Frontier supplier enquiry · GB300 NVL72 rack

Frontier Rack 20TB

A full-rack 20TB NVLink domain for the largest qualified deployments.

A quote-only GB300 NVL72 route with 72 Blackwell GPUs and 36 Grace CPUs. It is a complete rack-scale platform, not an ordinary GPU server.

Current quotation and acceptance plan required.

Price, exact parts, compatibility, warranty, delivery and workload results are indicative until confirmed in a written quotation and acceptance record.

Supplier reference / final build quoted
Official labelled NVIDIA GB300 NVL72 full-rack architecture reference
Inspect the platform
72 × 288GB GPU memory · 20736GB physical total, not automatically pooled GB300 NVL72 full-rack architecture reference platform GPU configuration 72 × NVIDIA GB300 288GB Image status GB300 NVL72 rack architecture reference · exact supplier quotation and facility design required Memory use 20,736GB is distributed across a rack-scale NVLink and NVSwitch fabric. It is specialist shared infrastructure, not one ordinary GPU memory space.

Key buying facts

Frontier Rack 20TB at a glance

Commercial status
Quote only current supplier quotation required
GPU configuration
72 × NVIDIA GB300 288GB 72 GPUs
Per-GPU memory
288GB physical VRAM per GPU
Physical GPU total
20736GB across 72 GPUs · not automatically pooled
Physical class
Rack server GB300 NVL72 full-rack architecture

How the memory works: 20,736GB is distributed across a rack-scale NVLink and NVSwitch fabric. It is specialist shared infrastructure, not one ordinary GPU memory space.

Power planning · Rack server

Up to 142kW full-rack power

Representative intended fit

A sovereign, research or large-enterprise programme

Buyer fit

Start with the reason to own it.

The maximum listed route for sovereign, research and large-enterprise programmes with specialist infrastructure and procurement.

Commercial status

Quote only

Final supplier quote required

Includes workload sizing, the configured system, AI software stack, security baseline, burn-in, agreed workload testing, documentation, remote onboarding and 30-day configuration-defect support.

Package figures exclude VAT and cover the stated GPU RIGS deployment scope. Every final order requires a written quotation confirming the specification, availability, delivery, warranty and workload acceptance plan.

A credible fit

  • A validated rack-scale model or training architecture
  • A sovereign, research or large-enterprise programme
  • A buyer able to accept full facility, network and support design

Choose another route when

  • A single HGX system meets demand
  • The programme lacks a named technical operator
  • Budget or value depends on unverified public performance claims

System evidence

What is confirmed, calculated and still unknown.

Vendor-documented values, calculations, assumptions and unknowns remain separate. A capability is marked as measured only when a reproducible benchmark exists.

Commercial status
quote-only
Reviewed
Review again
Open specification items
9
Measured / confirmedVendor documentedCalculatedExternal estimateAssumptionUnknownUnsupported

01 / model evidence

Exact checkpoint profiles

0 measured-supported profiles. A named model appears as supported only after the exact checkpoint, runtime, deployment and limits have a reproducible benchmark record.

Current boundary

Kimi K3 is shown only as a calculated capacity candidate. The rack exists for facility-scale throughput and concurrency, neither of which has been measured here; weight fit alone does not justify this platform.

Calculated capacity candidate

Moonshot AI moonshotai/Kimi-K3

Official checkpoint record
Repository state
Repository state retrieved 2026-07-28; pin an immutable commit before deployment
Licence
Licence must be rechecked and recorded before deployment
Weights
Official repository MXFP4 checkpoint; MXFP4 weights and MXFP8 activations
Dated artifact total
1,560,936,091,448 bytes across 96 safetensors files
Documented model context
Up to 1,048,576 tokens in the model specification
Runtime
No versioned runtime, container, driver or CUDA profile validated
Candidate deployment
72 GPUs; tensor and pipeline parallelism not yet fixed
Tested limits
Context, input, output, batch and concurrency are all unmeasured

GB300 NVL72's documented 20TB aggregate GPU memory is a capacity candidate for this checkpoint. Its commercial purpose is rack-scale throughput and concurrency, neither of which is measured here; checkpoint fit alone does not justify the platform.

02 / workload acceptance

Beyond language models

Each workload needs its own quality, latency, throughput, stability and resource test. The entries below define what would be measured; they are not performance claims.

