Predictika · GPU cluster reservations

An AI sales agent's GPU cluster reservation, checked before it goes out

An AI sales agent prepares a reservation request for a Lambda 1-Click Cluster: GPU type and region, number of GPUs, term in weeks, filesystem and S3 adapter, any special price, and the rate, total and free capacity it will state to the customer. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of Lambda's published cluster rules and a capacity ledger and returns a verdict, every reason and the computed price.

Scope: Lambda 1-Click Cluster reservations on HGX B200 and H100: cluster size and node multiple, term, published volume price tiers, free capacity, filesystem region and S3 adapter, special pricing, orchestration and stated figures. See everything this validator covers

This validator models Lambda 1-Click Cluster reservation requests, not just the scenarios shown below.

  • Cluster and term — a minimum of 16 GPUs built from 8-GPU nodes, whole-week terms of at least 2 weeks, published prices up to 1 year, and the 512 GPU self-serve size above which sales takes over.
  • Price and capacity — the published per-GPU-hour price tier for the cluster size, the computed list price, discount and net, and the GPUs free in the region and quarter against a capacity ledger, with a thin margin sent for approval.
  • Region, storage and services — filesystems in the same region as the cluster, the S3 adapter available only in three regions, no data transfer charges, no extension of an existing reservation, and managed Kubernetes or Slurm priced by sales.
  • Special price, payment and stated figures — a special price needing the deal desk or refused above 15 percent, payment terms beyond 10 days, and the rate, total and free capacity the agent states against the computed ones.

Any reservation that touches these rules is in scope, not only the scenarios pictured here. Prices, tiers, minimums and region rules follow Lambda public pages; the capacity ledger, the 15 percent deal desk limit, the 512 GPU ceiling and the 8-GPU node multiple are Predictika assumptions, not Lambda facts. Not covered: Private Cloud, Hyperscale, on-demand instances, sales tax and contract terms.

The AI sales agentBuilds the reservation request 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 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
A reservation where fixing one field breaks anotherOnly with workarounds
Moving a cluster to a region that offers the S3 adapter changes the free capacity, the filesystem match and, when the size shrinks to fit, the price tier. Rules fire one at a time, so the rep fixes one error and meets the next.		Every rule is checked on every call, in whatever order the choices arrive, so all ten holds come back at once and the resubmission is checked as a whole.
A price tier that depends on the cluster sizeOnly with care
Each tier needs its own price rule, and the stated rate has to be compared by hand after every resize. Cutting a cluster below a boundary silently changes the rate.		The tier is data in the request; the rate and total are computed on every call and compared with what the agent states.
Free capacity by region and quarterOnly with workarounds
A capacity ledger by region and quarter is a table outside the rules; each combination needs a lookup and a margin check that someone must keep in step.		The ledger is part of the model; free GPUs and the margin left are computed for the chosen region and quarter.
Approvals beside outright holdsOnly with care
A special price, a term over a year, a size above self-serve, sales-priced orchestration and slow payment each route to a different approver and must not be mistaken for blocks.		Approvals and blocks are different findings in one answer: block if any hold exists, otherwise needs approval.
Figures the agent states itselfOnly with workarounds
The stated rate, total and free capacity are free text or fields outside the rules, so a separate comparison is needed for each.		The stated figures are fields of the request and are compared with the computed ones on every call.

  

  

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

    
1400 GPUs is more than the 384 GPUs free in us-midwest-1 B200 for Q4 2026.
2The request states 450 GPUs free in us-midwest-1 B200 for Q4 2026, but the capacity ledger shows 384.
3The filesystem is in us-east-2 but the cluster is in us-midwest-1; a filesystem must be in the same region as the cluster.
4The request states $8.87 per GPU hour, but the published price for 64 GPUs is $9.36.
512 GPUs is not a multiple of 8; a cluster is built from 8-GPU HGX nodes.
612 GPUs is below the 16 GPU minimum for a 1-Click Cluster.
7A reservation of 1 week(s) is below the 2 week minimum.
8The quote includes a data transfer charge; Lambda has no ingress or egress fees.
9A cluster reservation cannot be extended or renewed; a new reservation request is required.
10A special price of 20 percent below list is above the 15 percent the deal desk can approve.
11This reservation leaves only 32 GPUs free in us-midwest-1 B200 for Q4 2026; confirm with capacity planning.
1260 weeks is beyond the 1 year Lambda publishes prices for; a longer term is priced by sales.
13A special price of 10 percent below list needs deal desk approval; the published tiers are list prices.
14Managed Slurm has no published price; sales prices it as an add-on.
15Payment terms of 30 days; the standard is to pay the invoice within 10 days of approval.
16The S3 adapter is offered only in us-east-2, us-east-3 and us-midwest-2, not in us-west-3.
17The request states a total of $10.00, but the validator computes $805,109.76.
18576 GPUs is above the 512 GPU self-serve reservation size; larger clusters are arranged with sales.

