Nvidia’s H100 GPU Rental Prices Are Up 40% in Six Months — and the Shortage Is Getting Worse

SemiAnalysis’s new H100 rental price index tells a story that Wall Street should be paying close attention to: AI compute demand is not slowing down, supply is running out, and Nvidia sits at the centre of it all.

Technology was supposed to get cheaper over time. Moore’s Law, commoditisation, competition — these are the forces that typically grind hardware prices downward. Yet in 2026, one of the most critical pieces of computing infrastructure on the planet is defying that logic entirely. Nvidia’s H100 GPU, a chip introduced back in 2022, is not only still in frenzied demand — its rental price has surged nearly 40% in just six months. For investors trying to understand where the AI trade is heading next, this is one of the most important data points of the year.

The SemiAnalysis Index: Real Data, Not Speculation

Semiconductor research firm SemiAnalysis has released its H100 One-Year Rental Contract Price Index — the first systematic, monthly tracker of this market, built from direct survey data gathered from over 100 cloud service providers, buyers, and sellers of compute resources. This isn’t anecdotal. It is the most rigorous price-tracking mechanism yet applied to the GPU rental market, and what it reveals is striking.

From a low of $1.70 per GPU per hour in October 2025, one-year H100 rental contracts climbed steadily through the end of the year, crossing the $2.00 threshold for the first time in late January 2026. Month-on-month increases in February and March both came in at 15–20%. By end of March, the market had settled at $2.35 per hour — a move of nearly 40% in five months. SemiAnalysis expects the pace of increases to remain elevated.

The Spot Market Has Effectively Ceased to Exist

As alarming as the contract price moves are, the spot market tells an even more urgent story. On-demand GPU rental capacity has been completely sold out across all GPU types. The twist: even as prices have risen sharply, customers who managed to secure on-demand instances are refusing to release that capacity back into the market pool. Holding compute has become a strategic asset in its own right — more valuable than the incremental cost of keeping it.

SemiAnalysis described the experience of searching for GPU compute in early 2026 as akin to booking a seat on the last flight out — high prices and virtually no availability. In practice, the analogy is even more extreme: the firm noted that trying to rent a compute cluster in this environment resembles trying to acquire a scarce commodity on short notice with willing buyers competing aggressively. Customers on AWS have reportedly fought to pay $14 per hour per GPU for Blackwell B200 spot instances — a price that would have seemed absurd just twelve months ago.

“Trying to find GPU compute in early 2026 has been like trying to book airplane tickets on the last flight out — high prices, and almost no availability.”

— SemiAnalysis, H100 Rental Price Index Report, April 2026

Why Are H100 Prices Rising When Newer Chips Exist?

This is the question that confounds conventional technology logic. Nvidia’s Blackwell architecture (H200, B200, GB300) is newer, more powerful, and more energy-efficient than the H100. Under normal circumstances, the arrival of a superior generation pushes older-generation prices downward as buyers upgrade and sellers discount legacy inventory. That is not what is happening here.

The H100 is seeing prices rise in absolute terms because demand for AI compute has completely overwhelmed available supply across every tier. New Blackwell deployments face lead times now stretching to June–July 2026, with production capacity through August–September already fully pre-booked. There is simply no slack in the system to absorb the surge in inference and training workloads. H100 and H200 contracts are being renewed at their original rates — signed two or three years ago — and in some cases operators are locking in four-year extensions through 2028. When a four-year-old chip commands multi-year forward commitments at rising prices, it says everything about the state of supply.

SemiAnalysis’s survey of suppliers found that half of all providers contacted were completely sold out of Hopper architecture GPUs (H100/H200), and most confirmed they have no contracts expiring soon that would release additional capacity into the market. Finding even a modest 8-node cluster of 64 H100s has become genuinely difficult.

What Is Driving Demand This Hard?

Multiple demand vectors are converging simultaneously, and each is self-reinforcing. The first wave is native media generation. Platforms from ByteDance, Google, and others have driven explosive growth in AI-generated video and image content, requiring enormous throughput of GPU compute just to serve existing users — let alone new ones.

