Request access
Private Preview

OperateOrchestrator Private Preview

Policy in.
Placement out.

The Fantasti Orchestrator is the control plane under every product. It checks each request against your policy, finds capacity and quota, places the work on one fabric, recovers it after a spot stop and meters it onto one statement. Every placement leaves a record of where it landed and why.

Request access Read the API

Opens with the first cohort. No account is open yet.

TAB 01 · Placement record Illustrative
TAB 01Placement record Illustrative
GPUs
Fabric
Region
Capacity
Policy 
placement: { fabric: infiniband, regions: ["us-*"] }
Request 
fx.clusters.create(gpu="H100:8", nodes=8,
    capacity=fantasti.OnDemand())
Placement record, illustrative. Placed in us-east, fabric a.
Pool Region Fabric Free GPUs Result
pool-b us-east · fabric b us-east InfiniBand · fabric b 32 Rejected: 32 free < 64
pool-c us-west · no InfiniBand us-west No InfiniBand 96 Rejected: no InfiniBand
pool-d eu-west · fabric d eu-west InfiniBand · fabric d 128 Rejected: region not allowed
pool-e eu-central · fabric e eu-central InfiniBand · fabric e 320 Rejected: region not allowed
pool-f us-central · fabric f us-central InfiniBand · fabric f 64 Rejected: no on-demand capacity
pool-g us-east · fabric g us-east InfiniBand · fabric g 128 Rejected: other GPU type
pool-a us-east · fabric a us-east InfiniBand · fabric a 96 Placed
Result

placed · us-east · fabric a · 64×H100 SXM5 · one fabric

Change a constraint to see which pools qualify.

The request was 64 × H100 SXM5 · InfiniBand · us-* · on-demand. 7 pools were checked. pool-b in us-east, fabric b was rejected: 32 free < 64. pool-c in us-west, no InfiniBand was rejected: no InfiniBand. pool-d in eu-west, fabric d was rejected: region not allowed. pool-e in eu-central, fabric e was rejected: region not allowed. pool-f in us-central, fabric f was rejected: no on-demand capacity. pool-g in us-east, fabric g was rejected: other GPU type. The job was placed in us-east, fabric a: 64 H100 SXM5 GPUs on one InfiniBand fabric.

§01 Policy policy.yaml

Constraints, written once.

A policy sets what a team can run, where, on which capacity and at what price. It applies to every request from that team, whether it comes from a person, a pipeline or an agent.

  • placementGPU types, the fabric and the regions
  • capacityon-demand, spot and reserved, with your ceilings
  • limitsGPUs, sandboxes and monthly spend
  • cost_centerthe statement line a request bills to
YAML
Preview API · subject to change 
# policy.yaml
# applies to every request from team "research"
team: research
placement:
  gpus: [H100, H200, B200]
  # multi-node jobs stay on one fabric
  fabric: infiniband
  regions: ["us-*"]
capacity:
  allow: [on_demand, spot, reserved]
  spot:
    # USD per GPU-hour, your ceilings
    max_price: { H100: 2.40, H200: 3.10 }
limits:
  gpus: { H200: 32 }
  sandboxes: { concurrent: 200, max_ttl: 60m }
  spend: { monthly_usd: 40000 }
cost_center: research-llm
§02 Broker 4 signals

Set the ceiling. The Broker works the market.

The Broker is the part of the Orchestrator that buys capacity for your work. It reads availability and spot prices across the regions, GPU types and fabrics of the capacity layer, and places each request where it fits under the max price you set.

What it reads

  • 01 Availability How much capacity of each GPU type is free in each region and fabric, and whether a request of your size fits on one fabric.
  • 02 Spot price Where the current spot price sits, and how it has moved.
  • 03 Reclaim history How often spot capacity of that type has been reclaimed recently, from the Orchestrator's own records.
  • 04 Your job GPU count, fabric requirement, checkpoint interval, deadline and budget.
§03 Spot 3 modes

Three ways to run on spot.

