Compute Futures
A Proposal for Standardized Markets in AI Compute
Why Futures?
Futures did not start as speculation. They started because growers, millers, refiners and utilities needed to agree on prices before goods existed or were delivered. Early grain trading was bilateral and messy: every lot differed in quality, every deal had its own terms, and nobody could see a reliable market price. Exchanges fixed this by standardizing the contract itself: a fixed size (the bushel, the barrel, the megawatt-hour), a defined grade, a named delivery point or hub, and a settlement rule. Once every contract meant the same thing, buyers and sellers who had never met could trade, hedge, and read a public price curve. Electricity markets added the idea of pricing a non-storable good by time block and location.
GPU compute today resembles grain before standard grades. SKUs are fragmented, "the same H100" prices differently across providers because you are buying a bundle (fabric/topology, node composition, commercial terms), hardware ages on 24–36 month cycles, and an idle hour can never be stored or recovered. Prices are mostly bilateral and opaque, so firms either over-commit capital or take spot volatility.
What a compute futures contract would standardize
- Unit: the Effective Compute Hour (ECH), verified throughput on an agreed reference workload during a one-hour block at a defined hub, with specified cluster/fabric characteristics. It prices performance, not a nominal GPU label-hour.
- Grade: a quality ladder defined by functional characteristics rather than chip model. Grade A means ≥8 H100/B-series SXM with NVSwitch/IB ≥3.2 Tb/s aggregate, and Grade B means ≥8 PCIe with stated fabric minima.
- Location: logical hubs (US-West, US-Central, US-East), with other regions, providers and hardware trading as a basis to the anchor.
- Settlement: cash settlement to an independently governed hub index, not physical delivery.
- Price discovery: a forward ECH curve that shows where and when scarcity bites, plus regional, hardware and calendar spreads.
The paper (Standardizing Compute for Futures Markets, Draft 5, Nov 17 2025) proposes a performance-anchored ECH unit, a published Hub ECH Index computed as a VWAP of validated contributors with outlier trimming and published fallbacks under IOSCO-style governance, and listed cash-settled futures. Its illustrative contract is US-East Grade-A ECH, monthly strip, 100 ECH per contract, cash-settled to the US-East Grade-A ECH Index over the last five business days of the month.
Note: All contract terms above are illustrative in the paper. This site is a research and education project, not a trading venue.
Executive Summary
Compute is becoming the most volatile, fast-depreciating input in modern industry. The current infrastructure imbalance is defined by a massive technological risk basis—the risk that a purchased GPU is made obsolete by a newer, more efficient chip—and a financial risk basis—the volatility of spot pricing and scarcity.
Top-tier GPUs can depreciate quickly, and unused capacity cannot be stored for later. Public pricing and commercial terms also vary, creating a planning and comparison challenge for buyers and providers.
This paper explores one possible path for compute markets: standardize a performance-anchored unit (Effective Compute Hour, or ECH), publish a robust hub-based index, and list cash-settled futures. The analogy to mature commodity markets is a design reference, not a claim that compute is already a fungible commodity.
The ECH path mirrors these mature commodity and power markets, emphatically moving beyond nominal “GPU hours” to emphasize performance, time, and location; to create a unified, transparent language for hedging, financing, and policy engagement.
Bottom Line
A workable design would need to test whether pricing can move from hardware labels toward measured throughput on reference workloads, segmented by time block and location. Any financial instrument would also require independent index governance, clear quality grades, regulatory review, and evidence that remaining basis risk is manageable.
Market Context and Milestones
Historical hardware releases and market developments that informed the original October 2025 proposal. This section is context, not a live news feed.
AI Accelerator Timeline
Sources: NVIDIA GTC announcements, manufacturer datasheets. GB200 pairs a Grace CPU with 2× B200 GPUs in a single module. Vera Rubin expected 2026 per NVIDIA roadmap.
