Compute surfaces · 2026-09-06
NVIDIA is the chip clerk. SSDF is not a landlord.
NVIDIA is the chip clerk. BIS is the export clerk. The neocloud is a landlord. SSDF publishes a machine-readable atlas of that surface — take-or-pay vs on-demand vs spot vs hyperscaler list.
Map the clerks. Credential the purchase. Do not rent GPUs, do not become an NVIDIA Cloud Partner, and do not remarket hours. That is the title-file rule on a different clerk.
- SSDF is a Florida C-Corp merchant of record. SSDF does not rent GPUs, does not operate a neocloud, and is not an NVIDIA Cloud Partner.
- NVIDIA is the chip clerk. BIS is the export clerk (EAR ECCN 3A090 and related .z / 4A090). The neocloud is a landlord. SSDF is none of those.
- A Passport attests that a buyer paid this atlas. It is not a cluster reservation, not a take-or-pay hour, and not a Cloud Partner seat.
- Published $/GPU-hr figures go stale. Re-verify the provider rate card. This file is a map, not a quote.
- Remote GPU access is still an export-control surface (cloud loophole / Remote Access Security Act). Mapping the rule is not a license.
- This SKU is commercial public process. It is not CUI, not ITAR technical data, and not a DARPA deliverable.
- AMAIMA and Genesis Engine do not run on this store. Hosted inference and live CLOB stay on their own ledgers.
NVIDIA (chip clerk)
chip clerk
H100 / H200 / B200 / GB200 NVL72 sit above the 3A090.a performance line. SSDF does not become a Cloud Partner by selling a map. The agentic-commerce blueprint is card checkout, not x402, and is not this atlas.
Not a public $/GPU-hr card for end buyers. Silicon is sold to landlords.
Cloud Partner is a nominated IaaS landlord. Inception is a free startup list. DGX Cloud is NVIDIA's own hosted stack. GPU revenue-share is for neoclouds, not file merchants.
BIS / EAR 3A090 (export clerk)
export clerk
The House Remote Access Security Act (369-22) targets the cloud-rental loophole: a restricted party renting a permitted-country GPU instead of importing the card. DOJ NSD Corporate Enforcement Policy (Mar 2026) adds voluntary-disclosure pressure. This atlas is not a classification opinion.
Not a rental market. A license is not a GPU-hour.
Worldwide licensing on 3A090.a advanced accelerators and systems (4A090 / .z). January 2026 H200 / MI325X China-Macau path is still a license, not free trade. IaaS used to train advanced models outside Tier 1 can be prohibited.
CoreWeave
neocloud
The enterprise-credible neocloud: InfiniBand, Kubernetes, contracts procurement will sign. SSDF does not resell CoreWeave hours.
Premium. MarkTechPost 21 Aug 2026: HGX H100 ~$6.16/GPU-hr node-implied; only public GB200 NVL72 list among the set (~$10.50/GPU). Re-verify.
Mostly 2–5 year take-or-pay. Backlog reported ~$99B as of 31 Mar 2026. ClusterMAX 2.0 Platinum. Customer concentration (Microsoft historically, plus OpenAI / Meta / Anthropic) is the risk, not a SKU.
Lambda
neocloud
Self-serve on-demand is the opposite of CoreWeave take-or-pay. Still a landlord. Not SSDF.
Often the published low print on B200 (~$6.69/GPU-hr on 21 Aug 2026 MarkTechPost). H100 SXM ~$3.99. Re-verify.
On-demand and reserved hourly, plus private cloud. Developer UX, academic penetration.
Crusoe
neocloud
Energy clerk is the differentiator. SSDF maps it. SSDF does not site a campus.
H100 HGX ~$3.90 on-demand (fixed, not auction) as of early Sep 2026 writeups. AMD MI300X / MI355X on the rate card. Newest GB200/B200 often sales-only. Re-verify.
Stranded-power / energy-sited campuses. Contracted power reported ~4.9 GW with a large pipeline. Hyperscaler offtake (Microsoft, Meta) is their book, not ours.
Nebius
neocloud
A 60% H100 spread vs CoreWeave is a procurement fact, not an SSDF quote.
HGX H100 ~$3.85 on-demand; only the set publishing B300 on-demand (~$7.85) as of 21 Aug 2026. Re-verify.
Committed MW deals at a premium to published base. NASDAQ: NBIS.
Groq
inference specialist
Do not compare Groq to CoreWeave as if both sold H100 hours. Different clerk. SSDF does not host LPUs.
Per-token. No HGX H100 card on the 21 Aug 2026 neocloud table.
LPU inference, not HGX rental. Licensed inference tech to NVIDIA (NVIDIA annual report). NVIDIA Cloud Partner after that deal. Per-token, not per-GPU-hr.
RunPod / Vast.ai (spot)
spot market
Fine for research bursts. Production inference that cannot interrupt should not live here. SSDF does not broker host GPUs.
Often the floor. Community H100 prints in the ~$1.2–2.5 band in 2026 roundups. Availability is not a reservation. Re-verify.
Community / spot / peer hosts. Interruption is the product. Vast.ai is a marketplace of independent hosts, not a single landlord.
AWS p5 / GCP A3 / Azure ND
hyperscaler
Coinbase Business is SSDF's treasury clerk, not AWS. Listing p5 is grammar, not an SSDF GPU SKU.
8x H100 on-demand list often $55–$98/node depending on cloud (intuitionlabs Sep 2026 comparison). Neoclouds undercut the list. Re-verify.
Capacity Blocks / reservations / savings plans. On-demand list is the high print; committed effective rate is the real one. Same take-or-pay shape at hyperscaler scale.