Technical Specifications of Compute Contracts and Their Impact on Pricing

[OVERVIEW]

Executive Summary

GPU compute offtake and reservation deals are priced across several distinct technical and commercial terms, but only a handful of them move the price by a meaningful, well-documented amount. Chip generation and model (A100 vs. H100 vs. H200 vs. B200 vs. AMD MI300X) is the single largest lever, commonly doubling the hourly rate with each generation step. Contract duration is the second-clearest lever: reserved terms of 3–12 months typically price 25–70% below on-demand rates, provided utilization stays high. Delivery timeline and capacity priority — getting guaranteed access now versus waiting months in a supply-constrained market — carries a real premium that most list prices do not even attempt to capture.

Several other terms that sound highly technical — memory bandwidth guarantees, interconnect type in isolation, uptime SLA tiers — are directionally real cost drivers but are almost never priced as standalone line items in public market data. They are typically bundled into a chip SKU or a cluster-level rate, or negotiated bespoke in custom offtake contracts. In an earlier paper, we discussed the differentiation in compute contracts with respect to fungibility - with the goal of creating a widely-used, tradeable contract. This paper works through all terms, gives the best available price-impact estimate for each, and is explicit about which figures are backed by published market data versus which are structural, directional reasoning.

[OVERVIEW]

Methodology and Caveats

The estimates in this paper are drawn from current (October 2026) public cloud and GPU-rental market pricing — on-demand and reserved rate cards from neocloud and hyperscale providers, cluster-cost breakdowns, and industry commentary on GPU cluster economics. They are not drawn from confidential offtake or bilateral supply agreements, which are privately negotiated and not published.

A central limitation applies throughout: real-world compute contracts bundle these terms together rather than pricing each in isolation. A move from H100 to H200, for example, changes chip generation, memory capacity, and often network bandwidth simultaneously, so it is not possible to cleanly attribute the resulting price change to any one variable. Where public data allows a clean isolation (chip generation, contract duration, form factor), this paper says so explicitly. Where it does not, the paper presents directional, structurally-reasoned estimates and flags them as such. Readers negotiating an actual contract should treat every figure here as an informed planning range, not a quoted price, and should confirm current rates with providers or counterparties directly.

[SECTION 01]

The Most Common Technical and Commercial Terms

Each term below is presented with its estimated price impact and a confidence label. "Confirmed" means the estimate is backed by published market pricing data. "Directional" means the relationship is real and structurally sound but the magnitude is inferred rather than directly observed in a published rate card.

1. Chip generation / model

Estimated price impact: Largest single lever — roughly 1.5–2x per generation step

Confidence: Confirmed (published market rates)

On-demand cloud rates cluster by generation: NVIDIA A100 80GB runs roughly $1.10–2.20/GPU-hr, H100 roughly $1.90–3.50/GPU-hr, H200 141GB roughly $3.00–4.50/GPU-hr, and B200 roughly $4.00–6.50/GPU-hr. AMD's MI300X undercuts H100 by 15–25% at budget neoclouds (as low as $1.71/hr) but can run as high as $6.00/hr at hyperscalers. Moving one generation (e.g., A100 to H100) roughly doubles the hourly rate; H100 to H200 adds another 50–70%. No other single spec change comes close to this magnitude.

2. Form factor: SXM vs. PCIe

Estimated price impact: Small direct premium — roughly 5–10% for SXM

Confidence: Confirmed, modest effect

Comparable H100 listings show PCIe around $1.40/hr versus SXM around $1.50/hr — a 5–10% premium. The real cost impact is indirect: SXM is what enables NVLink for multi-GPU scaling, so the form-factor choice matters far more for cluster-level training throughput than for the sticker price of a single card.

3. Memory capacity per chip (e.g., 80GB vs. 141GB)

Estimated price impact: Confounded with generation — roughly 50–70% when bundled

Confidence: Directional — hard to isolate

Memory capacity is usually a proxy for chip generation (H100 = 80GB, H200 = 141GB), so it is difficult to price in isolation. The ~76% memory increase from H100 to H200 correlates with a 50–70% jump in hourly rate, but no provider prices memory as a standalone line item on standard SKUs — it is bundled into the generation's price.

4. Memory bandwidth guarantees

Estimated price impact: Not separately priced on standard SKUs

Confidence: Unclear — no public benchmark

No published market pricing isolates memory bandwidth as its own term. On standard cloud rentals it rides along with the chip SKU. In custom offtake or colocation deals, bandwidth guarantees are typically paired with — and priced alongside — the interconnect tier rather than negotiated as a standalone clause.

5. Interconnect: NVLink/NVSwitch vs. InfiniBand vs. Ethernet

Estimated price impact: 15–25% of cluster capex for full InfiniBand fabric

Confidence: Directional — bundled into cluster rate

Public cluster-pricing data explicitly folds interconnect cost into the overall per-GPU rate rather than breaking it out. Directionally, building a full InfiniBand fabric (typically required past 8-GPU multi-node training) adds meaningfully to capex through NICs, switches, and cabling — industry estimates put this at roughly 15–25% of total cluster build cost versus a cheaper Ethernet/RoCE fabric. In rental and offtake deals this usually surfaces as a higher rate for "networked" versus "standalone" instances rather than as its own line item.

6. Node topology / cluster size

Estimated price impact: Minimal penalty at the node level; the real cost is in fabric

Confidence: Directional

Full 8-GPU nodes price close to the per-GPU on-demand rate (an H100 node at $15–28/hr implies roughly $1.90–3.50/GPU-hr — in line with single-GPU pricing), so there is little topology premium at the single-node level. The real cost inflection happens when nodes are joined into a multi-node cluster, which is where the InfiniBand fabric premium (item 5) applies.

