Independent engineering firm · Nvidia reference architectures
SPEC · PROCURE · INTEGRATE · VALIDATE

Private AI compute,
engineered to specification.

GPU Smith specifies, procures, integrates and validates private AI infrastructure, from a single inference node to a multi-megawatt hall. Every engagement is delivered against written acceptance criteria and an as-built documentation set.

Fixed-scope assessment · reference design in weeks · credited against the build

Reference Design R-04 — Private Inference PodDATASHEET · REV B
FIG. 1 — GENERAL ARRANGEMENT42U
IB spine · QM9790 · NDR 400G
NVLink switch tray
HGX B200 · 8× GPU · N01
HGX B200 · 8× GPU · N02
HGX B200 · 8× GPU · N03
HGX B200 · 8× GPU · N04
HGX B200 · 8× GPU · N05
HGX B200 · 8× GPU · N06
HGX B200 · 8× GPU · N07
HGX B200 · 8× GPU · N08
Mgmt · BlueField-3 DPU
NVMe storage · 1.9 PB raw
PDU A · 415 V 3Φ
PDU B · 415 V 3Φ
TABLE 1 — KEY PARAMETERS
ParameterValue
Accelerators64 × NVIDIA Blackwell
Aggregate HBM12.2 TB HBM3e
Fabric400G InfiniBand NDR
IT load≈ 100 kW
PrecisionFP8 / FP4
Serving stackTensorRT-LLM · vLLM
DOC GS-R04-DS
SHEET 1 / 4
SCALE 1 rack
VALIDATED
1–10kGPUs
single node to compute hall
one method at every scale
≥ 80%
target sustained utilization
measured at acceptance, not assumed
100%
procurement margin disclosed
no reseller quota, no vendor incentive
§ 01Scope

Six disciplines. One statement of work.

GS-0101/06

Hardware integration & commissioning

HGX, DGX and NVL72 systems specified, procured, rack-integrated, cabled and burned in. InfiniBand and Spectrum-X fabrics validated port by port against a link budget.

GS-0202/06

Cluster architecture & sizing

Topology, node count and serving stack derived from workload models and throughput targets. TensorRT-LLM, vLLM and Triton configured to measured requirements.

GS-0303/06

Inference build-out

Turnkey delivery from empty rack to first validated token. Procurement at a disclosed margin, integration on site, handover on acceptance.

GS-0404/06

Serving optimization

Quantization, batching, KV-cache and scheduler tuning to reduce cost per token and raise sustained utilization on installed hardware.

GS-0505/06

Datacenter operations

Capacity planning, telemetry, failure-mode analysis and operating procedures for existing facilities and colocation.

GS-0606/06

Sovereign & air-gapped systems

Disconnected deployments engineered to a defined constraint set, with local model registry and offline update paths.

NOTE — Electrical, mechanical and thermal engineering are delivered through vetted partner firms under GPU Smith coordination and a single statement of work.

§ 02Design rationale

Owned compute is a capital decision, not a belief.

At sustained load, per-token cloud pricing exceeds the amortized cost of owned hardware. The crossover is a calculation, not an opinion. GPU Smith locates it for a defined workload and states the result in writing.

  • Weights, prompts and telemetry remain within the client boundary
  • Amortized capital cost falls below per-token pricing at sustained load
  • No rate limits, deprecation schedules or policy dependencies in the serving path
  • Air-gap capable for regulated and classified environments
ParameterCloud APIOwned build
Model weightsvendor-heldclient-held
Prompts & telemetryvendor infrastructureclient boundary
Unit cost at scaleper token, indefinitelycapex + power, amortized
Model lifecycledeprecated on vendor schedulefrozen on validation
Rate limitsenforcednone
Air gapnot achievablesupported
The assessment quantifies each row for the specified workload.
§ 03Assurance

Engineered for environments where data cannot leave the boundary.

A-01

Air-gap capable

Clusters engineered to operate fully disconnected: local model registry, offline update path, no external dependency in the serving loop.

A-02

Data residency

Weights, prompts and telemetry remain on client-owned hardware, in the jurisdiction specified.

A-03

Chain of custody

Documented custody from procurement through installation, with serialized inventory and acceptance records.

A-04

Export-control assessed

International deliveries planned with export-control requirements evaluated at assessment, not after purchase.

A-05

As-built documentation

Architecture records, port maps, operating procedures and acceptance reports issued as standard, written for audit as much as operation.

A-06

Structural independence

No reseller quota, no cloud of our own. The engineering recommendation is the only product.

§ 04Engineering method

A defined sequence. Verification at every stage.

01

Assessment

Fixed scope. Workload modelling, total cost of ownership against cloud and API baselines, reference architecture, and a documented recommendation. Credited against the build.

Gate: build / no-build
02

Specification & procurement

The controlled build sheet: part numbers, link budget, lead times and pricing. Procurement executed through our channel at a disclosed margin.

Gate: released for build
03

Integration & validation

Rack integration, fabric bring-up, burn-in and acceptance testing on site or in colocation, against the written criteria.

Gate: acceptance test
04

Optimization & handover

Serving stack tuned to the utilization target, operating procedures issued, and the as-built documentation package delivered.

Gate: as-built issued
§ 05The engineering assessment

Every build begins with a written assessment.

A fixed-scope engagement that resolves a single question: build or do not build, and to what specification. Deliverables are yours regardless of outcome. The fee is credited in full against the build if it proceeds.

Request the assessment

Fixed scope, defined in one call · typically three weeks

DeliverablesGS-ASSESS.PDF
  1. 01Workload characterization: models, context length, tokens per day
  2. 02Total cost of ownership: owned hardware, cloud GPU and API baselines
  3. 03Reference architecture with itemized bill of materials
  4. 04Power, space, cooling and network requirement schedule
  5. 05Procurement plan with lead times and quoted pricing
  6. 06Build / no-build recommendation, issued in writing
§ 06Sectors

Three mandates. One requirement: control of the compute.

Government & regulated enterprise

Bounded inference

Healthcare, finance, legal, defence and public bodies that cannot transmit data to a third-party API. Frontier-class open weights, operated within the client boundary.

Scaling AI teams

Post-cloud economics

Teams whose API and rented-GPU expenditure has passed the ownership crossover. The point is located, then the system is built to it.

Investors & operators

Technical due diligence

Independent review of AI datacenter and GPU-cloud investments: architecture assessment, vendor-quote verification, and a capital plan reconciled with physical reality.

Specify before you spend

Describe the workload. Receive a written engineering assessment.

Build or do not build, and to what specification. Stated in writing, with an itemized bill of materials and a costed comparison against cloud and API baselines.