Technical / infrastructure

Infrastructure. Hardware, pricing, authorization evidence.

Hardware and cost orientation, with an evidence worksheet for evaluating an exact government AI offering. The worksheet does not authorize any vendor or assign impact levels.

SOURCE STATUS SEE EACH TABLE // TABLES 4 // SORTABLE CLICK ANY COLUMN HEADER // VERIFY BEFORE CITING

ON THIS PAGE

GPU Comparison for AI Workloads
Cloud GPU Pricing Snapshot
On-Prem Hardware Shopping List
Government AI offering: evidence to collect

TABLE // GPUs CHANGES YEARLY

GPU Comparison for AI Workloads

The hardware that runs models. Datacenter cards dominate training; consumer cards win for many small-model and dev workloads.

Earlier reference snapshot · Source verification date not recorded. Confirm current information before use.

GPU	VRAM (GB)	FP16 TFLOPS	NVLink	Typical Model Fit	Typical Use	Approx Price (USD)
NVIDIA H100 80GB SXM	80	989	yes	70B at 4-bit on one card; 70B at FP16 across 2; train 7B-13B comfortably	Datacenter training and high-volume inference	$30,000+
NVIDIA H200 141GB	141	989	yes	70B at FP16 on one card; 405B at 4-bit across 4	Memory-bound inference for long context and large models	$35,000+ _status as of 2026-06, verify before citing_
NVIDIA B200 192GB	192	2,250 _unverified_	yes	405B at 4-bit on 2 cards; 70B at FP16 with room to spare	Frontier training and inference; shipping in volume through 2026 _status as of 2026-06, verify before citing_	$40,000+ _unverified_
NVIDIA A100 80GB SXM	80	312	yes	70B at 4-bit; 13B at FP16	Workhorse for training and inference; widely available on spot markets	$10,000-$15,000
NVIDIA A100 40GB	40	312	yes	13B at FP16; 30B at 4-bit	Older training and mid-size inference; common in academic clusters	$8,000-$12,000
NVIDIA L40S	48	362	no	13B-30B at FP16; 70B at 4-bit	Inference-first datacenter card; cheaper than H100 for serving small/mid models	$8,000-$11,000
NVIDIA RTX 6000 Ada	48	364	no	30B at FP16; 70B at 4-bit	Workstation training and dev; quiet enough for an office	$7,000-$8,500
NVIDIA RTX 5090	32	419 _unverified_	no	13B at FP16; 70B at 4-bit (tight)	High-end consumer dev box; supply constrained through 2026 _status as of 2026-06, verify before citing_	$2,000-$2,500
NVIDIA RTX 4090	24	165	no	7B at FP16; 30B at 4-bit	Consumer dev workhorse; popular for fine-tuning small models	$1,800-$2,200
NVIDIA RTX 3090	24	71	yes	7B at FP16; 30B at 4-bit	Used-market budget option; NVLink lets you pair two for 48GB	$700-$900 (used)
AMD MI300X	192	1,307	no (Infinity Fabric)	70B at FP16 on one card; 405B at 4-bit on 2	ROCm-based training and inference; main NVIDIA alternative at scale	$15,000-$20,000 _unverified_

TABLE // CLOUD GPU $ CHANGES MONTHLY

Cloud GPU Pricing Snapshot

Hourly on-demand pricing for the most common AI GPUs across major clouds and neoclouds. Prices change every month. Cite the provider, not us.

Earlier reference snapshot · Source verification date not recorded. Confirm current information before use.

Provider	GPU	On-Demand $/hr	Spot $/hr	Per GPU?	Notes
AWS	H100 (p5.48xlarge, 8x)	$98.32 _unverified_	$30-$50 _unverified_	no	Priced per 8-GPU instance; per-GPU works out to ~$12.30/hr on-demand. Verify on AWS pricing page.
AWS	A100 80GB (p4d.24xlarge, 8x)	$32.77 _unverified_	$10-$15 _unverified_	no	Per-GPU ~$4.10/hr on-demand. Capacity reservation may be required.
AWS	L40S (g6e.xlarge, 1x)	$1.86 _unverified_	$0.55-$0.90 _unverified_	yes	Single-GPU instances available; good for inference.
Azure	H100 (ND H100 v5, 8x)	$98.00 _unverified_	n/a	no	8-GPU instance; spot capacity limited. Azure calls it 'Spot VM' but H100 availability is rare.
Azure	A100 80GB (ND A100 v4, 8x)	$27.20 _unverified_	$8-$12 _unverified_	no	8-GPU instance only; per-GPU ~$3.40/hr.
GCP	H100 (A3, 8x)	$88.49 _unverified_	$26-$40 _unverified_	no	Per-GPU ~$11/hr on-demand. Reservation discounts available.
GCP	A100 80GB (A2 ultra)	$5.07 _unverified_	$1.50-$2.20 _unverified_	yes	Per-GPU pricing available.
GCP	L40S (G2)	$1.85 _unverified_	$0.55-$0.85 _unverified_	yes	Inference-friendly; preemptible available.
Lambda	H100 SXM	$3.29 _unverified_	n/a	yes	Neocloud; no spot tier but cheaper baseline. Capacity comes and goes.
Lambda	A100 80GB	$1.79 _unverified_	n/a	yes	Cheaper than hyperscalers; reservation required for long jobs.
Lambda	L40S	$1.45 _unverified_	n/a	yes	Inference-grade card at a good price.
RunPod	H100 SXM	$2.99 _unverified_	$1.99 _unverified_	yes	Community Cloud tier is cheapest; Secure Cloud costs more.
RunPod	A100 80GB	$1.89 _unverified_	$1.19 _unverified_	yes	Community Cloud; bring-your-own image.
RunPod	RTX 4090	$0.69 _unverified_	$0.34 _unverified_	yes	Cheap dev/inference; consumer card.
Modal	H100	$3.95 _unverified_	n/a	yes	Serverless; billed per second. No spot tier.
Modal	A100 80GB	$2.10 _unverified_	n/a	yes	Per-second billing; good for bursty inference.
Modal	L40S	$1.95 _unverified_	n/a	yes	Per-second billing.
CoreWeave	H100 HGX	$4.76 _unverified_	n/a	yes	Reserved capacity is the main path; on-demand exists but contracts are typical.
CoreWeave	A100 80GB	$2.21 _unverified_	n/a	yes	Volume discounts standard.
DigitalOcean	H100 (1x)	$3.39 _unverified_	n/a	yes	Single-GPU droplets via Paperspace integration. _status as of 2026-06, verify before citing_

