. The Class A1 senior secured tranche is expected to be approximately 6 billion, backed by a first priority lien on the physical TPUs and a residual value guarantee from Broadcom. The much larger Class A2 senior secured tranche of roughly 25 billion shares the same collateral and 100 per cent loss coverage guarantee, giving the combined senior stack a total size of about 31 billion.
Because these senior notes benefit from both hard asset security and Broadcom’s investment grade balance sheet, target yields are expected to trade only 1.5 to 1.75 percentage points above relevant benchmarks, broadly in line with Broadcom’s own 4 to 5 per cent corporate bond yields.
By contrast, the subordinated Class B mezzanine notes, sized at around 4.5 billion and lacking the Broadcom backstop, rely entirely on Anthropics lease payments and carry significantly higher target coupons in the region of 8 to 8.5 per cent to compensate for their first loss position.
Anthropic’s financial profile is central to this structure. The laboratory’s run rate revenue reportedly exceeded 47 billion in May 2026, representing a fivefold increase from late 2025, supported by accelerating adoption of its Claude models and enterprise agent frameworks.
In parallel, Anthropic closed a separate 65 billion equity round at a reported 965 billion post money valuation, moving it ahead of OpenAI on private market metrics and reinforcing the rationale for preserving equity capital for strategic initiatives rather than hardware purchases.
The siting of the underlying assets matters as much as their financial structure. The TPUs financed by the SPV will be deployed across hyperscale data centres in New York, Texas, Louisiana, and Indiana, using geographic diversification to tap multiple regional grids and mitigate local power or permitting bottlenecks.
The Broadcom backstop: mitigating the technological death clock
At the heart of the senior tranches is an innovative credit enhancement. Broadcom, which co designs Google’s custom TPUs and provides advanced packaging and networking components, has agreed to provide a 100 per cent residual value guarantee on the Class A1 and A2 notes. In practical terms, this means that if Anthropic were to default and the collateral could not be sold for sufficient value, Broadcom would absorb the shortfall so that senior noteholders are made whole.
This mechanism directly addresses the “technological death clock” that haunts semiconductor backed financing. AI accelerators are subject to rapid generational cycles, with Nvidia and hyperscalers now refreshing key architectures on roughly annual intervals, each delivering major improvements in performance per watt. Once newer nodes deliver meaningfully higher compute output within the same power envelope, older chips can become uneconomic to operate well before their physical end of life, effectively collapsing their residual value.
Without structural mitigants, this obsolescence risk would be unacceptable to conservative private credit allocators.
The Broadcom guarantee converts an uncertain salvage value into a defined corporate obligation, allowing senior lenders to underwrite the transaction based on Broadcom’s investment grade credit profile rather than on speculative assumptions about secondary markets for custom Google TPUs.
By backstopping the senior debt of this 36 billion SPV, Broadcom has pioneered a vital structural innovation, effectively transforming rapid technological obsolescence into structured corporate credit risk.
It gains in return a multi year revenue stream from TPU design and supply, with analysts estimating that Anthropic related AI revenue for Broadcom could reach 21 billion in 2026 and 42 billion in 2027, contributing to a corporate target of more than 100 billion in AI chip revenue by 2027.
The guarantee also allows Broadcom to support the financing without consolidating the 31 billion of senior liabilities directly onto its own balance sheet, preserving its leverage profile while still anchoring the credit quality of the deal.
For institutional allocators, the result is a hybrid structure that behaves more like secured corporate credit than a purely asset backed bet on rapidly ageing silicon.
The megawatt monopoly: siting and grid realities
The compute funded by this facility is not an abstract capacity. It is embodied in roughly 3.5 gigawatts of next generation TPU deployments beginning in 2027, distributed across data centres in four United States states.
At that scale, power, not chip supply, becomes the dominant constraint. Modern AI campuses are graduating from megawatt sites to gigawatt complexes that compete directly with heavy industry for grid access.
Global data centre construction spending is projected to reach approximately 280 billion in 2026, potentially rising to around 330 billion in 2027 as AI workloads accelerate. Much of this capex is being channelled into regions where power markets are already stretched. Texas, for example, is emerging as a focal point. The Electric Reliability Council of Texas forecasts that data centre demand could increase by about 22 gigawatts by 2031, reaching approximately 36 per cent of total state electricity demand.
In this environment, grid connectivity, transmission queues, and permitting timelines are replacing silicon availability as the primary bottlenecks for AI expansion. Interconnection to high voltage networks can take four to ten years, far slower than the cadence of model development, creating a structural mismatch between compute demand and power infrastructure.