  1. Follow the market.

    No ceiling. You pay the spot price for each interval.

    fantasti.Spot(max_price="follow")

  2. Set a max price.

    You choose the most you will pay per GPU-hour. Above it, those nodes stop.

    fantasti.Spot(max_price=3.10)

  3. Let the Broker manage it.

    Planned

    Give the Broker a ceiling per GPU-hour and a budget for the run. It sets the max price each time your nodes start, at or below your ceiling, and waits when the market is above it. You pay the spot price, never the max and never more than your budget.

    fantasti.Spot(max_price="auto", ceiling=3.10)
Planned · subject to change 
import fantasti

fx = fantasti.Client()

run = fx.batch.create(
    input="s3://acme-evals/prompts.jsonl",
    gpu="H200:8",
    # never above 3.10 per GPU-hour
    capacity=fantasti.Spot(max_price="auto", ceiling=3.10),
    budget=fantasti.Budget(usd=1800, deadline="72h"),
    # optional: finish on on-demand, inside the budget
    fallback=fantasti.OnDemand(),
    start=False,
)

plan = run.broker.plan()  # nothing has started
print(plan.fits, plan.outlook, plan.windows)
run.start()
{ "capacity": { "mode": "spot", "max_price": "auto", "ceiling": "3.10",
                "currency": "USD", "unit": "gpu_hour" },
  "budget":   { "amount": "1800.00", "currency": "USD", "deadline": "72h" },
  "fallback": { "mode": "on_demand" } }
fantasti.broker_plan   { "gpu": "H200:8", "ceiling": "3.10",
                         "budget": "1800.00", "deadline": "72h" }

Output · illustrative

fits: true · outlook: low, low, medium · windows: 3
  • run.broker.plan()

    Before anything starts, the Broker shows its plan: the windows it expects to run in, the outlook for each, and whether the run fits your budget and deadline. If it does not fit, it says so and starts nothing.

  • max_price="auto"

    A max price on running nodes changes at a restart from a checkpoint, never in the middle of a step.

  • fallback=fantasti.OnDemand()

    If you allow it, the Broker finishes on on-demand capacity when spot cannot meet your deadline. On-demand hours count against the same budget.

§04 Interruption outlook 24 intervals

Interruption outlook.

Planned

For each GPU type, region and fabric, the Broker publishes an outlook for the coming hours: low, medium or high. It is an estimate from availability, price pressure and reclaim history. It is not a guarantee, and work under a low outlook can still be interrupted.

Use it to pick a window or to size your checkpoints, or let the Broker decide.

PLT 02Interruption outlook Illustrative series · not Fantasti market data
Scope
us-east · fabric a
Job
H200:8
Selected
14 h in 3 windows
Outlook
Low · Low · Medium
Plan
Fits
Interruption outlook for us-east · fabric a, illustrative. One row per GPU type, one column per hour.
GPU type Now+1 h+2 h+3 h+4 h+5 h+6 h+7 h+8 h+9 h+10 h+11 h+12 h+13 h+14 h+15 h+16 h+17 h+18 h+19 h+20 h+21 h+22 h+23 h
B300 High High High High Medium Medium Low Low Low Low Low Medium Medium High High High High High Medium Low Low Low Low Medium
B200 Medium Medium Low Low Low Low Low Low Medium High High High Medium Medium Low Low Low Low Low Low Medium Medium High High
H200 SXM Low, selected Low, selected Low, selected Low, selected Low, selected High High High Low, selected Low, selected Low, selected Low, selected High High Medium, selected Low, selected Medium, selected Medium, selected Low, selected High High High Low Low
H100 SXM5 Low Low Low Low Low Low Low Low Medium Medium Low Low Low Low Low Low High High Medium Low Low Low Low Low
RTX PRO 6000 Low Low Medium Low Low Low Low Low Low Low Low Low Low Medium Medium Low Low Low Low Low Low Low Low Low

Illustrative series, not Fantasti market data. Outlook for us-east · fabric a over the coming 24 hours, in three bands: low, medium, high. B300: high for 4 h, medium for 2 h, low for 5 h, medium for 2 h, high for 5 h, medium for 1 h, low for 4 h, medium for 1 h. B200: medium for 2 h, low for 6 h, medium for 1 h, high for 3 h, medium for 2 h, low for 6 h, medium for 2 h, high for 2 h. H200 SXM: low for 5 h, high for 3 h, low for 4 h, high for 2 h, medium for 1 h, low for 1 h, medium for 2 h, low for 1 h, high for 3 h, low for 2 h. H100 SXM5: low for 8 h, medium for 2 h, low for 6 h, high for 2 h, medium for 1 h, low for 5 h. RTX PRO 6000: low for 2 h, medium for 1 h, low for 10 h, medium for 2 h, low for 9 h. For a job on H200:8 that needs 14 hours under a ceiling of 3.10 per GPU-hour, the Broker selects 3 windows, 14 hours in all, with outlooks low, low, medium. The plan fits. An outlook is an estimate, not a guarantee.

