ASSAY-1: The Deliverable Compute Specification

Version 0.1 — Draft for Comment

A standard for describing, verifying, and comparing compute capacity as a deliverable good.

0. Why this document exists

As of October 2026, compute becomes a financially traded commodity. CME/NYMEX will list H100 and B200 rental index futures; ICE has announced a competing suite on a rival index; a third venue has announced perpetuals on the same benchmarks.

Every one of these instruments is cash-settled against a price index. None of them delivers anything. A buyer holding a long position in a shortage receives money, not GPUs.

That is a functioning financial market sitting on top of a physical market that has no specification, no delivery standard, and no public record of what capacity actually exists.

Every mature commodity has both layers. Crude has Brent and cargo specs, delivery points, and quality differentials. Power has a forward curve and firm-vs-interruptible tiers, scheduling, and transmission rights. Compute has an index and nothing underneath it.

ASSAY-1 is the missing layer.

0.1 The core error this specification corrects

A GPU-hour is not a fungible unit, and substitutability is not a property of a fleet.

Substitutability is a relation between two fleets with respect to a workload class. Eight H100s on an oversubscribed shared Ethernet fabric and eight H100s inside an NVLink domain with a rail-optimized non-blocking InfiniBand plane are:

• identical goods for batch inference,
• different goods for large-model serving,
• not comparable goods for frontier-scale training.

The existing indexes count chips. This specification describes deliverable capacity.

0.2 What ASSAY-1 governs

1. The Record — the field set describing one fleet (§2).
2. The Headline Metrics — four derived numbers that summarize deliverability (§3).
3. The Workload Classes — the reference workloads substitutability is judged against (§4).
4. The Substitutability Grade — the A/B/C/F relation between two fleets for a class (§5).
5. The Basis — the published spread between index price and delivered effective price (§6).
6. Assurance Tiers — how much any given record should be trusted (§7).
7. The Verification Harness — the open benchmark that earns a Verified badge (§8).

1. Definitions

2. The Record

Twelve field groups. Fields marked [R] are required for a record to be publishable; [V] are required for Verified assurance tier.

A. Identity & Provenance

B. Silicon

C. Memory & Capacity

D. Scale-Up Domain

E. Scale-Out Fabric

The field group no other catalog collects, and the reason this registry can exist.

F. Storage & Data Path

G. Power & Thermal

H. Site & Grid

I. Commercial & Delivery Terms

J. Reliability & Operations

K. Compliance & Sovereignty

L. Attestation

3. The Headline Metrics

Four derived numbers. These are the registry's public vocabulary — the thing quoted in contracts and press.

3.1 CSD — Coherent Scale Domain

scale_up_domain_size, and its aggregate HBM.

Why it matters: a model whose weights plus KV cache exceed the CSD's aggregate HBM cannot be served without crossing into the scale-out fabric, which is one to two orders of magnitude slower. This is a cliff, not a slope. A fleet with CSD=8 and a fleet with CSD=72 are not the same product at any price.

3.2 NBD — Non-Blocking Domain

Largest accelerator count reachable at 1:1 bisection bandwidth.

Why it matters: a training job requesting 2,048 accelerators on a fleet with NBD=512 will spend its life in incast collapse. Advertised fleet size is meaningless above NBD.

3.3 DCB — Delivered Collective Bandwidth

Measured all-reduce bus bandwidth (GB/s) at 1 GiB message size, across the largest allocation the operator will sell, under production tenancy conditions. Measured, not derived from line rate.

Why it matters: this is the single number that separates spec-sheet fabric from working fabric. Line rate is what you bought. DCB is what you got. The ratio between them is the truth about oversubscription, congestion control, cabling quality, and noisy neighbors — all at once.

DCB is the flagship metric of this registry. Everything else is context for it.

3.4 PFI — Power Firmness Index

firm_mw / contracted_load_mw, penalized by trailing-12-month curtailment hours and short notice periods.

PFI = (firm_mw / contracted_load_mw)
      × (1 − min(curtailment_hours_12mo / 8760, 1))
      × notice_factor

where notice_factor = 1.0 if notice ≥ 60 min, 0.9 if ≥ 10 min, 0.7 if < 10 min or undisclosed.

Why it matters: as grid-flexibility contracts spread, "1,000 MW of capacity" increasingly means "600 MW firm and 400 MW the utility can take back." A buyer signing a multi-week training run against interruptible power is buying a different good than they think. PFI makes that visible and priceable.

4. Workload Classes

Substitutability is judged per class. Five classes, chosen because each has a distinct binding constraint.

4.1 Class parameters

Each class declares the workload parameters a buyer supplies:

• W1: model_weights_gb, throughput_target_tok_s
• W2: model_weights_gb, tp_width, ttft_target_ms, user_metros[]
• W3: model_weights_gb, kv_cache_gb_at_target_concurrency, tp_width, ep_width
• W4: accelerator_count_requested, checkpoint_gb, checkpoint_interval_min, run_duration_hours
• W5: accelerator_count_requested, run_duration_hours, collective_intensity (low/med/high), max_tolerable_restarts

4.2 Reference Configurations

Throughput normalization requires a canonical fleet per silicon generation. Reference Configuration = the vendor's flagship reference architecture at 1:1 non-blocking, dedicated fabric, bare metal, liquid- or air-cooled per vendor spec, PFI = 1.0.

Seed reference values — VERIFY AGAINST CURRENT VENDOR DATASHEETS BEFORE PUBLICATION:

These seed values exist so the grader can run. They are not authoritative. Publishing a wrong number once destroys a benchmark business permanently — treat the reference table as the highest-scrutiny artifact in the company.

