Self-Optimising AI: What Automated Model Training Means for Small-Cap Developers

When the machine builds its own successor

There is an old joke in software development: the best program is the one that writes itself. What was long considered a thought experiment is now taking concrete shape with so-called self-optimising AI systems. The US start-up Recursive has published an early real-world test with its coding model Flint: according to the company, the model was developed largely by an in-house AI research system — not by human engineers writing code line by line.

At first glance, this sounds like a technical detail. For investors in small AI software providers, however, it is a structural signal: as the training of AI models itself becomes increasingly automated, the conditions under which smaller market participants can compete — or fail to — are changing fundamentally.

Automated training and the economics of model development

To appreciate the full significance, it is worth examining the existing cost curve. Training large language models (LLMs) has traditionally required three expensive resources: firstly, specialised hardware (Nvidia GPUs in data centres); secondly, high-quality training data; and thirdly — often underestimated — an enormous amount of manual research work. Researchers select architectures, define training strategies, debug model behaviour, and run ablation experiments.

It is precisely this third element that self-optimising approaches target. When an AI system independently formulates hypotheses about better training procedures, tests them, and implements the results, the need for human research capacity falls — at least in defined sub-areas. In economic terms, this corresponds to a decline in variable costs per model iteration: each new model version becomes cheaper than the previous one.

Smaller providers: opportunities, moats, and structural pitfalls

For small-cap developers in the AI space, this dynamic gives rise to two contrasting scenarios.

Scenario A – Democratisation: Cheaper model training lowers the barrier to entry for small software companies. A firm that today wants to build a specialised coding model for a niche sector — such as legal contract analysis or industrial process control — has until now required substantial capital. If self-optimising systems reduce the research burden, start-ups with modest capital could develop their own models that compete with large providers in narrow domains.

Scenario B – Erosion of technological advantage: At the same time, expertise in model training loses value as a differentiating factor once it becomes automated. A small cap whose central competitive advantage was training models with exceptional efficiency will see that advantage shrink as soon as rivals — including well-capitalised hyperscalers — achieve the same result automatically. The so-called moat shifts: proprietary data and domain-specific application expertise become increasingly decisive, rather than training capability alone.

An illustrative example from the recent past is the development of the open-source ecosystem. When Meta made its Llama models publicly available, many smaller providers that had built their business model on selling proprietary foundation models came under pressure. At the same time, those who used Llama as a basis for specialised fine-tuning applications benefited. The pattern repeats itself: standardisation at one layer creates new differentiation potential at the next.

What valuation models for AI small caps are not yet pricing in

For investors monitoring AI small caps, it is worth taking a critical look at the valuation logic. Many of these companies are valued on the basis of their technological unique selling points and expected revenue growth (Annual Recurring Revenue, ARR) — not on the basis of current profits, which are often simply absent.

This is precisely where the structural risk lies: if a company's primary claim to value is its ability to train models efficiently, and that ability becomes a commodity through self-optimising systems, the basis for its valuation can shift within a single investment cycle. Investors in such positions should ask themselves: Where is this company's genuine data moat, and how difficult is it to replicate?

There is also the capital question. Even where the technological approach is compelling, small AI companies without a positive cash flow base regularly require fresh capital. Each capital increase (share issuance) dilutes existing shareholders — a mechanism that is particularly painful during long development cycles or delayed market entry. The cash runway — the period a company can sustain operations with its available capital before requiring a new funding round — remains the central survival metric.

Automated model training as a seismograph for the next market phase

Self-optimising AI systems are still at an early stage. Recursive's Flint is a first public real-world test, not a mature mass-market product. Between a research prototype and a reproducible, scalable infrastructure for automated model training, there are typically several years of intensive development.

Nevertheless, the direction of travel is clear: the effort required to build a competitive AI model will decline over the medium to long term. For the market, this means an accelerated consolidation dynamic — similar to cloud infrastructure, where cheaper AWS services displaced many specialised hosting providers while simultaneously creating new category winners in the application layers.

For investors following the AI small-cap theme, an analytical framework is therefore advisable that evaluates not only the technology but explicitly the data and vertical strategy of a company. The decisive question is no longer: "Can this start-up train a good model?" — but rather: "Does this start-up have access to data and domain knowledge that cannot be automated?"

Key terms at a glance

Self-optimising AI system (recursive self-improvement)

An AI system capable of independently analysing and improving its own training process or architecture — without full human oversight of every iteration step.

Moat (competitive moat)

A structural competitive advantage that keeps imitators at a lasting distance. In the AI industry, this can take the form of proprietary datasets, network effects, or regulatory barriers to market entry.

Cash runway

The period a company can sustain operations with its current cash balance at a constant burn rate (monthly expenditure minus revenue) before new capital is required.

Dilution

The reduction in existing shareholders' percentage stake in a company through the issuance of new shares, for example via a capital increase. Each new tranche of shares distributes the company's value across a greater number of holdings.

ARR (Annual Recurring Revenue)

Annually recurring revenue from subscriptions or long-term customers. A key valuation metric for AI software companies — but not a profit figure and no guarantee of long-term profitability.

Fabless model

A business model in which a company designs chips but operates no manufacturing facility of its own — production is outsourced to specialised foundries. Analogy: AI start-ups that use foundation models from third-party providers rather than training their own.

Commodity risk

The risk that a technology or service becomes standardised, creating price pressure because many providers can deliver the same capability. Relevant in AI when model training becomes scalable and cheaper through automation.

⚠️ Important notice: This article is for informational and educational purposes only. It does not constitute investment advice, a recommendation, or a solicitation to buy or sell any security. Investments in small-cap exploration and mining companies carry a high risk, including the potential total loss of capital. Before making any investment decision, consult a registered financial advisor and conduct your own analysis. Aktienatlas-Redaktion is not responsible for decisions taken based on the content published here.

Educational content only, not investment advice. Small caps are highly speculative and total loss is possible.