GPU vs TPU: Architecture Fundamentals
GPU (Graphics Processing Unit)
- Architecture: Thousands of small cores optimized for parallel floating-point operations
- Flexibility: General-purpose; runs any CUDA/ROCm workload
- Memory: HBM (High Bandwidth Memory) with unified memory model
- Ecosystem: Mature software stack (PyTorch, TensorFlow, JAX)
- Best for: Training large models, mixed workloads, research
TPU (Tensor Processing Unit)
- Architecture: Systolic array design optimized for matrix multiplication
- Flexibility: Purpose-built for ML; requires JAX/TensorFlow
- Memory: HBM with large interconnect fabric for pod scaling
- Ecosystem: Tightly integrated with Google Cloud
- Best for: Large-scale training, inference at Google scale
ECH normalization allows comparison across architectures by benchmarking actual throughput on reference workloads, rather than comparing theoretical FLOPS which vary in real-world efficiency.
Compute Markets Landscape
Text-only nodes shown. Brand logos will be added only after verification from each company's press page. Ornn, OneChronos, and Compute Exchange omitted pending source verification.
Related Efforts in Compute Markets
The proposal is a standardized, performance-normalized (ECH), cash-settled futures contract on an independently governed hub index. The efforts below each address part of the same problem: pricing, hedging or financing compute.
SF Compute Company (SFC)
Source →GPU clusters and an order book where customers buy time on a cluster (priced per node-hour, 1 node = 8 GPUs) and resell unused allocation. Trades settle physically and atomically on SFC's own ledger ("Abel").
Compute Labs
Source →Tokenizes ownership of physical enterprise GPUs (GNFTs for whole GPUs and miniGPU tokens for fractions) held in "GPU RWA Vaults." Pays rental revenue to holders as monthly USDC yield and sells these as Reg D products to accredited investors.
Architect (Brett Harrison)
Source →AX exchange announced GPU-rental and DRAM perpetual futures (Jan 2026), and monthly, quarterly and yearly contracts on Nvidia H100/H200 prices (May 2026). Acquired a U.S. DCM (IMX Health) to launch the "American Innovation Exchange," pending regulatory review.
Architect's IOSCO-indexed, listed pathway now overlaps two of the paper's stated differentiators (governance and regulatory path).
AWS Reserved Instance Marketplace
Source →Lets AWS customers resell the remaining term of unused Standard EC2 Reserved Instances. Sellers set only an upfront price, AWS fills buyers lowest-price-first, and AWS takes a 12% fee.
USD.AI
Source →Onchain credit protocol that makes non-recourse, GPU-collateralized loans to neoclouds and operators (up to ~80% LTV, SPV structures, insurance). Depositors mint USDai from PYUSD and stake it as sUSDai to earn loan interest plus T-bill yield.
Compute Desk (Compute Index Inc.)
Source →Architect's ComputeConnect EFP (exchange-for-physical) partner, providing the physical settlement leg for Architect's futures contracts.
Secondary source; direct company source not verified.
Challenges to Consider
A compute futures market faces real obstacles. Hardware is not fungible: fabric, memory, location and provider terms change real throughput, so any single benchmark is an imperfect cross-hedge for a specific cluster, and basis risk may be large.
Fast generational turnover (H100 → B200/GB200 → Vera Rubin) can strand a reference grade before liquidity builds.
Price discovery is thin. Most capacity trades in private, multi-year bilateral contracts, and public list prices (like the ones this site collects) are not transaction prices, which leaves an index exposed to stale data or manipulation.
Finally, it is unproven that enough hedgers on both sides want a standardized contract rather than bespoke offtake agreements.
NVIDIA Blackwell Architecture
GTC Mar 18, 2024B200 was announced with 4,500 FP16 TFLOPS (2.3x H100). The GB200 "Superchip" pairs a Grace CPU with 2x B200 GPUs, while NVL72 systems were presented as a rack-scale architecture.
Google TPU v5p & Hypercomputer
Dec 7, 2023TPU v5p: 918 TFLOPS per chip, scales to 8,960-chip pods. v5e (Aug 2023) optimized for inference. Google's Hypercomputer architecture integrates TPUs, A3 VMs (H100), and custom networking.
AMD Instinct MI300X
Dec 6, 2023192GB HBM3 (2.4x H100 memory), 1,307 TFLOPS FP16. Now available on Azure, Oracle Cloud. Meta, Microsoft deploying at scale. First serious H100 alternative in enterprise market.