The second and more structurally significant driver is the rise of multi-agent AI workloads. As AI systems move from answering single queries to executing extended autonomous tasks — coding, research, financial modelling, data analysis — the token consumption per user interaction has grown parabolically. SemiAnalysis reported that their own firm consumed billions of tokens in a single week, at a cost of roughly $5 per million tokens. AI coding assistants, in particular, are a major driver: the firm projects that tools like Claude Code could represent more than 20% of all daily code commits by end-2026.

Behind all of this is an unprecedented wave of committed capital. The four largest hyperscalers — Alphabet, Microsoft, Meta, and Amazon — are collectively planning to spend approximately $700 billion on AI infrastructure in 2026 alone. These are not projections; they are disclosed capital plans. Microsoft has stated that around two-thirds of its capital expenditure is directed at GPUs and CPUs. With Nvidia commanding 85–90% of the GPU market, and even if chips represent only 20% of total AI infrastructure costs, this implies well over $140 billion in annual chip spending from just four customers. The math is not subtle.

The Memory Crisis Adding Fuel to the Fire

The supply crunch does not stop at the GPU level. A worsening upstream shortage in memory components is amplifying the problem throughout the AI infrastructure stack. According to SemiAnalysis’s memory pricing models, LPDDR5 contract prices in the first quarter of 2026 are on track to increase approximately fourfold year-on-year, with DDR5 prices rising fivefold. These are not rounding errors — they represent a memory market in genuine crisis.

Server OEMs, facing significant gross margin pressure from these component cost spikes, have raised AI server quotations by amounts that far exceed the underlying component increases — compressing the expected returns on new compute cluster deployments and forcing some operators to slow or abandon expansion plans. The shortage has become self-perpetuating: high costs slow new deployments, which keeps supply tight, which keeps prices high.

What This Means for Nvidia as an Investment

For investors, the H100 price surge is confirmation of something that the quarterly earnings numbers have been signalling for several quarters: Nvidia is operating in a demand environment with no historical precedent. The company reported Q4 revenues of $68.13 billion — up 73% year-on-year — and guided Q1 2026 revenue to $78 billion, beating analyst consensus by more than $5 billion. Revenue growth for fiscal year 2027 is currently expected at approximately 71%. Yet despite this, NVDA shares are down around 6.5% year-to-date, dragged by broader macro concerns around energy inflation and risk-off sentiment tied to geopolitical uncertainty.

The valuation picture is interesting. NVDA currently trades at approximately 15.7 times forward earnings — notably below its three-year average multiple of 19.4 times. Analyst consensus carries a price target of around $273.57, implying approximately 55% potential upside from recent levels. The combination of compressed multiple, accelerating revenue growth, and a supply-constrained market in its primary product is a rare alignment that patient investors historically find attractive.

On the product roadmap, Nvidia’s forthcoming Vera Rubin platform — initial shipments expected in the second half of 2026 — is projected to deliver ten times the performance-per-watt improvement over Blackwell, and roughly fifty times the token efficiency of the Hopper generation. This matters because it means the compute economics for AI applications continue to improve dramatically at the hardware level, even as rental prices rise. The bottleneck today is physical supply, not architectural capability.

Three Variables to Watch

SemiAnalysis identifies three key factors that will determine whether GPU rental prices continue their ascent or finally begin to stabilise. First is the pace of GB300 cluster deployment throughout 2026: whether new supply additions can outrun the growth in token demand will determine whether the current crunch eases or tightens further. Second is the semiconductor supply chain — particularly TSMC’s N3 advanced process node, which is under significant pressure, along with HBM and DRAM/NAND memory capacity. Any execution hiccup in these manufacturing processes can materially worsen the shortage. Third is the rate of AI adoption and the resulting growth in annual recurring revenue across major AI labs: the faster enterprise adoption accelerates, the more token consumption grows, and the more compute the market will need.

SemiAnalysis’s assessment, having surveyed over 100 market participants and tracked this market for months, is direct: GPU rental prices are highly likely to continue rising under current conditions. For investors, that is not a warning — it is a signal.