Predictika · Cloud GPU instance orders

An AI sales agent's GPU instance order, checked before it goes out

An AI sales agent prepares an order for Scaleway GPU and general-purpose Instances: an instance type, a zone, a number of instances, the workload the customer needs to run (GPUs, GPU memory, vCPUs, RAM, bandwidth, NVLink, MIG, scratch storage), a boot volume class and size, and a data volume. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the Instance catalog and returns a verdict, every reason and the totals it computed.

Scope: Scaleway Instances: L4, L40S, H100, H100-SXM and B300-SXM GPU types and six POP2 types, five zones and three block storage classes, with the workload fit, storage and totals checks. See everything this validator covers

This validator models Scaleway Instance orders, not just the scenarios shown below.

  • Zone and workload fit — which instance types are offered in which zone, and whether the type has enough GPUs, GPU memory, vCPUs, RAM and public bandwidth for the workload.
  • GPU features and storage — NVLink for multi-GPU jobs, MIG partitioning, scratch storage, the types documented as not made for graphics or video encoding, the boot volume class and size, 15K IOPS and the data volume limit.
  • Totals and approvals — the stated GPU total and the stated hourly total compared with the computed ones, orders of 16 or more GPUs sent for capacity confirmation, and an hourly total above 100 euros sent for sign-off.

Any Scaleway Instance order that touches these rules is in scope — not only the scenarios pictured here. Not covered: other Instance ranges, the fr-par-3 zone, Windows images, Elastic Metal, Kubernetes, networking, savings plans and taxes. Public versus modeled: instance specifications, zones and unit prices come from Scaleway's public documentation and product catalog as read in October 2026; the approval thresholds, the boot volume size limits, the NVLink requirement for multi-GPU jobs and the 15K IOPS rule are modeled by Predictika from comparable cloud products and are not published by Scaleway.

The AI sales agentBuilds the GPU instance order 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
A workload that has to fit the instance typeOnly with care
GPUs, GPU memory, vCPUs, RAM and bandwidth are separate limits per instance type, so each type needs its own set of comparisons.		The workload needs and the instance type are fields of one order and are compared on every call.
Features that only some GPU types haveOnly with care
NVLink, MIG, scratch storage and the graphics note each apply to a different list of types, so every rule needs its own table.		Each feature is checked against the chosen type and is a separate finding of the same call.
Types that are offered in some zones onlyOnly with care
A zone table per type has to be kept in step with the catalog.		The type and the zone are fields of one order and are checked together.
Stated GPU and hourly totals that must match the computed onesOnly with workarounds
Both totals are built from the quantity, the type, the volumes and the storage class, so extra formulas are needed to recompute them and compare them with the stated figures.		The stated figures are fields of the request and are compared with the computed totals on every call.
Orders that go to a person for capacity or spend sign-offOnly with care
A separate approval path with its own thresholds that must not overlap the blocking rules.		Blocks and approvals come back together; an approval-only order is not blocked.

  

  

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

    
1This instance type is not offered in the selected zone.
2The workload needs 8 GPUs per instance; this instance type has 4.
3The workload needs 400 GB of GPU memory per instance; this instance type has 192 GB.
4The workload needs 40 vCPUs per instance; this instance type has 32.
5The workload needs 500 GB of RAM per instance; this instance type has 384 GB.
6The workload needs 12,000 Mbps of public bandwidth; this instance type has 10,000 Mbps.
7The workload needs NVLink between GPUs and this instance type has none.
8The workload needs NVIDIA MIG partitioning and this instance type does not support it.
9The workload needs scratch storage and this instance type has none.
10This instance type is not made for graphics or video encoding workloads.
11This instance type boots on a low-latency (5K or 15K) block volume; the legacy SSD class is not allowed.
12The workload needs 15K IOPS and the quote uses a slower block volume class.
13The boot volume is 8 GB; the minimum is 10 GB.
14The data volume is 10,001 GB; a single block volume is limited to 10,000 GB.
15The order states 4 GPUs in total; the quoted instances give 8.
16The order states an hourly total of 12,305,281 (millionths of a euro) and the validator computes 12,305,280.
17The order totals 16 GPUs; a person should confirm capacity before the order goes out.
18The hourly total is above 100 euros; a person should sign off on the spend.
19The order totals 32 GPUs; a person should confirm capacity before the order goes out.
20The workload needs 1 GPUs per instance; this instance type has 0.
21The workload needs 8 GPUs per instance; this instance type has 2.
22The workload needs 600 GB of GPU memory per instance; this instance type has 160 GB.
23The workload needs 100 vCPUs per instance; this instance type has 48.
24The workload needs 900 GB of RAM per instance; this instance type has 480 GB.
25The order states 16 GPUs in total; the quoted instances give 4.