4.3 Quality dimension (reserved)

Every workload class carries a quality_basis field, null in all v0.x records. It reserves the slot for the unit-of-work layer (quality-adjusted delivered intelligence — the kWh step) so that when quality adjustment is defined, historical series remain valid rather than being rebuilt. A benchmark that breaks its own history loses its standing permanently; this field exists so ASSAY-1 never has to.

5. The Substitutability Grade

For an ordered pair (fleet A → fleet B) and a workload class W with parameters P, the grade answers: if my job runs on A, what happens when it runs on B?

5.1 Hard gates (any one → F)

• aggregate_hbm_per_scale_up_domain_gb < required weights + KV cache (W2, W3)
• non_blocking_domain_size < accelerator_count_requested × 0.5 (W5)
• export_control_regime prohibits the buyer
• grid_status ≠ energized and required-by date < energization_date
• resale_permitted = no and is_resale = true
• accelerator_model architecture incompatible with a compiled/quantized artifact (vendor-crossing: NVIDIA ↔ AMD is never grade A for W3–W5)

5.2 Soft penalties (accumulate, then map to grade)

Score 0 → A. 1–3 → B. 4–8 → C. >8 → F.

These weights are a first calibration and should be refit against observed transaction data as soon as there is any. Publishing them openly and revising them publicly is a feature, not a weakness — it is how a benchmark earns standing.

6. The Basis

6.1 The definition

Basis(fleet, class, tenor) = Delivered Effective Price − Index Price

Published per silicon generation, per region, per workload class, per tenor. Example output:

H100 · US-West · W5 · Nov-26: basis +$0.94/accel-hr

The futures contract settles on the index. The buyer's real cost is index + basis. Nobody publishes the basis. That gap is the entire commercial opening.

6.2 Delivered Effective Price

DEP = (contract_price_per_accel_hour / throughput_ratio_vs_reference)
      + egress_adder
      + storage_adder
      + idle_on_failure_adder
      + curtailment_risk_premium
      + provisioning_carry

• throughput_ratio_vs_reference — from §5 grading against the Reference Configuration. A fleet delivering 70% of reference throughput at 90% of reference price has a higher effective price.
• idle_on_failure_adder = job_interruption_rate × mean restart cost × run length. On a long W5 run this can exceed 15% of headline price.
• curtailment_risk_premium = (1 − PFI) × expected value of lost work.
• provisioning_carry = cost of capital over provisioning_lead_time_days.

6.3 Why this is defensible

The index publishers are financial data firms. Computing DEP requires knowing what a fabric actually delivers, what a checkpoint collapse costs, and what curtailment does to a run. That is a networking and systems problem wearing a finance hat. The people building the financial layer would have to hire the expertise; the registry has to hire the finance, which is far easier to buy.

7. Assurance Tiers

Every record carries a tier. Never publish a number without its tier. This is the honest answer to "operators will lie."

Records decay: Tier 3 → Tier 1 after 90 days without re-run. Tier 2 → Tier 0 after 180 days. Publish decay openly.

8. The Verification Harness

An open-source, MIT-licensed benchmark suite an operator runs on their own fleet. Output is signed and submitted; the operator earns a Assay Verified badge.

This solves the cold-start problem. The registry does not need access to clusters it does not own. Operators run the harness because verified fleets sell faster and at better terms — the same reason merchants sought Underwriters Laboratories listing and vendors publish SPEC results. Verification is a supplier-pull, not a buyer-push.

8.1 Required probes

8.2 Anti-gaming rules

• Probes run under production tenancy, not on a drained cluster. The harness records concurrent tenant count.
• Allocation size must equal the maximum the operator will actually sell — no benchmarking a golden 64-node slice of a fleet sold in 512-node blocks.
• Results signed with a run nonce issued by the registry; results older than the nonce window are rejected.
• Refusal to publish a completed run is itself recorded. An operator who runs and withholds is flagged.

9. Publication

The sequence is deliberate and is the same one every commodity price reporting agency ran: publish free and be right → become the reference → become the settlement standard → become the venue. Skipping steps kills you; there is no shortcut from newsletter to exchange.

10. Known weaknesses of this specification

Stated plainly, because a spec that hides its soft spots does not get adopted.

1. The soft-penalty weights in §5.2 are invented. They are an engineer's prior, not a fit to data. Until there are several hundred observed transactions, grades B and C are opinion.
2. DCB is gameable at the margin despite §8.2. A determined operator can schedule the run during a quiet window. Repeated sampling is the only real defense and it costs the operator goodwill.
3. Tier 0 is most of the world. Enterprise contracts are under NDA. The registry will over-represent the transparent long tail and under-represent the hyperscalers, which is exactly backwards from where the volume is.
4. resale_permitted may be no almost everywhere. If the major suppliers contractually forbid transfer, the deliverable secondary market in §6 is confined to neoclouds and brokers.
5. The whole thesis is levered to scarcity. If accelerator supply loosens and prices flatten, hedging demand thins, basis compresses toward zero, and this becomes a small trade-press business.
6. Silicon generations churn faster than specifications. Every 12–18 months the Reference Configuration table needs rebuilding, and each rebuild breaks historical series continuity.
7. The index publishers can hire fabric engineers. The moat is a head start plus a standard that gets adopted before they move. It is a race, not a wall.

ASSAY-1 v0.1. Draft for comment. Nothing in this document is investment advice or a representation about any named company or product.

Changelog

• v0.1.1 (2026-09-10, migrated to repo) — added §4.3 reserved quality dimension, per the design requirement that workload classes survive the move from delivered capacity to delivered work without breaking historical series.
• v0.1 (2026-09) — initial draft: record, headline metrics, workload classes, substitutability grade, basis, assurance tiers, harness.