TABLE // ON-PREM CHANGES YEARLY

On-Prem Hardware Shopping List

When the right answer is "buy the box", the shopping list looks like this. Three tiers: workstation, single-server, cluster.

Earlier reference snapshot · Source verification date not recorded. Confirm current information before use.

Tier	CPU	GPU Config	RAM	Storage	Power (W)	Approx Price (USD)	Typical Use
Workstation (entry)	AMD Threadripper 7970X (32-core)	1x RTX 4090 (24GB)	128GB DDR5	4TB NVMe + 8TB HDD	1,200	$8,000-$10,000	Solo dev box; fine-tune 7B-13B; serve small models locally
Workstation (pro)	AMD Threadripper 7980X (64-core)	1x RTX 6000 Ada (48GB) or 2x RTX 4090	256GB DDR5	8TB NVMe RAID-1	1,600	$15,000-$20,000	Team dev box; quiet enough for office; train 13B-30B
Single-server (4-GPU)	2x Intel Xeon Platinum 8568Y+ or 2x AMD EPYC 9554	4x H100 80GB PCIe (no NVLink) or 4x L40S	512GB-1TB DDR5 ECC	8TB NVMe RAID-10 + 30TB capacity	3,500	$120,000-$160,000 _unverified_	Small team inference + fine-tuning; 70B serving at FP8
Cluster (8-GPU node)	2x Intel Xeon Platinum 8568Y+ or 2x AMD EPYC 9654	8x H100 80GB SXM5 with NVLink (Dell XE9680 / Supermicro AS-8125GS)	2TB DDR5 ECC	30TB NVMe + 60TB capacity tier	10,200	$280,000-$350,000 _unverified_	Single-node training of mid-size models; production 70B-405B inference
Cluster (8-GPU, cleared / air-gapped)	2x AMD EPYC 9554 (no foreign-origin parts where required)	8x H100 80GB SXM5; no internet path; KVM-over-IP disabled or air-gapped console only	2TB DDR5 ECC	30TB NVMe RAID-10, all storage encrypted at rest with FIPS 140-3 module	10,200	$320,000-$400,000 _unverified_	SCIF deployment; screen approval required for display config; no outbound network; manual model loads via approved transfer media

TABLE // AUTHORIZATION EVIDENCE CHANGES QUARTERLY

Government AI offering: evidence to collect

A learning worksheet, not an approved-product list. Identify the exact offering and proposed use before asking the responsible organization to evaluate it.

Linked source facts checked 2026-10-02; no deployment approval or performance validation implied.

Question	Evidence to record	What it does not establish
What is the offering?	Service name, version, region, endpoint, enabled model/features and supporting services.	A vendor or cloud brand alone does not identify an authorized service boundary.
What is the use?	Workflow, data types, data flow, integrations and intended users.	Public-data retrieval and a sensitive operational workflow have different requirements.
What evidence exists?	Current applicable authorization package or assessment, scope, exclusions, customer responsibilities and source date.	An underlying cloud authorization does not automatically cover an added SaaS application.
Who can approve this use?	The agency or program decision-maker, system boundary and applicable handling requirements.	VetsLearn does not authorize products, classified workloads or organizational deployment.
What remains unresolved?	Missing evidence, owner, review deadline and an explicit no-use condition until resolved.	A model license, marketplace listing or vendor announcement is not approval for your use.

A source-check date applies only to the facts in that table. Tables without a date retain older orientation data and have no recorded source verification date. Verify current prices, versions and capabilities with primary sources before spending or citing them. Licenses and brand names do not establish authorization.