01

Language-model inference

Quality pass rate, TTFT, decode and aggregate throughput, latency percentiles, errors, memory, power and temperature at disclosed context, batch and concurrency.

Acceptance definition only. No performance, quality, capacity or suitability result has been measured for this system.

02

Fine-tuning

Completed steps, time, peak memory, loss and held-out evaluation result with exact base checkpoint, method, trainable parameters, dataset, sequence length and batch.

Acceptance definition only. No performance, quality, capacity or suitability result has been measured for this system.

03

Vision and OCR

Field accuracy or character error rate plus latency on a versioned, permission-safe image and document set.

Acceptance definition only. No performance, quality, capacity or suitability result has been measured for this system.

04

Image generation

Latency percentiles, images per second, peak memory, stability and accepted-output rate at exact checkpoint, resolution, steps, sampler, CFG and batch.

Acceptance definition only. No performance, quality, capacity or suitability result has been measured for this system.

05

Video generation

Clip latency, clips per hour, peak memory, stability and accepted-output rate at exact checkpoint, dimensions, frames, frame rate, duration, steps and batch.

Acceptance definition only. No performance, quality, capacity or suitability result has been measured for this system.

03 / configuration

Specification with its certainty attached

The current sources show the basis for each value. The final quotation names the exact supplier, parts, warranty and bill of materials for the ordered system.

Frontier Rack 20TB specification and evidence status
Field Public value Evidence state Qualification
Platform class GB300 NVL72 liquid-cooled full-rack architecture Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS. the selected supplier is the current enquiry route; exact authorised supply, facility integration and support terms remain unconfirmed.
Chassis class Full liquid-cooled rack Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
GPU route 72 × NVIDIA GB300 288GB Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS. Exact board part numbers and serials belong in the final bill of materials.
Physical GPU memory total 20736GB across 72 GPUs Calculated Calculated from disclosed inputs; not a measured system result. 20,736GB is distributed across a rack-scale NVLink and NVSwitch fabric. It is specialist shared infrastructure, not one ordinary GPU memory space.
Per-GPU memory 288GB Vendor documented Documented component capacity and the safer first sizing boundary before any supported multi-GPU test.
GPU interconnect NVLink/NVSwitch rack-scale fabric Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS. The commercial product route is the complete GB300 NVL72 liquid-cooled rack, not an independently supplied add-in GPU.
System topology NVLink/NVSwitch rack-scale fabric Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
CPU 36 NVIDIA Grace CPUs Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
System memory Exact system-memory capacity and population pending supplier BOM Unknown Not yet evidenced for the ordered system and must be resolved before acceptance.
Primary storage Exact NVMe capacity, model, endurance and layout pending supplier BOM Unknown Not yet evidenced for the ordered system and must be resolved before acceptance.
Network Exact network interfaces and fabric pending workload and supplier confirmation Unknown Not yet evidenced for the ordered system and must be resolved before acceptance.
Cooling Direct liquid-cooled rack; facility water and heat-rejection design required Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
Commercial status Quote only Unknown No public numeric offer is made. Exact hardware, facility, delivery, warranty and support scope require a written quotation.

04 / site and facility

The room stays inside the system boundary

Nameplate power, supplier cooling descriptions and aggregate GPU TDP are not a site approval. The exact build and customer facility need written acceptance.

Input power
Up to 142kW full-rack power Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS. The final facility design must use the exact ordered rack and deployment specification.
Measured heat output
Measured wall power and facility heat-rejection requirement pending reference build Unknown Not yet evidenced for the ordered system and must be resolved before acceptance.
Cooling route
Direct liquid-cooled rack; facility water and heat-rejection design required Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
Dimensions and handling
Full-rack delivery; exact dimensions and weight pending partner design Vendor documented Documented by the component or platform vendor; not independently measured by GPU RIGS.
Facility acceptance
Data-centre liquid cooling, high-capacity power and authorised integration required Unknown The customer site has not been surveyed. The final ordered configuration and qualified facility design control this decision.

