Nvidia's Toughest Competition in 2028 May Be the Chips It Already Sold

Nvidia's data-center revenue rose from $47.5B in fiscal 2024 to $115.2B in fiscal 2025 and $193.7B in fiscal 2026, while management expects roughly 70% growth in fiscal 2028. However, Microsoft, Alphabet and Meta depreciate AI servers over roughly 5.5-6 years, implying most hardware bought in the 2024-25 capex wave will remain operational through 2028 and limiting replacement demand. Amazon's decision to shorten the useful life of some AI servers to five years, adding $1.4B of 2025 D&A expense, is a counterpoint that could pull replacement demand forward; Nvidia shares trade near $230, or about 29x earnings.

Analysis

The investable issue is not accounting useful life itself but whether hyperscalers can monetize prior GPU cohorts at returns high enough to justify parallel purchases of new generations. Longer-lived fleets reduce replacement unit demand, but they also preserve scarce inference capacity and can improve cloud gross margins; this is more favorable to MSFT, GOOG and META than to NVDA if enterprise AI demand shifts from training toward cost-sensitive inference. A secondary market with transparent pricing would accelerate that transition by establishing residual values, potentially lowering customers' effective ownership cost while creating a cheaper substitute for new GPUs.

Near term (days to 3 months), useful-life disclosures are unlikely to move NVDA absent a capex-guidance revision; utilization, lead times, and hyperscaler capex remain the relevant checks. Over 6-18 months, watch for a widening gap between AI revenue growth and capex growth: that would indicate improving monetization and support continued buildouts, whereas sustained capex with weak AI revenue would make depreciation-policy changes a focal point and pressure cloud margins. AMZN is the key negative read-through if it continues shortening AI asset lives, because it implicitly raises the required return on AWS AI infrastructure and may force greater pricing discipline.

Consensus likely overstates a mechanical 2028 replacement cliff while understating architectural obsolescence. New models may require materially better memory bandwidth, networking and power efficiency, making older GPUs unsuitable for premium workloads even if functional; conversely, inference demand may be served adequately by depreciated hardware, constraining the economics of lower-end new GPU deployments. The falsifier for the bearish replacement argument is sustained hyperscaler AI capex acceleration alongside rising AI service revenue and stable cloud operating margins; the falsifier for NVDA's premium-growth case is any two consecutive quarters of capex deceleration or a material increase in disclosed idle/underutilized AI capacity.

Key Decisions for Investors

• Maintain NVDA core exposure but avoid adding on useful-life headlines alone; add only after hyperscaler capex guidance confirms acceleration. Hedge 6-12 month downside with NVDA put spreads if implied volatility is below post-earnings realized volatility, targeting protection against a 15-20% multiple reset.
• Express the monetization-over-hardware thesis as long MSFT and GOOG versus short NVDA in equal beta over 6-12 months. The pair benefits if AI shifts toward recurring software/cloud revenue and inference economics; exit if NVDA data-center growth reaccelerates while Azure and Google Cloud margins fail to expand.
• Watch AMZN's next depreciation and AWS margin disclosures rather than initiating a directional position. A further reduction in AI-server useful life or AWS margin compression despite strong revenue would support a 3-6 month AMZN underweight versus MSFT; stable margins and rising AWS backlog would invalidate it.
• Track disclosed remaining performance obligations, cloud growth, AI revenue commentary, and capex-to-revenue ratios for MSFT, GOOG, AMZN and META each quarter. Trigger a broader semiconductor de-risking only if capex growth slows materially before AI revenues demonstrate operating leverage; absent that evidence, depreciation assumptions are accounting noise rather than a standalone short signal.