Control Inference Margin, Compute Burn, and Runway as You Scale AI

Use Level observations as cross-trade context. Other cards are planning scenarios, not measured industry percentiles. Match your business model, period and cost definition before comparing.

R&D Credit Capture: Planning inputs to test, not measured industry percentiles: Illustrative share of qualified research expenses, not a statutory rate. Actual credit depends on method, base amount, Section 280C treatment and election limits. Payroll offset requires the gross-receipts/history tests, election, actual credit and employer tax liability.

GenAI Credit Programs: Planning inputs to test, not measured industry percentiles: Microsoft describes up to $150K of Azure credits and AWS up to $200K for eligible Portfolio applicants; other tiers differ. Confirm current provider eligibility, covered products, expiry and prior usage.

Burn Multiple Target: Planning inputs to test, not measured industry percentiles: Planning test, not an investor requirement: Net burn / net new ARR for the same period. Under 1.5x and above 2x are sensitivity points; zero or negative net new ARR makes this ratio unhelpful.

Inference COGS %: Planning inputs to test, not measured industry percentiles: Use 20-40% of revenue as a sensitivity, not a measured vertical-AI range. Reconcile usage, model and infrastructure invoices, data licensing and approved credits before interpreting feature or account margins.

Runway Standard: Planning inputs to test, not measured industry percentiles: Test 18-24 months and a separate 12-month downside case against expected receipts and milestone costs. These are planning horizons, not a documented VC requirement.

Treasury Yield: Planning inputs to test, not measured industry percentiles: Fictional yield sensitivity; confirm current rates, fees, eligible funds and liquidity before forecasting income.

We map engineering time and technical uncertainty to qualified expenses so your documentation holds up under review. Qualifying research requires documented technical uncertainty, experimentation, eligible activities and qualified expenses; custom software or AI work does not qualify automatically. The 6-14% range is an illustrative sensitivity, not a statutory credit rate. Eligible small businesses can elect up to $500K against employer payroll taxes, subject to actual credit, liability and election rules. Application starts with employer Social Security tax up to $250K per quarter; remaining credit can reduce employer Medicare tax, subject to liability and carryforward rules. The less-than-$5M gross-receipts test is joined by a five-tax-year receipts-history test and group rules. Review Form 6765 and Form 8974 requirements before claiming.

We track credit drawdown and model post-credit unit economics before your runway takes a hit. Startup credits are conditional provider offers, not cash and not guaranteed retroactive refunds. Microsoft describes up to $150K of Azure credits unlocked over time; AWS, as reviewed October 8, 2026, advertises up to $200K for eligible Portfolio applicants, with different Founders and invitation-only tiers. Confirm approved products, prior usage, expiry and provider access before forecasting savings. Do not assume a combined award; each provider sets its own tiers. We help document eligibility and application timing; approval belongs to each provider.

We separate production inference COGS from R&D training compute so you see real gross margin by feature and customer tier. Cost of goods sold (COGS), the direct costs assigned to serving revenue, may include inference and data licensing. Training compute may instead be research and development, depending on what the spending supports; agree the classification with your accountant. Standard SaaS metrics like gross margin and customer acquisition cost (CAC) payback need to be calculated differently for AI companies. Match usage and cloud invoices to the relevant customer or workload before comparing gross margin; blended revenue growth can hide an expensive inference-heavy feature.

This Level operating measure requires time records or a disclosed allocation and consistent treatment of engineering costs as direct costs or research and development. For a pre-revenue team, use cost per milestone instead. From approved books, payroll and cloud exports or supported connections, Level’s team reviews gross-profit ARR (annual recurring revenue adjusted by the chosen gross margin), burn multiple (net cash spent / net new ARR) and runway (how long the cash plan lasts).

Separate routine model subscriptions, ordinary configuration and production hosting from activities that actually satisfy Section 41. Then review the group, receipts-history and timely election tests before using Form 6765 and Form 8974; an AI product label alone establishes no credit.

We help assemble the application records and timing. For other model or cloud providers, confirm current offers directly, including covered products, expiry and prior-usage rules. Bedrock or Azure OpenAI coverage must be confirmed for your particular offer. Approval belongs to the provider.

We build runway models that account for usage-based COGS, training spike costs, and the gap between top-line ARR and gross-profit ARR.

A magic number compares recurring-revenue growth with sales-and-marketing spend; any adjustment for AI delivery costs must be stated. We prepare reporting around your actual investor requests and the source records available.

A 4.5-5% yield is a fictional sensitivity, not a current quote from Mercury, Ramp or Rho. At illustrative $5M invested for a full year at 4.8%, gross interest is $240K before fees and taxes; changing balances or rates changes the result. A $10M balance at 4.5-5% would yield $450K-500K on the same assumptions. Check current terms and access before moving cash; income is not a guaranteed engagement payback.

It requires qualifying original-issue C-corp stock, an eligible business and shareholder, applicable asset tests and the required holding period. For stock acquired after July 4, 2025, the phased exclusion is 50% after 3 years, 75% after 4 years and 100% after 5 years; the per-issuer exclusion limit is the greater of the applicable $15M dollar limit or 10 times qualifying basis, with prior-exclusion coordination and other limits. Earlier stock uses different rules, including the older $10M limit. Formation alone does not secure eligibility; we coordinate records and tax review before issuance or an exit. Founder review can also include an 83(b) election (recognizing income on qualifying nonvested property at transfer, with a filing deadline no later than 30 days after property transfer), incentive stock options (ISO) and alternative minimum tax (AMT), as your legal and tax advisers determine.