§05 Broker record 5 decisions

Every choice, with its reason.

The Broker record sits beside the placement record. A written rule makes each choice. A language model writes the sentence that explains it.

Planned

TAB 03Broker record for one run Illustrative
Broker record H200:8 · spot One run No new starts
Job
H200:8
Ceiling
3.10 USD per GPU-hour
Budget
1,800 USD
Deadline
72h
  1. 01. Running.
    Signal
    Outlook low · spot under ceiling
    Decision
    Start · max set under ceiling
    Rule
    Allowed: inside ceiling 3.10
  2. 02. Waiting.
    Signal
    Spot above ceiling
    Decision
    Wait
    Rule
    Not allowed: above ceiling
  3. 03. Stopped.
    Signal
    Reclaim notice
    Decision
    Checkpoint, stop
    Rule
    Notice received
  4. 04. Running.
    Signal
    Outlook medium · spot under ceiling
    Decision
    Resume from checkpoint
    Rule
    Allowed: inside ceiling
  5. 05. Held.
    Signal
    Budget nearly used
    Decision
    No new starts
    Rule
    Budget guard

Illustrative. The Broker record of one spot run of H200:8, with a ceiling of 3.10 USD per GPU-hour, a budget of 1,800 USD and a deadline of 72h. 5 decisions in order. 1: signal, Outlook low · spot under ceiling; decision, Start · max set under ceiling; rule, inside ceiling 3.10. 2: signal, Spot above ceiling; decision, Wait; rule, above ceiling. 3: signal, Reclaim notice; decision, Checkpoint, stop; rule, Notice received. 4: signal, Outlook medium · spot under ceiling; decision, Resume from checkpoint; rule, inside ceiling. 5: signal, Budget nearly used; decision, No new starts; rule, Budget guard.

Sheet
01 / 03
Title
Broker record
Scope
One spot run
Stamp
Illustrative
§06 One fabric Two fabrics, one pool

One job, one fabric.

InfiniBand does not cross fabrics. The Orchestrator places every multi-node job inside a single fabric, rejects pools that cannot hold the whole job, and never presents two fabrics as one. Data moves between fabrics over object storage, and the placement record says so.

  • pool-aus-east · fabric a. 96 GPUs free, 64 needed. Placed.
  • pool-bus-east · fabric b. Passed over: 32 free < 64.
  • pool-cus-west. Passed over: no InfiniBand.

GPU Clusters

One job, one fabric. Illustrative.

Illustrative. A request for 64 H100 SXM5 GPUs, 8 nodes of 8, on one InfiniBand fabric. It is placed on pool-a in us-east, fabric a, where 96 GPUs are free: all 8 nodes sit inside that one fabric. pool-b, fabric b in us-east, is passed over: 32 free < 64. pool-c in us-west is passed over: no InfiniBand. The fabrics are joined only by object storage. InfiniBand does not cross from one fabric to another. Connections: 8 nodes, 64 GPUs to Leaf and spine switches, 8 links; pool-b to Leaf and spine, 4 links; InfiniBand to Object storage; Pool c to Object storage.

