Finance & Unit Economics for AI

Software companies got fat on 80% gross margins because the marginal cost of the next customer was a rounding error. AI companies do not get that gift. Every request your product serves burns real compute, and that compute lands in cost of goods sold, not in the R&D line where investors forgive it. This chapter is about the one number that quietly decides whether your beautiful AI product is a business or a science project: the margin left after you pay for the tokens.

13.1AI COGS: inference, not just servers#

In a classic SaaS profit-and-loss statement, cost of goods sold (COGS) is hosting, a support team, and some third-party tooling — a thin, mostly fixed layer under a torrent of near-free revenue. In an AI company, COGS grows almost linearly with usage, because the dominant term is inference: the compute you spend generating each response. The industry has re-learned this the hard way. The average AI product shipped a 52% gross margin in 2026 versus the ~80% that defined mature SaaS, and inference alone consumes roughly 23% of revenue at scaling-stage AI B2B companies.¹ Traditional SaaS runs total COGS of 10–25% of revenue; AI puts a single line item at the top of that range before you have paid for anything else.²

Your AI COGS has five buckets, and founders who only budget the first one get blindsided by the other four:

The mental shift is this: inference is a variable input cost, like flour for a bakery, not a fixed asset like the oven. You cannot amortize it away with scale. If anything, it grows with success. Two structural choices determine its size, and both live in other chapters: which provider and model you buy (see Chapter 7) and how you architect routing, caching, and context (see Chapter 9). Here we only turn those choices into numbers.

13.2Unit economics: cost and margin per action#

Seat-based pricing is a lie for AI products, because a “seat” that fires 5,000 agent runs a month costs you fifty times more to serve than one that fires 100. The only honest unit of analysis is the action: one resolved ticket, one generated document, one completed agent task. Compute your fully-loaded cost per action, compare it to the price you charge for that action, and you have your true unit economics.

Cortex, our autonomous support agent for mid-market SaaS companies, sells on exactly this unit: price per resolved ticket. This is the emerging standard. Intercom’s Fin charges $0.99 per resolved conversation;⁶ Zendesk launched outcome pricing at $1.50 per automated resolution on committed volume ($2.00 pay-as-you-go); Sierra charges roughly $1.50 per resolved interaction and rode that model from $100M to $200M ARR and a $15.8B valuation; Decagon pairs a ~$50K platform fee with a custom per-resolution rate.⁴⁵ Cortex prices at $1.20 per resolved ticket — cheap enough to undercut a human ticket (which costs the customer $15–$35 fully loaded) by an order of magnitude, expensive enough to have a business, if the cost side cooperates.

Now the cost side. A single resolved ticket is not one model call — it is a conversation. Cortex’s naive v1 architecture, running everything through a frontier model, looked like this:

Cortex cost per resolved ticket — naive (frontier-only)
  intent classification        1 call   ~2K tok   $0.03
  retrieval + rerank           2 calls  ~1K tok   $0.02
  reasoning / tool selection   3 calls  ~18K tok  $0.34
  answer drafting + revision   2 calls  ~9K tok   $0.21
  in-prod eval (judge)         1 call   ~4K tok   $0.08
  ─────────────────────────────────────────────────────
  inference subtotal                              $0.68
  retrieval infra + tool APIs                     $0.09
  observability                                   $0.02
  human escalation (18% of tickets × $0.22 amort) $0.04
  ─────────────────────────────────────────────────────
  fully-loaded COGS per resolved ticket           $0.83

At $1.20 price and $0.83 cost, Cortex’s contribution margin is $0.37, or 31%. That is a company that dies at scale: every dollar of new revenue drags 69 cents of cost, sales and G&A eat the rest, and the burn multiple never converges. The founders’ instinct — “we’ll fix it with volume” — is exactly wrong, because there is no volume discount on a variable input.

13.3The gross-margin reality of AI products#

Here is the uncomfortable table every AI founder should stare at. Gross margin is not a vanity metric; at exit it is a valuation multiplier, because a 70%-margin dollar of revenue is worth far more than a 50%-margin dollar.

Investors have caught on. In 2026 they check gross margin “more carefully than before to make sure AI inference costs are not quietly eating your profit margin”; a GPT-wrapper at 50–60% invites questions about defensibility, while proprietary or fine-tuned systems that clear 70%+ get the premium.¹³ For a SaaS company bolting AI onto an existing product, the typical hit is 12–17 points of gross margin — enough to reset a company’s valuation category.³

The way out is not to charge more (the market caps your price) but to engineer your COGS down. The single highest-leverage move is model routing: send the easy majority of requests to a small, cheap model and reserve the frontier model for the hard minority. Route the easy ~60% down and you save roughly 38% of inference spend with no measurable accuracy penalty; aggressive setups with semantic caching report 50–70%.⁹¹⁰ (The how — cascades, confidence thresholds, routers — is Chapter 9’s territory. Here we only bank the savings.)

