January 2027: Monthly average compute price of NVIDIA's H100
$0.00
1
9
January 2027: Monthly average compute price of NVIDIA's H100
$0.00
1
9
AI Analysis
Trader mode: Actionable analysis for identifying opportunities and edge
About This Event
January 2027 If the average value of NVIDIA H100 compute per hour is above X in January 2027, then the market resolves to Yes. The market resolves based on the average value of NVIDIA H100 compute per hour in January 2027, calculated as the arithmetic mean of hourly values reported by Ornn. Revisions to the underlying made after expiration will not be accounted for. The "USD" iteration of the index will be used unless explicitly stated otherwise. Values are rounded to two decimal places unless
What Prediction Markets Are Forecasting
Traders on Kalshi are betting heavily that the average hourly cost of renting NVIDIA's RTX 5090 compute will stay above $2.00 in January 2027. The market currently gives this about a 93% chance, which is roughly a 13 in 14 odds. That's a strong consensus, though not a certainty. In plain terms, if you asked fourteen traders, thirteen would say yes, the price holds above that line.
This isn't a prediction about the GPU's sticker price. It's about the rental market for GPU compute, where people pay per hour to use someone else's hardware for AI training, rendering, or other workloads. The RTX 5090, NVIDIA's flagship consumer card, has become a workhorse for small AI startups and independent researchers who can't afford data-center GPUs like the H100.
Why the Market Sees It This Way
Three forces are pushing the odds up.
First, demand for consumer-grade AI compute keeps climbing. OpenAI, Meta, and others are building massive data centers, but that doesn't trickle down to the hobbyist or small-lab market. The RTX 5090, with its 32GB of VRAM, is the sweet spot for running local models like Llama or fine-tuning smaller neural nets. As AI tools get more accessible, more people want this hardware.
Second, supply is constrained. NVIDIA's manufacturing capacity is focused on high-margin data-center chips. Consumer GPUs get leftover wafer allocation. This has been true for years, and there's no sign of change by early 2027.
Third, electricity costs are creeping up globally. Renters of GPU compute typically pay for power separately or bundled into the hourly rate. If energy prices stay elevated, that pushes the floor price higher.
Key Dates and Events to Watch
Watch NVIDIA's quarterly earnings calls. Any mention of shifting more production to consumer cards could soften prices. Also pay attention to AMD's next GPU launch, if it offers comparable VRAM at a lower price, that could pressure the rental market. And keep an eye on cryptocurrency mining demand, which sometimes soaks up consumer GPUs and tightens supply further.
How Reliable Are These Predictions?
Prediction markets have a decent track record on commodity prices, but they're not infallible. The $2.00 threshold is fairly low, which is why the odds are so high. If the market were betting on $3.00, it'd probably be closer to a coin flip. The main risk is a sudden technological shift, like a new compression technique that makes smaller GPUs viable, or a major economic downturn that slashes AI spending. Neither looks likely right now, but that's why the market leaves that 7% gap.
Current Market Outlook
Kalshi traders currently price a 74% chance that NVIDIA's H100 compute will average above $2.00 per hour in January 2027. That is a strong but not overwhelming bet. The market expects H100 pricing to hold above that threshold roughly three years from now, despite the inevitable arrival of newer GPU architectures and potential supply expansions.
A 74% probability translates to implied odds of roughly 3-to-1 in favor. The market is pricing in a base case where H100 demand remains sticky enough to sustain pricing above $2/hour through early 2027.
Key Factors Driving the Odds
The H100 launched in late 2022 at roughly $3.50 per hour on cloud instances. Prices have fallen since, but the decline has been slower than typical GPU depreciation cycles. Two structural factors explain this.
First, AI training workloads have exploded. OpenAI, Meta, and Google collectively ordered hundreds of thousands of H100s in 2023-2024. The supply chain constraints, particularly CoWoS packaging capacity at TSMC, limited total H100 output to roughly 2-3 million units through 2024. That scarcity kept cloud pricing elevated.
Second, the H100's successor, the B100 and B200 Blackwell chips, face their own production challenges. Blackwell's launch has been delayed, and initial yields are low. If Blackwell supply remains constrained through 2026, H100 pricing could stay above $2.00 as hyperscalers continue renting out older hardware to meet insatiable demand.
What Could Change These Odds
The bear case centers on two catalysts. If TSMC significantly expands CoWoS capacity by 2026, total H100-equivalent compute could flood the market. A 2025 ramp to 400,000 wafers per month would roughly double available supply. That would push spot pricing below $2.00.
