Independent technical underwriting evidence for financed AI infrastructure
Doku determines which material technical assumptions inside the underwriting model are supportable, which must change, and what evidence, conditions, covenants, or monitoring closes the gap.
Confidential 20-minute fit review. No sensitive materials required. A standard remote pilot targets an initial committee package in 7–10 business days after scope confirmation and complete agreed evidence access.
Why Doku exists
Conventional diligence often confirms inventory, diagrams, contracts, and commissioning documents. Doku asks whether the integrated system can produce the output, timing, resilience, and operating evidence assumed by the financing.
Accelerator count, vendor specifications, installed power, and nominal peak benchmarks.
Sustained useful output under the intended workload after topology, throttling, failures, availability, and acceptance.
Pipeline, reservations, customer concentration tables, and commercial forecasts.
Workload compatibility, onboarding readiness, portability, technical switching constraints, and forecast-to-observed conversion.
N+1 diagrams, duplicate components, vendor warranties, and documented recovery paths.
Shared dependencies, correlated failures, recovery time, operator response, runbook quality, and successful re-entry to service.
Delivery, installation, component commissioning, and facility readiness documentation.
Integrated workload acceptance, security and operations readiness, evidence closure, ramp timing, and repeatable production performance.
The Doku Evidence-to-Credit Framework
The engagement is bounded by the lender’s decision, model, and date. Technical work is selected because it can support, revise, condition, or monitor an underwriting assumption, not because it is interesting in isolation.
Financing structure, asset boundary, decision date, baseline model, lender policy, critical assumptions, and permitted reliance.
Capacity, workload, price, utilization, ramp, availability, acceptance, capex, opex, recovery, and collateral assumptions.
Documents, telemetry, tests, owners, access, dependencies, evidence class, cut-off date, and explicit pass or closure criteria.
Architecture analysis, representative-load testing, telemetry integrity, FMEA, operating-envelope review, acceptance, and recovery evidence.
Baseline versus supportable assumptions, sensitivities, timing effects, mitigants, owners, unresolved evidence, and materiality classification.
Technical posture, evidence-backed conditions, possible covenant logic, reporting triggers, re-validation, and committee readout.
World-class technical diligence
Every material conclusion is traceable to a defined assumption, evidence source, validation method, materiality class, model implication, owner, and closure criterion.
Every material technical assumption is named, defined, tied to a model variable, assigned an owner, and given an evidence requirement.
Observed, system-generated, third-party, vendor, management, inferred, and unverified sources are distinguished and versioned.
Completeness, time alignment, source-of-truth, extraction method, sampling, exclusions, and selective-presentation risk are assessed.
Test scope, workload mix, duration, environment, pass criteria, repeatability, variance, failures, and recovery are documented.
Failure modes, shared dependencies, severity, likelihood, detectability, mitigation, recovery, and financial exposure are connected.
Every condition specifies the owner, required evidence, threshold, deadline, permitted interim treatment, and re-test trigger.
The finance bridge
Doku converts technically supportable capacity, timing, reliability, and operating evidence into explicit changes to revenue, opex, capex, liquidity, DSCR, covenant headroom, and recovery assumptions.
| Potential technical finding | Model variables affected | Possible lender treatment |
|---|---|---|
| Demand concentration or weak workload portability | Utilization, renewal probability, downside revenue, concentration sensitivity, and residual demand. | Revised downside case, concentration reporting, trigger, customer covenant, or resized exposure. |
| Topology or dependency bottleneck | Usable capacity, ramp date, capex efficiency, scalability, availability, and stranded-asset risk. | Stage the draw, require remediation, revise capacity, create a holdback, or add a completion condition. |
| Sustained goodput below plan | Performance-realization factor, billable output, revenue per installed accelerator, unit economics, EBITDA, and DSCR. | Revise the base case, require re-test, resize a tranche, define a minimum performance trigger, or monitor goodput. |
| Power or cooling operating-envelope gap | Derating, availability, opex, outage exposure, ramp timing, accepted capacity, and expansion capacity. | Operating limits, resilience work, reserve, staged activation, evidence deadline, or insurance review by the lender. |
| Acceptance or runbook evidence incomplete | Acceptance factor, start date, interest carry, revenue recognition, operating risk, and completion confidence. | Condition precedent, holdback, owner assignment, evidence deadline, re-acceptance test, or post-close monitoring. |
Illustrative only. Actual variables, thresholds, covenants, and credit actions depend on the transaction, lender policy, legal advice, evidence, and engagement scope.