Pre-order site checks

  • A suitable 19-inch rack, rail depth, handling route and secure operating location
  • A qualified electrical design based on the final PSU population and measured load
  • Cooling and heat-rejection capacity for sustained accelerator operation
  • Appropriate switching, cabling, remote management and network segmentation
  • A named operational owner or contracted support route

05 / delivery and support

A handover boundary, not an implied managed service

The final quotation must name who owns procurement, facility work, acceptance, warranty, remote support, recurring operations and any on-site service.

Hardware warranty
Exact whole-system, GPU, DOA, RMA, spares and on-site warranty route pending Unknown Not yet evidenced for the ordered system and must be resolved before acceptance. No supplier or GPU RIGS warranty term is promised by this planning record.
Availability and lead time
Build-to-order availability and lead time pending written supplier confirmation Unknown Not yet evidenced for the ordered system and must be resolved before acceptance. This record does not mean the product is stocked or orderable.
GPU RIGS support baseline
Remote onboarding and 30-day configuration-defect support Assumption The written quotation and contract confirm this service boundary. It is not a 24-hour managed service or an implied on-site warranty.

Standard service boundary

  • Documented workload and site-fit review
  • Confirmed bill of materials before procurement
  • Configuration, burn-in and agreed smoke-test evidence
  • Asset schedule, admin notes and user quick-start material
  • Collection or the quoted kerbside or pallet-delivery route
  • Remote onboarding and 30-day configuration-defect support

Separate scope or customer responsibility

  • Building electrical work, rack, UPS, cooling or structured cabling
  • Nationwide on-site installation unless separately quoted
  • Migration of customer data, every integration or every application
  • Continuous managed operations, security monitoring or a 24-hour support agreement
  • Third-party model, API, marketplace or software charges
  • A compliance certificate, performance guarantee or income guarantee

06 / sources and review status

Sources and unresolved items

Review dates show evidence freshness, not future availability. A current quote, serialised bill of materials and accepted test record control the order.

Commercial state
quote-only
Reviewed
2026-07-28
Review again
2026-08-28

Unresolved before acceptance

  • System memory: Exact system-memory capacity and population pending supplier BOM
  • Primary storage: Exact NVMe capacity, model, endurance and layout pending supplier BOM
  • Network: Exact network interfaces and fabric pending workload and supplier confirmation
  • Commercial status: Quote only
  • Measured heat output: Measured wall power and facility heat-rejection requirement pending reference build
  • Facility acceptance: Data-centre liquid cooling, high-capacity power and authorised integration required
  • Hardware warranty: Exact whole-system, GPU, DOA, RMA, spares and on-site warranty route pending
  • Availability and lead time: Build-to-order availability and lead time pending written supplier confirmation
  • GPU RIGS support baseline: Remote onboarding and 30-day configuration-defect support
  1. Product record

    Current product specification and commercial status

    GPU RIGS

    Retrieved 2026-07-28 · reviewed 2026-07-28

    The current product record defines the specification and commercial status shown on this page.

  2. Product record

    Current product specification and commercial status

    GPU RIGS

    Retrieved 2026-07-28 · reviewed 2026-07-28

    The current product record defines the specification and commercial status shown on this page.

  3. Official component/platform record

    NVIDIA GB300 NVL72

    NVIDIA

    Retrieved 2026-07-28 · reviewed 2026-07-28

    Use the linked source to confirm the dated component, platform or checkpoint facts stated above.

  4. Supplier reference

    Dated supplier configuration evidence

    System supplier

    Retrieved 2026-07-28 · reviewed 2026-07-28

    Dated supplier evidence. Configuration, stock, price and warranty require confirmation in the final quotation.

  5. Official model record

    Kimi K3 official model card and repository

    Moonshot AI

    Retrieved 2026-07-28 · reviewed 2026-07-28

    Use the linked source to confirm the dated component, platform or checkpoint facts stated above.

  6. Official model record

    Kimi K3 repository file inventory

    Moonshot AI / Hugging Face

    Retrieved 2026-07-28 · reviewed 2026-07-28

    Use the linked source to confirm the dated component, platform or checkpoint facts stated above.

  7. Official component/platform record

    NVIDIA NVL72 AI factory reference architecture components

    NVIDIA

    Retrieved 2026-07-28 · reviewed 2026-07-28

    Use the linked source to confirm the dated component, platform or checkpoint facts stated above.