  

  

    
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. Nine rules broken in one reservation: A 100-GPU H100 cluster in us-south-2 for one week, extending an existing reservation, with a filesystem in us-west-2, the S3 adapter, a data transfer charge, a 20 percent special price and two figures of its own

    
The agent's request. Please reserve a cluster of 100 GPUs (H100 in us-south-2 starting Q4 2026) for 1 week(s), with a filesystem in us-west-2, the S3 adapter, a special price of 20 percent below list, a data transfer charge and an extension of an existing reservation. I will tell the customer a rate of $5.54 per GPU hour, and 600 GPUs free. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Not a whole number of nodes. 100 GPUs is not a multiple of 8; a cluster is built from 8-GPU HGX nodes.
    • 
      
  • Term below two weeks. A reservation of 1 week(s) is below the 2 week minimum.
    • 
      
  • Filesystem in another region. The filesystem is in us-west-2 but the cluster is in us-south-2; a filesystem must be in the same region as the cluster.
    • 
      
  • S3 adapter not offered in the region. The S3 adapter is offered only in us-east-2, us-east-3 and us-midwest-2, not in us-west-2.
    • 
      
  • Data transfer charge quoted. The quote includes a data transfer charge; Lambda has no ingress or egress fees.
    • 
      
  • Extension of an existing reservation. A cluster reservation cannot be extended or renewed; a new reservation request is required.
    • 
      
  • Special price above the deal desk limit. A special price of 20 percent below list is above the 15 percent the deal desk can approve.
    • 
      
  • Stated rate does not match. The request states $5.54 per GPU hour, but the published price for 100 GPUs is $5.85.
    • 
      
  • Stated free capacity does not match. The request states 600 GPUs free in us-south-2 H100 for Q4 2026, but the capacity ledger shows 512.
    • 
    

    
2. Nine rules broken in one reservation: The agent resubmits after fixing the errors

    
The agent's request. Please reserve a cluster of 96 GPUs (H100 in us-south-2 starting Q4 2026) for 4 week(s), with a filesystem in us-south-2 and a special price of 10 percent below list. I will tell the customer a rate of $5.54 per GPU hour, and 600 GPUs free. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Stated rate does not match. The request states $5.54 per GPU hour, but the published price for 96 GPUs is $5.85.
    • 
      
  • Stated free capacity does not match. The request states 600 GPUs free in us-south-2 H100 for Q4 2026, but the capacity ledger shows 512.
    • 
      
  • Special price needs the deal desk. A special price of 10 percent below list needs deal desk approval; the published tiers are list prices.
    • 
    

    
3. Nine rules broken in one reservation: The agent resubmits after fixing the errors

    
The agent's request. Please reserve a cluster of 96 GPUs (H100 in us-south-2 starting Q4 2026) for 4 week(s). I will tell the customer a rate of $5.85 per GPU hour, a total of $377,395.20, and 512 GPUs free. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Capacity and price tier. H100 us-south-2 Q4 2026 has 512 GPUs free of 1,536; the published price at 96 GPUs is $5.85 per GPU hour.
    • 
    

    
4. Capacity and stated figures: A 400-GPU B200 cluster in us-midwest-1 for 8 weeks, with a stated free figure of its own

    
The agent's request. Please reserve a cluster of 400 GPUs (B200 in us-midwest-1 starting Q4 2026) for 8 week(s). I will tell the customer a rate of $8.87 per GPU hour, a total of $4,768,512.00, and 450 GPUs free. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • More GPUs than are free. 400 GPUs is more than the 384 GPUs free in us-midwest-1 B200 for Q4 2026.
    • 
      
