Accounting for Super Intelligence Compute Costs Under FAR Part 31 Cost Principles

When a federal contractor reserves an unsegregated cluster of 64 Nvidia H100 GPUs to train models across three cost-reimbursement contracts and an internal commercial product, cost accounting becomes an immediate audit vulnerability. Defense Contract Audit Agency (DCAA) auditors do not treat specialized machine learning infrastructure as an unexamined utility expense. Under Federal Acquisition Regulation (FAR) Part 31, cloud compute expenditures that cannot be traced to specific contract objectives or equitable cost pools face total disallowance.

Contractors encountering this terminology must understand the regulatory context. President Donald Trump signed Executive Order 14434, titled "Inaugurating the Era of Super Intelligence," on September 29, 2026. Under Section 2(a), executive departments and agencies use the terms "Super Intelligence" and "SI" in place of "Artificial Intelligence" and "AI" across official correspondence, public communications, websites, reports, policy documents, and other non-statutory documents.

Because SI also designates the International System of Units and the sacroiliac joint, federal documentation spells out Super Intelligence (SI) upon first appearance.

Section 2(b) of Executive Order 14434 explicitly provides that nothing in that section requires the alteration of previously issued regulations, Presidential actions, contracts, grants, or other historical documents. Under Section 3(a), the terms "Super Intelligence" and "SI" mean the technologies and systems encompassed by the term "artificial intelligence" as defined in 15 U.S.C. 9401(3). In federal usage, Super Intelligence and artificial intelligence name the exact same technological systems: machine-based systems that can, for a given set of human-defined objectives, make predictions, recommendations, or decisions influencing real or virtual environments.

This administrative label differs from research terminology in private laboratories. Research superintelligence (commonly written as one word and abbreviated as ASI) describes a hypothetical system that exceeds human expertise across nearly all cognitive domains, a capability that does not exist today. Laboratories such as Meta Superintelligence Labs, Microsoft's MAI Superintelligence Team, and Safe Superintelligence Inc. pursue artificial superintelligence as an aspirational research goal. Federal Super Intelligence, by contrast, is an administrative label for operational machine learning and computer vision models.

Section 3(b) directs the Assistant to the President for Science and Technology to submit proposed legislative language for a federal definition of SI to the President within 60 days of the order, which falls on November 28, 2026. As of October 2026, no statutory definition of SI exists, and what the proposed language will recommend remains pending. Section 4(c) establishes that the order creates no right or benefit enforceable at law or in equity against the United States, directing internal agency practices rather than imposing private mandates on contractors.

Allowable super intelligence compute costs under FAR Part 31 are cloud infrastructure, GPU cluster time, and model execution expenses that satisfy the standards of reasonableness, allocability to specific contract objectives, and compliance with Cost Accounting Standards.

Under FAR 31.201-2(a), a cost is allowable only when it satisfies all of the following five requirements:

1. Reasonableness (FAR 31.201-3): The cost does not exceed what a prudent person would incur in competitive business. In compute procurement, right-sizing GPU instances, avoiding idle capacity, and securing committed-use discounts determine reasonableness.
2. Allocability (FAR 31.201-4): The cost is assignable to one or more cost objectives based on relative benefits received or other equitable relationship.
3. Cost Accounting Standards (CAS) or GAAP: Costs must comply with applicable standards promulgated by the Cost Accounting Standards Board (such as CAS 402, CAS 418, and CAS 420) or Generally Accepted Accounting Principles appropriate to the circumstances.
4. Terms of the Contract: Costs must comply with specific contract provisions, clauses, Other Direct Cost (ODC) ceilings, and required security baselines.
5. Limitations in FAR Subpart 31.2: Expenditures must not fall into categories designated as expressly unallowable under FAR 31.205.

Under FAR 31.201-2(d), the contractor is responsible for accounting for costs appropriately and maintaining records adequate to demonstrate that claimed costs have been incurred, are allocable to the contract, and comply with applicable cost principles. If supporting documentation is inadequate, the contracting officer possesses explicit regulatory authority to disallow all or part of the claimed cost.