Fabric aInfiniBandus-east
Fabric bInfiniBandus-east
Pool cus-west
  • Job8 nodes, 64 GPUsH100 SXM5
    HGX H100 8-GPU baseboard, line drawing8 GPUs per node
    pool
    pool-a
    free
    96
    needs
    64
    result
    placed
    8 nodes on one fabric
  • Leaf and spine switchesOne fabric
  • pool-b32 free
    Free 32Needs 64
    32 free < 64
  • Leaf and spine
  • pool-c96 free
    fabric
    none
    no InfiniBand
  • Between fabricsObject storageInfiniBand does not cross fabrics
Fabric aInfiniBandus-east
Fabric bInfiniBandus-east
Pool cus-west
  • Job8 nodes, 64 GPUsH100 SXM5
    HGX H100 8-GPU baseboard, line drawing8 GPUs per node
    pool
    pool-a
    free
    96
    needs
    64
    result
    placed
    8 nodes on one fabric
  • Leaf and spine switchesOne fabric
  • pool-b32 free
    Free 32Needs 64
    32 free < 64
  • pool-c96 free
    fabric
    none
    no InfiniBand
  • Between fabricsObject storageNo InfiniBand across fabrics
  • Job8 nodes, 64 GPUsH100 SXM5
    HGX H100 8-GPU baseboard, line drawing8 GPUs per node
    pool
    pool-a
    free
    96
    needs
    64
    result
    placed
    8 nodes on one fabric
  • Leaf and spine switches
  • pool-b32 free
    Free 32Needs 64
    32 free < 64
  • pool-c96 free
    fabric
    none
    no InfiniBand
  • Between fabricsObject storage
Fabric a · InfiniBandFabric b · InfiniBandPool c
  • The placed job
  • Data only, never InfiniBand
  • Passed over
§07 Capacity and quota 3 limits

Your quota grows with you.

On most clouds a quota increase is a form and a wait. On Fantasti it is a calculation the Orchestrator runs before you need the answer.

A quota is a control on pay-as-you-go accounts. Prepaid credit has no instance quota: it is limited by its balance and by available capacity. Enterprise quotas are set by contract.

Early access

Three different limits

  • 01 Account quota On pay-as-you-go, how much of each GPU type you can run at once. Granted by Fantasti. It grows automatically, in steps. You can cap it or turn growth off. fx.quotas.list()
  • 02 Team limits Set by your admins. Automation never raises them. Agents and API keys act inside them. gpus: { H200: 32 }
  • 03 Spend controls Spend caps, credit balance and spend threshold are separate. A larger quota does not raise what you may spend. monthly_usd: 40000
§08 Increase 5 steps

How an increase happens.

DWG 05Quota headroom Illustrative
Quota gpu.h200 Account acme Raised
At review
quota 16peak 13headroom 3
Now
quota 24peak 18headroom 6

Decisiondec_4k9t…

Hiro Quota analystAI agent

raised · 16 → 24desk · inside the limits

Illustrative. Demand for H200 GPUs on the account acme rises in steps toward its quota of 16. A dashed review line sits under the quota. When demand reaches the review line, at a peak of 13, the quota is raised one step to 24 and a decision record is written. Demand passes the old quota later and never reaches a ceiling.

  • A quota is permission, not a reservation. It does not guarantee that capacity is free at the moment you ask. To guarantee capacity, reserve it.
  • A refusal or a reduction always goes to senior review.
  • Every change is in your audit log with its decision record.

Who decides.

A high-stakes decision goes to senior review, on the orchestrator tier. The reviewer is a different agent from the one that proposed the decision, and it records its reasoning. Everything else, agents decide inside their playbooks and log.

Who decides a quota change. Illustrative.
Trace a decision

Up to the limits of the Direct plan there is no sales step. The Quota analyst raises your quota in steps and writes a decision record.

Illustrative. The Orchestrator compares peak use with the quota for each GPU type and opens a review near the limit. At the desk, Hiro · Quota analyst · AI agent reads usage, reservations, payments and account standing, and decides alone: quota increases in steps, up to the limits of the Direct plan. Leila · Capacity broker · AI agent decides alone: asking the capacity layer for availability and quotes. Senior review, by the orchestrator tier, is given by Ingrid · Head of Review · AI agent, and takes what they send up: increases above those limits; any refusal or reduction; a spend-threshold step outside the ladder; capacity requests. The reviewer is a different agent from the one that proposed the decision, and it records its reasoning. The owner sets the rules and the hard limits, and performs the few acts only a legal person can perform. Decision: Raise your account quota. Desk agent: Acts, in steps, up to the limits of the Direct plan. Senior review or the owner: Senior review, above those limits. Connections: Orchestrator to Hiro · Quota analyst; Orchestrator to Leila · Capacity broker; Hiro · Quota analyst to Ingrid · Head of Review; Leila · Capacity broker to Ingrid · Head of Review; Desk to The owner; Senior review to The owner.

