Why Nvidia and Broadcom Are Mobilising $535bn for AI Infrastructure

Nvidia and Broadcom are taking on a new role in the artificial-intelligence build-out: not just selling chips, but helping create the financial structures that will pay for the data centres and computing power those chips require. This week, Nvidia announced a programme to mobilise up to $500bn of third-party capital for AI infrastructure, alongside Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. Broadcom has already moved in the same direction, setting up a $35bn computing-capacity financing platform with Apollo and Blackstone in June.

The numbers matter because AI infrastructure has so far been financed largely by the giant cloud companies — the hyperscalers — which Barclays estimates have shouldered roughly $1.5tn of AI capital expenditure since 2023. The bank's analysts, Ross Sandler and Tom O'Malley, argue that model is reaching natural limits: hyperscalers can carry only so much debt and can secure only so many energy supply agreements.

Barclays expects the new "guaranteed" capital-expenditure structures to account for more than 20% of sector AI capex next year and up to half of it by 2028. The core idea is to split long-lived data-centre assets, often expected to operate for 25 to 30 years, from shorter-lived computing capacity, using separate financing vehicles. That, the analysts say, could give the largest buyers of AI compute more control over their economic models than traditional cloud-service agreements allow.

How Guaranteed Financing Splits Data Centre Longevity From Compute Risk

What the Shift Means for Nvidia and Broadcom

The financing push turns the two chipmakers from component suppliers into organisers of AI infrastructure capital. For Nvidia, the $500bn figure is a demand-side lever: by helping buyers fund clusters, it reduces the chance that customer balance sheets become the bottleneck for future chip orders. The same logic applies to Broadcom's smaller $35bn platform. The trade-off is that both companies are attaching their names and guarantee mechanisms to assets Barclays describes as the "most expensive and potentially riskiest" part of AI deployment.

Why Hyperscalers Are Hitting Natural Limits

Barclays' argument is that hyperscalers have not run out of appetite, but are approaching practical constraints on debt levels and energy-supply agreements around 2027. If buyers can now build capacity outside hyperscaler data centres, the cloud providers may lose some exclusivity over AI workloads. At the same time, they would offload a significant financial burden, so the shift is not purely adversarial: it could allow hyperscalers to keep growing services while external capital absorbs part of the asset-heavy expansion.

The Data Centre/Compute Split Changes Who Takes the Risk

The new structures separate assets with very different risk profiles. Data centres are familiar, long-lived real assets with standardised construction and operating finance. Compute capacity, however, depreciates faster and is more exposed to technological change and utilisation. By putting the two into distinct vehicles, the arrangement can attract real-estate and infrastructure investors to the data-centre layer while concentrating the more specialised risk of the computing layer in vehicles tied to chipmakers and their financing partners.

Who Stands to Gain — and Where the Control Shifts

The clearest gainers are the large compute buyers — AI labs and other capacity purchasers — that get deeper control over infrastructure without having to fund the full cost on their own balance sheets. Nvidia, Broadcom and finance partners such as Apollo, KKR and Blackstone gain fee and financing opportunities. Hyperscalers face the most ambiguity: they may cede some buyer control, but they also get relief from the capital intensity that Barclays sees as unsustainable at current pace.

What Buyers, Hyperscalers and Financiers Should Watch Next

  • Watch the SPV terms, not the headline financing size. Barclays expects guaranteed structures to amount to more than 20% of sector AI capex next year; the contracts will show whether Nvidia and Broadcom are taking first-loss, residual-value or utilisation risk on shorter-life compute assets.
  • Model 2027 as the constraint point. The bank points to debt levels and energy-supply agreements as likely limits for hyperscalers; capacity buyers and data-centre developers should use that window to assess whether buyer-owned compute becomes economical.
  • Separate data-centre and compute investments. Investors should evaluate 25- to 30-year data-centre assets differently from rapidly depreciating compute capacity, because the new financing vehicles are deliberately split along those lines.
  • Reconsider cloud-service lock-in at renewal. AI labs seeking more control over compute can compare traditional cloud agreements against direct-owned capacity structures now being scaled by Nvidia and Broadcom.

Risk & Opportunity Assessment

Commercial RiskHighNvidia and Broadcom are attaching guarantees to $500bn and $35bn compute-financing structures, while Barclays describes compute capacity as the most expensive and riskiest part of AI deployment, creating potential liabilities if utilisation or residual values fall short.
Competitive RiskHighBarclays expects guaranteed capex to exceed 20% of sector AI spending next year and reach up to half by 2028, enabling buyers to build outside hyperscaler walls and weakening cloud providers' control over AI workloads.
Regulatory RiskMediumNo regulatory action is reported, but the growth of specialised vehicles that separate long-lived data centres from shorter-life compute assets could attract scrutiny if guaranteed structures grow to half of sector capex.
Reputation RiskMediumNvidia and Broadcom are placing their names and guarantee mechanisms behind financing for compute capacity that Barclays calls the riskiest part of AI infrastructure, exposing them to reputational damage if projects underperform.
Technology DisruptionHighSplitting 25- to 30-year data-centre assets from shorter-life compute capacity and allowing small-tranche buyer-owned builds could accelerate AI infrastructure deployment beyond hyperscaler balance-sheet and energy constraints.
Commercial OpportunityHighBarclays projects guaranteed structures will represent more than 20% of sector AI capex next year and as much as half by 2028, expanding financing opportunities and supporting continued demand for Nvidia and Broadcom chips.