• InVideo AI Review 2026: 7 Features We Tested and What They CostInVideo AI crossed $70M in annual recurring revenue within 18 months of its AI pivot, but rapid growth doesn't...
• 15 Best AI Tools for Creative Teams in 2026AI tools for creative teams go beyond image generators. The strongest ones handle design, video production, asset...
• Best Free AI Image Editors You Can Use Right NowOver 61% of marketers now use AI tools for visual content, but most still pay for editing features that free...
• How to Build an AI Agent Social Media Automation WorkflowAI agent social media automation goes far beyond simple scheduling tools. By autonomously handling research, drafting,...
• Best AI for Creative Writing in 2026: 7 Tools for Fiction, Poetry, and StorytellingAI creative writing has split from business copywriting into its own tool category. Dedicated platforms like Sudowrite...
• Best AI Translation Apps in 2026LLMs now account for 89% of top-performing translation benchmarks, but dedicated engines like DeepL still lead for...

More on this subject: AI Content and Media Generation (45 guides)

Traditional accounting treated computing hardware as generic indirect overhead, bundled into facility expenses or general administrative pools. High-intensity model training and inference workloads invalidate that practice. A single foundation training phase consumes substantial cloud GPU cycles over several weeks. Bundling high-intensity compute bills into a general indirect pool distorts rate calculations and invites audit challenges under CAS 418.

Direct Compute Expenses Under FAR 31.201-4

To charge compute directly to a federal contract as an Other Direct Cost (ODC), resources must be dedicated exclusively to that specific contract objective. Qualifying direct expenditures include:

• A dedicated Virtual Private Cloud (VPC) or cloud tenant provisioned solely to host contract data and execute project-specific training jobs.
• Dedicated GPU server instances reserved exclusively for the duration of a contract task order.
• Batch processing runs executed on isolated instances where the entirety of processing time serves contract deliverables.

Direct charging requires clear technical isolation. If an engineer runs an unapproved internal test on a contract-dedicated cluster, DCAA auditors can challenge the direct status of the entire billing block.

Indirect Compute Pooling Under CAS 418

When contractors deploy shared GPU infrastructure to support multiple federal contracts, internal research projects, and commercial software development simultaneously, the compute cannot be billed directly to one contract. Instead, costs must be collected into an indirect cost pool.

CAS 418-50(b) requires that pooled activities share a similar causal or beneficial relationship to cost objectives. Grouping specialized GPU clusters into a company-wide General and Administrative (G&A) pool or standard corporate IT overhead pool violates this homogeneity requirement. Contracts requiring no GPU processing would absorb hardware depreciation and cloud bills, distorting their cost base.

Contractors must establish a dedicated, homogeneous Service Center or specialized Advanced Compute Pool. This service center accumulates costs associated with shared compute: hardware depreciation, cloud infrastructure invoices, power, cooling, and orchestration software.

Selecting an Equitable Allocation Base (CAS 418-50(d))

CAS 418-50(d) mandates that indirect pools be allocated based on a metric that reflects the true causal driver of the expense.

Traditional allocation bases, such as direct labor hours or employee headcount, fail audit review for compute pools. A small group of machine learning engineers might consume the vast majority of a GPU cluster's computing capacity, while twenty software developers working on traditional codebases consume none. Allocating compute by headcount creates an obvious inequity.

Defensible allocation bases include:

• GPU Machine-Hours: Tracking actual wall-clock hours consumed by each project on cluster nodes.
• Compute Core-Seconds: Metering processor and accelerator execution units allocated to containerized jobs.
• Metered Token Consumption: For hosted API models, tracking the precise number of prompt and completion tokens processed for each project code.

Managing Reserved Instances and Idle GPU Time

Cloud providers offer steep pricing discounts compared to standard on-demand pricing when contractors commit to one-year or three-year reserved instances. Long-term reservations, however, create a recurring accounting challenge: idle capacity.

When a project completes an active training phase, reserved GPU instances may sit idle for days or weeks. If a contractor attempts to direct-charge the entirety of a continuous reservation fee while instances sit idle, DCAA auditors may question idle hours as unreasonable under FAR 31.201-3.