Developers and AI operators are responding by adopting off grid, behind the meter solutions. These include on site natural gas turbine generation and hybrid arrangements that combine grid power with dedicated gas fired plants, providing resilience against congestion, curtailment, or regulatory interventions.
For Anthropic and Google, siting TPU campuses in multiple jurisdictions across New York, Texas, Louisiana, and Indiana provides not only diversification but also optionality in negotiating with utilities and regulators as load grows.
The implication for private credit allocators is that “location risk” now encompasses not just real estate and local taxation, but also grid adequacy, transmission policy, and state level attitudes toward high density compute.
Deals of this kind therefore sit at the intersection of technology finance and energy infrastructure underwriting.
Custom silicon versus Nvidia
The Apollo-Blackstone transaction crystallises a broader realignment in the semiconductor landscape. Custom AI accelerators such as Google’s Trillium and Ironwood TPUs are emerging as viable alternatives to Nvidia’s general purpose GPUs for certain workloads, particularly inference at scale. These ASICs are architected specifically for matrix multiplication, stripping away functionality that is unnecessary for neural network operations and thereby improving performance per dollar and thermal efficiency.
Nvidia’s Hopper, Blackwell, and forthcoming Rubin families remain the benchmark for highly flexible, general purpose training of frontier models, benefiting from a vast ecosystem and software tooling.
However, as models mature and the emphasis shifts from training to inference, hyperscalers are increasingly motivated to deploy custom silicon that can deliver lower marginal costs for serving workloads.
Broadcom sits at the heart of this custom silicon trend. It provides design services that translate Google’s TPU architecture into manufacturable ASIC layouts, handles advanced packaging, and supplies high speed optical interconnects that allow up to 9,216 accelerators to operate coherently within a single cluster.
The roadmap includes TPU v8 “Sunfish” training and “Zebrafish” inference architectures targeted at TSMC’s 2 nanometre node in the late 2027 timeframe, alongside other custom inference chips under discussion with partners such as Marvell.
Both Nvidia GPUs and Broadcom enabled custom ASICs ultimately depend on Taiwan Semiconductor Manufacturing Company for leading edge wafer fabrication and advanced packaging, creating a shared geopolitical and capacity bottleneck. Competition for wafer starts is intensifying, and there are indications that Google is pursuing Apple style customer owned tooling arrangements to lock in priority access to 2 nanometre and sub 2 nanometre production windows.
From a market structure perspective, custom AI silicon is unlikely to displace Nvidia entirely in the near term.
Instead, it is positioned to capture an expanding share of cost sensitive, high volume inference and specialised workloads, while Nvidia continues to dominate flexible training and heterogeneous compute environments.
The Apollo backed TPU financing thus represents a capital market expression of this technological bifurcation, with private credit underwriting the scale out of proprietary silicon inside walled garden cloud platforms.
The sovereign capex trap
The 36 billion Anthropic TPU financing sits within an unprecedented capital expenditure cycle for hyperscalers and AI labs. Aggregate capex for Microsoft, Amazon, Alphabet, and Meta is projected to approach roughly 725 billion in 2026, including substantial allocations to data centres, networking, and custom silicon. Alphabet alone is expected to invest between 175 and 190 billion in 2026, while Meta has signalled AI related infrastructure capex in the 125 to 145 billion range.
At a global level, this level of spend is comparable to the investment required to maintain worldwide oil and gas output, effectively creating a second, digital industrial complex focused on steel, silicon, and electrons.
The financing of this complex is being shared between retained earnings of megacap technology firms, public bond markets, and a rapidly expanding private credit universe.
Such sustained capex has macroeconomic consequences.
By absorbing vast quantities of capital, AI infrastructure investment acts as a structural anchor for long term interest rates, supporting higher neutral rate estimates and limiting the scope for materially lower yields absent a collapse in investment or profitability. At the same time, it channels global dollar liquidity into a relatively narrow cluster of counterparties, raising concentration risk in both equity and credit portfolios.
Public equity markets have already begun to reflect this shift. Hardware and semiconductor partners such as Broadcom and Alphabet saw their shares appreciate by approximately 1.9 and 1.2 per cent respectively in the wake of the deal’s announcement, as investors extrapolated stronger AI related revenue visibility.
In contrast, traditional enterprise software equities, as proxied by the S&P North American Technology Software Index, have suffered valuation multiple compression, with the index down about 19 per cent year to date as IT budgets are cannibalised by infrastructure commitments.
For fixed income allocators, the picture is bifurcated. Public investment grade spreads have compressed toward multi decade tights, reflecting intense demand for high quality liquid paper, while private credit yields remain compelling, particularly in junior tranches of structured deals where coupons in the high single digits are still achievable.
