# Mercatus

> Financial intelligence for the AI compute market. Real-time GPU rental price comparison across 30+ cloud providers, LLM API token pricing, hardware specs and benchmarks, and AI infrastructure news with market-impact sentiment scoring.

## Core Products

- [GPU Index](https://www.mercatus-ai.com/gpu-index): Live GPU rental price comparison across 30+ cloud providers, updated daily.
- [Token Index](https://www.mercatus-ai.com/token-index): LLM API token-pricing tracker across major model providers.
- [Compute Atlas](https://www.mercatus-ai.com/compute-atlas): Searchable GPU and LLM model database with specs, benchmarks, and pricing.
- [News](https://www.mercatus-ai.com/news): AI infrastructure news with AI-generated market-impact sentiment scoring.
- [Model Market](https://www.mercatus-ai.com/model-market): Featured AI models with pricing and metadata.
- [Predictions](https://www.mercatus-ai.com/predictions): GPU price trend forecasts and market predictions.
- [Weekly](https://www.mercatus-ai.com/weekly): Weekly digest of AI compute market movements.

## Company

- [About](https://www.mercatus-ai.com/about)
- [Blog](https://www.mercatus-ai.com/blog)
- [Contact](https://www.mercatus-ai.com/contact)
- [Status](https://www.mercatus-ai.com/status)

## Legal

- [Privacy](https://www.mercatus-ai.com/privacy)
- [Terms](https://www.mercatus-ai.com/terms)
- [Security](https://www.mercatus-ai.com/security)
- [Compliance](https://www.mercatus-ai.com/compliance)

## GPU Model Pages

- [A100_80GB](https://www.mercatus-ai.com/gpu/a100-80gb)
- [H100](https://www.mercatus-ai.com/gpu/h100)
- [H200](https://www.mercatus-ai.com/gpu/h200)
- [B200](https://www.mercatus-ai.com/gpu/b200)
- [B300](https://www.mercatus-ai.com/gpu/b300)
- [RTX_4090](https://www.mercatus-ai.com/gpu/rtx-4090)
- [RTX_5090](https://www.mercatus-ai.com/gpu/rtx-5090)
- [RTX_PRO_6000](https://www.mercatus-ai.com/gpu/rtx-pro-6000)

## Blog Posts

- [NVIDIA H200 Price 2026: $32K GPU, $370K HGX Server](https://www.mercatus-ai.com/blog/h200-price): NVIDIA H200 costs $32K to $40K per GPU and $320K to $420K for an 8-GPU HGX server in 2026. Compare OEM, cloud, and total ownership cost.

