What are unit economics in AI, and how do they determine the return on investment?

Unit economics answer a simple question: how much does a company earn and spend on each unit of a product, such as a query, user, transaction or subscription? In traditional SaaS, for example, this unit is inexpensive. As each additional user adds almost nothing to the company’s costs, the gross margin typically ranges from 75 to 85%. With AI products, however, things are different: each unit of consumption requires actual computation at the time of the request.

The cost of an AI product comprises several layers, each of which is consequential for the company’s valuation.

• Inference refers to the cost incurred when the model generates a response, i.e. the computational resources consumed when a user submits a request. Unlike model training, which takes place only once, inference scales with every new customer and every new query. This means it behaves as a variable cost rather than a one-off capital expenditure.
• Companies either rent computing power from hyperscalers or invest in their own infrastructure for cloud and GPU capacity. Both options are expensive: renting ties margins to the supplier’s prices, while in-house clusters require capital expenditure that is difficult to justify without guaranteed workloads.
• Many AI products marketed as 'fully automated' actually rely on people to annotate data, check response quality, or handle complex cases the model cannot resolve independently. This labour is a hidden cost that is rarely highlighted in marketing materials.
• Quality monitoring: model output can be unpredictable, so companies allocate resources to systems that detect biases and errors, as well as to continuous testing of response quality in production.
• Retraining: models become outdated, competitors catch up and user behaviour changes. Consequently, companies regularly retrain or refine the model, with each iteration incurring new computing and data costs.

This is precisely why the question ‘How much does it cost to serve one additional user?’ is far more complex in AI than in traditional software, and why it must be at the heart of any investment assessment.

It is also important to understand the value that the company derives from each unit of consumption. In SaaS, the subscription price remains fixed regardless of usage intensity. In contrast, in AI products, revenue and costs increase with each request.

It creates a situation in which revenue growth says nothing about the quality of the business: a company could double its revenue while simultaneously tripling its operating losses if its pricing fails to keep pace with the actual cost of the service provided. Investors should therefore focus not on absolute growth figures but on the dynamics of the gap between the price customers pay and the costs the company incurs for each request.

Signs of an AI bubble

Overheating, capital concentration, multiples, and the share of AI in indices

The first sign of overheating is the concentration of capital in a small group of companies. Over the past two years, a significant proportion of all venture capital investment in the technology sector has been directed towards AI. The lion’s share of this money has gone to a handful of laboratories and infrastructure players rather than being spread across the wider market. This is a classic sign of the late stage of the cycle, when capital chases a narrative rather than a diversified set of business models.

The second sign is revenue multiples that have become detached from fundamentals. Early-stage startups are raising funding rounds at valuations of twenty to thirty times annual revenue or more. Historically, however, a healthy, publicly listed SaaS company with strong growth would trade at 8 to 15 times revenue. Expectations of future growth explain this difference. However, when these expectations become the main asset on the balance sheet, the market becomes vulnerable to any disappointment in the pace of progress of these business models.

A third sign is the proportion of artificial intelligence-related companies in the market capitalisation of leading stock indices. When a small group of companies accounts for the majority of an index's gains and their valuations are based on the expectation of a technology that has yet to demonstrate consistent profitability through mass adoption, the market becomes concentrated in areas that were previously diversified. This does not mean that a crash is inevitable. Still, the system’s vulnerability has increased, and any investment thesis today must take this structural risk into account rather than ignoring it in pursuit of growth.

The fourth factor, often underestimated, is the speed at which capital changes hands, leaving no time for verification. Funding rounds are now closed in days rather than weeks, with due diligence reduced to vetting the team and reviewing a technology demonstration. Meanwhile, the company’s financial model takes a back seat. When the speed of a deal takes precedence over the depth of analysis, the market systematically undervalues the risks associated with unit economics. It is precisely these underestimated risks that usually surface first during any market correction.

A checklist to help investors identify genuine unit economics

Business health indicators: burn multiple, gross margin, net revenue retention (NRR), payback

The burn multiple is the ratio of capital spent to the increase in annual revenue over a given period. A figure below 1.5 is considered a strong result, while a figure between 1.5 and 2 is considered acceptable for an early-stage company. The above three, however, signal that the company is buying growth at an unjustifiable price.

Gross margin for an AI company is a separate matter. While in traditional SaaS, a margin below 70% raises questions, for an AI product with active inference, a margin of 50–60% can be considered an acceptable starting point. However, if the margin remains negative or near zero for years, it indicates problems with the business model.

Net revenue retention (NRR) shows the extent to which existing customers increase their spending over time. An NRR above 120% indicates genuine product-market fit and the product's ability to integrate more deeply into customers' workflows. In contrast, an NRR below 100% means the company is losing revenue faster than it generates new one, so any growth is solely due to new sales.

The payback period is the time it takes for a company to recoup the costs of acquiring a customer. For AI products, where costs continue to accrue after the contract is signed, a healthy payback period should be shorter than for traditional SaaS products — ideally within 12 months. Additional inference costs erode the contract's margin throughout its term.

These four metrics should be considered together, as strong performance in one area can mask weakness in another. For instance, a company may demonstrate an excellent NRR thanks to existing customers increasing their product usage. However, if the company incurs more costs than it earns per additional request, growth in NRR will accelerate cash burn rather than improve the business's economics. Therefore, rather than a one-off snapshot at the time of the investor presentation, the correct approach is to require the company to provide quarterly trends for all four metrics simultaneously.

Red flags: negative margins, 'AI hype,' and a lack of a data moat

The first warning sign is a consistently negative gross margin without a clear plan to rectify it. If a company is losing money on every additional user request and cannot demonstrate a clear path to achieving positive margins through model optimisation, changing its pricing policy or transitioning to its own infrastructure, then revenue growth will only accelerate the burning of capital.

The second sign is the so-called 'AI wrapper': a product that is essentially just a thin interface over a third-party large-language-model API. It has no logic, data or technological advantages of its own. These companies are vulnerable in two ways: the model provider may change prices or release a competing feature for free, and competitors can easily copy the interface.

The third warning sign is the absence of a 'data moat' — that is, a unique dataset or user feedback that improves the product over time and that competitors cannot easily replicate. Without this, a company's advantages depend solely on marketing and speed to market rather than on the business's structural resilience.

Green flags: first-party data, workflow lock-in and vertical focus

The first positive sign is proprietary data collected through customers' actual use of the product. This data cannot be bought or copied from outside sources. This creates a cycle in which every new user improves the product for everyone else, while competitors are left playing catch-up for years.

The second positive sign is the product’s deep integration into clients’ workflows. Switching suppliers requires more than just signing a new contract; it also involves restructuring internal processes, training the team, and migrating historical data. This workflow lock-in gives the company pricing power and revenue predictability, which translates directly into a healthy NRR.

The third positive sign is a vertical focus on a specific industry or function rather than attempting to be a one-size-fits-all solution. Vertical AI companies have a better understanding of the relevant data, regulatory requirements and customer pain points. This enables them to build higher barriers to entry and command a premium price compared to their horizontal competitors.

The fourth characteristic is the founders' transparency regarding the cost structure. Teams that show investors a breakdown of costs by line item and openly discuss the risks of dependence on model suppliers usually have a deep understanding of their product’s economics. Conversely, evasive answers to specific questions about the per-unit margin usually mean that the founders have either not calculated these figures or are not satisfied with the results.

Why are profit margins lower for AI companies than for SaaS companies, and might this change?

A key factor: The cost of inference, as well as dependence on hyperscalers and chip manufacturers

Each time an AI model is queried, computing power is consumed, which must either be rented from a cloud provider or run on hardware purchased from a limited number of chip manufacturers. This dependence leaves AI companies at the mercy of pricing. If a supplier increases the cost of computing or chips, the product company’s margin is instantly squeezed. Meanwhile, the ability to pass on these costs to customers is limited by competition.

An additional risk is that access to the most powerful computing resources is often allocated not only on price but also by priority. Large research laboratories and hyperscaler strategic partners receive guaranteed quotas. Consequently, smaller product companies may face bandwidth constraints precisely at times of peak demand, which directly affects the quality of service they provide to their customers.

Human-in-the-loop and the costs of AI quality control

Even the most advanced models make mistakes, and the higher the cost of an error for the customer — in finance, medicine, or law, for example — the more human oversight needs to be built into the product. This means that part of the 'AI product' is actually a service business, with people at the heart of the process. Service businesses have historically had lower margins and found it harder to scale up than pure software businesses.

Factors driving margin growth: Model optimisation, in-house GPU clusters and falling token prices

At the same time, several structural forces can gradually increase margins in the AI business. Firstly, the cost of generating tokens at a comparable level of quality is steadily falling due to competition among model providers and architectural improvements.

Secondly, sufficiently large companies are moving away from renting cloud computing power and towards using their own or long-term-leased GPU clusters, thereby reducing the marginal cost of computation.

Thirdly, optimisation techniques such as model distillation, response caching, and routing queries between models of different sizes based on task complexity enable reductions in inference costs without compromising end-user quality.

Companies that invest in all three of these areas simultaneously have a real chance of bringing their margins closer to those of traditional SaaS within a few years. In contrast, those that simply resell someone else’s API have virtually no chance of doing so.

Who actually makes money in AI? An analysis of the value chain

'Pickaxes and shovels': chip manufacturers and hyperscalers

History has shown us that during tech bubbles, it is not the prospectors who make the most money, but the sellers of pickaxes and shovels. In the current AI cycle, this role is played by manufacturers of specialised chips for training and inferencing models, as well as major cloud providers that supply the necessary infrastructure. These companies generate revenue regardless of which AI application company prevails in a particular niche — they profit from the industry's investment in computing power.

Next in the chain are the laboratories that create foundational models. Their business model depends on their ability to monetise their research through APIs and subscriptions more quickly than they spend on training the next generation of models. Only at the top of the chain are the applied product companies, which build the interface, business logic, and distribution on top of third-party or in-house models.

