Confidential

Compute Futures Market

Standardized Pricing for GPU Compute

Joseph "Bear" Januszewski · Patent Pending · February 2026

The Problem

$800B+ Market. Zero Price Standardization.

$800B+

Annual Cloud GPU Spend

40%+

Price Volatility YoY

Hedging Instruments

Enterprise AI budgets are projected to exceed $1T by 2027 — yet there is no standard unit of GPU compute and no financial instruments to hedge exposure. Every provider prices differently.

The Analogy

Before WTI/Brent, Every Barrel Was Different

Compute Today

No standard unit of measure
Opaque, provider-specific pricing
Zero hedging instruments
Each provider is its own siloed market

Energy After Benchmarks

BTU and barrel as universal standards
Transparent, exchange-traded pricing
$15T+ annual derivatives volume
Hedging available to all market participants

Compute needs its WTI moment.

The Solution

Compute Units — One Number, Any GPU

CU normalizes any GPU to a single score. H100 = 1.00 (the reference point).

TCU

Training Compute Unit — weighted toward FP throughput and VRAM for large-model training

ICU

Inference Compute Unit — weighted toward memory bandwidth and latency-sensitive workloads

CU

Unified CU — harmonic mean of TCU and ICU for general pricing

Methodology

Patent-Protected Scoring System

Score = r_fp^α_fp × r_bw^α_bw × r_vram^α_vram × r_ic^α_ic

where r_i = ratio of GPU spec to H100 reference, α_i = optimized exponent

FP Throughput

α = 0.43 (TCU)
α = 0.60 (ICU)

Memory BW

α = 0.22 (TCU)
α = 0.28 (ICU)

VRAM

α = 0.55 (TCU)
α = 0.73 (ICU)

Interconnect

α = 0.11 (TCU)
α = 0.05 (ICU)

Exponents calibrated via differential evolution across 19 enterprise GPUs. TCU mean error: 6.3%. ICU mean error: 13.7%.

IP Protection

Provisional Patent Filed January 9, 2026

Coverage includes:

System for normalizing heterogeneous GPU hardware into standardized units
Immutable reference architecture (H100 SXM = 1.0)
Workload-specific categorization (TCU, ICU, unified CU)
Geometric scoring methodology with optimized exponents

Live Data

Real-Time Spot Aggregation Is Operational

Aggregating live GPU spot pricing from AWS, Azure, GCP, Lambda, CoreWeave, and others
Converting raw $/GPU-hr to normalized $/CU-hr using calibrated CU scores
Building the observable transaction dataset that IOSCO Principle 7 requires for benchmark construction

The data pipeline is live. The methodology is calibrated. The IP is protected. What's missing is the financial layer.

Index Finding

The Index Reveals Structural Pricing Gaps

When we normalize every GPU to $/CU-hr, architecture-level inefficiencies become quantifiable.

GPU	TCU	ICU	Spot $/hr	$/TCU-hr	$/ICU-hr
H100 SXM	1.00	1.00	$2.50 - $5.00	$2.50 - $5.00	$2.50 - $5.00
MI300X	2.17	2.82	$1.71 - $1.99	$0.79 - $0.92	$0.61 - $0.71
MI325X	2.61	3.60	$1.85 - $2.25	$0.71 - $0.86	$0.51 - $0.63
H200 SXM	1.48	1.67	$3.50 - $4.50	$2.36 - $3.04	$2.10 - $2.69

Inference on MI325X costs ~80% less per compute unit than H100. This is the kind of structural signal a published index makes visible to the entire market.

This is where the index gets interesting beyond just benchmarking — it reveals actual pricing inefficiencies. When you normalize everything to a single compute unit and compare spot pricing, AMD's MI300X and MI325X deliver dramatically more compute per dollar than the NVIDIA equivalents. An MI325X scores 2.61 on training and 3.60 on inference — that's 2.6x and 3.6x the compute of an H100. But it costs less than half the per-hour price. The result: inference on MI325X costs roughly 80% less per compute unit than the same workload on H100. This isn't a theoretical argument — it's what the data shows when you have a rigorous normalization methodology. I'm not making a trade recommendation. I'm pointing out that the index, once published, will make these efficiency gaps visible to every market participant — hedge funds, AI companies, infrastructure investors. Your research desk may find this kind of data compelling.

Emerging Infrastructure

AMD-Native Cloud Is Undercapitalized

A new class of AMD-only providers is emerging — pricing significantly below NVIDIA-equivalent capacity.

