Key Takeaways
- Rental costs for Nvidia’s H100 GPUs have jumped nearly 40% from October levels, climbing from $1.70 per hour to $2.35 per hour, driven by market-wide capacity constraints.
- Supply limitations extend to Nvidia’s latest Blackwell GPU lineup, where delivery timelines now reach into the middle of 2026, contrary to predictions that newer technology would reduce demand for previous-generation hardware.
- Tech giants Alphabet, Microsoft, Meta, and Amazon are projected to allocate approximately $700 billion toward AI infrastructure investments during 2026, with Nvidia capturing 85-90% of GPU market share.
- Recent regulatory clearance for H100 chip sales in China, with multiple customer licenses granted, creates an estimated $25 billion yearly revenue stream not reflected in current company projections.
- NVDA shares trade at 15.7x forward earnings — beneath the 3-year average of 19.4x — while analyst consensus targets average $273.57, suggesting approximately 55% potential upside.
The cost trajectory for Nvidia’s H100 GPUs continues upward rather than declining, with rental rates experiencing a dramatic rise. Since October, hourly rental fees have increased nearly 40%, escalating from $1.70 to approximately $2.35, based on SemiAnalysis data compiled from surveys covering over 100 industry stakeholders.
Available GPU capacity has essentially disappeared from the market. Organizations that locked in access during earlier periods are maintaining their positions despite escalating costs. Some enterprises are turning to premium-priced spot instances through services like AWS to secure necessary computing resources.
Supply constraints aren’t confined to legacy hardware. Nvidia’s latest Blackwell GPU architecture faces identical challenges, with procurement lead times now extending to mid-2026. This contradicts initial forecasts suggesting that advanced, more efficient processors would diminish interest — and pricing pressure — for products like the H100.
The surge in requirements stems from diverse AI implementation scenarios, spanning media-generation platforms at enterprises such as ByteDance and Google to expanding adoption of models including Anthropic’s Claude.
Cloud Infrastructure Commitments Establish Baseline Revenue
Underpinning the supply shortage is a massive influx of dedicated investment capital. Alphabet, Microsoft, Meta, and Amazon are projected to deploy collectively around $700 billion toward AI infrastructure during 2026 alone. These represent confirmed capital allocation plans, not speculative forecasts.
Microsoft has indicated approximately two-thirds of infrastructure expenditure targets GPUs and CPUs. Given Nvidia’s commanding 85-90% dominance of the GPU sector, the majority of these dollars flow to Nvidia. Even assuming semiconductors represent merely 20% of comprehensive AI infrastructure expenses, this translates to over $140 billion in annual chip procurement from these four clients alone.
Nvidia’s Q4 performance delivered $68.13 billion in revenues, marking 73% year-over-year expansion, while its Q1 forecast of $78 billion surpassed analyst expectations by more than $5 billion. Revenue growth projections for fiscal 2027 currently stand at 71%.
Nevertheless, NVDA has declined roughly 6.5% year-to-date, pressured by broader macroeconomic headwinds related to energy inflation and defensive market positioning. The stock presently commands 15.7x forward earnings — trailing both its 3-year historical norm of 19.4x and AMD’s forward multiple of 18.9x, despite Nvidia’s substantially superior market position, profit margins, and expansion trajectory.
Chinese Market Access and Vera Rubin Launch Provide Additional Catalysts
A potentially significant catalyst absent from current financial models: Chinese authorities have authorized Nvidia’s H100 chip sales, with regulatory licenses granted to several clients. Wells Fargo projects this opportunity could generate $25 billion or more in annual revenue. This revenue stream wasn’t incorporated into Nvidia’s latest guidance.
Regarding product development, the forthcoming Vera Rubin platform provides ten times superior performance per watt versus Blackwell and approximately 50 times enhanced tokens per watt relative to the preceding Hopper architecture. Distribution is anticipated to commence during the second half of 2026.
Nvidia additionally completed a $2 billion equity stake in Marvell Technology (MRVL) on March 31, integrating its NVLink infrastructure with Marvell’s customized AI processors — deployed by Amazon, Alphabet, and Microsoft. NVDA appreciated more than 5% following that disclosure. Marvell surged 13%.
Wall Street maintains a Strong Buy rating consensus on NVDA, featuring 41 Buy recommendations, one Hold, and one Sell across the previous three months. The mean price target stands at $273.57.
