The focus of artificial intelligence is shifting toward smaller models, not just LLMs
The artificial intelligence market is going through a quiet but pretty significant transformation. And the ones signaling this shift are the very people putting money on the table.
For years, all eyes were on the big LLMs — those massive models that promised to solve everything with a single solution. GPT, Gemini, Claude, and other household names dominated headlines, attracted billions in investments, and shaped the way the world sees the potential of artificial intelligence. It was almost as if the size of the model was directly proportional to its value.
But the game is changing, and anyone following the industry closely has already noticed that things are heading in a different direction. Smaller, specialized models are steadily gaining ground, both among developers and investors. It’s no surprise that key players in the Silicon Valley ecosystem have already started turning their attention to this niche.
This pivot is exactly what Lu Zhang, founder of Fusion Fund and one of the leading investors in tech startups, has been highlighting. During an appearance on CNBC’s Inside India, she shared the main investment themes she’s been tracking in Silicon Valley. In her view, the market is mature enough to realize that bigger isn’t always better — and investment flows already reflect that.
Why smaller models are gaining momentum
The logic behind this shift isn’t hard to understand, but it represents a real break from the thinking that dominated recent years. When the big LLMs burst onto the scene, the narrative was clear: more parameters meant better performance, and better performance meant more value for companies and users. But the day-to-day reality of actual applications revealed that this equation has very real limits — and that not every task needs the heaviest cannon on the shelf to be handled efficiently.
Smaller models, also known as small language models or SLMs, come with a different proposition: do less, but do it better, within a specific context. A healthcare company, for example, doesn’t necessarily need a model trained on virtually all of the internet’s content to automate patient triage or organize medical records. What it needs is a model that deeply understands medical vocabulary, clinical workflows, and the sector’s regulatory specifics. In that scenario, a smaller model well-trained for that domain will perform far more consistently and safely than a generic giant.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Beyond contextual performance, there’s a factor that can’t be ignored: operational cost. Running a large LLM for every request in a corporate application is expensive, resource-heavy, and demands considerable computing power. Smaller models, on the other hand, can run on more affordable hardware — even on local devices — which reduces latency, improves data privacy, and cuts costs significantly. For companies that need to scale AI solutions without burning through cash, that makes all the difference.
What investors like Lu Zhang are seeing in the market
Lu Zhang, at the helm of Fusion Fund, has been one of the most active voices in this conversation within Silicon Valley’s venture capital ecosystem. Her perspective starts with a very practical observation: the artificial intelligence market is entering a maturity phase where the question shifts from which model is the most impressive to which model solves the right problem, in the right context, at the lowest cost with the highest possible return. This mindset shift is already directly impacting investment flows into sector startups.
Startups showing up with smaller, verticalized models focused on niches like legal, financial, industrial, or educational are finding a much more receptive fundraising environment than two or three years ago. That’s because investors, after the initial hype around large LLMs, started demanding more tangible, predictable results. A specialized model that can demonstrate clear ROI for a specific segment tells a much more compelling story than a generic platform promising to do a little bit of everything.
Another relevant point raised by Lu Zhang is the competitive dynamic between the United States and China in the AI landscape. According to her, at this point, the two countries aren’t exactly competing for the same markets. This suggests there’s room for different approaches in each ecosystem, with the Americans betting heavily on purpose-built models and cutting-edge infrastructure, while the Chinese market follows its own development and application paths. This coexistence, rather than a head-to-head battle, could actually benefit the global advancement of the technology, since each side ends up exploring different frontiers.
Specialized models versus generalist models
The difference between a specialized model and a generalist one goes far beyond parameter count. The central point lies in targeted training. While a traditional LLM is fed with absurd volumes of data from virtually every domain imaginable, a smaller, specialized model goes through a training process focused on curated datasets from a specific sector. This means it makes fewer mistakes within that context, generates more relevant responses, and needs fewer fine-tuning adjustments after deployment.
In practice, this approach brings some very concrete advantages:
- Lower hallucination rate — models trained on specific domains tend to fabricate less information because their knowledge scope is more controlled
- More accurate responses — since the model deeply understands its area of expertise, its outputs are better aligned with what users actually need
- Faster deployment — putting a smaller model into production is significantly simpler and less costly than dealing with the infrastructure required by a full-blown LLM
- Greater control over updates — retraining or fine-tuning a small model is more agile, allowing companies to keep up with regulatory or market changes more quickly
The data sovereignty and privacy question
Another factor that weighs heavily in this equation is data sovereignty. With regulations like Brazil’s LGPD and Europe’s GDPR increasingly on companies’ radar, solutions that allow AI models to run locally — without sending sensitive data to third-party servers — hold enormous appeal.
Smaller models fit perfectly into this scenario since they can be deployed on-premise or in private cloud environments far more easily than their giant counterparts. A financial services company that needs to analyze customer data, for instance, can do so with an SLM running internally, without any sensitive information leaving its controlled environment. This isn’t just a technical advantage — it’s a regulatory necessity that’s becoming increasingly strict across markets around the world.
The AI investment landscape is diversifying
The movement that Lu Zhang and other Silicon Valley investors are leading reflects a structural shift in how venture capital views the artificial intelligence sector. In previous years, the bulk of funding flowed toward companies trying to build the next big language model, competing directly with OpenAI, Google, and Anthropic. This created a monumental barrier to entry, since training a state-of-the-art LLM can easily cost hundreds of millions of dollars.
Now, with smaller models validated as a viable and profitable alternative, the ecosystem opens up to a much larger pool of startups. Smaller teams with more realistic budgets can build high-impact solutions for specific verticals without needing to compete in the arms race of large models. This is healthy for the market as a whole because it distributes innovation more evenly and allows creative solutions to emerge from unexpected places.
On top of that, diversifying investments reduces systemic risk across the sector. When all the capital is concentrated in a few companies betting on the same approach, any paradigm shift can cause significant losses. With a more diversified portfolio that includes both large-model and specialized-model companies, investors build more resilient positions with multiple paths to returns.
What this means for the future of AI
This trend doesn’t mean that large LLMs will disappear or lose relevance. They still play a fundamental role in tasks that require broad reasoning, creativity, and the integration of knowledge from multiple domains. What’s happening is a smarter division of labor within the AI ecosystem, where each type of model finds its most fitting role.
Think of it like a well-assembled team: you don’t put the tax law specialist to write production code, just like you don’t ask the senior developer to draft a legal brief. Everyone in their lane — the results are way better.
For the market as a whole, this evolution opens up interesting opportunities on multiple fronts:
- For companies consuming AI — the diversification of options means more choice, solutions better aligned with real needs, and more controllable costs
- For those building AI — it opens the door for smaller startups to compete on equal footing with major players, since building a high-quality specialized model doesn’t require the same volume of resources as training an LLM from scratch
- For investors — it creates a more varied pipeline of opportunities with clearer return theses and shorter paths to monetization
What becomes clear from all this movement is that artificial intelligence is truly maturing — not just in size, but in depth and applicability. And maturity, in the tech world, tends to come hand in hand with a natural shift toward specialization and efficiency.
Smaller models aren’t gaining ground because they’re a passing trend or a reaction against the industry’s giants. They’re establishing themselves because they solve real problems more precisely, more affordably, and more safely. And as Lu Zhang rightly observed, the next big wave of artificial intelligence might not be the largest one of all — but rather the smartest. 🚀
