TL;DR
Nvidia’s artificial intelligence capital expenditure thesis has shifted from a story of passive hardware sales to one of active financial and technical engineering. Hyperscalers are accelerating their infrastructure outlays toward historic peaks, while Nvidia is directly underwriting its customers' physical footprints through massive off-balance-sheet guarantees and co-designing custom high-bandwidth memory to break physical supply bottlenecks.
Parabolic Cloud Spending Decouples from Near-Term Revenue
Big Tech's commitment to physical infrastructure is decoupling from near-term revenue realities as hyperscalers aggressively expand their capital outlays.
"The five largest US cloud and AI infrastructure providers – Microsoft, Alphabet, Amazon, Meta, and Oracle – have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026, nearly doubling 2025 levels." — Big Tech's AI Capex Parabolic Run
(originally sourced from Futurum Group)
This aggressive posture is driving overall hyperscaler capital expenditures toward an estimated $750 billion in 2026, representing a massive 67% year-over-year increase [big-tech-capex-reaches-710b-in-2026]. The long-term horizon is even more expansive, with Goldman Sachs projecting that the four largest hyperscalers (excluding Oracle) will deploy a combined $5.3 trillion in capital expenditures between fiscal 2025 and fiscal 2030 [big-tech-capex-reaches-710b-in-2026
]. This massive capital deployment signals that the industry is prioritizing the strategic risk of under-investing over the financial risk of over-building, effectively guaranteeing a prolonged demand runway for high-performance silicon.
What to watch: Whether the climbing capex-to-revenue ratios among the top five hyperscalers trigger a broader public market correction or if building ahead of demand remains the consensus strategy.
Financial Underwriting as a Customer Acquisition Strategy
Nvidia is shifting from a hardware supplier to a structured financial underwriter, guaranteeing the physical footprints of its largest customers to secure future orders.
"Nvidia has extended US$105bn of guarantees to SB Energy in an unprecedented deal... Under the latest agreement, if OpenAI files for insolvency or is unable to make payments to SB Energy, Nvidia will commit to paying any shortfall..." — Off-Balance-Sheet Joint Ventures
+1 (originally sourced from IFRE)
By committing to a $105 billion residual value guarantee for SB Energy's PORTS-Pike Technology Campus in Ohio, Nvidia is using its balance sheet to directly de-risk the real estate and power infrastructure that OpenAI requires [off-balance-sheet-ai-infrastructure-financing-models+1]. Nvidia is also injecting $3 billion directly into SB Energy, effectively acting as a capital allocator to construct the physical environments where its graphics processors will live [off-balance-sheet-ai-infrastructure-financing-models
+1]. This aggressive integration of finance and technology allows Nvidia to secure long-term compute commitments while keeping the debt off its own balance sheet, brushing off Wall Street's growing concerns over "circular financing" by pointing to $100 billion in contracted customer demand [off-balance-sheet-ai-infrastructure-financing-models
+1].
What to watch: The market reception of the SB Energy IPO, which will test public investor appetite for an infrastructure developer with zero operational data center revenue [off-balance-sheet-ai-infrastructure-financing-models+1].
Deeply Integrated Custom Memory Partnerships
Hardware bottlenecks are forcing a shift from off-the-shelf component sourcing to deeply integrated, custom memory co-design.
"Samsung Electronics is building a custom 8-layer version of its HBM4E memory specifically for Nvidia, according to a report from Seoul Economic Daily... The chip, described in the reporting as part of a program called NVHBM, targets pin speeds of 17-18 Gbps..." — Samsung Custom NVHBM Sourcing
(originally sourced from Shattered.io)
To resolve the high-bandwidth memory shortages that have capped graphics processor shipments, Nvidia is co-designing a custom "NVHBM" part with Samsung [samsung-custom-nvhbm-and-nvidia-memory-sourcing]. This custom NVHBM design utilizes an 8-layer stack, representing a deliberate engineering trade-off: sacrificing extreme stack capacity in exchange for higher yields, lower thermal output, and reliable volume production [samsung-custom-nvhbm-and-nvidia-memory-sourcing
]. By targeting pin speeds of 17-18 Gbps—a 20% increase over earlier samples—Nvidia is engineering a highly specialized secondary supply chain to bypass physical scaling limits and reduce its reliance on primary supplier SK Hynix [samsung-custom-nvhbm-and-nvidia-memory-sourcing
].
What to watch: How quickly Samsung can ramp production of the custom NVHBM parts to ease the broader server price hikes currently squeezing the infrastructure ecosystem.
What surprised us
- An IPO Propped Up Entirely on Promises: SB Energy filed for an IPO aiming for a $50 billion valuation despite having zero operational data center revenue [off-balance-sheet-ai-infrastructure-financing-models
- The 8-Layer Trade-off: While the industry has been pushing toward taller 12-layer and 16-layer high-bandwidth memory stacks to maximize capacity, Nvidia chose to co-design an 8-layer custom NVHBM part with Samsung [samsung-custom-nvhbm-and-nvidia-memory-sourcing
- Jensen Huang's Aggressive Defense: Rather than downplaying "circular financing" criticisms, Nvidia's CEO openly mocked them at a major Wall Street conference, telling investors, "If that is, let's do more of that" [off-balance-sheet-ai-infrastructure-financing-models