The Anthropic transaction exemplifies this divergence, with tight senior pricing and high yield mezzanine exposure concentrated in a single AI tenant and custom silicon ecosystem.
Re-engineering the family office portfolio
Ultra high net worth individuals and family offices sit at the confluence of these trends. Surveys such as the UBS Global Family Office Report indicate that around 65 per cent of family offices now identify AI as their top thematic investment priority, but the preferred expression of that theme is changing.
Rather than chasing late stage venture rounds or crowded public AI software trades, sophisticated allocators are deliberately rotating toward the physical and structural layers of the stack.
Family office allocations to private credit have risen materially, with the share of under allocated investors falling from 36 per cent to 26 per cent and allocations roughly doubling in key European markets.
Investors are attracted by the prospect of equity-like returns in the 10 to 12 per cent range combined with senior secured positioning and contractual cash flows, particularly in infrastructure adjacent segments such as data centres, power, and structured compute leases.
Within this context, the Anthropic TPU financing offers a vivid case study of how private wealth capital can link into the AI infrastructure build out through institutional vehicles rather than direct project ownership. Participation can occur indirectly via commitments to senior direct lending funds managed by platforms such as Apollo and Blackstone, via asset backed securities referencing data centre real estate and contracted hyperscaler leases, or via utility and power infrastructure strategies that supply long duration energy to AI campuses.
For globally aware private investors, Bancara’s multi asset lens helps frame AI infrastructure through listed equities, private credit, commodities, indices, and risk management tools, highlighting both the opportunity set and the structural correlations embedded in these exposures.
This approach emphasises capital preservation, cross asset diversification, and liquidity discipline over speculative attempts to time individual AI equity narratives.
Two risk vectors require particular attention.
- First, hidden public megacap concentration, where portfolios heavily exposed to the largest technology names may already be implicitly levered to AI capex cycles. If AI monetisation fails to justify current spending trajectories, drawdowns could simultaneously impair public holdings and private credit commitments linked to the same ecosystem.
- Second, liquidity mismatches in semi liquid private credit funds, where the promise of periodic redemptions is structurally at odds with the illiquidity of underlying assets such as compute backed loans and long dated infrastructure debt.
The systemic horizon
While the structural protections embedded in the Anthropic TPU financing are substantial, the transaction carries a non-trivial risk matrix that reaches beyond standard corporate credit considerations. At the top of the list is collateral illiquidity.
Unlike commodity Nvidia GPUs that can be resold to smaller cloud providers or research institutions, Google TPUs are tightly integrated with Google Cloud’s proprietary network and software stack, rendering their resale market almost non-existent outside that ecosystem.
Chip obsolescence risk compounds this illiquidity. As next generation architectures such as TPU v8 and subsequent 2 nanometre designs arrive, older generations risk falling below the economic viability threshold where their power consumption per unit of compute makes them uneconomic to operate, even if they remain physically functional. This dynamic underpins the emphasis on accelerated amortisation schedules and on Broadcom’s residual value guarantee for the senior tranches.
Credit and counterparty risk are heavily concentrated. Senior noteholders ultimately rely on Broadcom’s capacity and willingness to honour its 100 per cent guarantee, while mezzanine investors are exposed directly to Anthropics ability to monetise Claude and associated products at the scale implied by its 47 billion revenue run rate and near trillion dollar valuation.
A slowdown in enterprise AI adoption, pricing pressure on model APIs, or regulatory interventions affecting AI deployment could all weaken lease coverage ratios and trigger covenant stress.
Systemic factors extend the risk perimeter further. Energy and grid bottlenecks could delay data centre commissioning, leaving expensive hardware idle and pushing back lease commencement dates. Regulatory or political pushback on high density data centre clusters, including carbon pricing or targeted taxation, could increase operating costs or constrain expansion in key jurisdictions.
Geopolitical supply chain risk is also non-negligible. The advanced nodes on which both custom TPUs and Nvidia GPUs depend are concentrated in Taiwan, meaning that serious disruption in the Taiwan Strait or tighter export controls could impair the availability of replacement hardware or upgrades, affecting both the economics and timelines of AI infrastructure projects.
In a stress scenario, refinancing risk for delayed draw facilities could rise if higher neutral interest rates coincide with weaker AI revenue growth or diminished investor appetite for complex private credit structures.
For private credit allocators, the key discipline is to view this transaction not as a one off anomaly but as the leading edge of a broader structural trend, where AI compute is securitised and syndicated across global balance sheets.
