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

Nvidia generated $115.2 billion in data-center revenue in fiscal 2025 (+142%) and $193.7 billion in fiscal 2026 (+68%), while guiding for roughly 70% revenue growth in fiscal 2028. The central risk is that hyperscalers' AI servers have estimated useful lives of roughly five to six years, meaning most hardware purchased in 2024-25 will remain operational in 2028 and Nvidia's demand must come primarily from new capacity rather than replacements. Meta's 5.5-year server-life extension reduced 2025 depreciation by about $2.9 billion, while Amazon's selective reduction to five years added $1.4 billion of depreciation, underscoring how useful-life assumptions could signal either slower replacement demand or faster AI hardware obsolescence.

Analysis

The relevant variable for NVDA is not accounting useful life by itself, but whether customers can earn an acceptable inference return on prior-generation capacity while funding frontier-training clusters. Longer depreciation schedules support hyperscaler reported EPS and capex tolerance near term, yet they also reduce the urgency of replacement purchases; Amazon's treatment is more economically informative because it implies performance obsolescence can outrun physical failure. A broad useful-life extension at the next filing cycle would be a negative read-through for NVDA's 2028 revenue mix and could compress its premium growth multiple before it affects reported orders.

The underappreciated second-order risk is inference commoditization. As older GPU fleets become transferable or leasable, their marginal cost falls toward power, colocation and networking, pressuring GPU-cloud pricing and favoring operators with low-cost power, utilization software, and customer distribution rather than hardware vendors. This is selectively negative for smaller GPU lessors and cloud intermediaries, while hyperscalers can retain value through proprietary models, enterprise distribution, and internal workload aggregation. NVDA offsets this only if new architectures create sufficiently large cost-per-token improvements to induce voluntary upgrades, rather than merely higher revenue per installed gigawatt.

Consensus is likely too focused on supply constraints and too little on capex composition. The next 1-3 months should remain governed by backlog, lead times and hyperscaler capex guidance; the more material 6-18 month catalyst is whether incremental spend shifts from new GPU capacity toward networking, power, memory, custom silicon and depreciation-supported maintenance. Watch quarterly cloud capex alongside GPU utilization, inference pricing, and useful-life disclosures—not aggregate capex alone—as the falsification set for the replacement-risk thesis.

Key Decisions for Investors

• Maintain NVDA exposure only as a tactical 1-3 month long into hyperscaler capex updates; use a defined-risk collar or trim if management commentary shifts from supply limitation to utilization normalization. The thesis fails if aggregate hyperscaler capex guidance rises while NVDA forward revenue expectations stop moving higher.
• Prefer a relative-value long META / short AMZN over 3-6 months, sized modestly: META retains near-term EPS support from lower depreciation while AMZN's shorter-life assumption embeds a recurring D&A headwind. Exit if Amazon demonstrates materially superior AWS AI monetization or Meta's capex-to-revenue trajectory deteriorates.
• Set an event-driven alert for 2026-27 annual-report useful-life changes at MSFT, GOOG and META. A further extension should trigger a review of NVDA long exposure and a potential NVDA/QQQ underweight hedge; a broad shortening would instead validate accelerated refresh demand.
• Avoid chasing GPU-cloud intermediaries until there is verifiable evidence that secondary-fleet inference pricing clears above power, colocation, and financing costs. The key missing data are realized utilization, contract duration, resale values, and price-per-token trends.