Export Controls & Supply Chain
Context through Oct 2023The original research considered US BIS export controls issued in October 2022 and updated in October 2023, along with the resulting product and supply-chain responses. Consult current BIS guidance for the rules now in effect.
The Problem
The AI compute market faces critical inefficiencies that hinder both innovation and equitable access:
⚠️ Market Opacity
Lack of transparent pricing mechanisms and standardized units makes it difficult for consumers to compare offerings or plan budgets effectively.
💰 Access Barriers
Small researchers and organizations face prohibitive costs and uncertain availability, creating a two-tier system that favors well-capitalized incumbents.
📊 Investment Risk
Producers struggle to justify massive capital expenditures without forward visibility into demand and pricing, leading to boom-bust cycles.
🌍 Sustainability Concerns
No systematic mechanism exists to price in environmental or social values, resulting in suboptimal allocation of scarce computational resources.
From the Whitepaper
The current AI compute market operates without standardized units, transparent pricing mechanisms, or systematic ways to incorporate environmental and social values. This creates several critical inefficiencies:
For Consumers: Without standardized benchmarks, comparing offerings across providers is nearly impossible. A “GPU hour” from one vendor may deliver vastly different performance than from another, yet both claim similar specifications. This opacity makes budgeting difficult and creates information asymmetries that favor large, sophisticated buyers.
For Producers: Massive capital expenditures for AI infrastructure—often billions of dollars—must be made without forward visibility into demand or pricing. This leads to boom-bust cycles, with periods of chronic shortage followed by overcapacity and stranded assets.
For Society: There is no systematic mechanism to price in environmental externalities (energy consumption, water usage, carbon emissions) or to prioritize socially valuable workloads (academic research, healthcare AI, climate modeling) over purely commercial applications.
The Solution
A standardized compute futures market built on transparent benchmarks and flexible overlays
Core Innovation: The ECH Unit
We propose a standard unit of compute called the ECH (Effective Compute Hour), benchmarked against industry-standard MLPerf scores. This creates a universal language for pricing, trading, and allocating AI compute resources.
Standardization
Clear benchmarks and tiering (A/B/C) enable apples-to-apples comparisons across diverse hardware configurations
Financial Instruments
Futures, forwards, and options provide price discovery, risk management, and budget certainty for all market participants
Value Overlays
GreenCAC, OpenCAC, and sector-specific overlays embed social values directly into market mechanisms
Core Proposal
We propose establishing a standardized compute futures market built on three pillars:
1. The ECH (Effective Compute Hour) Standard Unit: A benchmarked unit of computational throughput that enables apples-to-apples comparisons across diverse hardware configurations. One ECH represents the computational capacity of a reference system (e.g., 8× NVIDIA H100 GPUs in a standard cluster configuration) for one hour, normalized by industry-standard MLPerf benchmarks.
2. Financial Instruments: Standard derivatives—futures, forwards, and options—adapted for compute markets. Cash-settled futures enable price discovery without physical delivery complexity. Forwards provide certainty for critical workloads. Options offer flexibility for variable or burst usage patterns.
3. Value-Added Overlays: Parallel markets that embed social and environmental values directly into pricing mechanisms. GreenCAC credits certify low-carbon compute. OpenCAC subsidizes research with open-source outputs. Sector-specific overlays (Health, Education) provide verified access for socially valuable applications.
Market Precedent: Oil Benchmarks
Commodity markets have long used standardized benchmarks with location-based pricing. The oil market's Brent-WTI basis spread demonstrates how delivery location affects pricing—a direct analog for compute node differentials.
WTI (West Texas Intermediate) trades at Cushing, Oklahoma, while Brent trades in the North Sea. The spread between them reflects transportation costs, regional supply/demand, and infrastructure constraints—exactly the factors that would drive compute node basis spreads.
Market Precedent: Electricity LMP
Electricity markets pioneered Locational Marginal Pricing (LMP), where prices vary by node based on transmission constraints, losses, and local demand. This is the closest existing analog to what compute markets need.
