Predictika · NVIDIA AI factory hall design

An AI design agent's AI factory hall design, checked before it goes out

An AI design agent drafts an AI factory hall from NVIDIA rack-scale systems: racks, liquid cooling, power, storage and a leaf/spine fabric. Before the design goes out, the agent sends it to Predictika, which checks it against a declarative model built from NVIDIA's public product information, compares the GPU count and capex the agent states with the figures it computes, and returns every finding together with the line-item bill of materials.

Scope: NVIDIA AI factory hall design from four rack-scale systems (GB300 NVL72, Vera Rubin NVL72, Vera Rubin Ultra NVL576, Rubin Ultra NVL144 Kyber) with in-row CDUs, 800 VDC sidecars, storage racks and a leaf/spine fabric. See everything this validator covers

This validator models the design of an AI factory hall built from NVIDIA rack-scale systems. It is an illustrative model built from NVIDIA's public product information; every rating, weight, price and limit in it is a fictional illustrative value.

Cooling and power — coolant temperature at the racks after the CDU, liquid and air cooling capacity, CDU capacity with a spare unit, 800 VDC architecture and sidecar sizing per row, row busway and the site power budget.
Structure and layout — floor loading per rack, positions per row, hall rows, and NVLink-domain rack counts.
Fabric — leaf downlinks against GPU ports, spine ports against leaf uplinks, and InfiniBand versus Ethernet matching.
Delivery and money — rack ship year against the hall go-live year, GPU target, capex against the budget (with a finance sign-off band), and the GPU count and capex the agent states.

Within this scope the validator handles any hall design built from these rack systems, at any row count, power budget or fabric size, not just the scenarios shown here. A hall uses one rack system in all of its compute rows. It does not cover NVIDIA products outside these rack-scale systems, other vendors' racks, or building structure beyond floor loading.

The AI design agentBuilds the hall design from the conversation and says it is ready.

Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

7 scenarios in 4 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI design agent says

Proposing agent: AI design agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator

Waiting for the call.

What this means

Run the validator to see the outcome

The explanation appears here once the validator returns.

Where rules-based CPQ gets it wrong

Click to expand

Our assessment of what a typical rules-based CPQ (product rules, price rules and validation rules, as in Salesforce CPQ) does with the same cases. It is a judgment about effort and reliability, not a claim that nothing can be built.

What the validator caught	Typical rules-based CPQ	Predictika
Coolant temperature after the CDU	Only with care The limit belongs to the rack system and the water temperature rise belongs to the CDU, so the check needs a figure derived across two catalogs. A validation rule can do it once someone builds the calculated field.	Both are in the model; the temperature at the racks is computed and compared on every call.
Power sidecars sized to the row	Only with care Per-row power against sidecar capacity, only for racks that take 800 VDC. It takes a rule per rack system, and a hall that adds a row later is easy to miss.	The row draw and the sidecar capacity are computed together for whatever rack system is chosen.
Fabric ports across both layers	Only with workarounds Leaf downlinks against GPU ports and spine ports against leaf uplinks are arithmetic across layers that a product-rule engine does not carry; usually done outside the tool.	Port totals for both layers are computed and checked together, so one short layer cannot hide behind the other.
The agent's stated GPUs and capex against the computed ones	Only with care A validation rule can compare a stated figure with a calculated field, but only if the stated figure is a field on the design, which it rarely is.	The stated figures are compared with the computed ones at every step of processing the design.

What this validator catches: findings from its test cases

Click to expand

Every rule is checked on every call, in any order the choices arrive; one problem never hides another. Blocking findings stop the design; approval findings need a person to sign off; everything else is computed.

1	GB300 NVL72 needs coolant at or below 42 C; 40 C facility water through the In-row CDU 1.4 MW reaches the racks at 43 C.
2	Rubin Ultra NVL144 Kyber takes 800 VDC power only; this hall distributes 415 VAC.
3	Each row of Rubin Ultra NVL144 Kyber draws 2,400 kW and needs 800 VDC power sidecars; the sidecars in a row give 0 kW.
4	Rubin Ultra NVL144 Kyber loads the floor at 4,250 kg/m2 (3,400 kg per rack); the floor is rated 2,500 kg/m2.
5	5,184 GPU fabric ports; 150 leaves give 4,800 downlinks.
6	4,800 leaf uplinks; 70 spines give 4,480 ports.
7	The agent stated $363,690,000; the validator computes $361,620,000.
8	Estimated capex $363,690,000 is above 95 percent of the $375,000,000 budget; finance must sign off.
9	Rubin Ultra NVL144 Kyber is not expected to ship until 2027; the hall goes live in 2026.
10	The design delivers 3,456 GPUs; the target is 4,000.