Known limitations

  • This is a complete 72-GPU liquid-cooled rack, not an the selected OEM 4U product or ordinary office delivery.
  • 20TB is the rounded marketing class; 72 × 288GB equals 20,736GB aggregate GPU memory.
  • This is a complete 72-GPU liquid-cooled rack, not an the selected OEM 4U product or ordinary office delivery.
  • 20TB is the rounded marketing class; 72 × 288GB equals 20,736GB aggregate GPU memory.
  • Pricing is quote-only.
  • 20TB is aggregate rack accelerator memory and not a generic one-process promise.
  • Aggregate GPU memory is not one universal memory pool.
  • No model, speed, context, concurrency or multi-GPU efficiency promise exists without a controlled fit profile.
  • Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.

Do not buy this route when

  • Kimi K3 weight fit alone is the requirement; HGX B300 is the more proportionate capacity candidate.
  • The buyer lacks the liquid cooling, power, data-centre and authorised support infrastructure.
  • A referral/channel project and hosted capacity do not pass commercial comparison.

Inspect the evidence behind Frontier Rack 20TB.

The system view shows the selected hardware reference where available. Technical diagrams explain memory, data boundaries, queues, power and handover. The final bill of materials and workload test confirm the ordered system.

Diagram combining model weights, context, cache and active requests into a memory headroom check
Diagram combining model weights, context, cache and active requests into a memory headroom check
Memory fit depends on the workload, context, cache and simultaneous demand, then needs a representative test. GPU Servers technical illustration.
Memory fit depends on the workload, context, cache and simultaneous demand, then needs a representative test. GPU Servers technical illustration.
Diagram showing jobs entering a queue, being assigned to independent workers and producing measured outputs
Diagram showing jobs entering a queue, being assigned to independent workers and producing measured outputs
Queued rendering, batch and coding work can be divided between workers, with waiting time and failures measured. GPU Servers technical illustration.
Queued rendering, batch and coding work can be divided between workers, with waiting time and failures measured. GPU Servers technical illustration.
Diagram showing an approved request, a local service, an approved store and a policy-controlled data path
Diagram showing an approved request, a local service, an approved store and a policy-controlled data path
A private deployment starts with the permitted data path, access policy and logging boundary. GPU Servers technical illustration.
A private deployment starts with the permitted data path, access policy and logging boundary. GPU Servers technical illustration.
Diagram of an evidence pack containing an asset schedule, burn-in record, health readings, workload test and admin guide
Diagram of an evidence pack containing an asset schedule, burn-in record, health readings, workload test and admin guide
A credible handover records the supplied assets, checks, operating evidence, instructions and unresolved items. GPU Servers technical illustration.
A credible handover records the supplied assets, checks, operating evidence, instructions and unresolved items. GPU Servers technical illustration.

Software and security

The usable product is more than the chassis.

The final stack stays deliberately small. Versions, licences, access and recurring ownership are recorded so the customer is not left with an opaque collection of containers.

Software baseline

  1. Ubuntu LTS on a recorded operating-system version
  2. NVIDIA driver, CUDA components and container support validated for the ordered hardware
  3. Docker Engine and NVIDIA Container Toolkit
  4. One primary model server selected from Ollama, vLLM, SGLang, TensorRT-LLM or llama.cpp for the accepted workload
  5. Open WebUI or another reviewed browser interface
  6. Named authentication, TLS and reverse-proxy approach
  7. GPU, node and service monitoring with an agreed log-retention period
  8. Pinned versions, a software bill of materials and a model source and licence record

Security ownership

GPU RIGS baseline
Initial operating-system state, named administrator handover, host firewall baseline, agreed access route, secrets transfer and documented update state.
Customer or contracted operator
User lifecycle, network and VPN policy, backups, monitoring review, patch approval, incident response, data governance and lawful use.
Shared before acceptance
Model and software licence checks, retention and logging choices, recovery test, acceptance criteria and a named owner for every recurring task.

Local infrastructure can reduce disclosure to external AI APIs. It does not automatically make the service secure, accurate, confidential or UK GDPR compliant.

Testing and acceptance

The evidence pack is part of the machine.

Performance is confirmed against the accepted workload, exact build and disclosed test conditions. No throughput, latency or quality figure is claimed before that test.

  1. 01

    Record the final bill of materials, serial numbers and firmware versions.

  2. 02

    Run at least 24 hours of GPU, CPU, memory and storage stress testing.