CAISO (California ISO) data shows how prices fluctuate hourly and spatially. Northern California (NP15) often trades at different prices than Southern California (SP15) due to transmission congestion—just as compute prices should vary between East Coast, Central, and West Coast nodes.
Observed GPU Pricing Data
The collected GPU cloud-pricing observations show meaningful variation across providers and regions, making like-for-like comparison and budgeting difficult.
The visualizations below demonstrate three critical perspectives on GPU pricing: raw price ranges, normalized cost-efficiency, and regional basis spreads. Together, they illustrate the research question behind standardized benchmarks such as the proposed ECH unit.
Methodology note: charts use automated observations of publicly available cloud prices. Collection dates are stored with the data. Quotes may omit negotiated discounts, commitments, taxes, availability, and other commercial terms, and should not be treated as executable prices.
💡 Use the floating control panel on the right to switch between USD/GPU and $/ECH pricing views, and filter by instance size
What the Data Suggest
- Observed public prices differ across providers and configurations
- Commitments, availability, networking, and service terms limit direct comparability
- A normalized index could make the remaining dispersion easier to analyze
- Sample public quotes vary by geography as well as hardware and provider
- Capacity, power, data residency, and latency can all contribute to regional differences
- Location could therefore be an important parameter in any standardized index
- Spot and low-priority discounts compensate users for interruption or preemption risk
- Provider reservations and commitments address some, but not all, planning needs
- Additional standardized instruments are one concept for further research
Market Overview: Latest Collected GPU Pricing
Distribution of GPU pricing across all providers, showing median, quartiles, and range for each hardware type. Illustrates overall price variability before examining individual factors.
Pricing Input Parameters: Individual Factor Analysis
GPU compute pricing is determined by multiple independent factors. Below, we isolate each parameter to understand its individual impact on cost:
🖥️ Parameter 1: Hardware Type Impact
How GPU architecture affects cost-efficiency. Older GPUs (V100) may offer better $/TFLOP for certain workloads, while newer GPUs (H100, A100) provide superior absolute performance but at premium pricing.
☁️ Parameter 2: Cloud Provider Impact
How provider choice affects pricing for identical hardware. Azure consistently commands 34-40% premiums over AWS, while GCP and CoreWeave offer competitive alternatives. Price ranges reveal provider lock-in costs.
🌍 Parameter 3: Geographic Region Impact
How datacenter location affects pricing (basis spreads). Similar to Brent-WTI oil differentials or electricity LMP nodes. Brazil commands 119% premium over US West for Azure A100, while US regions cluster within 10-15% of anchor prices.
AWS Regional Basis (A100)
Azure Regional Basis (A100)
Multi-Factor Synthesis: Combined Impact
While individual parameters show clear patterns, real-world pricing is determined by all factors simultaneously. The heatmap below synthesizes hardware type, geographic zone, and provider into a unified view, revealing that even when standardizing for these variables, significant pricing variance persists—demonstrating market fragmentation and the need for liquid futures contracts to enable efficient price discovery.
🔥 Multi-Dimensional Pricing Matrix
Hardware × Geographic Zone heatmap with normalized regions across providers. Each row represents a different GPU type (independent color scale per row), showing how mature hardware (V100, P100) exhibits consistent cross-region pricing, while newer hardware (H100, A100) shows high variance—signaling pricing dispersion across the sample.
Key Market Insights
Wide spreads between min/max prices within same GPU type reveal lack of price transparency and standardization
Cost-per-compute varies dramatically—older GPUs sometimes more cost-efficient for specific workloads
Geographic location creates significant basis spreads, similar to oil and electricity commodity markets
ECH normalization reveals true cost-efficiency, enabling apples-to-apples comparisons across hardware
Quantitative Market Analysis
Statistical analysis of observed pricing dispersion across providers and regions. Interactive visualizations and data tables illustrate the variance patterns relevant to research on standardized compute indices.
📈 Price Trends Over Time
Track how GPU pricing evolves across providers. Select different hardware types to compare pricing trajectories and identify market trends.
🔵 Cost-Efficiency by Performance (ECH Factor)
Each bubble represents a hardware SKU aggregated to median price. X-axis shows relative performance (ECH factor, H100 = 1.0); Y-axis shows cost per GPU-hour. Bubble size reflects typical cluster configuration (GPU count). Vertical spread at similar ECH factors indicates pricing variance across providers.