Managing portfolio level exposure to this theme, monitoring early warning indicators such as GPU and TPU rental rate compression, hyperscaler capex revisions, and Broadcom’s non AI earnings trajectory, and maintaining adequate liquidity buffers will be central to preserving capital through future cycles.
Works cited
- https://aiweekly.co/alerts/apollo-backs-36b-record-debt-deal-for-anthropic-tpus
- https://www.gurufocus.com/news/8890022/debt-financing-deal-for-anthropic-pbc-involves-broadcom-and-alphabet
- https://www.investing.com/news/stock-market-news/apollo-and-blackstone-ready-record-36-billion-loan-to-power-anthropics-ai-push-93CH-4715562
- https://news.futunn.com/en/post/73804744/36-billion-largest-ever-private-credit-and-chip-financing-deal
- https://www.jdsupra.com/legalnews/debt-markets-evolve-as-data-center-boom-5417330/
- https://finimize.com/content/apollo-and-blackstone-line-up-36-billion-for-anthropic-chips
- https://www.spglobal.com/ratings/en/regulatory/article/ai-infrastructure-buildout-weighs-credit-risks-and-rewards-s101666157
- https://247wallst.com/investing/2026/05/28/despite-hitting-a-1-trillion-valuation-micron-remains-my-favorite-stock-for-2026/
- https://www.tomshardware.com/tech-industry/big-tech/big-techs-ai-spending-plans-reach-725-billion
- https://m.economictimes.com/news/company/corporate-trends/billionaires-fomo-ultra-rich-families-pouring-money-into-ai-stack/articleshow/131371258.cms
- https://www.dakota.com/resources/blog/european-family-office-allocation-trends-2026-dakota
- https://futurumgroup.com/insights/anthropics-gigawatt-scale-tpu-deal-with-broadcom-creates-a-structural-advantage/
- https://blog.citp.princeton.edu/2025/12/18/ai-chip-lifespans-a-note-on-the-secondary-market/
- https://www.stanleylaman.com/signals-and-noise/gpus-how-long-do-they-really-last
- https://www.weforum.org/stories/2026/05/electricity-data-grid-connectivity-strategic-bottleneck-ai-transformation/
- https://www.ft.com/content/fd0aec4a-50d1-4594-b489-7420bd0b4268?syn-25a6b1a6=1
- https://thenextweb.com/news/google-marvell-ai-chips-inference-tpu-broadcom
- https://manufacturingdigital.com/articles/broadcom-anthropic-google-developing-custom-ai-chips
- https://whtc.com/2026/05/28/apollo-blackstone-work-on-36-billion-debt-deal-for-anthropic-bloomberg-news-reports/
- https://www.techinasia.com/news/apollo-blackstone-shop-36b-ai-deal-for-anthropic
- https://economictimes.indiatimes.com/tech/artificial-intelligence/apollo-blackstone-work-on-36-billion-debt-deal-for-anthropic-bloomberg-news-reports/articleshow/131380476.cms
- https://www.apollo.com/insights-news/insights/2026/05/private-credit-now-in-an-etf
- https://www.spglobal.com/ratings/en/regulatory/article/ai-disruption-worries-spill-over-to-private-credit-markets-s101670132
- https://www.apollo.com/content/dam/apolloaem/documents/insights/outlook/apollo-global-2026-credit-outlook.pdf
- https://www.gurufocus.com/news/8889816/apollo-global-management-apo-and-blackstone-bx-lead-36b-debt-financing-for-anthropic
- https://www.apollo.com/strategies/asset-management/real-assets/infrastructure
- https://www.investing.com/analysis/googles-underrated-ai-weapon-200671096
- https://futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/
- https://www.csis.org/analysis/electricity-supply-bottleneck-us-ai-dominance
- https://247wallst.com/investing/2026/05/28/mark-zuckerbergs-145-billion-ai-gamble-could-spawn-metas-next-monster-business/
- https://www.ropesgray.com/en/insights/viewpoints/102mvfl/data-center-investment-in-2026-ai-demand-power-constraints-and-private-equity
- https://www.spglobal.com/ratings/en/regulatory/article/scenario-analysis-global-data-center-abs-ratings-are-holding-steady-s101685173
- https://www.fool.com/investing/2026/04/25/the-720-billion-capex-trap-2-artificial-intelligen/
- https://www.goldmansachs.com/pressroom/press-releases/2025/2025-family-office-investment-insights-report-press-release
- https://prestelandpartner.com/why-family-offices-still-overlook-private-credit.html
- https://www.diamond-hill.com/insights/a-899/articles/ais-securitized-footprint-the-rise-of-data-center-financing/
- https://www.belfercenter.org/research-analysis/ai-data-centers-us-electric-grid