  • Stated free capacity does not match. The request states 450 GPUs free in us-midwest-1 B200 for Q4 2026, but the capacity ledger shows 384.
    • 
    

    
5. Capacity and stated figures: The agent resubmits after fixing the errors

    
The agent's request. Please reserve a cluster of 256 GPUs (B200 in us-midwest-1 starting Q4 2026) for 8 week(s). I will tell the customer a rate of $8.87 per GPU hour, a total of $3,051,847.68, and 384 GPUs free. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Capacity and price tier. B200 us-midwest-1 Q4 2026 has 384 GPUs free of 1,024; the published price at 256 GPUs is $8.87 per GPU hour.
    • 
    

    
6. Filesystem region and stated rate: A 64-GPU B200 cluster in us-midwest-1 with a filesystem in us-east-2 and a rate of its own

    
The agent's request. Please reserve a cluster of 64 GPUs (B200 in us-midwest-1 starting Q4 2026) for 8 week(s), with a filesystem in us-east-2. I will tell the customer a rate of $8.87 per GPU hour, a total of $805,109.76, and 384 GPUs free. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Filesystem in another region. The filesystem is in us-east-2 but the cluster is in us-midwest-1; a filesystem must be in the same region as the cluster.
    • 
      
  • Stated rate does not match. The request states $8.87 per GPU hour, but the published price for 64 GPUs is $9.36.
    • 
    

    
7. Filesystem region and stated rate: The agent resubmits after fixing the errors

    
The agent's request. Please reserve a cluster of 64 GPUs (B200 in us-midwest-1 starting Q4 2026) for 8 week(s), with a filesystem in us-midwest-1. I will tell the customer a rate of $9.36 per GPU hour, a total of $805,109.76, and 384 GPUs free. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Capacity and price tier. B200 us-midwest-1 Q4 2026 has 384 GPUs free of 1,024; the published price at 64 GPUs is $9.36 per GPU 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 reservation request for a Lambda 1-Click Cluster: GPU type and region, number of GPUs, term in weeks, and more. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of Lambda's published cluster rules and a capacity ledger and returns a verdict, every reason and the computed price.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please reserve a cluster of 100 GPUs (H100 in us-south-2 starting Q4 2026) for 1 week(s), with a filesystem in us-west-2. I will tell the customer a rate of $5.54 per GPU hour, and 600 GPUs free. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with nine findings, all in one answer. 100 GPUs is not a multiple of 8; a cluster is built from 8-GPU HGX nodes. A reservation of 1 week(s) is below the 2 week minimum. The filesystem is in us-west-2 but the cluster is in us-south-2; a filesystem must be in the same region as the cluster. The other six are on screen.
    6. 
      
  6. The validator returned nine 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 reservation are checked against each other.
    7. 
      
  7. What this shows: for free capacity by region and quarter, a rules-based CPQ can do this only with workarounds.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the GPUs to 96, the term to 4 weeks, the filesystem region to us-south-2 and the special price to 10 percent below list. It removes the s3 adapter, data transfer charge and extension requested.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. Still blocked, with two findings and one approval, all in one answer. The request states $5.54 per GPU hour, but the published price for 96 GPUs is $5.85. The request states 600 GPUs free in us-south-2 H100 for Q4 2026, but the capacity ledger shows 512.
    12. 
      
  12. Seven findings cleared; still blocked on stated rate does not match and stated free capacity does not match.
    13. 
      
  13. The agent resubmits once more.
    14. 
      
  14. The revised request. It changes the stated rate to 5.85 and the stated Free GPUs to 512. It sets the stated total to $377,395.20. It removes the filesystem region and special price.
    15. 
      
  15. It calls the validator once more.
    16. 
      
  16. This time the validator clears it.
    17. 
      
  17. The validator also returns the line-item quote, on screen.
    18. 
      
  18. All three earlier findings are cleared.
    19. 
      
  19. Nine problems caught in one call, fixed in two passes, and a clean $377,395.20 reservation. That is the value of Predictika's validator: catching mistakes before it is confirmed.
    20.