The second risk is architectural substitution. If Blackwell or the rumored Rubin architecture deliver 3-4x performance per dollar, hyperscalers might retire H100s faster than expected, dumping excess capacity onto the secondary market.
The key date to watch is NVIDIA's GTC 2025 in March, where Blackwell volume shipment timelines will become clearer. If Blackwell is shipping in volume by Q3 2025, the current 74% probability looks too high. If Blackwell remains constrained through 2026, that probability is too low.
AI-generated analysis based on market data. Not financial advice.
Overview
NVIDIA's H100 Tensor Core GPU, launched in 2022, has become the dominant hardware for training and running large-scale artificial intelligence models. Its compute price per hour is a critical metric for the AI industry, reflecting supply and demand dynamics, technological shifts, and the broader economics of cloud computing. This prediction market focuses on the average price of an H100 compute hour in January 2027, as measured by the Ornn index in US dollars. The H100 is not sold directly to end users at a fixed price; instead, its compute is rented through cloud providers like AWS, Google Cloud, Microsoft Azure, and specialized GPU rental services. Prices vary based on contract length, instance type, and market conditions. During the AI boom of 2023 and 2024, demand for H100s far outstripped supply, pushing rental prices to peaks of over $4 per hour for on-demand instances. By late 2024, supply had improved, and prices began to decline, with some long-term contracts falling below $2 per hour. The January 2027 price will depend on several factors: NVIDIA's next-generation Blackwell architecture (B100/B200), which is expected to offer higher performance and could reduce demand for H100s; the overall growth of AI workloads; the pace of data center construction; and potential competition from AMD and custom chips from companies like Google and Amazon. The Ornn index aggregates hourly data from multiple providers, providing a transparent benchmark. This market is a proxy for the health and maturity of the AI hardware ecosystem, as well as the profitability of NVIDIA and its customers. Investors, AI startups, and cloud providers all have a stake in this metric, as it influences capital expenditure decisions and the cost of AI research and deployment.
Historical Context
The H100 was announced at NVIDIA's GTC conference in March 2022 and began shipping in late 2022. Its predecessor, the A100, launched in 2020 and was widely used for AI training. The A100's on-demand price on AWS started around $3.06 per hour for a single GPU instance and remained relatively stable until the AI boom in late 2022. The public release of ChatGPT in November 2022 triggered a massive surge in demand for AI compute. By early 2023, H100s were sold out for months, with lead times extending to 8-12 months. On-demand H100 prices spiked to over $4 per hour on some platforms, while reserved instances with 1-3 year commitments were priced around $2.50-$3.00 per hour. In 2024, NVIDIA ramped production, and new competitors like AMD's MI300X entered the market. Prices began to fall. By mid-2024, long-term contracts for H100s were reported at $1.50-$2.00 per hour, and on-demand prices dropped to $2.50-$3.00. The introduction of NVIDIA's H200 (a memory-upgraded H100) in late 2024 created a two-tier market, with H100s becoming the lower-cost option. The historical pattern from the A100 shows that GPU compute prices decline over time as newer architectures appear. The A100's price dropped roughly 30-40% over its 3-year lifecycle. If the H100 follows a similar trajectory, its price in January 2027 could be 50-60% below its 2023 peak, assuming the Blackwell architecture is widely available and demand growth moderates. However, the AI market is growing faster than any previous technology cycle, which could sustain higher prices for longer.
Why It Matters
The price of H100 compute is a direct input cost for virtually every major AI company, from OpenAI and Anthropic to Google and Meta. A higher price means higher costs for training and running AI models, which can slow down innovation and reduce the number of startups that can afford to compete. Conversely, falling prices democratize access to AI, enabling more research and smaller players to train state-of-the-art models. This metric also affects NVIDIA's revenue and stock price, which has become a bellwether for the tech sector. In 2024, NVIDIA's market capitalization briefly exceeded $3 trillion, making it one of the most valuable companies in the world. A sustained decline in H100 prices could signal that supply is catching up with demand, potentially reducing NVIDIA's profit margins and growth rate. For cloud providers, H100 pricing influences their own margins and competitive positioning. AWS, Azure, and Google Cloud all offer H100 instances, and their pricing strategies affect customer acquisition and retention. The broader economic impact extends to energy consumption, as H100s are power-hungry. Lower prices could lead to more widespread deployment, increasing data center energy use. Investors in AI startups, cloud infrastructure, and semiconductor companies all watch this number closely, as it provides a real-time measure of the cost of AI progress.
Educational content is AI-generated and sourced from Wikipedia. It should not be considered financial advice.