Level 01DeskAI agents
Level 02Senior reviewThe orchestrator tier
Level 03The ownerRules and hard limits
  • 01OrchestratorH200
    Peak use 13Quota 16
    review
    opened
    trigger
    usage near
    ceiling
  • 02Hiro · Quota analystAI agent
    • Usage
    • Reservations
    • Payments
    • Standing

    Decides alone: quota increases in steps,up to the limits of the Direct plan.

    files
    Decision record · per change
  • 05Leila · Capacity brokerAI agent

    Decides alone: asking the capacity layerfor availability and quotes.

    files
    Capacity report · daily
  • 04Ingrid · Head of ReviewAI agent
    Increases above those limits
    Any refusal or reduction
    A spend-threshold step outside the ladder
    Capacity requests

    The reviewer is a different agent from theone that proposed the decision, and itrecords its reasoning.

    Hears every appeal.

    files
    Review record · per decision
  • RuleDecision rightsSet by the owner
    decision
    Raise your account quota
    decision
    Raise your account quota
    decision
    Refuse or reduce a quota
    decision
    Quote reserved capacity or a rack-scale system
    desk agent
    Acts, in steps, up to the limits of the Direct plan
    desk agent
    Acts, in steps, up to the limits of the Direct plan
    desk agent
    Never
    desk agent
    Drafts, and collects quotes from the capacity layer
    above it
    Senior review, above those limits
    above it
    Senior review, above those limits
    above it
    Senior review
    above it
    Senior review, for every reservation or capacity request
  • The owner

    The rules and limits themselves, and the fewacts only a legal person can perform.Once the desks open, the owner does not workcases. Agents do, under the rules the ownersets.

Level 01DeskAI agents
Level 02Senior reviewThe orchestrator tier
Level 03The ownerRules and hard limits
  • 01OrchestratorH200
    Peak use 13Quota 16
    review
    opened
    trigger
    usage near ceiling
  • Hiro · Quota analystAI agent
    • Usage
    • Reservations
    • Payments
    • Standing

    Decides alone: quota increases insteps, up to the limits of the Directplan.

    files
    Decision record · per
    change
  • Leila · Capacity brokerAI agent

    Decides alone: asking the capacitylayer for availability and quotes.

    files
    Capacity report · daily
  • 04Ingrid · Head of ReviewAI agent
    Increases above those limits
    Any refusal or reduction
    A spend-threshold step outside the ladder
    Capacity requests

    The reviewer is a different agent from the one that proposed the decision, and itrecords its reasoning.

    Hears every appeal.

    files
    Review record · per decision
  • RuleDecision rightsSet by the owner
    decision
    Raise your account quota
    decision
    Raise your account quota
    decision
    Refuse or reduce a quota
    decision
    Quote reserved capacity or a rack-scale system
    desk agent
    Acts, in steps, up to the limits of the Direct plan
    desk agent
    Acts, in steps, up to the limits of the Direct plan
    desk agent
    Never
    desk agent
    Drafts, and collects quotes from the capacity layer
    above it
    Senior review, above those limits
    above it
    Senior review, above those limits
    above it
    Senior review
    above it
    Senior review, for every reservation or capacity request
  • The owner

    The rules and limits themselves, and the few acts only a legal person can perform.Once the desks open, the owner does not work cases. Agents do, under the rules theowner sets.

  • 01OrchestratorH200
    Peak use 13Quota 16
    review
    opened
    trigger
    usage near ceiling
    to
    Leila · Capacity broker
  • 02Hiro · Quota analystAI agent
    • Usage
    • Reservations
    • Payments
    • Standing

    Decides alone: quota increases in steps, up tothe limits of the Direct plan.

    files
    Decision record · per change
    to
    Ingrid · Head of Review
  • 05Leila · Capacity brokerAI agent

    Decides alone: asking the capacity layer foravailability and quotes.

    files
    Capacity report · daily
  • 04Ingrid · Head of ReviewAI agent
    Increases above those limits
    Any refusal or reduction
    A spend-threshold step outside the ladder
    Capacity requests

    The reviewer is a different agent from the onethat proposed the decision, and it records itsreasoning.