TensorWave

$1.71/hr

MI300X (192GB)

AMD-native infrastructure. Sales-driven. Early-stage, no public API.

Vultr

$1.85/hr

MI300X & MI325X

32 data centers globally. Full REST API. MI325X at $2.00/hr.

Hot Aisle

$1.99/hr

MI300X (192GB)

Full REST API. Swagger-documented. Bare-metal and managed.

These providers are early-stage and largely invisible to institutional capital. The index makes the compute-per-dollar gap quantifiable — and positions them as the supply side of a structurally underpriced market.

The index doesn't just reveal the efficiency gap — it points to where the infrastructure is being built. A new generation of AMD-only cloud providers has emerged — TensorWave, Vultr, Hot Aisle — pricing MI300X compute at $1.71 to $1.99 per hour. For context, comparable NVIDIA H100 capacity is $2.50 to $5.00 on the hyperscalers. These are real companies, live today, serving real customers — but they're early-stage, undercapitalized, and largely invisible to institutional capital. Once the CU index is published and compute-per-dollar becomes a standard metric, these providers become the obvious supply side of a structurally underpriced market. For Goldman's principal investment or infrastructure teams, this may be interesting. I'm not suggesting a coordinated strategy — I'm just noting what the data reveals. The index makes market structure visible. What Goldman's various desks do with that information is Goldman's business.

Market Structure

Natural Buyers and Sellers Already Exist

Short Side — Compute Providers

AWS, Azure, GCP, CoreWeave, Lambda Labs
$200B+ in depreciating GPU fleet CapEx
Need to lock in forward revenue against rapid hardware obsolescence

Long Side — Compute Consumers

OpenAI, Anthropic, Meta AI, enterprise ML teams
GPU compute is 60-80% of total operating cost
Need cost certainty for budgeting and investor commitments

Benchmark

The CBOE/VIX Model — But for Compute

1. Methodology

Patent-protected CU normalization — calibrated, reproducible, hardware-agnostic

2. Benchmark Administrator

IOSCO-compliant index publisher with governance, audit trail, and oversight

3. Financial Products

Futures, swaps, options — traded on CME/ICE against the published index

Just as VIX standardizes volatility, CU standardizes compute — enabling the same derivatives infrastructure.

Partnership

Two Ways This Can Work

Path A — Benchmark Publisher Model

License CU to a benchmark data provider — they publish the IOSCO-compliant index
Goldman accesses a ready-made benchmark for product construction
Open market — every bank trades against the same index
Active conversations underway — this path has a timeline

Path B — Goldman Exclusive License

Goldman licenses the CU methodology exclusively — first-mover in compute derivatives
Build or acquire benchmark publication capability internally
Own the index, control the product pipeline, capture all licensing revenue
Exclusivity window closes once Path A signs — decision point is now

Let me be direct about where things stand. Path A conversations are active — I'm in discussions with a benchmark data provider that already has Bloomberg terminal distribution, six million data points, and institutional backing from DRW and Jump Trading. If I sign with them, they publish the index, every bank gets access, and Goldman trades against a ready-made benchmark. Zero build cost, but zero exclusivity. Path B is the bigger play for Goldman. You license the entire CU stack exclusively, build or acquire the benchmark publication capability, and own the market. Think about what ICE did with LIBOR — owning the benchmark is owning the infrastructure that the entire market depends on. But these two paths are mutually exclusive. Once Path A signs, Path B is permanently off the table. I'm not creating artificial urgency — I'm telling you where the conversations are. If Goldman wants to explore Path B, that conversation needs to happen before Path A closes.

Timeline

Path to Market

Benchmark beta launch, historical data validation, partnership term sheets finalized

IOSCO governance framework, benchmark begins daily publication, industry advisory board

CME/ICE contract design, clearing house engagement, market maker recruitment

First listed compute futures contracts, OTC swap documentation standardized

Next Steps

30 Minutes With the Right Team

30-minute deep-dive with your commodities structuring or custom hedging team
Walk through the full CU methodology and the efficiency data — quants welcome
Discuss Path A vs. Path B licensing structure and economics
Happy to share the full calibration report and live data pipeline
Compute hedging is coming. The question is whether Goldman is building it — or trading against someone else's benchmark.

Joseph 'Bear' Januszewski

bearjanuszewski@gmail.com

Patent Pending — Filed January 9, 2026 | Confidential