14 August 2026 · AI & Valuation · Research Article

When model quality converges, what are investors actually valuing?

If model-level differentiation can weaken faster than conventional forecasts assume, where does the duration of value move?

By Luis Fontana / Mjolnir Capital

Informational only — not investment advice, an offer, or a solicitation.

I came across two AI announcements recently, published seven weeks apart, and the contrast between them got me thinking.

On February 4, 2026, ElevenLabs announced a US$500M funding round at a US$11B valuation. The company said it had closed 2025 with more than US$330M in annual recurring revenue (ARR).

On March 23, Mistral released Voxtral TTS, a 4B-parameter text-to-speech model available through an application programming interface (API), with a version released as open weights under a non-commercial license.

In Mistral’s own multilingual, zero-shot voice-cloning evaluation, Voxtral received 68.4% of listener preferences against ElevenLabs Flash v2.5. Mistral launched the API with introductory pricing of US$0.016 per 1,000 characters. ElevenLabs listed Flash and Turbo at US$0.05 per 1,000 characters on its public API pricing page, although effective pricing can vary by plan and volume.

The comparison has limits. It was Mistral’s benchmark and it covered a specific use case. There are also two different forms of access here: the API can compete commercially on price, while the downloadable weights make experimentation easier but carry a non-commercial license.

On May 5, ElevenLabs reported that ARR had passed US$500M. Given the timing, that figure mostly reflects commercial momentum established before Voxtral had time to affect customer behavior. It tells us little about the competitive impact of Mistral’s release. It does show that ElevenLabs had built a rapidly growing business around the technology, which makes the valuation question more interesting than a simple comparison between two models.

At first, I treated this as an interesting voice-model comparison. Then I went back through the announcements, the benchmark methodology, the pricing, and the later revenue update. I also kept bringing it up in conversations because each answer seemed to create another question.

Those conversations tended to go in two directions. One was that model capabilities are commoditizing faster than many valuations assume. The other was that the model may never have been the whole moat in the first place.

Both readings can fit the same facts. There is no clean winner-or-loser story here. The more I looked into it, the less it seemed like a voice-model comparison and the more it seemed like a valuation question.

A part of the technical advantage was challenged quickly and at a lower public price. At the same time, ElevenLabs was reporting rapid revenue growth. Perhaps the model was only one part of what investors were valuing. Perhaps enterprise distribution, customer relationships, embedded workflows, brand, infrastructure, and execution already mattered more.

Or perhaps the market is still working out what any of those things are worth.

One example cannot establish a permanent change in valuation. A private funding round, a competitor’s benchmark, and a later revenue update are still only one sequence of events. But they were enough to make the assumptions behind that valuation worth examining.

What becomes scarce?

Software has always been copied. AI changes how quickly and cheaply it can happen.

A feature that took years to build can now face a credible alternative within a product cycle. A model can lead one benchmark and lose another after the next release. A capability that once justified a standalone company can appear inside a larger platform as part of an existing subscription.

One question kept returning as I read and talked it through: what is the customer actually buying?

If that continues, what exactly becomes scarce?

This does not mean that building a company has suddenly become easy. Capital, compute, proprietary data, enterprise distribution, security, compliance, and execution are still unevenly distributed. The cost of recreating a capability can fall without the cost of creating the business around it falling by the same amount.

Is it still the technology? Is it access to customers? Permissioned data? Trust? The workflow around the model? The ability to move from one model to another without the customer noticing?

And if the scarce asset changes, should the way companies are valued change with it?

That would affect more than the multiple. It would change how long investors can reasonably forecast pricing power, retention, and margins, and how much value they assign to the years beyond the forecast.

A familiar valuation question

Trying to make the question more concrete led me to an idea I have been playing with: “moat half-life.”

The underlying idea is not new. Investors have long thought about the competitive advantage period, meaning how long a company can sustain returns above its cost of capital, and the rate at which those excess returns fade.

“Moat half-life” is not a replacement for that work or a formal accounting metric. I find it useful for a narrower question: how long can the specific advantage being valued today support better pricing, retention, margins, or growth before competition, bundling, or cheaper substitutes begin to weaken it?

AI adds another complication. The advantage may not simply fade. Its source may move from the model to the workflow, distribution, data, trust, or infrastructure around it.

The concept leads to a different set of questions:

What remains if competing models reach similar quality?
Does the company own the customer relationship, or does another platform control access?
What would a customer have to rebuild, retrain, migrate, or reapprove to switch?
Who owns the data, the permissions, and the feedback generated through use?
Do compliance, audit, identity, consent, or trust requirements become harder to reproduce at scale?
Is there a marketplace, transaction network, developer ecosystem, or physical operation that code alone cannot copy?
Can the company change model providers, reduce inference costs, and improve the product without disrupting customers?

None of these questions provides a complete answer. Together, they help separate a temporary product lead from something that may last longer.