Hypothetical $75M term facility financing 4,096 accelerators. Observed performance realization and incomplete acceptance reduce presently supportable usable capacity by approximately 13% versus the baseline.
| Model variable | Baseline | Supportable | Delta |
|---|---|---|---|
| Performance realization | 92% | 84% | −8 pts |
| Usable capacity | 32.5M hrs | 28.3M hrs | −13.0% |
| Revenue capacity | $84M | $73M | −13.1% |
| DSCR | 1.52× | 1.23× | −0.29× |
What the lender receives
A versioned technical record showing the evidence, method, finding, model delta, materiality, condition, owner, and closure test.
Founding lender design partners
Doku is selecting a small founding cohort of lenders with live AI-compute decisions and willingness to provide structured feedback on the evidence standard, materiality framework, report architecture, and monitoring model.
Energized or firmly contracted facility; standard or mostly standard hardware; a real funding, draw, takeout, collateral-addition, refinancing, or deployment-milestone decision.
No success fee, approval fee, hardware economics, capacity economics, financing economics, or other outcome-linked compensation.
For the initial committee package after confirmed scope and complete agreed evidence access. On-site or materially bespoke work is separately scoped.
Evidence register, validation record, model deltas, exceptions, conditions, possible monitoring logic, limitations, and versioned supporting appendix.
Early participation hardens the standard against real lender decisions, not invented abstractions.
The fit review requires only nonconfidential context. NDA, conflicts, permitted users, access, retention, reliance, and scope are aligned before evidence transfer.
Independence and information governance
Before evidence transfer, Doku and the lender define conflicts, scope, permitted users, access, retention, reliance, evidence cut-off, and human accountability.
Doku does not sell hardware, compute capacity, financing, placement, or implementation services tied to its conclusion. Compensation is not contingent on approval, funding, or a particular posture.
No sensitive documents are required for the initial fit review. Evidence access begins only after NDA and scope alignment, preferably through a lender-approved room or agreed least-privilege channel.
Sources are classified, versioned, time-bounded, and tied to permitted users. Retention, deletion, backups, and any exception are agreed for the engagement rather than assumed.
Conclusions are conditional on defined evidence and date. Doku does not replace legal counsel, credit judgment, or engineering certification. Material judgments remain human-reviewed and principal-signed.
Accountable principal
Doku uses software, structured evidence, and repeatable protocols to improve rigor and traceability. They do not obscure who is responsible for the technical conclusion or model bridge.
Computer-science and ML-infrastructure practitioner with prior roles spanning Google Brain, Meta AI, RunPod, Brium, and Fluidstack. The relevant work crossed workload deployment, production ML systems, compiler infrastructure, cloud operations, reliability, and the interfaces where technical performance becomes commercial capacity.
Questions before confidential access
Doku assesses the material technical assumptions that support a specific financing decision. Depending on scope, that can include workload compatibility and demand assumptions, architecture and dependencies, sustained useful output, telemetry, power and cooling operating envelope, recovery behavior, acceptance, security and operations readiness, and the model variables those findings affect.
No. Doku provides an independent technical record, a technical posture, model deltas, evidence-backed conditions, and possible monitoring logic. The lender retains full responsibility for credit policy, structuring, legal analysis, approval, and portfolio decisions.
The target applies to a standard remote founding pilot only and begins after scope confirmation, NDA and conflicts alignment, and complete access to the agreed evidence and owners. On-site work, materially bespoke tests, extensive remediation analysis, or delayed evidence can change the schedule and fee.
Only nonconfidential context: institution, transaction type, approximate exposure, decision timing, asset stage, primary technical concern, and available stakeholders. No contracts, credentials, telemetry, system diagrams, or customer-identifying materials are required to determine fit.
Doku can evaluate technical compatibility, concentration, dependence, portability, onboarding readiness, and how commercial assumptions interact with the system. Legal enforceability, contract interpretation, and legal diligence remain the responsibility of counsel and the lender.
The founding fee covers a bounded standard remote review and agreed remote validation. Travel, on-site testing, unusually large evidence reconstruction, custom-hardware work, or material scope expansion is separately agreed before it is performed.
Permitted users, purpose, reliance, scope, evidence cut-off, limitations, and any third-party use are defined in the engagement terms. The public specimen is illustrative and cannot be relied upon for any transaction.
Doku prefers a lender-approved data room or an agreed least-privilege channel. Access, permitted users, retention, deletion, and any model or software use are agreed before transfer. Confidential engagement data is not used to train public or third-party models.
Software and automated tools may assist evidence organization, comparison, calculation, and traceability. Material findings, model translations, conditions, and written conclusions remain human-reviewed and principal-signed. Tool output is not treated as evidence merely because a model produced it.
The lender receives the final report, supporting appendix, committee readout, open-item register, and agreed closure or monitoring design. Any post-close re-validation, recurring monitoring, additional assets, or broader relationship is optional and separately scoped.
Doku will first determine fit, define the evidence boundary, and explain what must be true before technical diligence begins.
Cornell University
Google Brain
AI
Fluidstack