    Hears every appeal.

    files
    Review record · per decision
  • RuleDecision rights
    decision
    Raise your account quota
    decision
    Raise your account quota
    decision
    Refuse or reduce a quota
    decision
    Quote reserved capacity or a
    rack-scale system
    desk agent
    Acts, in steps, up to the
    limits of the Direct plan
    desk agent
    Acts, in steps, up to the
    limits of the Direct plan
    desk agent
    Never
    desk agent
    Drafts, and collects quotes
    from the capacity layer
    above it
    Senior review, above those
    limits
    above it
    Senior review, above those
    limits
    above it
    Senior review
    above it
    Senior review, for every
    reservation or capacity
    request
  • The owner

    The rules and limits themselves, and the fewacts only a legal person can perform.Once the desks open, the owner does not workcases. Agents do, under the rules the ownersets.

Level 01 · DeskLevel 02 · Senior reviewLevel 03 · The owner
  • The traced decision
  • A review opens
  • Reads the rules
  • Goes to senior review

  1. Compare

    The Orchestrator compares your recent peak use with your quota for each GPU type.

  2. Read

    Near the limit, the Quota analyst, an AI agent, reads your usage, reservations, payment history and account standing.

  3. Raise

    Up to the limits of the Direct plan there is no sales step. The Quota analyst raises your quota in steps and writes a decision record. You get a notice with the old value, the new value and the reason. The risk rules check every increase before it applies.

  4. Review

    Above those limits, the Quota analyst prepares the case and senior review decides. You see the analysis either way.

  5. Capacity

    Behind your quota, the Orchestrator requests capacity from the capacity layer ahead of demand, and places work where capacity is free. A reservation or a capacity request goes to senior review before it is placed.

§09 Review record 1 review

One review, on record.

Yours to set
Spend caps and budgets belong to you. A quota review never raises them. Lock any quota and it will not change without you.
Ask any time
You can still request more. An agent answers with the analysis, not a queue number, and tells you what would change the result.
Goes to senior review Planned
The Capacity broker requests racks and reservations for a customer through a shared channel with the capacity layer, and lines up both sides. A reservation or a capacity request goes to senior review before it is placed.
Leila Capacity brokerAI agent

The published billing-model rules still apply. They are on the Pricing page.

TAB 07Quota review record Illustrative
Record dec_4k9t… Quota review Raised

Read

account
acme
resource
gpu.h200
trigger
usage near ceiling
commitment
rsv_… · 6 months
payments
in good standing
risk
clear

Decided

decision
raised · 16 → 24
decided by
Hiro Quota analystAI agent
level
desk · inside the limits
review
none
notice
email · Slack

Illustrative. Quota review dec_4k9t for account acme, resource gpu.h200. Trigger: usage near ceiling, at a peak of 13 against a quota of 16. Commitment: a 6-month reservation. Payments: in good standing. Risk: clear. Decision: raised from 16 to 24, decided at the desk by Hiro · Quota analyst · AI agent, inside the limits. Senior review: not needed. Notice sent by email and Slack.

Sheet
02 / 03
Title
Quota review record
Record
dec_4k9t…
Stamp
Illustrative
§10 Controls Preview API

Cap it, lock it, or ask for more.

Read your quota, recent peak and headroom for each GPU type. Set a cap, turn automatic growth off, or request an increase with a reason. Each change emits an event your systems can listen for.

Events

  • quota.increased
  • quota.review_opened
  • quota.request_decided
Python
Preview API · subject to change 
# quota, recent peak and headroom per GPU type
fx.quotas.list()

fx.quotas.auto_growth(
    enabled=True,
    cap={"H200": 64, "B200": 32},
)

fx.quotas.request(
    gpu="B200", to=48,
    reason="January fine-tuning run",
)

Output · illustrative

quotas.list    H200 · quota 24 · peak 18 · headroom 6
auto_growth    on · cap H200 64 · B200 32
request        B200 to 48 · review opened
§11 Limits and cost centers 6 controls

Limits people set and agents cannot raise.

Agents and API keys act inside these limits. Changing a limit requires an organization admin.