The questions also started connecting with three parts of a broader framework I have been developing through Mjolnir Capital: demand quality, operational durability, and capital readiness. A company may have demand today while the durability and valuation of that demand remain uncertain.

Four ways this could play out

The next step was working out how the idea might enter a valuation. The survival-or-failure framing seemed too crude. Companies rarely move from relevance to zero in one step. A business can keep growing while losing pricing power. It can retain customers while margins fall. It can remain useful without producing the returns implied by its last valuation.

I ended up sketching four paths that could be weighted by probability rather than assuming any one of them is certain:

Rapid commoditization. The central feature is replicated or bundled, prices fall, and retention weakens.
Workflow durability. Model performance converges, but integrations, approvals, data migration, and operating habits make the product difficult to replace.
Technical compounding. The leader keeps improving model performance, reliability, tooling, and cost efficiency fast enough to preserve or extend the technical gap.
Business compounding. Distribution, proprietary data, trust, or network effects become stronger as the product scales, even if visible model performance converges.

These are not predictions. They are a way to force different assumptions into revenue retention, gross margins, reinvestment, and terminal value instead of burying all of the uncertainty inside one discount rate.

For AI applications exposed to rapid feature competition, that may mean beginning with a three-to-five-year explicit forecast and extending the duration only where the evidence supports it. It also means avoiding the same risk being counted twice through both lower projected cash flows and a higher discount rate.

But there are still open questions.

Does terminal value shrink, or does it move from the model to the workflow and distribution around it? Should investors accept lower multiples for technical revenue and higher ones for embedded customer relationships? How should a model account for a company that may lose its product advantage but gain from falling inference costs?

There is no settled convention for this yet.

What might last

Across the four paths, the same possible sources of durability kept appearing:

Proprietary and permissioned data that improves through product use
Workflows that are deeply embedded in how customers operate
Direct distribution and customer relationships
Compliance, auditability, identity, consent, and trust infrastructure
Marketplaces, liquidity, transaction networks, and developer ecosystems
Hardware, logistics, regulated delivery, and other physical-world integration
The ability to change models, lower costs, and ship improvements quickly

None is automatically a moat. Data may be non-exclusive. Integrations can be replaced. Distribution can be rented. A network can look active while incentives are doing most of the work.

Whether any of them is a moat would still have to show up in customer behavior and economics.

Simply having one of these assets does not make it a moat. The question is whether it changes customer behavior or company economics: switching, pricing, acquisition costs, retention, margins, or the ability to keep improving the product.

Does retention hold when a cheaper competitor arrives? What happens to gross margins after inference, support, and delivery costs? How concentrated are customers? Are renewals getting stronger? How difficult is replacement in practice? Could a larger platform include the same feature at little or no additional price?

ARR shows that customers are paying today. But the public figure does not show retention cohorts, customer concentration, contract duration, usage mix, gross margins, or how much growth came from expansion rather than new customers. That is not a criticism of ElevenLabs. It is simply the limit of what an external reader can infer from the headline number.

A parallel with digital assets

At some point, this began to feel familiar from digital assets. A protocol can report activity and attract capital while incentives are high. The more difficult question is what remains when emissions fall, competitors appear, tokens unlock, or attention moves elsewhere.

The evidence looks different, but the valuation problem is similar. Recurring use after incentives, credible value accrual, holder concentration, unlock schedules, and liquidity under selling pressure all help test whether present demand can persist.

In both cases, current activity can be real without telling us how long the economics supporting it will last.

The connection opens a much larger rabbit hole: the x402 protocol and HTTP-native payments, agents paying for API calls, data, and compute with stablecoins, programmatic wallets and spending permissions, the possibility of agent-to-agent commerce, and whether blockchains become a settlement layer for an underlying AI-agent economy.

That is a whole other conversation.

What I still don’t know

After reading through the evidence, sketching the different cases, and talking the idea through, I still have more questions than answers. My predictions will probably change as more evidence appears.

My current guess is that model-level differentiation will become harder to value over long periods, while distribution, workflows, data rights, trust, and networks will receive more attention. Terminal value may not disappear. It may migrate to whatever keeps compounding after the technology becomes widely available.

But that is still a thesis.

Frontier model providers may preserve more scarcity than expected. Regulation and compute requirements may concentrate the market rather than open it. Some companies may keep extending their technical lead faster than competitors can close it. Others may discover that strong revenue was evidence of timing rather than durability.

Nobody really knows how these forces will settle, and I think that is perfectly fine.

That uncertainty is what makes this period so interesting. This example does not prove that valuation has entered a new regime. It has made one question feel worth asking more explicitly: where does duration sit, what evidence supports it, and how quickly can its source move?

If model quality becomes broadly available, what will investors still pay a premium for?

Continue the research

For adjacent questions in digital-asset markets, continue with How to Read a Vesting Schedule Before It Changes the Float and Understanding Secondary-Market Discounts for Locked Tokens.

Primary sources

Originally published on LinkedIn.

Informational only — not investment advice, an offer, or a solicitation.

← All research