Why traditional SaaS multiples are insufficient
A SaaS company with $5M ARR and 80% gross margins commands a very different multiple than an AI company with $5M ARR and 40% gross margins because GPU and inference costs consume a large portion of revenue. The multiple-on-ARR framework penalises AI companies unless their unit economics are adjusted. Investors increasingly look at gross profit multiples rather than revenue multiples for AI companies — a 10x gross profit multiple on a 40% margin company implies a 4x revenue multiple, which looks "cheap" but actually reflects the cost structure.
Token economics — the new unit economics
For API-based AI companies, the fundamental unit is the token. Revenue per million output tokens, cost per million tokens (compute), and gross margin per token are the core metrics. A company generating $20 per million tokens at $8 compute cost has a 60% token margin. As model efficiency improves (smaller models, faster inference), cost per token drops — which means margin expansion is a structural tailwind for efficient AI companies. Investors model token margin trajectory as a proxy for long-term profitability.
ARR per GPU-hour: capital efficiency for AI
How much revenue does the company generate per unit of compute consumed? This metric captures whether the AI product is capital-efficient at scale — or whether compute costs will compress margins as usage grows. A high ARR/GPU-hour ratio suggests the product is delivering enough value per inference to justify the compute cost. A low ratio suggests the business model requires either significant price increases or dramatic compute cost reduction to be viable at scale.
Agentic task completion rate
For AI agent companies, the percentage of tasks completed autonomously without human intervention is a direct proxy for product maturity and scalability. A rate of 95% means 5% of tasks require human review — at scale, that's a manageable cost. A rate of 70% means the product is still semi-manual and labour costs grow with revenue. Investors track this metric quarter-over-quarter as an indicator of whether the autonomy curve is improving. Combined with task volume growth, it tells the story of whether the company is building a scalable business or an expensive human + AI hybrid.
Data moat as valuation driver
Proprietary training data is increasingly the durable competitive advantage in AI — not the model architecture, which can be replicated. Companies with exclusive access to high-quality, domain-specific data (clinical records, legal documents, financial transactions, proprietary user behaviour) command valuation premiums that don't show up in standard metrics. Investors assess: How was the data obtained? Is the access exclusive or time-limited? Does the data compound (does more usage generate better training data)? A strong data flywheel justifies a higher multiple than the current revenue metrics alone would suggest.