- [DeepSeek V4 Pro API Pricing: Rates After the Increase](https://www.mercatus-ai.com/blog/deepseek-v4-pro-api-pricing): DeepSeek V4 Pro now costs $0.66 to $1.32 per million input tokens after the August 16 increase. New peak/off-peak rates, cache math, and what changed for buyers.
- [NVIDIA H100 Resale Value 2026: Used Sells at $22K–$25K](https://www.mercatus-ai.com/blog/h100-resale-value): What used NVIDIA H100s resell for in 2026 ($22K–$25K at 18–30 months), where to sell them, and how residual value swings your real cost of ownership.
- [GLM-5.2 API Pricing: The 30x Spread Just Collapsed to 3.5x](https://www.mercatus-ai.com/blog/glm-5-2-api-pricing): GLM-5.2 now costs $0.40 to $1.40 per million input tokens across 21 hosts. Two weeks ago the spread was 30x. What repriced, who got burned, and the new rates.
- [H100 vs H200: Is the Upgrade Worth the 25% Premium? (2026 Decision Guide)](https://www.mercatus-ai.com/blog/h100-vs-h200): H100 vs H200 comparison across memory, bandwidth, pricing, and inference performance. See when upgrading to H200 makes sense in 2026.
- [Hidden Cloud GPU Costs in 2026: The 7 Charges Behind the Sticker $/hr](https://www.mercatus-ai.com/blog/hidden-cloud-gpu-costs): Hidden cloud GPU costs add 30 to 50 percent to a typical managed bill. The seven charges behind the gap, with 2026 rates and a formula to price your own.
- [GPU Rental Prices: H100, H200, B200 Cost Per Hour by Provider](https://www.mercatus-ai.com/blog/gpu-rental-prices): H100 rents for $3.84 an hour on average, H200 for $4.43, B200 for $6.39, with a 4x to 6x spread between the cheapest and most expensive provider on each. Current hourly rates by provider from the Mercatus GPU Index, what an 8-GPU node costs per month, and how the hourly rate sets the floor under token prices.
- [Why GPU Prices Differ 30%+ for the Same Hardware (and What It Says About the Market)](https://www.mercatus-ai.com/blog/why-gpu-prices-differ): GPU prices can vary dramatically across providers for the same hardware. This guide explains why the spread exists, what drives the premium, and what it says about the AI compute market.
- [GPU Utilization: The Most Important Metric in AI Infrastructure (and Why Most Teams Measure It Wrong)](https://www.mercatus-ai.com/blog/gpu-utilization): GPU utilization is the biggest controllable cost lever in AI infrastructure. Learn what utilization actually means, why most teams measure it wrong, and how idle GPU capacity can now be monetized in 2026.
- [The Cheapest GPU Cloud Providers in 2026: Where AI Compute Is Actually Lowest](https://www.mercatus-ai.com/blog/cheapest-gpu-cloud): Looking for the cheapest GPU cloud providers in 2026? This guide compares H100, A100, and H200 pricing across hyperscalers, specialty GPU clouds, long-tail providers, and decentralized networks — including spot, reserved, and on-demand rates.
- [LLM API Pricing Comparison: 6 Models, Every Provider](https://www.mercatus-ai.com/blog/llm-api-pricing-comparison): Official rates, cheapest hosts, cache terms, and spreads for DeepSeek, GLM, Kimi, and Qwen, compared in one table. Updated from the Mercatus model pricing pages.
- [DeepSeek V4.1 Flash Pricing: $0.15 Input, $0.60 Output per 1M Tokens](https://www.mercatus-ai.com/blog/deepseek-v4-1-flash-api-pricing): DeepSeek V4.1 Flash costs $0.15 per million input tokens and $0.60 output off-peak, double at peak, cache hits $0.003. V4 Flash is retired; V4 Pro stays live after DeepSeek dropped its plan to route it here. Official rates, every host, and the workload math.
- [How to Lock In Token Prices: The DeepSeek 12x Case Study](https://www.mercatus-ai.com/blog/lock-in-token-prices): DeepSeek raised token prices up to 12x with ten days' notice. What locking in with a forward contract would have looked like, and how token hedging works now.
- [A100 vs H100: Should You Pay for Hopper or Stick with Ampere? (2026)](https://www.mercatus-ai.com/blog/a100-vs-h100): A100 vs H100 comparison across cost, performance, FP8 support, and real AI workloads. See when H100 justifies the premium and when A100 still makes sense in 2026.
- [DeepSeek V3.2 API Pricing: Official Rates vs the Cheapest Providers](https://www.mercatus-ai.com/blog/deepseek-v3-2-api-pricing): DeepSeek V3.2 listed at $0.28 per million input tokens before DeepSeek retired it from the official API. Fifteen third-party hosts now serve it from $0.21. Full provider comparison, cache math, and blended cost.
- [Compute Derivatives: The CFTC Just Asked How to Build Them](https://www.mercatus-ai.com/blog/cftc-compute-derivatives): The CFTC is asking how compute derivatives should work, and CME plans GPU futures for October. What the RFC says, the settlement problem, and who fills the gap.
- [Kimi K3 API Pricing: $3 Input, $15 Output, 11 Providers](https://www.mercatus-ai.com/blog/kimi-k3-api-pricing): Kimi K3 costs $3 per million input tokens and $15 output on the official API. All 11 providers, the 1.2x spread, cache math, and how it compares to DeepSeek V4.
- [Same Model, 3x the Price: Why LLM Token Pricing Varies So Much](https://www.mercatus-ai.com/blog/why-token-prices-differ): The same open-weight model costs up to 3x more depending on the provider. Where the spread comes from, what a fair token price is, and how buyers should compare.
- [DeepSeek V4 API Pricing: New Peak and Off-Peak Rates](https://www.mercatus-ai.com/blog/deepseek-v4-api-pricing): DeepSeek V4 Pro costs $0.66 to $1.32 per million input tokens since the August 16 increase. Before it, DeepSeek was the cheapest of 16 hosts. Here is the provider table now, and who wins at each hour.
- [Financing AI Compute in 2026: Buy, Lease, Loan, or Rent](https://www.mercatus-ai.com/blog/financing-ai-compute): A 100 H100 cluster runs $3.3M to $4.2M fully built. Compare the five GPU financing paths in 2026 and the formula that decides which one wins.
- [NVIDIA H100 Price 2026: $25K GPU, $285K HGX Server](https://www.mercatus-ai.com/blog/h100-gpu-cost): NVIDIA H100 costs $25K to $30K per GPU and $250K to $320K for an 8-GPU HGX server in 2026. Compare OEM, cloud, and total ownership cost.