💱 Cross-Provider Price Dispersion
Compares like-for-like public list prices for the same hardware across providers to quantify dispersion relevant to research index design. These are catalog observations, not executable spreads.
🔄 Multi-Vendor Efficiency Comparison
Compare cost-efficiency across NVIDIA GPUs, Google TPUs, and AMD Instinct using the ECH normalization metric. Reveals which hardware offers the best value per unit of effective compute.
TPU vs GPU Efficiency Comparison
Comparing cost-efficiency across NVIDIA GPUs, Google TPUs, and AMD Instinct using the ECH (Effective Compute Hour) metric. Lower $/ECH = better value. Prices are estimated cloud on-demand rates.
NVIDIA GPUs
| Model | TFLOPS | ECH | $/hr | $/ECH |
|---|---|---|---|---|
| B200 | 4,500 | 2.27 | $6.00 | $2.64 |
| B100 | 3,500 | 1.77 | $5.00 | $2.82 |
| L40S | 733 | 0.37 | $1.20 | $3.24 |
| H100 | 1,979 | 1.00 | $3.50 | $3.50 |
| L4 | 242 | 0.12 | $0.50 | $4.10 |
| H200 | 1,979 | 1.00 | $4.20 | $4.20 |
| A100 | 624 | 0.32 | $2.00 | $6.35 |
| T4 | 65 | 0.03 | $0.35 | $10.61 |
| V100 | 125 | 0.06 | $1.20 | $19.05 |
Google TPUs
| Model | TFLOPS | ECH | $/hr | $/ECH |
|---|---|---|---|---|
| TPU v5e | 393 | 0.20 | $1.20 | $6.03 |
| TPU v5p | 918 | 0.46 | $4.20 | $9.05 |
| TPU v4 | 275 | 0.14 | $3.22 | $23.17 |
| TPU v3 | 123 | 0.06 | $2.40 | $38.71 |
| TPU v2 | 45 | 0.02 | $1.35 | $58.70 |
AMD Instinct
| Model | TFLOPS | ECH | $/hr | $/ECH |
|---|---|---|---|---|
| MI300X | 1,307 | 0.66 | $3.80 | $5.76 |
| MI250X | 383 | 0.19 | $2.50 | $12.89 |
Understanding $/ECH
ECH (Effective Compute Hour) normalizes compute power relative to H100 (ECH = 1.0). A TPU v5p with ECH 0.464 delivers ~46% of H100 performance. If it costs $4.20/hr, $/ECH = $4.20 / 0.464 = $9.05 to match one H100-hour of compute. Lower $/ECH = better cost-efficiency for equivalent work.
Detailed Data Tables
Raw statistical analysis from collected pricing data. These tables provide granular insights into provider spreads, regional variance, and pricing tier structures.
💡 Analysis Summary
Provider dispersion: Public list prices vary across providers, configurations, and commercial terms.
Regional dispersion: Location appears to be a meaningful pricing variable, but the dataset does not isolate every capacity, latency, or service-level difference.
Interruption risk: Spot and low-priority discounts indicate that preemption and commitment terms matter to pricing.
Conclusion: The observations support continued research into a normalized benchmark. They do not, by themselves, establish market readiness, product suitability, or expected economic benefits.
Market Architecture
The compute futures market is built on a layered architecture, flowing from standardized units through financial instruments to value-added overlays. Explore the interactive diagram below:
The ECH Unit
ECH = Effective Compute Hour
A standardized Effective Compute Hour (ECH), defined as performance-normalized throughput over a time block, serves the role of the bushel or barrel in commodity markets.
Formal Definition
One ECH equals delivering a minimum verified throughput on an agreed reference workload during a one-hour block at a defined hub, with specified cluster/fabric characteristics.
Where ThroughputConfig is the measured performance of the specific configuration and ThroughputBase is the reference baseline (e.g., 8× H100 SXM with NVSwitch).