  3. 03

    Capture temperature, fan, error, health and wall-power evidence under the agreed load.

  4. 04

    Test cold boot, restart and the available remote-management route.

  5. 05

    Check drive health, network throughput and the container and GPU runtime.

  6. 06

    Run model and workload smoke tests against the written acceptance set.

  7. 07

    Record any measured speed only with the model, quantisation, context, concurrency and runtime disclosed.

  8. 08

    Test an agreed fault or recovery route and provide the resulting handover record.

What is not claimed today

No fixed users, tokens per second, latency, model size, accuracy, availability, savings or marketplace contribution is stated without the missing configuration and test conditions.

See how systems are tested

Commercial reality

Tax and spare capacity are supporting questions.

Neither belongs in a guaranteed saving or payback claim. The purchase must stand on the accepted workload, control case and operating plan.

Finance, VAT and capital allowances

  • The displayed figure is for a complete GPU RIGS private AI deployment, not unconfigured hardware. The final written quotation confirms the exact specification, delivery, warranty and accepted workload scope.
  • Prices exclude VAT. VAT recovery depends on the buyer, its taxable activities and the normal evidence rules.
  • Qualifying equipment may be plant and machinery for capital-allowance purposes. The buyer's accountant decides eligibility and timing.
  • A third-party lease or hire-purchase route may be explored after partner validation and credit approval. GPU RIGS is not presented as a lender.
  • There is no generic capital-gains advantage and no automatic research and development relief because the equipment supports AI.
Read the guarded UK buyer notes

Optional idle capacity

  • Marketplace mode is off by default and is excluded from the purchase case.
  • A separate environment, no customer data mounts, network controls and a local kill switch would be required.
  • The customer, insurer, supplier warranty and marketplace terms must permit the intended use.
  • Vast.ai, Render, Golem and direct batch work do not guarantee acceptance, demand, rate or income.
  • Any dated estimate must identify its source date; a pilot must report achieved utilisation, fees, electricity, cooling, faults and operator time.
Review the marketplace decision gate

Limits and alternatives

A good specification leaves room for “no”.

The fit check can recommend a smaller system, hosted service, hybrid route, custom build or no purchase. That is preferable to making this package carry a workload it has not proved.

Package boundaries

  • Aggregate GPU memory is not one universal memory pool.
  • No model, speed, context, concurrency or multi-GPU efficiency promise exists without a controlled fit profile.
  • Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.
  • This is a complete 72-GPU liquid-cooled rack, not an the selected OEM 4U product or ordinary office delivery.
  • 20TB is the rounded marketing class; 72 × 288GB equals 20,736GB aggregate GPU memory.
  • Pricing is quote-only.
  • 20TB is aggregate rack accelerator memory and not a generic one-process promise.
  • Aggregate GPU memory is not one universal memory pool.
  • No model, speed, context, concurrency or multi-GPU efficiency promise exists without a controlled fit profile.
  • Rack, electrical work, UPS, cooling, cabling, colocation and on-site installation require a separate scope.

Questions answered

Frontier Rack 20TB questions that affect the order

The written quotation and acceptance plan confirm any price, specification, warranty or workload assumption discussed below.

Does Frontier Rack 20TB provide 20736GB as one memory pool?

Only a single-GPU system provides its stated GPU memory on one card. Multi-GPU totals are aggregate physical capacity; usable sharding depends on the exact model, runtime and topology.

Which models are supported?

Only models with a reproducible fit profile are listed as supported. Capacity candidates are labelled separately; a memory calculation alone is not a support claim.

Is the displayed price final?

No. An indicative price is a transparent planning value based on dated inputs. Quote-only platforms and the final order require a current supplier quote, accepted bill of materials and delivery scope.

Can spare capacity earn income?

It may be evaluated as an optional isolated secondary use. Marketplace acceptance, utilisation, rates, fees, energy and income can change and are never guaranteed or included in the purchase case.

Prepare the next decision

Specify the work before the parts.

Record the workload, data boundary, users, site and acceptance test. No confidential documents or credentials are needed for the first brief.

Package figures exclude VAT and cover the stated GPU RIGS deployment scope. Every final order requires a written quotation confirming the specification, availability, delivery, warranty and workload acceptance plan.