AI factory architecture compiler, aligned to NVIDIA's reference architectures
Never guess what belongs in a billion-dollar architecture.
Send us the RFP. In ten working days you get a complete AI factory architecture package aligned to NVIDIA's reference architectures, and you keep the compiler that produced it.
- NVIDIA reference architectures
- Deterministic compiler
- Ten working days
- You keep the compiler
Who built this
We have built for the leading GPU manufacturer in the world.
The people who built this have designed compliant architectures for leading hyperscalers and leading model builders.
We cannot name them. In GPU as a service, nobody can.
01
See it run before you talk to anybody.
We host an instance and give you access. Nothing to install, no security review, no ticket with your own IT department. You put the parameters in and walk the designs yourself.
The demo is a throwaway environment and one platform. The licensed build covers the full range. Please do not put a live customer RFP into it.
02
For operators with a bid in flight
You run five hundred to ten thousand GPUs. You have a live RFP, a customer waiting, and one solutions architect who is already the bottleneck. The bid is due in weeks and the design is the part that has not started.
If you build one cluster once, this is not for you.
03
Scope
Where we fit in
Neon is not a data centre design tool. It designs and validates the digital infrastructure of the AI factory, the IT stack that makes it operate, and it hands the facility side the numbers that stack implies. See the public release briefing.
The facility · Not us
- Site and facilities
Land, buildings, rooms, racks, aisles, access, structural, fire safety
- Power infrastructure
Utility interface, substations, switchgear, UPS, generators, PDUs, power distribution
- Cooling and mechanical
Chillers, CRAH and CRAC, cooling towers, piping, HVAC, airflow, cooling plant
- BIM and digital twin
3D, 4D and 5D models of buildings and infrastructure
Facilities, electrical and mechanical engineering firms and their tools. Neon does not replace BIM, electrical, mechanical, civil or building design platforms.
The IT stack · We design, model and validate
- GPU compute infrastructure
GPU servers, racks, power to rack, compute capacity modeling
- High speed network fabric
Leaf and spine, InfiniBand and Ethernet, cabling, optics, network performance modeling
- Storage and data infrastructure
Object, file and block storage, tiers, data paths, backup and replication
- IT power modeling and capacity
IT load profiling, power headroom, what-if analysis, PUE contribution
- Infrastructure middleware
Leak detection, environmental monitoring, sensor data integration
The operating software · We support
- Management and orchestration
DCIM and MCS, resource orchestration, policy, automation
- Monitoring, telemetry and analytics
Telemetry collection, observability, analytics, alerts, dashboards
- Automation and closed loop operations
Workflows, remediation, capacity automation, predictive operations
We also hand the facility side its requirements: rack counts, IT power demand, cooling load, cabling demand and the environmental monitoring interfaces.
04
The test
An AI factory RFI is not a request for hardware pricing
It is a test of whether you can read hundreds of requirements, design an architecture that holds up against NVIDIA's reference architectures, produce a complete bill of materials, build a credible delivery plan, define the operating model, and defend every decision under technical scrutiny.
Four ways a bid dies
- One missed requirement.
You are disqualified, and you never find out which one.
- An unbuildable promise.
You win, and the win is the disaster.
- Documents that do not reconcile.
The diagram, the BOM and the proposal disagree, and the reviewer notices.
- An architecture that looks unready.
Not wrong. Unready. The customer concludes you are not who they thought.
05
Compiled, not written
The architecture is produced by a deterministic compiler, not by a language model. The same inputs produce the same design, every time, and every decision traces to the reference architecture rule behind it.
- The same inputs produce the same design. Every time.
Ask a language model the same low level design question twice and you get two architectures. Ask a different model and you get a third. You do not want to risk a fifty million dollar contract on a hallucination.
- Several aligned options, not one answer.
Swap the storage vendor, the storage fabric or the compute fabric and get another aligned design with a different bill of materials and a different cost.
- Alignment is what keeps your NVIDIA support entitlement.
Deviate on Spectrum-X, InfiniBand or certified storage, and when the GPUs underperform the answer you get is that you used the wrong kit. This is a reason to align that has nothing to do with winning the bid.
- Maker and checker are separated.
An immutable candidate, a decision log and an evidence package, for when someone asks how you arrived at this.
- The knowledge does not leave with your architect.
It is in the system.
06
The artefact
This is the package. Read it before you decide anything.
A complete architecture package for a 20,736 GPU GB300 NVL72 build on Quantum-3 InfiniBand with WEKA storage. Redacted, and otherwise exactly what lands on day seven.
07
The evidence
Do not take our word for any of this. Take NVIDIA's.
We extracted every requirement an NVIDIA Cloud Partner has to satisfy from NVIDIA's published corpus. Each one carries its real requirement ID, so you can check any line against the source document rather than against us.
Score your design against them
Free, and the report is yours whether or not you ever talk to us.
08
Delivery
Ten working days
Four deliveries. The gaps between them are to scale. Six months to the first requirement shows where bids wait.
Up to 10xfaster to submission
Our AI Architect's own figure, from doing this work by hand and then with a language model. It is a comparison of effort, not a promise about the outcome of your bid.
- 01
Requirements assessment and compliance matrix. Every requirement in your RFP extracted, traced and rated.
- 03
Two or three candidate architectures with the tradeoffs made explicit, plus the risk register and the open questions list.
- 07
The full package. Architecture model, topology diagrams, logical BOM, port and connectivity budget, power and rack envelope, requirements traceability matrix, decision record, evidence manifest.
- 10
Handover. The licensed compiler, your configuration already loaded, and a walkthrough so your architect can regenerate and re-cost it without us.
Day one and day three exist on purpose. Visible progress against a bid clock is worth as much as speed.
09
Our risk
Ten working days, or you pay nothing and keep the package
You cannot risk your bid clock on a vendor you have not used. So we carry that risk instead.
- Miss the ten working days and you pay nothing.
You keep the architecture package we delivered, and you keep the compiler licence. There is no clause underneath that sentence.
- Every decision traces to a rule and its evidence.
Any that does not, we fix free.
- If your reviewer finds a disqualifying gap in what we delivered,
we work it until it closes, at no charge.
10
Ways in
Three ways in
Send us a design. Three working days later you have a scored compliance report naming every gap. Pro rated against whatever you buy next.
The compiler, perpetual. You do the work.
Send the RFP. Ten working days later you have the package and the compiler.
The Bid Pack price rises with every one we deliver. We are telling you that rather than pretending otherwise.
Four Bid Packs a quarter
Send us the RFP.
One person delivers them, and that is the real constraint. If the quarter is full we will tell you the date rather than take the work.
Request a call about your bid- 10
- working days
- $75,000
- bid pack
- 4
- per quarter