7. Quantity / scale (volume commitment)

Estimated price impact: Reserved tiers typically require 16+ GPU minimums; larger deals negotiate further

Confidence: Directional — deal-specific

Reserved capacity pricing tiers typically require a minimum commitment (commonly 16 or more GPUs) before they unlock at all. Beyond that threshold, commitments scaling into the hundreds or thousands of GPUs attract additional negotiated discounts, but these are deal-specific and not published in any public benchmark.

8. Performance / FLOPS guarantees and minimum utilization SLAs

Estimated price impact: Not priced as a premium — negotiated as a risk-transfer / credit mechanism

Confidence: Directional

No public rate card prices delivered-FLOPS guarantees separately. In custom deals this is typically negotiated as a contractual protection for the buyer (guaranteed delivered performance versus theoretical peak) backed by service credits or discounts if unmet, rather than as a spec that itself commands a price premium.

9. Uptime / availability SLA (e.g., 99% vs. 99.9% vs. 99.99%)

Estimated price impact: Roughly 5–15% per "nine," steepening above 99.9%

Confidence: Directional — no published price table

Sources confirm the qualitative relationship (moving from 99.9% to 99.99% "requires significantly more infrastructure investment") but none publish the exact price delta. By analogy to enterprise cloud SLA pricing generally, expect roughly a 5–15% premium per additional nine, growing steeply past 99.9% as redundant power, cooling, and failover infrastructure becomes necessary.

10. Storage attached (type and capacity)

Estimated price impact: Billed separately per GB/month; premium tiers roughly 2x budget tiers

Confidence: Confirmed, but a distinct line item

Storage is usually billed independently of GPU-hours rather than folded into chip pricing. Ratios commonly range from 2TB per GPU (large pretraining workloads) up to 12TB per GPU (multimodal workloads), with high-performance storage tiers running roughly double the cost of budget tiers. Small per-GPU, but a meaningful share of total contract value at scale.

11. Network bandwidth (e.g., 400G vs. 800G per GPU)

Estimated price impact: Bundled with interconnect tier; tracks chip generation

Confidence: Directional

No isolated public pricing exists for network bandwidth alone — it is bundled with the interconnect tier (item 5). Directionally, doubling fabric bandwidth (400G to 800G) tracks with the price jump between H100-class and H200/B200-class cluster offerings, since bandwidth requirements scale with chip generation.

12. Contract duration (on-demand vs. reserved terms)

Estimated price impact: 25–70% below on-demand for 3–12 month reserved terms

Confidence: Confirmed — best-documented lever after chip generation

This is the second clearest lever in the data. Reserved and committed terms of 3–12 months typically run 25–45% below on-demand at cluster scale; some sources cite savings of 40–70% for longer commitments. A representative H100 example: on-demand $2.50–5.00/hr versus reserved $1.00–2.00/hr. The caveat repeated across every source: these savings assume high utilization. Idle reserved capacity can erase the discount entirely, which is why duration terms are typically paired with utilization or take-or-pay clauses.

13. Deployment model (multi-tenant cloud vs. bare metal / colo vs. on-prem)

Estimated price impact: Bare metal commonly 10–20% above multi-tenant for the same chip

Confidence: Directional

Dedicated bare-metal access (no noisy-neighbor risk) typically carries a 10–20% premium over multi-tenant virtualized instances of the same chip. On-premises or colocation can be cheaper per-GPU-hour when amortized over a long enough term, if the buyer already controls the facility — effectively substituting capital expense for operating expense.

14. Data center location / region

Estimated price impact: Roughly 10–20% between low-cost and premium/restricted regions

Confidence: Directional — less standardized than other terms

A secondary factor driven by regional power and land costs, plus compliance-driven markups for data-sovereignty-restricted regions. Commonly cited in the 10–20% range between cheap regions and premium or regulation-constrained ones, though this is far less standardized across providers than chip generation or contract duration pricing.

15. Delivery timeline / capacity priority (guaranteed vs. spot/preemptible)

Estimated price impact: Spot: 40–60%+ discount with preemption risk; guaranteed near-term delivery: unpriced premium

Confidence: Confirmed direction, deal-specific magnitude

Spot or interruptible instances are dramatically cheaper — often 40–60% or more below on-demand — but carry preemption risk. Conversely, in the current supply-constrained market, guaranteed near-term delivery (capacity now versus a multi-month waitlist) commands its own premium that is not captured in any list price at all. This is currently one of the most heavily negotiated non-technical terms in real offtake deals.

[SECTION 02]

Summary Comparison Table

[SECTION 03]

Negotiation Takeaways

Three terms account for most of the real, documented price movement in a compute deal: chip generation, contract duration, and delivery timeline or capacity priority. Buyers and sellers should spend the bulk of their negotiating effort there.

Interconnect type, memory bandwidth, uptime SLA tier, and storage performance are real cost drivers but rarely appear as isolated line items — they are usually embedded in the chip SKU or the cluster-level rate. Even so, they should be specified precisely in the contract, since ambiguity here is a common source of post-signing disputes about whether delivered performance matched what was promised.

Reserved-pricing discounts (25–70% below on-demand) assume high utilization throughout the term. Buyers should pair any long-duration commitment with a realistic utilization forecast, and sellers should expect utilization or take-or-pay clauses to be a standard part of these deals — the discount is compensation for the seller's reduced flexibility, not a free lunch for the buyer.

In the current supply-constrained market (as of November 2026), guaranteed near-term delivery is itself a heavily negotiated, largely unpriced term. Buyers who need capacity now should expect to pay a premium that will not show up in any public rate card, and sellers holding available capacity have real pricing power on timing alone.