Cortex did exactly this. It added a router: a fast, cheap model (DeepSeek-V4 class, ~$0.14/$0.28 per 1M tokens⁸) handles intent, retrieval, and the 62% of tickets that are routine password-reset-and-billing questions end to end. Only genuinely ambiguous tickets escalate to a frontier model (Claude Opus 4.8 class, $5/$25 per 1M⁷). Semantic caching kills repeat questions outright. The rebuilt cost stack:

Cortex cost per resolved ticket — routed + cached
  cheap-model path (62% of tickets)      blended  $0.11
  frontier path (38% of tickets)         blended  $0.29
  ─────────────────────────────────────────────────────
  blended inference per resolved ticket           $0.18
  retrieval infra + tool APIs                     $0.06   (cache cuts calls)
  observability (sampled 20%, not 100%)           $0.01
  human escalation (11% × $0.22 amort)            $0.02
  ─────────────────────────────────────────────────────
  fully-loaded COGS per resolved ticket           $0.27

Same $1.20 price, new $0.27 cost: contribution margin jumps to $0.93, or 78%. Cortex went from a doomed 31% product to a fundable 78% product without touching its price or its accuracy target — purely by treating inference as an engineering cost to be optimized, not a bill to be paid. That is the whole game.

13.4Burn, runway and compute budgeting#

AI startups burn differently from SaaS, and the difference is compute. GPU spend consumes 40–60% of technical budgets in the first two years, and AI startups burn roughly 2× what SaaS companies spend on infrastructure.² Your runway math has a term SaaS never had: a compute line that can spike with a viral week or a single runaway agent loop.

Compute budgeting splits into two very different problems. Inference compute scales with revenue and should be modeled as a variable COGS (dollars per action × actions). Training / fine-tuning compute is lumpy capex-like spend you incur ahead of revenue. If you self-host, an H100 runs a market-median $2.29–$3.12/hr in 2026 — as low as $1.38/hr on verified neo-clouds, but $6.88/hr on AWS p5 and ~$12/hr on Azure.¹¹¹² A single H100 running 24/7 is $1,500–$3,000/month; an eight-GPU fine-tuning cluster is $12,000–$24,000/month.¹² Provider choice alone is a CFO decision, not just an infra one (see Chapter 7).

Three disciplines keep compute from eating your runway:

• Underwrite COGS at build time. Before you ship a feature, estimate its cost per action at expected volume. A feature that is margin-negative at scale should be re-architected or repriced before launch, not after the invoice arrives.
• Set token budgets and hard caps per request and per customer. An agent with no step limit is an unbounded liability; one pathological ticket looping through a frontier model can cost more than a hundred normal ones. Cap steps, cap tokens, alert on anomalies.
• Exploit credits, but do not build on them. GPU credit programs and provider startup deals routinely offer $100K–$350K in compute; take them, but model your unit economics at list price, because credits expire and your margins must survive the cliff.

13.5Financial modeling and the metrics investors probe#

Build your model bottom-up from the unit, never top-down from a TAM percentage. Start with cost and price per action, layer in volume growth, model the COGS curve bending down as routing and caching mature (a defensible, dated assumption you must justify), then stack fixed costs to get to burn and runway. The single most important output is the trajectory of gross margin over time — investors want to see a credible path from today’s 50s to the 70s, driven by specific engineering levers, not hope.

The metrics an AI-literate investor will probe in 2026, and where you need to land:

Two AI-specific traps hide in these numbers. First, CAC payback must be computed on gross-margin dollars, not revenue — at 50% margin, a customer takes twice as long to pay back as the top-line suggests, so thin AI margins quietly wreck otherwise-healthy sales efficiency. Second, watch NDR versus COGS growth: usage-based AI revenue expands beautifully, but if a customer’s usage doubles and your cost to serve them doubles with it, net-dollar retention flatters a P&L that is not actually getting more profitable. The good news: AI-native companies are already posting burn multiples of 0.8–1.2×, beating traditional SaaS at nearly every stage, precisely because automation holds headcount down while revenue compounds.¹³

For Cortex, the model that raises a Series A is legible: 78% and rising gross margin (with a stated path to low-80s as the router improves), cost per resolved ticket falling ~15% per quarter, revenue per employee north of $400K because the product is the labor, and a burn multiple under 1.0×. Every one of those numbers traces back to the same discipline — engineering inference down instead of paying it as a bill. That is what makes it an AI company worth funding rather than a SaaS company with a token problem. The fundraising story built on top of these numbers is Chapter 14.