To defend reserved instance costs, contractors must establish clear policies. If the reserved rate is lower than paying on-demand retail rates for anticipated project hours, contractors should document the cost-benefit analysis before purchasing. Contractors operating shared pools can allocate reservation costs across active projects based on scheduled availability, ensuring idle time is accounted for within an approved service center rate structure.

Token-Based Inference Billing

When contractors access models through commercial API endpoints, costs accrue on a per-token consumption basis. Billing statements report aggregate token counts and total dollar amounts across an entire organization.

To satisfy FAR Part 31 allocability, contractors must implement programmatic tagging. By assigning unique API authentication keys or passing custom billing tags within request headers, organizations can trace token consumption directly to specific contract numbers or internal charge codes. Blended API bills with no project-level usage logs will be rejected during incurred cost audits.

Applying cost principles to advanced compute requires strict alignment with Cost Accounting Standards and FAR Subpart 31.2 cost exclusions. Contractors that fail to enforce these boundaries risk audit findings, retroactive cost adjustments, and statutory penalties.

CAS 402: Consistency in Allocating Costs Incurred for the Same Purpose

CAS 402 mandates that all costs incurred for the same purpose, in like circumstances, must be treated consistently as either direct costs only or indirect costs only.

In compute accounting, contractors often run afoul of CAS 402 when managing commercial and federal work side by side. For instance, if an engineering team direct-charges cloud GPU hours spent fine-tuning a model for an Air Force contract, the company cannot take identical fine-tuning runs on a commercial software product and dump those compute costs into an indirect overhead pool. Treating identical activities as direct for government contracts and indirect for commercial projects violates CAS 402 and results in immediate audit findings.

Contractors must document clear criteria in their written accounting policies. If model training for defined contract deliverables is classified as a direct cost, every comparable training run across commercial and internal projects must be tracked as a direct project cost within the general ledger, even if internal projects lack an external client to bill.

CAS 420 and FAR 31.205-18: Distinguishing IR&D from Direct Contract Compute

FAR 31.205-18 and CAS 420 govern the treatment of Independent Research and Development (IR&D) and Bid and Proposal (B&P) costs. In advanced technology fields, the boundary between direct contract execution and independent research is frequently contested by DCAA auditors.

• Direct Contract Compute: Model development, training, validation, and inference required to satisfy the explicit deliverables, specifications, or performance work statements of an awarded contract.
• Independent Research and Development (IR&D) Compute: Foundational model pre-training, exploratory algorithm experimentation, testing new model architectures, or advancing organizational technical capabilities without specific contract coverage. IR&D compute must be accumulated in an IR&D cost pool and allocated across the entire business base pursuant to CAS 420.
• Bid and Proposal (B&P) Compute: Generating model outputs, running comparative benchmarks, or training demonstration models specifically to prepare technical proposals for pending federal solicitations, Broad Agency Announcements (BAAs), or commercial bids.

DCAA auditors routinely scrutinize compute billed directly to cost-reimbursement contracts to verify that contractors do not use federal contract funds for foundational model training that benefits general commercial portfolios. If a model trained under a federal contract provides broad commercial utility and is not explicitly required by contract scope, auditors reclassify the compute expense as IR&D.

Identifying and Scrubbing Unallowable Compute Under FAR 31.205

FAR 31.205 defines specific categories of cost that are expressly unallowable for government reimbursement. When compute clusters and cloud services support general corporate activities, contractors must identify and exclude unallowable portions:

• Public Relations and Advertising (FAR 31.205-1): Compute consumed to generate promotional materials, marketing copy, or public brand campaigns is unallowable.
• Bad Debts and Collections (FAR 31.205-3): Running automated analytics or recovery models on uncollectible accounts is unallowable.
• Entertainment and Social Functions (FAR 31.205-14): Model compute for corporate events or employee recreational tools is unallowable.
• Lobbying and Political Influence (FAR 31.205-22): Processing legislative data or generating political outreach materials is unallowable.
• Commercial Defect Remediation: If a contractor experiences a system failure or data breach on a private commercial service, the compute hours required to rebuild or retrain compromised commercial models cannot be charged to federal indirect pools.