Benchmarked
MLPerf training and inference scores provide objective performance metrics across diverse hardware
Tiered
Tier A (premium), Tier B (standard), Tier C (budget) classifications match different performance needs
Scalable
Standard cluster sizes (Train-8, Train-64, Train-512) support everything from research to frontier models
Quality Grades
Quality is defined by Grades to minimize quality basis risk. Each contract specifies the grade, with different grades trading at basis (premium/discount) to the anchor.
| Quality Grade | Description (Quality Ladder) | Example Hardware/Fabric |
|---|---|---|
| Grade A | High-density training cluster, emphasizing low-latency fabric | ≥8 H100/B-series SXM, NVSwitch/IB ≥3.2 Tb/s aggregate |
| Grade B | Cost-optimized training/large-scale inference cluster | ≥8 H100/B-series PCIe, IB/Ethernet with defined minima |
Each other region/provider/hardware class then trades as a basis (premium/discount) to the anchor.
ECH Technical Specification
The ECH standard addresses a fundamental challenge in AI compute markets: how to compare the computational capacity of vastly different hardware configurations. Current approaches rely on vendor-specific metrics (FLOPS, memory bandwidth, interconnect topology) that are difficult to translate across platforms.
Benchmarking Methodology: We leverage MLPerf—the industry-standard benchmarking suite developed by MLCommons—to normalize performance across hardware. MLPerf provides objective scores for both training (large language models, computer vision, recommendation systems) and inference (real-time serving, batch processing) workloads.
Tiering System: To accommodate diverse use cases, ECH units are classified into grades: Grade A represents high-density training clusters (SXM GPUs with NVLink fabric), Grade B covers cost-optimized configurations at lower cost, enabling different use cases and price points while maintaining standardized performance benchmarks.
Cluster Configurations: Standard sizes (Train-8, Train-64, Train-512) reflect common scaling patterns in distributed training, enabling standardized contracts without excessive fragmentation.
Financial Instruments
Standard financial instruments adapted for the compute market enable price discovery, risk management, and capital efficiency:
Futures (Cash-Settled)
Standardized contracts for future delivery of ECH units at predetermined prices. Cash-settled to avoid physical delivery complexity.
Forwards (Physical Delivery)
Customized contracts for actual compute resource delivery. Provides certainty for critical workloads.
Options (Priority & Burst)
Right (but not obligation) to access compute at specific prices. Ideal for variable or burst workloads.
Ancillary Services
Service Parameters
Contracts can be customized along multiple dimensions to match specific workload requirements:
Performance / Latency
Reliability / Preemptibility
Hardware Specifications
Geographic Location
Efficiency, Energy & Societal Utility
The ECH market framework provides the transparency necessary to incentivize energy efficiency and steer compute usage toward high-utility workloads.
TP$·W: The North Star Metric
Our North-star metric is Tokens per Dollar per Watt (TP$·W), the practical synthesis of cost, energy, and capability. Optimizing this aligns provider incentives (lower opex/capex), buyer incentives (lower $/token), and public goals (lower energy intensity per useful output).
TP$·W is used here as a proposed analytical shorthand. It is not presented as an industry standard, and any practical benchmark would require independent testing, governance, and stakeholder review.
TPW (Tokens per Watt)
Measures energy efficiency
TP$ (Tokens per Dollar)
Measures cost efficiency
TP$·W (Unified Metric)
The ultimate efficiency frontier
⚠️ Jevons Paradox and Market Design
Efficiency lowers the unit cost of compute, which, rather than reducing overall consumption, often unlocks massive new demand that can increase total aggregate energy consumption as delineated in the well-documented Jevons Paradox.
In compute, this means that as TP$·W improves, the marginal cost of training new, ever-larger foundation models drops, leading to a surge in total tokens consumed. To address this inevitable surge, markets must be designed to channel the throughput toward socially valuable ends and cleaner energy sources.
Mitigations via Market Design
We propose market-based solutions, primarily through optional carbon-aware hubs/tenors and peak/off-peak blocks. By aligning contract delivery periods to regional clean-energy windows (e.g., high solar/wind output), we reward clean supply without mandating technology choices. This creates a powerful financial incentive for demand response and helps internalize the carbon externality.
