Break 01
The §41 credit needs project-level proof.
Form 6765's new Section G requires you to tie qualified compute to specific business components. Lump-sum AWS billing won't pass review.
$90K trained the model. $60K served customers. $30K was idle capacity. Compute is the single largest variable cost for most AI companies — and it gets booked the worst. Generalist firms drop the entire bill into "software subscriptions," which kills your gross margin reporting, your §41 R&D study, and your investor diligence package at the same time.
Compute spend used to be a small line item. For modern AI companies, it's often 30 to 60 percent of total operating expense. That changes how the accounting needs to work.
Form 6765's new Section G requires you to tie qualified compute to specific business components. Lump-sum AWS billing won't pass review.
ASU 2023-08 and ASC 606 expectations sharpened for AI revenue models. Hybrid pricing and usage components don't recognize the way subscription does.
Venture investors started asking gross margin questions on AI deals nobody asked SaaS companies five years ago. Compute cost per dollar of revenue is now standard diligence.
Companies that haven't segmented their compute spend usually find out at the worst possible moment, which is mid-diligence.
The same AWS line item can fall into any of these depending on what it ran. The tagging is the work.
Inference compute that runs the product for paying customers. Production GPU hours, hosted model endpoints, customer-attributable storage and bandwidth.
Training runs, evaluation pipelines, ablation studies, hyperparameter sweeps, and experimentation. Has to meet the §41 four-part test.
Internal tooling, dev environments, CI/CD compute, and engineering productivity workloads.
Multi-month reserved instances, committed-use discounts, and prepaid GPU capacity. Typically prepaid and amortized over the commitment period.
Most AI companies need this segmentation by Series A. Without it, gross margin is a guess and the R&D credit is undocumented.
AWS, GCP, Azure, Modal, RunPod, CoreWeave, Lambda Labs.
COGS, §41 R&D, OpEx, prepaid — applied before sync.
Each provider's bill flows in already classified.
We work with your engineering team to set up cost allocation tags directly in your cloud accounts. Tags map to four categories: production, training, evaluation, internal tooling. Each provider's bill flows into QBO or Xero pre-segmented, not as a single line.
We build your gross margin schedule monthly from segmented compute. Compute cost per customer, per product line, per dollar of revenue. Investor-ready GM that ties directly to the GL.
Tied directly to source data, not a separate spreadsheet.
The unit-economics number investors actually ask for.
Customer-level cost basis when usage data supports it.
Project-level documentation tying compute to components.
Each business component gets traceable QRE.
Up to $500K against quarterly payroll liability.
Form 6765 with project-level allocation. Section G compliant. Training compute, evaluation pipelines, and experimentation tagged against specific components, so the credit holds up to review.
Reserved instances and prepaid GPU as balance-sheet prepaid expenses, amortized over the commitment. Schedule rebuilt monthly so capacity utilization shows up in the right period.
Multi-month commits booked as asset, not expense.
Capacity utilization hits the right period every month.
RIs, savings plans, and prepaid GPU credits handled.
Most AI companies we onboard get clean compute segmentation within sixty days, faster if your engineering team's already using cost allocation tags.
Most venture-backed AI companies run thirty to fifty percent compute as a share of revenue at scale. Below twenty percent often signals under-investment. Above seventy, a pricing or architecture problem.
We work with engineering teams to set up cost allocation tagging directly in your cloud accounts. The segmentation isn't a spreadsheet exercise after the fact — it's wired into the bill itself.
QBO, Xero, AWS Cost Explorer, GCP Billing, Azure Cost Management, CloudZero, Vantage, and CUDOS Intelligence.
We tie qualified research compute to §41 business components the way the IRS now expects under Section G of Form 6765.
Direct Slack channel. Response under four hours during business days.
Every workpaper reviewed by a senior team member. We don't outsource your books.
Compute-specific capabilities most generalist startup CPAs don't run.
Comparison reflects our read of public information about each firm's published service descriptions as of April 2026. Happy to walk through specifics on a call.
Here are some of the frequently asked questions and their best answers for your clarification.
Under the new Section G of Form 6765, you need project-level documentation of business components, research activities, and qualified employee wage, supply, and contract research expenses. The rules apply to refund claims postmarked after June 18, 2024.
Still have questions? Contact our Customer support.
Different shapes of compute spend. Same accounting work underneath.
If your last firm couldn't tell you what portion of your AWS bill belongs in cost of revenue, you're in the right place.