  

  

    
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. Eight-GPU training job: An eight-GPU NVLink training order on two H100 PCIe 2-GPU instances

    
The agent's request. Please quote 2 H100-2-80G instances in fr-par-2 for a workload that needs 8 GPUs, 600 GB of GPU memory, 100 vCPUs, 900 GB of RAM, 15,000 Mbps of public bandwidth, NVLink, and 15K IOPS, with a 100 GB Low Latency 5K boot volume and a 2,000 GB data volume. The order states 16 GPUs in total and an hourly total of 12.012 euros. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Not enough GPUs per instance. The workload needs 8 GPUs per instance; this instance type has 2.
    • 
      
  • Not enough GPU memory. The workload needs 600 GB of GPU memory per instance; this instance type has 160 GB.
    • 
      
  • Not enough vCPUs. The workload needs 100 vCPUs per instance; this instance type has 48.
    • 
      
  • Not enough RAM. The workload needs 900 GB of RAM per instance; this instance type has 480 GB.
    • 
      
  • Multi-GPU job needs NVLink. The workload needs NVLink between GPUs and this instance type has none.
    • 
      
  • 15K IOPS needed. The workload needs 15K IOPS and the quote uses a slower block volume class.
    • 
      
  • Stated GPU total does not match. The order states 16 GPUs in total; the quoted instances give 4.
    • 
    

    
2. Eight-GPU training job: The agent resubmits after fixing the errors

    
The agent's request. Please quote 1 H100-SXM-8-80G instance in fr-par-2 for a workload that needs 8 GPUs, 600 GB of GPU memory, 100 vCPUs, 900 GB of RAM, 15,000 Mbps of public bandwidth, NVLink, and 15K IOPS, with a 100 GB Low Latency 15K boot volume and a 2,000 GB data volume. The order states 8 GPUs in total and an hourly total of 25.7025 euros. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Order summary. 1 instance(s), 8 GPUs in total, 8 GPUs and 640 GB of GPU memory per instance, hourly total 25,702,500 (millionths of a euro).
    • 
    

    
3. Memory-optimized instances in Milan: Three POP2 high-memory instances in Milan with an 8 GB boot volume

    
The agent's request. Please quote 3 POP2-HM-8C-64G instances in it-mil-1 for a workload that needs 8 vCPUs, 60 GB of RAM, and 1,000 Mbps of public bandwidth, with an 8 GB Low Latency 5K boot volume and a 500 GB data volume. The order states an hourly total of 1.23912 euros. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Instance type not offered in the zone. This instance type is not offered in the selected zone.
    • 
      
  • Boot volume too small. The boot volume is 8 GB; the minimum is 10 GB.
    • 
      
  • Stated hourly price does not match. The order states an hourly total of 1,239,120 (millionths of a euro) and the validator computes 1,434,120.
    • 
    

    
4. Memory-optimized instances in Milan: The agent resubmits after fixing the errors

    
The agent's request. Please quote 3 POP2-HM-8C-64G instances in nl-ams-1 for a workload that needs 8 vCPUs, 60 GB of RAM, and 1,000 Mbps of public bandwidth, with a 20 GB Low Latency 5K boot volume and a 500 GB data volume. The order states an hourly total of 1.4388 euros. It is ready to send.

    
Verdict: Allowed.