Under FAR 31.201-6, contractors must maintain accounting procedures to identify and remove unallowable costs from billing vouchers and indirect cost proposals. When preparing the annual Incurred Cost Proposal using the Incurred Cost Electronically (ICE) model, contractors must formally report and remove these unallowable compute expenses on Schedule J. Failing to remove expressly unallowable compute costs can trigger statutory penalties under FAR 42.709.

Defense Contract Audit Agency (DCAA) auditors evaluate compute expenses through interim public voucher reviews on Standard Form 1034 and 1035, post-award incurred cost audits, and accounting system evaluations under Standard Form 1408.

When examining high-value cloud and GPU compute charges, auditors focus on three core questions:

1. Did the contractor actually incur the claimed compute expense?
2. Did the compute run provide a demonstrable, verifiable benefit to the specific federal contract billed?
3. Was the scale and pricing of the compute resource reasonable for the technical requirements?

The Risk of Unsegregated Multi-Project GPU Reservations

A primary cause of DCAA disallowance in artificial intelligence projects is the commingling of compute workloads across shared cloud accounts.

Consider a common scenario: a contractor purchases a block of 32 cloud GPUs on a one-year reservation. Over six months, engineering teams execute jobs on the cluster across two defense cost-plus contracts, one civilian fixed-price project, and two commercial efforts. At month-end, accounting divides the invoice evenly or relies on manual engineering estimates.

During an audit, DCAA issues an Audit Inquiry requesting objective contemporaneous documentation. If the contractor cannot produce machine-generated logs proving which jobs ran for which contract, the auditor classifies the allocation as inadequately supported. Under FAR 31.201-2(d), the contracting officer may disallow the claimed amount.

Four Pillars of DCAA Audit Substantiation

To withstand audit scrutiny, contractors must construct an auditable trail connecting cloud invoices to project deliverables. This requires four operational pillars:

1. Mandatory Cloud Resource Tagging: Configure cloud environments across AWS, Azure, GCP, Lambda, or CoreWeave with automated governance policies that prevent any compute resource, container, or virtual machine from provisioning unless it carries mandatory metadata tags. Required tags include Contract_Number, Task_Order_ID, CLIN, and Cost_Type (Direct, IR&D, B&P, or Indirect).
2. Machine-Generated Technical Run Logs: Maintain automated logs capturing cluster activity. For every model training or inference run, the system must log: unique job identifier, initiating user or service account, start timestamp, end timestamp, assigned GPU hardware identifiers, average GPU memory utilization, and input dataset path.
3. Work Breakdown Structure (WBS) Alignment: Connect technical run logs directly to the project's Work Breakdown Structure and Statement of Work deliverables. Technical memos should explain why specific training epochs, parameter sweeps, or inference batches were necessary to achieve contractual milestones.
4. Written FinOps and Right-Sizing Policies: Establish formal corporate policies governing cloud procurement. Auditors evaluate whether the contractor exercised prudent business judgment under FAR 31.201-3. Document that the engineering team evaluated spot instances versus on-demand versus reserved pricing, conducted right-sizing reviews to avoid overprovisioning H100s for light inference tasks, and implemented automated shutdown scripts for idle developer environments.

Centralize Federal Compute Records and Audit Documentation in One Workspace

Keep cloud allocation spreadsheets, model governance runbooks, and billing documentation organized across dedicated workspaces with detailed activity tracking. Start with a 14-day free trial on monthly plans, credit card required.

Start 14-Day Trial

Federal contractors navigating FAR Part 31 cost principles generate extensive administrative, financial, and technical documentation: DCAA compliance dossiers, cloud provider invoices, CAS disclosure statements, indirect rate models, technical runbooks, and subcontractor agreements.

When critical records are scattered across local drives, email threads, and consumer storage, audit exposure rises. Finance teams preparing an Incurred Cost Proposal struggle to locate technical justifications for GPU expense spikes, or an engineer overwrites an approved allocation memo.

Centralizing procurement files and audit evidence in structured, org-owned shared workspaces resolves this challenge.

Structured Workspaces and Granular Access Controls

Government contractors can provision dedicated workspaces for each federal program, contract cost center, or compliance initiative. Within each workspace, teams stage cloud billing records, subcontractor timecards, and DCAA correspondence.