- [The Case for Open Price Indices in AI Compute](https://www.mercatus-ai.com/blog/the-case-for-open-price-indices-in-ai-compute): AI compute is becoming a commodity. Its benchmark indices should be public goods — open, transparent, and free — just like every other mature market.
- [The 3-Year TCO of Owning 100 H100 GPUs: A Full Capital Allocation Breakdown](https://www.mercatus-ai.com/blog/100-h100-cluster-tco): A full breakdown of the 3-year total cost of owning 100 NVIDIA H100 GPUs, including hardware, power, colocation, networking, operations, depreciation, and utilization economics.
- [Qwen API Pricing: Every Qwen3.8 Model and Host](https://www.mercatus-ai.com/blog/qwen-api-pricing): Qwen3.8 Max costs $2 input and $6 output per million tokens. Its open-weight twin costs the same at every host, while Qwen3.8 27B spans 3x across 13 hosts.
- [NVIDIA B200 Server Price in 2026: What an 8-GPU HGX System Costs](https://www.mercatus-ai.com/blog/b200-server-price): An 8-GPU HGX B200 server costs $400,000 to $500,000 in 2026, if you can get an allocation. What's inside the price, the 7x cloud spread, and the math vs H200.
- [GLM-5.3 Flash API Pricing: $0.15 In, $0.50 Out per 1M Tokens](https://www.mercatus-ai.com/blog/glm-5-3-flash-api-pricing): GLM-5.3 Flash costs $0.15 per million input tokens and $0.50 output on Z.ai's API, with cache hits at $0.03. The 50% launch promo ended September 9, but 11 of 30 hosts still sell below list and three still charge the promo rate. Every host, the workload math, and what a discount with an end date means for buyers.
- [DeepSeek API Pricing: Official Rates vs Third-Party Hosts](https://www.mercatus-ai.com/blog/deepseek-api-pricing): The DeepSeek API runs two models: V4.1 Flash at $0.15 in and $0.60 out per million tokens, and V4 Pro at $0.66 in and $1.98 out, both double at peak. Every DeepSeek model, every third-party host, and what changed on August 16 and September 10.
- [GPU ROI in 2026: Payback Period, IRR, and NPV for AI Infrastructure](https://www.mercatus-ai.com/blog/gpu-roi): How institutional buyers model the return on owning AI compute. Payback period, IRR, and NPV for a 100 H100 cluster in 2026, with current pricing, formulas, and sensitivity tables.
- [US vs China Token Prices: A 7x Gap Fell to 4x in 30 Days](https://www.mercatus-ai.com/blog/us-vs-china-token-prices): US inference costs $4.56 per blended million tokens. Chinese inference costs $1.12. A month ago the gap was 7x. TPI data on the fastest convergence in AI pricing.
- [Buy vs Rent GPUs: The 2026 Decision Framework for AI Infrastructure](https://www.mercatus-ai.com/blog/buy-vs-rent-gpus): A detailed 2026 framework for deciding whether to buy or rent GPUs for AI infrastructure. Compare H100 ownership economics, reserved cloud pricing, utilization thresholds, and how capacity monetization changes the break-even point.
- [Colocation Economics for AI Compute: From $/kW/Month to GPU Cost per Hour](https://www.mercatus-ai.com/blog/colocation-economics): Colocation pricing for AI compute ranges from $80/kW/month wholesale to $300/kW/month in premium metro facilities. Here’s how colocation economics translate into real GPU cost per hour, what drives pricing variance, and where institutional AI operators can optimize infrastructure costs.
- [NVIDIA H200 Server Price in 2026: What an 8-GPU HGX System Costs](https://www.mercatus-ai.com/blog/h200-server-price): An 8-GPU HGX H200 server costs $320,000 to $420,000 in 2026. Full breakdown: what's inside the price, OEM differences, cost per GPU-hour, and buy vs rent math.
- [H100 Depreciation: How Fast NVIDIA H100s Lose Value (and What It Means for TCO)](https://www.mercatus-ai.com/blog/h100-depreciation): How fast do NVIDIA H100 GPUs lose value? This guide breaks down H100 depreciation rates, Blackwell’s impact on residual value, secondary market pricing, and how depreciation affects real AI infrastructure TCO in 2026.
- [A100 vs H100 vs H200: Complete Cost & Performance Comparison for AI Workloads (2026)](https://www.mercatus-ai.com/blog/a100-vs-h100-vs-h200):  A100 vs H100 vs H200 comparison across cost, performance, and real workloads. See which GPU makes sense for training, inference, and scale.