MCP server

TAB 08Limits and who sets them Source · Fantasti
Limits and who sets them
ControlWhat it limitsSet by
Team GPU limitlimits.gpusGPUs per type, per teamOrganization admins
Sandbox capslimits.sandboxesConcurrent sandboxes, maximum time-to-live, allowed GPU typesOrganization admins
Spend caplimits.spendMonthly spend per team or projectOrganization admins
Max pricecapacity.spot.max_priceThe most a team pays per GPU-hour on spot, per GPU typeTeam policy
Agent key budgetWhat one agent key may spend before new launches stopWhoever creates the key
Cost centercost_centerWhich statement line a request bills toTeam policy
§12 One statement 3 cost centers

One statement, split by cost center.

Every product on one statement, metered by the second, billed hourly. Each line names its cost center, listing, region and rate, so finance can reconcile without a spreadsheet.

Pricing and billing

TAB 09Statement for one month Illustrative

Statement · Illustrative

Account
acme-research
Period
One month · 730 h
Terms
NET 30 · metered per second, billed hourly
Statement for one month, illustrative. Total $37,173.66.
Cost center Product Listing Pool Metered GPU·s Billed GPU·hr Rate Amount
cc-4102research GPU Clusters, on-demand H100 SXM5 us-east 21,024,000 5,840.00 $5.40 $31,536.00
cc-4102research GPU Instances, on-demand H200 SXM us-east 1,843,200 512.00 $5.94 $3,041.28
Subtotal cc-4102 $34,577.28
cc-2207evals Batch Inference, spot H100 SXM5 us-central 1,476,000 410.00 $2.11 $863.56
cc-2207evals Sandboxes L40S us-east 301,968 83.88 $1.86 $156.02
Subtotal cc-2207 $1,019.58
cc-3310serving Serverless, on-demand RTX PRO 6000 us-central 2,628,000 730.00 $2.16 $1,576.80
Subtotal cc-3310 $1,576.80
Total $37,173.66
TAB 10Also on this statement Illustrative
Also on this statement: usage against a reservation and flat lines. Illustrative.
Cost center Line Billed GPU·hr Rate Amount
cc-4102 GPU Clusters, reserved · B200 · rsv_7d2k 11,680.00 Order form Per order form
account Support plan · Direct Monthly Flat $1,000.00
§13 Identity 4 controls

Identity

  • 01 Early access

    Single sign-on

    SAML and OIDC connections per organization.

  • 02 Early access

    Role-based access

    Roles per organization and project, applied to every product.

  • 03 Private Preview

    API and agent keys

    Scoped, revocable and shown once at creation.

  • 04 Planned

    Audit log

    Account, role, key and resource actions, with export.

Sheet
03 / 03
Title
Limits, statement and identity
Reviewed
2026-10-10
§14 Questions 13 questions

Questions about the Orchestrator.

Placement

Does the Orchestrator move running jobs?

No. It places a job once. If a spot job is stopped, Batch Inference and managed jobs resume on a new placement from the last checkpoint.

Can I keep a workload in one region?

Yes. Restrict regions in the policy.

What happens when no pool qualifies?

The request stays in placing until its timeout, then returns an error listing the closest pools and the constraint each one failed.

The Broker

Does the Broker compete against other accounts for me?

No. The market sets one spot price per GPU type. The Broker decides when your work runs and what max price it accepts.

Can the Broker spend more than my budget?

No. It stops new starts when the budget is reached, and it never sets a max price above your ceiling.

How accurate is the outlook?

It is an estimate in three bands. We will publish how each band performed against what happened once there is enough history to do that honestly.

Is a language model choosing my price?

No. The max price comes from a written rule over market data.

Quota

What does a quota review look at?

Your peak use against your quota, your reservations and prepayments, your payment history and your account's standing with the risk rules. It does not look inside your workloads.

Can I stop my quota from growing?

Yes. Set a cap per GPU type, or turn automatic growth off.

Does a higher quota change my bill?

No. You pay for what runs. Spend caps and your credit balance or spend threshold still apply.

Why did my increase go to senior review?

It was above the limits of the Direct plan, the account is new, or a risk rule asked for a check.

Does my spend threshold rise the same way?

The spend threshold rises automatically, on its published ladder. A step outside the ladder goes to senior review.

Who works on my quota?

Hiro, the Quota analyst, and Leila, the Capacity broker. Both are AI agents. Every agent identifies itself as an AI in its first message and in its signature.
END Request Private Preview

Bring the constraints. We will show you where they place.

Request access Talk to the solutions desk

Next