From High-Margin SaaS to the AI Compute Tax

Over the past two decades, investors have developed a near-standard playbook for understanding software companies: recurring subscription revenue, high customer retention rates, low marginal costs, and gross margins sustainably above 70–80%. As long as growth was sufficiently rapid, even short-term losses could be justified using the 'Rule of 40' or revenue multiples. This was the most alluring aspect of the SaaS era: once software was built, it could theoretically be sold infinitely, with near-zero incremental cost per new customer.

With the emergence of AI, this narrative has suddenly become less clean-cut. Traditional SaaS sells features; AI software sells each inference, each generative output, each model call. The more active users are, the higher the user stickiness—but also the higher the compute costs borne by the company. The cost structure of legacy software companies resembled a toll bridge already built and paid for; today’s AI applications resemble a busy restaurant—every additional guest requires more ingredients, gas, and labor. Revenue and costs are no longer decoupled, marking the starting point where valuation models may need rewriting.

Markets have already seen early signs in recent years. Cloud leaders are benefiting from surging AI demand while simultaneously acknowledging that AI infrastructure investments are compressing gross margins. This isn’t short-term accounting noise—it reflects a fundamental shift in business models. If AI capabilities are merely bundled into existing flat-rate subscriptions, companies effectively absorb variable costs with fixed monthly fees; the more heavily users engage, the lower the gross margin contribution may become. Investors used to focus solely on ARR growth; now they must also ask: behind every dollar of ARR, how much GPU, electricity, memory, vector database usage, and model licensing fees are consumed?

This doesn’t mean AI software is inherently a bad business. On the contrary, if a product genuinely saves enterprises time, boosts revenue, or reduces errors, customers will pay a premium—and AI companies can still create immense value. The issue is that valuation methodologies can no longer cut corners. Traditional SaaS valuations could rely roughly on revenue multiples because gross margin structures were similar across companies. AI software, however, demands scrutiny of inference cost curves, usage frequency, model choices, pricing models, and cost pass-through capability. Two companies each generating $100 million in annual revenue—one with 75% gross margins, the other with only 45%—should logically command vastly different valuations.

More subtly, AI costs are driven by two opposing forces. On one hand, improvements in model efficiency, advances in chips, and maturation of open-source models continuously drive down the cost per inference. On the other hand, as costs fall, enterprises tend to use AI more extensively, deeply, and automatically—evolving from basic chatbots to multi-step agents, long-context processing, real-time voice, image, and video generation. Total token consumption may not rise linearly but explode exponentially. This mirrors the 'cloud bill shock' of the cloud computing era: the cheaper the unit price, the more uncontrollable total usage becomes.

Thus, in the AI era, the most valuable software companies may not be those with the flashiest models, but those that know how to turn computing power into billable value. They will impose more precise limits on free trials, migrate heavy users to usage-based pricing, deploy low-cost models for simple tasks, reserve expensive models for high-value scenarios, and even build proprietary small models to handle vertical workflows. In other words, product managers and CFOs must sit at the same table—because every button click, every auto-complete suggestion, and every long-text analysis could represent a tiny cost item.

Investors, too, must shift from a 'revenue multiple mindset' to a 'unit economics mindset.' The core question for AI software isn't just whether it can grow, but whether it becomes more profitable as it scales. Companies worthy of revaluation should exhibit three characteristics: first, their AI features enable clear price premiums rather than merely serving as retention gimmicks; second, inference costs as a share of revenue decline with scale; and third, deeper customer usage leads to higher renewal and expansion revenue—sufficient to cover compute expenses. Without these three traits, high growth may simply be another term for heavy subsidization.

$Dropbox (DBX.US)$$Microsoft (MSFT.US)$$Applovin (APP.US)$$Canadian Nexus Team Ventures Corp (TEAM.CA)$$NVIDIA (NVDA.US)$

228K Views