- [H200 Buy vs Cloud: When Does Owning H200 Make Financial Sense? (2026)](https://www.mercatus-ai.com/blog/h200-buy-vs-rent): H200 buy vs cloud comparison across utilization, fleet size, and infrastructure costs. See when owning H200 becomes cheaper than renting.
- [B200 vs H100: When to Buy Blackwell in 2026](https://www.mercatus-ai.com/blog/b200-vs-h100): B200 ships in early-cohort volumes through 2026, mostly to hyperscalers. The decision framework for when Blackwell actually beats H100 on cost in 2026.
- [Qwen3.8 Max Prime Pricing: Same Model, 2x the Price](https://www.mercatus-ai.com/blog/qwen3-8-max-prime-api-pricing): Qwen3.8 Max Prime costs exactly 2x Qwen3.8 Max on Alibaba's API for the same weights, in exchange for 1.5 to 2x the output speed. Official rates by region, what Prime is, and what the speed premium costs on a real workload.
- [Cloud GPU Pricing in 2026: What You Actually Pay Across All Major Providers](https://www.mercatus-ai.com/blog/cloud-gpu-pricing): Cloud GPU pricing varies widely across providers. This guide breaks down H100, A100, H200, reserved pricing, hidden costs, and what buyers actually pay in 2026.
- [How a Token Exchange Works: Spot, Forward, and Price Discovery for AI Inference](https://www.mercatus-ai.com/blog/how-a-token-exchange-works): How a token exchange works for AI inference: spot markets, forward contracts, order books, and public price discovery. What it changes for compute buyers locking costs and GPU operators monetizing spare capacity.
- [NVIDIA H100 Server Price in 2026: What an 8-GPU HGX System Costs](https://www.mercatus-ai.com/blog/h100-server-price): An 8-GPU HGX H100 server costs $250,000 to $320,000 in 2026. What's inside the price, what resale value does to the math, and when renting beats owning.
- [DeepSeek Peak and Off-Peak Hours: When Tokens Cost Half](https://www.mercatus-ai.com/blog/deepseek-peak-off-peak-hours): DeepSeek's peak hours converted to your time zone, the savings math ($3,131 a year on one workload), and the batch jobs worth moving. Weekends are always off-peak.