Unit economics are the least predictable at this level, as these companies simultaneously pay for infrastructure-tier computing power and compete for customers in a crowded market. For investors, this has a clear implication: the infrastructure tier has historically offered lower but more predictable returns. In contrast, the application tier offers higher potential returns but comes with a significantly higher risk of business failure.

An intermediate level in the chain is often overlooked: companies that provide tools for developing, testing, and deploying AI products, such as model orchestration platforms, quality monitoring systems, and tools for fine-tuning models using client data. Rather than monetising the model itself or the finished product, these companies monetise the infrastructure surrounding the process of building AI solutions. Their business model is often closer to classic B2B SaaS, as they sell a tool rather than charging for inference for each client’s end users.

What will happen if the AI bubble bursts? Scenarios and portfolio protection

Deflation versus detonation: the most likely correction scenarios

It is important to distinguish between two fundamentally different market correction scenarios. The first is 'detonation': a sharp, synchronised collapse in valuations caused by a sudden reduction in access to capital or technological disappointment. This occurs when the pace of model improvement slows more quickly than the market anticipates. This scenario resembles the dot-com crash, leading to the mass closure of companies trapped by negative margins and a lack of cash reserves.

The second, more likely scenario is 'slow deflation': multiples gradually decline over several years as the market learns to distinguish between companies with genuine unit economics and those sustained by narrative alone. In this scenario, infrastructure-level companies and those with proven margins continue to grow. At the same time, weaker players gradually lose access to funding and/or are consolidated or disappear from the market without a spectacular collapse. Historically, the second scenario occurs more frequently than the first, as markets rarely correct themselves in unison without an external macroeconomic shock.

How investors can prepare: Diversification, a focus on cash flow, and enhanced due diligence

Regardless of the eventual outcome, three practical steps can be taken right now to reduce portfolio risk. Firstly, diversification is key, not only across companies, but also across levels of the value chain. Combining infrastructure assets with a limited number of carefully vetted, application-based companies reduces the portfolio’s reliance on a single narrative.

The second step is to deliberately focus on companies that can demonstrate a path to positive cash flow in the near future rather than merely focusing on revenue growth. A company that can demonstrate a reduction in its burn rate from one quarter to the next is far more resilient to a sudden closure of the funding window than a company that relies on constant new rounds of funding to cover operating costs.

The third step is to conduct more rigorous due diligence, specifically at the unit-economics level rather than merely at the narrative and team levels. This means demanding access to actual cost structure data, rather than just a presentation with general growth figures, and asking tough questions about reliance on specific suppliers, models or calculations, even before signing the term sheet.

In an era where unit economics are essential, how should one value an AI startup?

Burn multiples and ‘tiny teams’ as the new valuation standard

The market is gradually shifting from a growth-at-any-cost model to one where capital efficiency is valued as highly as growth rate. In this context, each employee's productivity is becoming an increasingly popular metric: small teams using in-house AI tools can generate revenue comparable to that of teams ten times their size.

These 'tiny teams' have lower operating costs and demonstrate that the founders understand the product's economics in great detail rather than relying on team size to compensate for inefficient processes. For investors, a high revenue-per-employee figure is a valuable, albeit indirect, indicator of a disciplined decision-making culture.

A team accustomed to achieving more with fewer resources during the product development phase will typically apply the same discipline to managing infrastructure costs after launch.

Due diligence questions regarding the economics of an AI product

When evaluating a specific AI startup, an investor should seek clear answers to several structured questions. The first is: what is the exact cost of goods sold (COGS) breakdown? Specifically, how much is attributable to inference, human oversight and data storage and processing, and how has this changed over the last few quarters?

Secondly, what is the specific path to achieving a positive gross margin? Which technical and commercial steps are planned, and what is the expected impact of each, expressed as a percentage?

Thirdly, what is the extent of the business’s reliance on a single supplier for a model or cloud infrastructure, and how would the product’s economics be affected if that supplier’s prices rose by 20 or 30%?

Fourthly, how secure is the company’s data? Is there a legally enshrined right to use customer data to improve the model? Is there a real barrier preventing a competitor from gathering a similar dataset within a reasonable timeframe?

In addition to these four questions, we should ask how the product’s cost price would change if the company’s user base were to grow tenfold within a year. While linear scaling of costs alongside revenue is acceptable, super-linear cost growth means the business model breaks down at the moment of success, when demand for the product is highest. Investor confidence is most vulnerable to disappointment.

The answers to these questions provide a far more accurate picture of a company’s future sustainability than any presentation. The AI wave will transform entire industries — this is no longer a matter of debate but of capital. If it is invested in companies without sound unit economics, it will not survive this transformation alongside the winners.

Rather than opting out of the biggest technological shift of the decade, investors must learn to distinguish between companies building sustainable businesses and those merely seizing the moment, with the market willing to pay more for storytelling than for profits.

Discipline in assessing unit economics today is not a constraint on portfolio growth; it is a tool that will enable you to remain in the market until the cycle's winners become clear to everyone.