Regulatory Alignment
Voluntary disclosures and structure should align with FERC/RTO constructs. This is necessary because the fundamental nature of compute hinges on locational, time-sensitive, and non-storable parameters which parallels the physics and finance of electricity grid management.
Index Governance
Adoption of IOSCO-style benchmark governance is non-negotiable for the Compute Index. This framework ensures transparency, reliability, and non-manipulability, which are preconditions for widespread financial and regulatory adoption of the ECH as a primary economic indicator.
Value-Added Overlays
Overlay markets enable value-aligned compute allocation without sacrificing market efficiency:
GreenCAC
Compute hours certified for low carbon footprint and sustainable water usage. Premium pricing reflects environmental value.
OpenCAC
Subsidized compute for research that commits to open-source outputs and reproducibility. Accelerates scientific progress.
Sector-Specific CAC
Health, education, and public sector allocations managed through registrar verification to ensure equitable access.
Market Design Philosophy
The overlay system represents a key innovation in market design: embedding social and environmental values directly into pricing mechanisms without distorting core price signals.
GreenCAC operates as a premium market where providers voluntarily certify compliance with environmental standards (renewable energy usage above threshold, PUE below benchmark, water consumption limits). Third-party auditors verify claims, and premium pricing creates market incentives for sustainable infrastructure investment. This approach avoids regulatory mandates while channeling capital toward environmentally responsible capacity.
OpenCAC addresses the public goods problem in AI research. Academic institutions and researchers often generate socially valuable outputs (open-source models, published methodologies, reproducible experiments) but lack resources to compete with well-capitalized commercial entities. OpenCAC provides subsidized compute conditional on open-source commitments, funded through a mix of government grants, philanthropic capital, and cross-subsidies from commercial users.
Sector-specific overlays (HealthCAC, EduCAC) use registrar-based verification to ensure that discounted compute reaches intended beneficiaries. This model draws on established patterns in software licensing (educational discounts) while adding cryptographic verification to prevent misuse.
Stakeholder Benefits
Producers
- ▸Hedge-to-Build: Forward contracts enable capital investment with demand certainty
- ▸Utilization: Improved capacity planning reduces idle infrastructure
- ▸Price Discovery: Transparent market signals guide capacity expansion
Consumers
- ▸Budget Certainty: Fixed-price contracts eliminate cost uncertainty
- ▸Access: Standardized units lower barriers for smaller organizations
- ▸Flexibility: Options and spot markets support diverse workload patterns
Society
- ▸Price Signals: Efficient allocation of scarce computational resources
- ▸Equitable Access: Overlay markets prioritize social benefit
- ▸Sustainability: GreenCAC incentivizes environmental responsibility
Implementation Challenges
While the vision is compelling, several technical and market challenges must be addressed:
Hardware Heterogeneity
GPU architectures, memory configurations, and interconnects vary significantly. Normalizing performance requires sophisticated benchmarking and continuous recalibration.
Verification Complexity
Ensuring providers actually deliver contracted performance levels requires trusted attestation mechanisms without excessive overhead.
Liquidity Fragmentation
Too many contract variants (different tiers, locations, parameters) could split liquidity and reduce market efficiency.
Regulatory Ambiguity
Compute futures may face commodity regulations, securities laws, or entirely novel regulatory frameworks depending on jurisdiction.
Equity vs Efficiency
Market mechanisms naturally favor those with capital. Overlay systems must be carefully designed to promote access without distorting price signals.
Path Forward
The AI compute market is at an inflection point. As models grow larger and compute becomes the critical bottleneck for AI progress, we need market infrastructure that matches the sophistication of the technology itself.
A standardized futures market for AI compute can deliver price discovery, risk management, and equitable access—all while embedding sustainability and social values directly into market mechanisms.
Next Steps
- Benchmark Validation: Work with MLCommons to refine ECH standardization and tier definitions
- Pilot Exchange: Launch limited OTC market with select producers and consumers
- Overlay Prototypes: Demonstrate GreenCAC and OpenCAC with philanthropic partners
- Regulatory Engagement: Proactively work with CFTC, SEC, and international bodies
- Ecosystem Building: Recruit market makers, clearinghouses, and risk management providers
For more information, contact: