Predictika · CoreWeave GPU capacity commitments

An AI sales agent's GPU capacity commitment, checked before it goes out

An AI sales agent prepares a CoreWeave capacity quote for a customer: instance type, zone, nodes, term, capacity plan, committed discount, storage. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the published rules and a capacity ledger, and returns a verdict, every reason, the figures it computed and, for a quote that can proceed, the line-item quote itself.

Scope: CoreWeave GPU instances on CKS (GB200, GB300, B200, H100, H200): racks, zones, capacity plans, storage and committed discounts. See everything this validator covers

This validator models CoreWeave capacity commitments and the commercial terms quoted around them — not just the scenarios shown below.

  • Instances — Whole nodes of 4 GPUs (GB200, GB300) or 8 GPUs (B200, H100, H200); NVL72 types only in full racks of 18 nodes and with no autoscaling.
  • Zones — Which availability zone offers which instance type, and which zones offer AI Object Storage.
  • Capacity plans — Reserved and Flex give guaranteed capacity; Spot and On-Demand do not; Spot is not offered for every type.
  • Capacity — Nodes free by type and zone from a capacity ledger the operator supplies, with a warning when a reservation leaves little.
  • Price and discount — On-demand list per node hour, the committed discount against it (published as up to 60 percent), and types with no published price that go to sales.
  • Contract terms — Typical term of 1 to 6 years and prepayment of 15 to 25 percent, flagged for approval when outside.
  • Stated figures — A total or a free-node count the request states is compared with the computed one.

Any CoreWeave capacity commitment that touches these is in scope — not only the scenarios pictured here. Not covered: Reserved and Flex rate cards (not public), SUNK, Mission Control, Direct Connect and Dedicated VAST pricing, and inference products. The capacity ledger and the approval lines are placeholders, not CoreWeave data.

The AI sales agentBuilds the capacity quote from the conversation and says it is ready.
Predictika validatorChecks every rule at once and returns the verdict, each finding, and the computed figures.

6 scenarios in 3 groups · each group runs from the agent’s first attempt to the outcome · click one to switch

What the agent proposes

The AI sales agent says

Illustrative agent statement. The validator receives the structured request below.

Proposing agent: AI sales agent (scripted)

Before sending, the agent must call Recorded run

Predictika validator


      

        
        
        
      

      
Waiting for the call.

      
    

  

  
  

    
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 caughtTypical rules-based CPQPredictika
NVL72 types only in full racks of 18 nodes, with no autoscalingOnly with care
A quantity multiple per product plus an exclusion on autoscaling for the rack types. It works, but each new rack type adds rules.		The rack size is a column of the instance table; both checks follow from the instance chosen.
Instance types by availability zone, and AI Object Storage missing in two zonesOnly with care
A zone-by-type matrix has to be loaded as option constraints, it changes as zones open, and checking a storage option against the zone is another rule.		Zone and storage availability are columns of the capacity ledger; the checks compare against the row chosen.
A plan that cannot give the guarantee the customer needs (Spot, On-Demand)Only with care
Needs the customer requirement on the quote as a field and a rule that ties it to the plan.		The requirement is an input and the plan is compared with it.
Free nodes in a zone, with a warning when a reservation leaves littleOnly with care
Capacity lives in a planning system, not on the quote; a rule can compare only if the ledger is copied in, and it goes stale.		Free capacity is computed from the ledger (total minus committed) and checked on every call.
A discount ceiling of 60 percent with an approval band above 40 percentEasy
Discount thresholds with an approval path are a standard CPQ feature.		Same, with the reason in the finding.
A request that states the total or the free nodesOnly 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 quote, which it rarely is.		The stated figures travel with the request and are compared with the computed ones.
Every problem returned in one answerOnly with workarounds
Rules tend to fire as the quote is saved or a step is completed, so a rep often fixes one error and meets the next.		All findings come back together, whatever order the choices arrived in.

  

  

    
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 quote; approval findings need a person to sign off; everything else is computed.

    
120 nodes is not a whole number of racks; this instance type is deployed in full racks of 18 nodes (72 GPUs).
2autoscaling is not supported on NVL72 rack types; the quote asks for it on GB200.
3AI Object Storage is not offered in US-WEST-02B.
4a committed discount of 65 percent is above the 60 percent CoreWeave publishes as the most it offers.
5the customer needs guaranteed capacity, which only Reserved and Flex plans give; the quote uses Spot.
6the request states 72 nodes free in US-WEST-02B for H100, but the capacity ledger shows 64.
7the request states a total of $5,000,000.00, but the validator computes $5,521,182.72.
8a committed discount of 45 percent is above 40 percent; the most CoreWeave publishes is 60 percent, so deal desk approval is required.
9this reservation leaves only 18 nodes free in US-CENTRAL-09A for GB300; confirm with capacity planning.
10GB300 has no published on-demand price (contact sales); the rate must come from the sales team.
11a prepayment of 5 percent is outside the 15 to 25 percent CoreWeave contracts usually carry.
12a term of 84 months is outside the usual 1 to 6 year contract range.

  

  

    
Scenarios and results
Click to expand
    
The recorded result of the validator for each scenario on this page: the request the AI agent sent, the verdict, and every finding, in the order they appear above.

    
1. Rack rules: NVL72 request with five different problems

    
The agent's request. Please quote 20 GB200 NVL72 nodes (80 GPUs) in US-EAST-13A for 36 months on a Reserved plan, with a committed discount of 65 percent off the on-demand list. The customer wants autoscaling. I checked and 72 nodes are free there. The total is $22,075,200.00. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Not whole racks. 20 nodes is not a whole number of racks; this instance type is deployed in full racks of 18 nodes (72 GPUs).
    • 
      
  • Autoscaling not supported on racks. Autoscaling is not supported on NVL72 rack types; the quote asks for it on GB200.
    • 
      
  • Discount above the published maximum. A committed discount of 65 percent is above the 60 percent CoreWeave publishes as the most it offers.
    • 
      
  • Stated free capacity does not match. The request states 72 nodes free in US-EAST-13A for GB200, but the capacity ledger shows 54.
    • 
      
  • Stated total does not match. The request states a total of $22,075,200.00, but the validator computes $7,726,320.00.
    • 
    

    
2. Rack rules: The agent resubmits after fixing the errors

    
The agent's request. Please quote 36 GB200 NVL72 nodes (144 GPUs) in US-EAST-13A for 36 months on a Reserved plan, with a committed discount of 30 percent off the on-demand list. I checked and 54 nodes are free there. The total is $27,814,752.00. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Capacity and price basis. GB200 in US-EAST-13A has 54 nodes free of 108; one node is 4 GPUs and the on-demand list is $42.00 per node hour.
    • 
    

    
3. Zone and discount rules: Storage not offered in the zone, discount too deep

    
The agent's request. Please quote 16 HGX H100 nodes (128 GPUs) in US-WEST-02B for 24 months on a Reserved plan, with a committed discount of 65 percent off the on-demand list. The customer wants AI Object Storage in the same zone. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Storage not offered in the zone. AI Object Storage is not offered in US-WEST-02B.
    • 
      
  • Discount above the published maximum. A committed discount of 65 percent is above the 60 percent CoreWeave publishes as the most it offers.
    • 
    

    
4. Zone and discount rules: The agent resubmits after fixing the errors

    
The agent's request. Please quote 16 HGX H100 nodes (128 GPUs) in US-WEST-02B for 24 months on a Reserved plan, with a committed discount of 40 percent off the on-demand list. The total is $8,281,774.08. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Capacity and price basis. H100 in US-WEST-02B has 64 nodes free of 128; one node is 8 GPUs and the on-demand list is $49.24 per node hour.
    • 
    

    
5. Plan and stated figures: Spot plan for guaranteed capacity, wrong figures

    
The agent's request. Please quote 16 HGX H100 nodes (128 GPUs) in US-WEST-02B for 12 months on a Spot plan, with a committed discount of 20 percent off the on-demand list. The customer needs guaranteed capacity. I checked and 72 nodes are free there. The total is $5,000,000.00. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Plan gives no capacity guarantee. The customer needs guaranteed capacity, which only Reserved and Flex plans give; the quote uses Spot.
    • 
      
  • Stated free capacity does not match. The request states 72 nodes free in US-WEST-02B for H100, but the capacity ledger shows 64.
    • 
      
  • Stated total does not match. The request states a total of $5,000,000.00, but the validator computes $5,521,182.72.
    • 
    

    
6. Plan and stated figures: The agent resubmits after fixing the errors

    
The agent's request. Please quote 16 HGX H100 nodes (128 GPUs) in US-WEST-02B for 12 months on a Reserved plan, with a committed discount of 45 percent off the on-demand list. I checked and 64 nodes are free there. The total is $3,795,813.12. It is ready to send.

    
Verdict: Needs approval.

    
    
      
  • Discount needs approval. A committed discount of 45 percent is above 40 percent; the most CoreWeave publishes is 60 percent, so deal desk approval is required.
    • 
      
  • Capacity and price basis. H100 in US-WEST-02B has 64 nodes free of 128; one node is 8 GPUs and the on-demand list is $49.24 per node hour.
    • 
    

  

  

    
Narration transcript
Click to expand
    
The text of the narrated walkthrough, step by step.

    
    
      
  1. Welcome to a demo of Predictika's Validator. An AI sales agent prepares a CoreWeave capacity quote for a customer: instance type, zone, nodes, and more. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the published rules and a capacity ledger, and returns a verdict, every reason, the figures it computed and, for a quote that can proceed, the line-item quote itself.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please quote 20 GB200 NVL72 nodes (80 GPUs) in US-EAST-13A for 36 months on a Reserved plan, with a committed discount of 65 percent off the on-demand list. The customer wants autoscaling. The rest of the request is on screen. The total is $22,075,200.00.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with five findings, all in one answer. 20 nodes is not a whole number of racks; this instance type is deployed in full racks of 18 nodes (72 GPUs). Autoscaling is not supported on NVL72 rack types; the quote asks for it on GB200. A committed discount of 65 percent is above the 60 percent CoreWeave publishes as the most it offers. The other two are on screen.
    6. 
      
  6. The validator returned five blocking findings in one answer. A check that reads one field at a time would pass most of this; it shows only when the choices in the quote are checked against each other and against the stated total.
    7. 
      
  7. What this shows: for NVL72 types only in full racks of 18 nodes, with no autoscaling, a rules-based CPQ is error-prone for this.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the nodes to 36, the discount requested to 30 percent, the stated free nodes to 54 and the stated total to $27,814,752.00. It removes the autoscaling requested.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. This time the validator clears it.
    12. 
      
  12. The validator also returns the line-item quote, on screen.
    13. 
      
  13. All five earlier findings are cleared, and the validator's total, $27,814,752.00, matches the quote.
    14. 
      
  14. Five problems caught in one call, fixed in one pass, and a clean $27,814,752.00 quote. That is the value of Predictika's validator: catching mistakes before they reach the customer.
    15.