    
    
      
  • Order summary. 3 instance(s), 0 GPUs in total, 0 GPUs and 0 GB of GPU memory per instance, hourly total 1,438,800 (millionths of a euro).
    • 
    

    
5. Large H100-SXM cluster: Four H100-SXM 8-GPU instances with a stated total of 16 GPUs

    
The agent's request. Please quote 4 H100-SXM-8-80G instances in fr-par-2 for a workload that needs 8 GPUs, 600 GB of GPU memory, 100 vCPUs, 900 GB of RAM, 15,000 Mbps of public bandwidth, NVLink, MIG partitioning, scratch storage, and 15K IOPS, with a 200 GB Low Latency 15K boot volume and an 8,000 GB data volume. The order states 16 GPUs in total and an hourly total of 107.1288 euros. It is ready to send.

    
Verdict: Blocked.

    
    
      
  • Stated GPU total does not match. The order states 16 GPUs in total; the quoted instances give 32.
    • 
      
  • GPU capacity to confirm. The order totals 32 GPUs; a person should confirm capacity before the order goes out.
    • 
      
  • Hourly spend sign-off. The hourly total is above 100 euros; a person should sign off on the spend.
    • 
    

    
6. Large H100-SXM cluster: The agent resubmits after fixing the errors

    
The agent's request. Please quote 4 H100-SXM-8-80G instances in fr-par-2 for a workload that needs 8 GPUs, 600 GB of GPU memory, 100 vCPUs, 900 GB of RAM, 15,000 Mbps of public bandwidth, NVLink, MIG partitioning, scratch storage, and 15K IOPS, with a 200 GB Low Latency 15K boot volume and an 8,000 GB data volume. The order states 32 GPUs in total and an hourly total of 107.1288 euros. It is ready to send.

    
Verdict: Needs approval.

    
    
      
  • GPU capacity to confirm. The order totals 32 GPUs; a person should confirm capacity before the order goes out.
    • 
      
  • Hourly spend sign-off. The hourly total is above 100 euros; a person should sign off on the spend.
    • 
      
  • Order summary. 4 instance(s), 32 GPUs in total, 8 GPUs and 640 GB of GPU memory per instance, hourly total 107,128,800 (millionths of a euro).
    • 
    

  

  

    
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 an order for Scaleway GPU and general-purpose Instances: an instance type, a zone, a number of instances, and more. Before it goes out, the agent sends it to Predictika, which checks it against a declarative model of the Instance catalog and returns a verdict, every reason and the totals it computed.
    2. 
      
  2. I will start with the first request by the agent.
    3. 
      
  3. The agent says. Please quote 2 H100-2-80G instances in fr-par-2 for a workload that needs 8 GPUs, 600 GB of GPU memory, 100 vCPUs. The order states 16 GPUs in total and an hourly total of 12.012 euros. The rest of the request is on screen.
    4. 
      
  4. Before it goes out, the agent calls the validator.
    5. 
      
  5. Blocked, with seven findings, all in one answer. The workload needs 8 GPUs per instance; this instance type has 2. The workload needs 600 GB of GPU memory per instance; this instance type has 160 GB. The workload needs 100 vCPUs per instance; this instance type has 48. The other four are on screen.
    6. 
      
  6. The validator returned seven 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 order are checked against each other and against the stated total.
    7. 
      
  7. What this shows: for stated GPU and hourly totals that must match the computed ones, a rules-based CPQ can do this only with workarounds.
    8. 
      
  8. Now the agent resubmits.
    9. 
      
  9. The corrected request. It changes the instance type to H100-SXM-8-80G, the instances to 1, the boot volume to 100 GB, Low Latency 15K, the stated Total GPUs to 8 and the stated hourly total to 25.7025.
    10. 
      
  10. It calls the validator again.
    11. 
      
  11. This time the validator clears it.
    12. 
      
  12. All seven earlier findings are cleared.
    13. 
      
  13. Seven problems caught in one call, fixed in one pass, and a clean order. That is the value of Predictika's validator: catching mistakes before they reach the customer. Now you can try the other scenarios yourself.
    14.