Bharath Srivatsan·September 2026

Tokenomics

AI is changing software economics. LLMs are increasingly both a flywheel for engineering productivity and a new source of non-negligible marginal costs (as businesses build AI-enabled features that require large amounts of compute).

These changes threaten to upend the traditional market structure for software companies. What might these new dynamics imply for firm sizes and competition? What might they mean for the distribution of surpluses between consumers and producers (model creators like Anthropic, inference providers like AWS, and application layer businesses like Salesforce)? How might they change as regulatory shocks or compute scarcity take effect? This multi-part series will use simulations and surveys of recently-published literature to explore these questions.

Part 1: AI and the application layer

1.1 Background

While physical businesses experience economies of scale as they sell to more users (their average cost per unit falls as fixed costs are spread and production and distribution get more efficient), they still incur a real cost to build each incremental unit.

In fact, physical businesses can even experience dis-economies of scale, as extraordinarily large firms struggle with coordination costs and higher prices for increasingly scarce inputs.¹ For businesses past that point, there is eventually a size at which it is cheaper for smaller entrants (on the other side of the cost curve) to serve each marginal user. This can be a natural brake against monopolization that doesn't exist in software markets.

Software is different. Most software costs a lot to initially develop, but has almost zero marginal cost; serving each additional user is close to free. This means that average costs decrease even more substantially with scale: spreading fixed costs over a billion users is cheaper than over a hundred.²

Many other factors³ also disadvantage potential entrants and encourage increased concentration in software markets:

• Bundles: Bundling products deters new entrants into a market: competitors may need to rebuild the whole suite to displace users. Because adding another product to a bundle costs ~nothing for software firms, incumbents are more incentivized to create large bundles
• Free tiers: zero marginal costs allow software firms to build free tiers that remove the potential strategy of new entrants offering initially worse but cheaper products (in order to gain a foothold)
• Switching costs: Moving from one product to another incurs costs as well, benefiting incumbents. Studies have found that migration, integration, and training can cost anywhere from 1.3 to eleven times the cost of the software itself.⁴
• Network effects: If a product is worth more to each user as more people use it, the leader's product is better because it leads. This can lead to accelerated consolidation, as the prospect and momentum of a winner leads to buyers preferring them⁵

Taken together, this is why software markets produce Googles and Microsofts and Metas rather than fifty mid-sized competitors. Even in new categories, companies understand the immense returns to scale and thus invest in massive, aggressive user land grabs.

1.2 Nonzero marginal costs and scale advantages

With AI, the era of zero-marginal-cost software may be over. An increasing number of applications are building AI features for their users that cost them non-negligible amounts in inference⁶, and the introduction of nonzero marginal costs raises the lowest price they can charge to break even on each unit.

Counterintuitively, this may advantage smaller firms. We can use a simple simulation to show how non-negligible marginal costs can decrease the advantage that larger firms have over smaller ones.

But marginal costs aren't the only way AI is changing businesses' unit economics. Tokens are also being consumed by application layer firms as an input into their fixed costs (since engineering teams can use AI to speed up development).

Together, these simulations suggest that the gap between large and small firms is smaller than it was before AI.

• When tokens are expensive, larger firms have a smaller advantage over new entrants, because fixed costs play a smaller role in the overall cost of the software and amortizing them over more users matters less
• When tokens are cheap, the product is cheaper to build in the first place, so there is less fixed cost to spread at all⁸

1.3 AI and anti-incumbency effects

AI may be weakening some of the other pro-incumbency levers of software as well:

Anecdotally, it does seem like we're seeing more competition in software markets. New features seem to percolate through competitors extremely rapidly, and new (software) business formation is way up.¹⁰

It isn't a given that this will remain true, though. AI might eventually introduce powerful new vectors for concentration:

• Data-enabled user customization may become more important, as AI-enabled software and custom models fit users' needs better and better¹¹
• With a new wave of software, attracting customers and getting them to change behavior may become harder than before. Distribution may then become more important, benefiting incumbents
• Keeping up with feature wars and a more expensive (risky) security landscape may dramatically increase the cost of building winning software (even with cheaper token costs). This could yield benefits to high capitalization as wealthier firms outcompete startups
• Many on Twitter seem to believe that AI spells the death of SaaS, as users rebuild all software in house. In this world, the model/inference layer captures the value of the application layer and consolidates the overall software ecosystem

These factors don't yet seem to have taken effect. New apps are exploding, and Ramp's AI Index shows AI budgets growing alongside software budgets rather than in place of them. For now, at least, we are in the golden age of software.