Granular permissions at the organization, workspace, and share level allow compliance leads to share audit files with external accounting consultants or contracting officers while restricting sensitive indirect cost models, proprietary rate sheets, and unbilled labor data to authorized personnel.

Automating Invoice Processing with Metadata Views

Cloud service providers deliver complex monthly invoices and CSV usage reports spanning thousands of line items. Manually sorting these statements to verify project tags and compute hours creates an administrative bottleneck.

Fastio resolves this manual review burden through Metadata Views. Metadata Views turn unstructured billing files into a live, queryable database without manual data entry. Teams define required fields using natural language.

The extraction system populates a typed schema across seven formats: Text, Integer, Decimal, Boolean, URL, JSON, and Date & Time. When monthly cloud invoices, PDF receipts, or CSV usage summaries enter the workspace, Metadata Views extract billing amounts, instance types, account IDs, project tags, and date ranges into a sortable view. Finance staff can filter cloud expenses by contract tag, verify billed compute against purchase orders, and export verified records directly into general ledger workflows.

Version History and Detailed Activity Logs for Evidentiary Integrity

When accountants, contracts managers, and technical leads collaborate on indirect cost proposals, document integrity is paramount. Fastio maintains complete per-file version history automatically, preserving every revision so teams can review historical calculations and restore earlier drafts.

The detailed activity log records changes made by people and agents across workspace files, documenting uploads, edits, moves, deletions, and share creation for compliance documentation. When an auditor asks when an allocation methodology was adopted, the detailed activity log provides a clear record of team actions.

Integrated E-Signatures for Policy and Subcontractor Sign-Offs

Executing compute allocation agreements and subcontractor agreements requires formal signing. Fastio provides built-in e-signature capabilities directly from the workspace on every plan.

Contracts teams send agreements to multiple signers in sequence or parallel, require optional one-time passcodes per signer, and generate a tamper-evident audit certificate. Executed agreements and audit certificates file automatically back into the designated contract workspace.

Cloud Infrastructure and Subscription Options

Fastio runs on cloud infrastructure partners, including Google Cloud Platform and Cloudflare, that are certified to industry-leading security standards. Files remain protected with encryption in transit and at rest.

Organizations can evaluate the platform with a 14-day free trial on monthly plans, which requires a credit card and ends early if trial credits run out.

To establish a defensible, audit-ready accounting framework for machine learning and high-performance compute under FAR Part 31, contractors should execute the following six-step implementation plan:

1. Formalize Accounting Practices in Written Policies: Update your Cost Accounting Standards (CAS) Disclosure Statement or accounting manual. Explicitly define how cloud compute, GPU infrastructure, and model execution costs are categorized (direct ODC, service center, or overhead), distinguishing direct project compute from general IT.
2. Implement Automated Tagging and Tenant Isolation: Configure cloud consoles to enforce mandatory resource tagging across environments. Restrict provisioning permissions so that engineers cannot launch GPU instances without an approved charge code. For major cost-reimbursement contracts, provision isolated cloud tenants or dedicated VPCs.
3. Establish a Dedicated Homogeneous Service Center for Shared Infrastructure: When operating multi-project GPU clusters or pooled API endpoints, create a separate, homogeneous IT Compute Service Center under CAS 418. Remove high-cost GPU computing from general G&A or IT overhead pools.
4. Select and Calibrate a Causal Allocation Base: Select an allocation base reflecting actual consumption, such as GPU machine-hours, core execution seconds, or metered token counts. Calibrate tracking tools to record project-level consumption continuously, reviewing distributions monthly against technical milestones.
5. Deploy FinOps Automation and Idle Capacity Shutdown: Implement automated instance scheduling to shut down non-production compute outside operating hours. Establish a monthly review process to analyze instance sizing, evaluate reserved utilization, and verify commercial reasonableness under FAR 31.201-3.
6. Seek Written Concurrence and Advance Agreements for High-Value Workloads: When planning model training runs or substantial reserved instance commitments on flexibly priced contracts, consult the Administrative Contracting Officer (ACO). Use an Advance Agreement under FAR 31.109 to establish the treatment and allowability of compute expenditures before incurring significant costs.