Cloudability AI Hub

See AI spend across cloud model platforms, direct APIs, and AI embedded in SaaS. Assign it with business dimensions and attribution. Prove it with unit economics and value signals. Every widget shows its data lineage; select the info icon on any card.

Data provenance: Billing (CUR, invoices, exports) AI telemetry (logs, admin APIs, usage views) Instrumentation (gateway OTEL, traces) Derived in Cloudability

Top insight

$29.4K per month sits in accepted-quality optimizationsRouting, caching, and batch recommendations below the eval quality bar are ready to apply. See FinOps → Optimization.

Cost trend

AI COGS reached 3.1% of ARR, up from 2.4% in Q1Growth is concentrated in agent workloads. Unit cost per request is falling, but volume is growing faster.

Capacity risk

2 commitments on pace to saturate this monthGemini GSUs at 91% and Bedrock MUs at 82%. Overflow bills at on-demand rates.

AI spend MTD

All providers

Forecast EOM

Intelligent forecast

Budget variance

vs $600K monthly AI budget

Allocation coverage

Spend mapped to teams

Month to date spend vs budget

Cumulative actuals through Jul 8, forecast to end of month, against the $600K budget

Spend by provider MTD

Billed and metered

Budget and showback by business unit and team

Burn vs budget, with chargeback statements at the level finance bills. Model-level detail lives in the FinOps zoom.

Spend by consumption type

Tokens, provisioned capacity, seats, credits

AI COGS

Share of ARR
3.1%
▲ from 2.4% in Q1
Serving costs attributed to revenue products

Cost per 1K requests

Blended, trailing 7d
$2.61
▼ 14% over 30 days
Requests from telemetry, cost from reconciled billing

Cost per active customer

AI serving cost / MAU
$0.84
▼ 9% MoM
Customer dimension via requestMetadata

Assistant seat efficiency

Active seats / paid seats
78%
54 seats inactive 60+ days
GitHub Copilot, M365 Copilot, Figma, Cursor

Unit economics trend

Cost per 1K requests, trailing 30 days

Cost to serve by product feature

Feature dimension via gateway attributes

AI cost to serve by customer

Top customers by AI serving cost, with revenue and margin signal

Gateway coverage

Request volume through instrumented paths
62%
▲ 11 pts since OTEL endpoint launch
Uninstrumented traffic falls back to key-level attribution

Dimension policy compliance

Spend carrying required dimensions
87%
Team required · Product required · Customer optional
Enforced at ingest by business mappings

Shadow AI exposure

Unsanctioned AI spend discovered
$4.9K/mo
4 unsanctioned tools found
SSO, expense, and network discovery

Policy guardrails

Spend under an active quota or budget policy
92%
9 policies active · 6 breaches caught MTD · 2 blocked by the customer's gateway
Blocking happens in your gateway. Cloudability detects, notifies, and reconciles.

Quality gates

Eval-gated routing changes
3 of 6
routes passing or in canary · est $6.8K/mo held by failing gates
Savings only ship when the eval score delta clears the gate

Shadow AI findings

Discovered outside sanctioned providers

Dimension coverage by source

Share of spend carrying each dimension

Allocation coverage

Spend mapped to a team owner

Unallocated spend

Needs an owner before chargeback

Gateway vs billed delta

Rate reconciliation, MTD

Showback and chargeback by team

Team, provider, model, MTD spend, allocation basis

Rate reconciliation

Gateway estimated cost vs provider billed cost

Attribution basis

How spend gets an owner, share of MTD spend

Slice AI spend by business dimension

The same key-value business context used across Cloudability, applied to AI

Dimension breakdown detail

MTD spend, requests, and cost per 1K requests

Where dimensions come from

Mapping inputs per source family

Optimization recommendations

Ranked by savings, gated by eval quality where routing is involved

Routing policies

Declarative task-to-route rules. The gateway executes them; the eval gate decides what is allowed to shift.

Model routing arbitrage

Same model, different paths: effective cost after discounts, cache, and commitments

Provisioned throughput utilization

Committed capacity MTD

Commitment planner

Coverage, overflow, and break-even for AI capacity commitments

Quota and budget policies

Token and spend limits across cloud platforms and frontier labs. The gateway enforces; Cloudability governs.
Cloudability is the system of record for policy. It detects, notifies, and exports gateway config. It does not sit in the request path and it does not write to your gateway. Where a limit blocks, your gateway blocks. New policies start in shadow mode: they match and record, but do not block.

Member consumption vs quota

Token quotas per member per model, MTD consumption

Policy events

Breaches, blocks, and near-limit warnings

Failure spend MTD

Retries, timeouts, rejected outputs
$3.2K
2.2% of MTD spend bought nothing
Largest driver: the support-triage retry loop

Requests traced

Share of volume with full trace detail
62%
▲ 11 pts since OTEL endpoint launch
Untraced requests still get cost at key or account grain

Median cost per request

Traced requests, trailing 7 days
$0.011
▼ 8% since classify-flash routing shipped
P95 is $0.42. Distributions beat averages here.

Request and trace explorer

Sampled from the AI usage store. Every row carries cost, tokens, cache, latency, quality, and attribution. Select a flagged row to expand its trace.

Agent and session economics

Cost per session, retries, cache behavior, by agent workload

Model consumption detail

Tokens, cache efficiency, and cost per model MTD

AI assistant seats

Per-member activity across GitHub Copilot, M365 Copilot, Figma, Cursor

Eval suites and quality gates

Quality scores from LangSmith and Langfuse runs. A cheaper route goes live only when its score delta clears the gate.

Quality regressions

Score drops correlated with deploys and drift

Anomalies and signals

Near real time from telemetry, daily from billing

Cloud AI platforms

Billing plus per-request telemetry

Direct model APIs

Admin APIs for usage and cost

AI in SaaS

Seats, credits, and usage reports

Gateways and LLM observability

The instrumentation layer: near real time attribution and quality signal

Data requirements

Every source, its grain, freshness, status, gaps, and the widgets it feeds. Build items first, then by widgets fed.