1.4 Token costs and market-level dynamics

So far, we've discussed individual firms and their responses to varying token costs. We can go one step further, and model how token costs impact market-wide dynamics like firm counts and gross margins.

For a given software vertical, we can model the market as a Salop (1979) circular city. Consumer preferences are spread uniformly, and they each buy from their "nearest" product, bearing a cost that models how much they care which product they get. Firms freely enter so long as they can cover their fixed costs (free entry is more likely to be true given the weaker incumbency effects discussed in 1.3). This allows us to back into closed-form formulas for both the number of firms and the price each customer pays.

The specific numbers here are illustrative, but we can see some takeaways as a function of token costs:

• As token costs rise, gross margins and the equilibrium number of firms fall (the latter because the market shrinks)
• Higher token prices raise the fixed costs of building software and shrink the number of users in the category. The latter spreads costs among fewer people, meaning pricing rises faster than the token price increase alone
• As applications become more token hungry (as users use more tokens each month), these effects become more pronounced

1.5 Pricing dynamics

For a long time, seat-based subscription pricing was the natural model for software, since each additional user cost the same (negligible) amount to support. With the introduction of material token costs and wide variance in usage patterns, flat seat rates no longer make sense.¹²

Instead, we're seeing new pricing structures take off:

• Usage-based billing, where firms charge users per token or per action. Sometimes this is modeled as hybrid pricing, where users are charged a flat fee plus a usage-based overage rate
• Outcome-based pricing, where firms charge for successful outcomes (resolved customer support cases, etc.)

We can build a simulation to show how each pricing model performs under varying user conditions. We first model the following variables:

• θ, preferred usage: the number of tokens a given consumer would use if tokens were free. This is a lognormal model (with a long tail), as power users can burn extraordinary numbers of tokens
• k, efficiency: the number of tokens needed for a given result. This is also lognormal, since some users have more complex tasks (coding tasks in a large, messy codebase needs more tokens than doing the same in a small one)
• w, preference: what each result is worth to a user, in dollars. This is modeled independently of θ - some users' workloads may be low value but burn lots of tokens (e.g. using gpt-6-astra to grep for code)

We can now see¹³ how each pricing regime responds to rising token costs:

• Flat / seat-based: The flat price must rise to cover heavy users as token costs rise, which then pushes out light users and raises the average bill of those remaining.
• Metered billing: Metering never loses money. Users for whom the results are not worth the token costs do not use the product at all, and usage goes down for those who do use the product since they stop at the marginal point beyond which usage is no longer worth it. This is what similar case studies like usage-based broadband pricing show.¹⁴
• Outcome pricing serves the most users and makes the most money when tokens are cheap (as users are paying for something that costs less to produce), but drops margin fastest as tokens get more expensive. The firm is paying for its least efficient users, who burn many tokens to achieve few results
• Hybrid pricing: The flat fee adds a floor to pricing, excluding some users who would have paid just for the tokens-consumed. On the other hand, the fee extracts a surplus from users who value the product above what metering alone would cost them. In other words, the strategy is to charge fewer users, more. Hybrid pricing is never worse than metering here, because the fee chosen by the model maximizes profits, and a hybrid price with a $0 fee is a metered price

The main takeaway is that all of these approaches dominate flat / seat-based pricing, and as token costs remain volatile, firms can protect their margins by shifting pricing structures.¹⁵

1.6 Summary

AI is changing the economics of application-layer software firms:

• AI re-introduces nonzero marginal costs to software, while reducing the upfront price for building software. Both of these effects disproportionately benefit small firms over large (though they do not even the playing field on their own)
• AI may also reduce or reverse the other software-specific factors that previously benefited incumbents
• As token costs rise, gross margins, market size, and the number of firms supported in each vertical all fall
• Nonzero marginal costs encourage businesses to shift from seat-based pricing to outcome-based, metered, or hybrid approaches

Part 2 introduces the labs and inference providers, models how token prices are set, and explores how various levers affect the creation and distribution of consumer surpluses. Subscribe to receive an email when it's published.