Amazon Looks to Offload $8B of Nvidia AI Chips From Its Balance Sheet

Amazon is reportedly considering transferring about $8 billion of Nvidia Blackwell AI chips into an SPV and leasing the equipment back, removing depreciating assets from its balance sheet while retaining their use in data centers. The structure highlights Big Tech's increasing use of off-balance-sheet financing to fund AI infrastructure while protecting margins and credit ratings; nine major U.S. tech firms held roughly $3 trillion in such commitments as of August. Amazon and Nvidia shares each rose about 2%, with Amazon up 10% and Nvidia up 27% year-to-date in 2026.

Analysis

The relevant equity question is not whether AMZN can reclassify hardware exposure, but whether third-party capital will accept AI-server residual-value risk at a cost below AWS's incremental return on invested capital. If financing is non-recourse and economically transfers obsolescence risk, AWS can preserve reported capital efficiency and sustain a faster deployment cadence; if Amazon retains guarantees, purchase obligations, or residual support, investors and rating agencies will likely look through the structure. Under U.S. GAAP, VIE/consolidation and lease-accounting treatment remain material gating items, so the reported balance-sheet benefit may prove narrower than the economic narrative implies.

This is incrementally constructive for NVDA only at the margin: it broadens the financing channel for demand already committed, but does not by itself establish incremental GPU consumption or pricing power. The larger second-order beneficiary is the equipment-finance/data-center ecosystem able to underwrite high-utilization GPU collateral, while lenders accepting concentrated hyperscaler and technology-obsolescence exposure bear the downside. A successful structure could lower the hurdle rate for AI capacity across hyperscalers over the next 6-18 months, extending demand for networking, power, cooling and data-center infrastructure; a failed placement would instead expose how dependent the buildout has become on off-balance-sheet leverage.

Consensus may reward apparent margin and FCF improvement too quickly. Lease expense can be smoother than owned-asset depreciation, but fixed lease commitments reduce flexibility if AI pricing falls or utilization disappoints; that is especially relevant in a 12-36 month window when Blackwell-generation assets face successor-product depreciation. The near-term catalyst is disclosure of lease duration, guarantees, residual-value support, and financing spread; widening AMZN credit spreads, a higher-than-expected fixed-charge burden, or AWS growth failing to outpace infrastructure commitments would falsify the asset-light thesis.

Key Decisions for Investors

• Maintain a neutral-to-modest long AMZN only after structure terms are disclosed; add if the SPV is demonstrably non-recourse with limited residual guarantees and management indicates no material fixed-charge drag. Reassess over 1-3 months on AWS operating-margin/FCF guidance rather than headline accounting presentation.
• Do not add directional NVDA exposure solely on this development. Treat it as a demand-financing watch signal; add only if subsequent hyperscaler capex guidance or NVDA backlog commentary shows financing expands orders beyond already installed capacity. Risk: financing substitutes for direct purchases and signals lower customer returns on AI capital.
• Monitor AMZN senior CDS/bond spreads versus MSFT and GOOGL over the next quarter. A sustained relative widening alongside larger lease/other-commitment disclosure would support a tactical AMZN underweight or AMZN/MSFT relative short; close if spreads normalize after full recourse and maturity terms are released.
• Watch listed data-center and digital-infrastructure beneficiaries such as EQIX, DLR, VRT and ETN for evidence that hyperscaler financing constraints are shifting spend toward leased capacity and power/cooling buildouts. Favor exposure only on confirmed order/backlog acceleration; these equities are vulnerable if the transaction reveals capacity is being financed because underlying GPU utilization is weaker than expected.