Line-item billing rows inspected across 50 cloud services and 598 SKUs (2025-09-01 to 2026-11-10).
Quality of Earnings · AI & Cloud COGS
What happens to gross margin
when the cloud credits expire?
Startup and scale credit grants routinely offset 93%–99% of native cloud and model inference on an invoice-basis P&L while leaving partner marketplace LLM SKUs at 0% coverage. Test a target's reported gross margin, run-rate EBITDA, and credit-cliff runway in your browser.
Free. No signup. CSV parsing and arithmetic run locally in this browser—nothing you enter is uploaded or transmitted.
Promotional credit coverage observed on first-party Cloud Run, Compute Engine, Gemini API, and native Vertex AI SKUs.
Promotional credit coverage on third-party partner LLM and vector database SKUs ($3,474 billed 100% in cash).
Commercial extraction finding drop (27 → 0 synergies) when routing recall-heavy extraction off frontier models without passing gate.
First-Party Operational Telemetry (n = 1)
What 3.53 million cloud & AI billing rows reveal about post-close margin risk
To ground this check in real billing mechanics rather than synthetic estimates, we audited OloLand's own production BigQuery billing export (3,531,284 rows, $81,383.28 gross spend across 50 services and 598 SKUs from 2025-09-01 through 2026-11-10) alongside our production runtime-substrate verification records. Two diligence traps stood out immediately.
Finding 1: The marketplace SKU split inside a “credit-covered” cloud account
Startup and scale credit pools absorbed 93.4%–100.0% of native cloud compute and first-party Gemini inference. Inside the exact same billing account, third-party partner models on Vertex AI and marketplace database subscriptions received 0.0% promotional credit offset—billing 100% cash out-of-pocket from day one. Meanwhile, a separate $10,000 cloud grant hit its hard calendar cutoff and forfeited $4,273.81 unspent.
| Service / SKU group | Workload category | Gross spend | Promo credits | Net cash billed | Credit treatment |
|---|---|---|---|---|---|
| Cloud Run | Native serverless containers | $16,126.71 | −$16,023.37 | $103.34 | 99.4% · Credit-covered |
| Compute Engine | Native VM & cluster nodes | $14,515.27 | −$14,466.85 | $48.42 | 99.7% · Credit-covered |
| Vertex AI (service aggregate, 135 SKUs) | First-party tuning/inference + partner pass-through | $12,735.48 | −$11,896.39 | $839.09 | 93.4% · Mixed |
| Gemini API (84 SKUs) | First-party frontier & flash token inference | $8,139.38 | −$8,029.30 | $110.08 | 98.6% · Credit-covered |
| Cloud Text-to-Speech API | First-party audio synthesis | $861.09 | −$860.76 | $0.33 | 100.0% · Credit-covered |
| Qdrant Vector Database (Marketplace) | Third-party vector database subscription | $1,054.32 | $0.00 | $1,054.32 | 0.0% · 0% cash-trap |
| Claude Opus 4.7 (Vertex Partner API) | Third-party frontier LLM tokens | $874.43 | $0.00 | $874.43 | 0.0% · 0% cash-trap |
| Claude Opus 4.6 (Vertex Partner API) | Third-party frontier LLM tokens | $617.97 | $0.00 | $617.97 | 0.0% · 0% cash-trap |
| Claude Sonnet 4.5 (Vertex Partner API) | Third-party mid-tier LLM tokens | $502.90 | $0.00 | $502.90 | 0.0% · 0% cash-trap |
| Claude Sonnet 4.6 (Vertex Partner API) | Third-party mid-tier LLM tokens | $424.04 | $0.00 | $424.04 | 0.0% · 0% cash-trap |
Finding 2: Why “we'll swap to cheaper models post-close” requires a verification gate
Management presentations often argue that expiring credits or high frontier LLM bills can be mitigated post-close by routing workloads to lower-cost models. When we evaluated moving production diligence workflows onto our credit-covered harness, seven ordinary deal-chat templates retained identical deterministic boundary-gate coverage (without a separate comparative quality study), whereas our full-corpus commercial extraction soak suffered a 78% finding collapse and missed every synergy item.
| Workflow class | Candidate lower-cost route | Verification gate | Observed outcome | Production decision |
|---|---|---|---|---|
| 7 interactive deal-chat templates | Credit-covered Gemini harness | Deterministic boundary citation & math verifier | Boundary-gate coverage unchanged; 7 of 7 default routes flipped (comparative quality unperformed) | Switched |
| Full-corpus commercial diligence extraction | Credit-covered Gemini harness | Finding-count & synergy recall gate against baseline | 24 findings vs. 108 baseline; 0 vs. 27 synergies (failed gate) | Rejected |
| IC coordinator & live web-enabled turns | Credit-covered Gemini harness | Multi-agent orchestration & live web routing contract | Managed-only runtime capabilities required | Retained on paid |
Methodology, data provenance, and n = 1 scope boundary
Single-operator dataset (n = 1): All telemetry figures in the tables above come from OloLand's own multi-cloud billing export and internal runtime engineering records—zero customer or tenant financial data is included. They illustrate concrete cloud-billing mechanics on an AI-intensive software platform and must not be cited as cross-industry medians or peer benchmarks.
Promotional grants vs. enterprise discounts: Startup, accelerator, and migration credits expire on a fixed schedule or dollar cap. Contractual committed-use discounts (GCP CUDs, AWS Savings Plans / EDP) represent ongoing volume pricing commitments and should remain in normalized run-rate COGS only to the extent the post-close entity retains the contract.
Move from quick-check to source lineage
Audit every adjustment back to the underlying ledger.
Inside OloLand, deal teams connect data-room documents, assumption registers, forensic red-flag screens, and human-governed approvals—keeping every citation and calculation inspectable.