As someone who has been an investor and an operator for most of your career and who has been looking at AI from both of those vantage points, how would you say the market has changed in recent years?
That’s an interesting question. Back when I was working at a company called Quid and was trying to sell NLP-powered applications to end users, our customers really weren’t all that concerned about AI or ML. At the time, they wanted products that met their immediate needs and had clearly defined budgets to buy software for specific use cases regardless of whether or not they had AI or ML. Today’s customers are obviously far more focused on adopting AI and have much bigger budgets to not only invest in it but also to experiment with it. Unlike SaaS budgets, AI vendor budgets aren’t limited to software but extend themselves to operational budgets, which meaningfully expands the market for AI products and services.
It’s also important to note that as interest in AI has grown, so too has people’s understanding of what it actually is. Just a few years ago, it was almost exclusively thought of as tools and infrastructure. At that time, AI was a luxury that only larger companies with enough data scientists or big enough ML platform teams could afford. That dynamic created a big market for cloud providers who could package up different solutions and services at the infrastructure layer. It also led to a lot of demand for SaaS solutions on the commercial and technical side, where people were looking for full-stack solutions tailored to specific use cases.
So, from an investor perspective, if you were an AI-focused fund a few years ago, the core areas where you could invest were business intelligence, data infrastructure, and ML infrastructure. On the application layer, the only thing categorized as AI at the time was deep learning applications like computer vision. Everything else was either SaaS at the front end or ML optimization and recommendations on the back end.
So how would you contrast that with the market today?
With the surge in unstructured data, AI has rapidly advanced, leading to a fragmented AI stack divided between traditional machine learning and modern generative AI. The GenAI stack has multiple layers that includes the compute layer, followed by cloud infrastructure and data infrastructure, foundation models, AI tools, and finally the application layers. Beyond that, there are full-stack application companies building their own domain-specific and general models.
Of course, not all of those layers are ideal for early-stage investing. At Zetta, we focus our investments on data and cloud infrastructure, AI tooling, and application-layer companies, largely because other layers are highly capital-intensive and it’s uncertain whether foundation models will achieve meaningful differentiation.
What about investor sentiment? How has that changed?
With the sharp rise in public and private market valuations of AI companies like NVIDIA, Meta, Palantir, and Databricks, it has become clear that investing in market leaders across different layers of the AI stack offers significant upside potential. The prevailing narrative in AI follows scaling laws, which suggest that increasing compute power, model size, and high-quality data lead to predictable improvement in model performance. This belief has driven up the cost of early-stage funding, especially in the compute and foundation model layers. Any application company that’s trying to build its own model wants more compute. Investors understand that and are becoming increasingly open-minded and flexible about investing at the seed round, even at unreasonable valuations.
As an early-stage investor, I have concerns about the sustainability of this trend. We’ve witnessed companies like Adept, Character, and Inflection secure massive funding rounds, only to deliver underwhelming exits. However, it seems that the market is gradually returning to fundamentals — at least to some extent — and DeepSeek’s approach serves as a testament to this shift.
Honestly, after a couple of years where there’s been a lot of noise in the market and really frothy rounds, particularly in the data and AI space, we’re hoping for a correction.
You mentioned DeepSeek, which has been highly disruptive to say the least since it debuted back in January. What do you see as its impact?
Whether or not its numbers are accurate, DeepSeek has challenged existing ideas about how much capital it actually takes to build successful models. That says to me that there are going to be a lot more model brokers going forward, even if the market eventually consolidates into just a handful of players.
One of the most interesting things about DeepSeek is that it’s open source. Going forward, I think that means the big question won’t be who owns the models (unless it’s a model built on proprietary data), but rather who owns the end customer experience and the data? We continue to believe that the real value resides in the application and data layers, both of which are model independent and will benefit from having multiple players in the market.
Value resides in model-independent applications and data layers.
The collapse of the dominant narrative is also a positive shift for AI researchers, proving that small, agile teams with novel ideas can drive meaningful advancements. It also enriches the broader AI research ecosystem by fostering exploration across diverse directions, many of which may originate in academic labs and remain open — offering a stark contrast to the increasingly closed nature of research within major AI labs today.
Finally, many of the founders who have been opting to skip over the seed round up until now are starting to actively rethink whether they need to invest so much in compute or if they can instead focus more of their attention on training and designing infrastructure. Ultimately, that will allow researchers to work in a more nimble and less capital-intensive way.
Let’s unpack something you just said a bit more. Why else do you see big opportunities in the application and the data layers?
This AI shift is benefiting both incumbents and startups alike. Not by replacing deterministic incumbent products and workflows, but by enabling AI models and agents to function like employees or service providers — enhancing efficiency, decision-making, and overall workflow automation. By doing so, AI helps transform bespoke, human-driven processes into AI-native applications, whether as products or services.
At its core, AI serves as an augmentation tool, capturing both tribal knowledge and documented processes while also making nuanced decisions on behalf of users. In the application layer, you not only control the end-user experience but also gain significant advantages as inference and training costs decline while reasoning capabilities improve. This holds true regardless of whether the application is built on an open-source or closed-source model, ultimately making it easier to develop innovative and impactful AI-driven applications for end users.
Meanwhile at the data layer, even if you take the best reasoning traces from today’s models or download the open weights and train your own model, you’re still going to need access to curated or newly generated data for post training, fine tuning or model alignment. That is one incredibly important tailwind for the data layer.
What are some of the most striking technical innovations that have come out of DeepSeek?
The fact that DeepSeek’s R10 model was built end-to-end using reinforcement learning is certainly one of the biggest innovations. Up until now the common thinking has been that to train a workable model you need a lot of compute, highly curated data, supervised fine-tuning, and reinforcement learning from human feedback (RLHF). That’s a pretty capital-intensive process. DeepSeek has shown that by using reinforcement learning and rewarding the right behavior, you can build something from the ground up that’s smarter, has better reasoning capabilities, and that is able to separate reasoning from outcomes. That’s going to have a big impact on both model providers and people who are fine-tuning existing models for their specific use cases.
In addition, DeepSeek, uses two key optimization techniques to push the boundaries of efficiency and scalability in AI models. The first is known as Mixture of Experts (MoE) and is an approach that optimizes both training and inference by selectively activating only a subset of model parameters during computation. By dynamically routing different inputs to specialized expert subnetworks, MoE significantly improves computational efficiency while maintaining high performance.
The second technique is called Multi-Head Latent Attention (MLA) and is a critical innovation for handling longer context windows. MLA ensures that memory consumption remains manageable even when processing extensive sequences of information. This technique optimizes how attention is allocated across different segments of input data, preventing memory overload while preserving the ability to capture long-range dependencies.
Looking ahead, what are some of your predictions for AI in the years to come?
At the data layer, I think we will see an unbundling of the stack. Right now, it’s the big players like Databricks and Snowflake, and not the customers, that have all of the data. That’s largely been the case for the sake of scalability and efficiency, but the reality is not only do customers not like that, it’s also become incredibly expensive to put all of their data into full-stack systems, particularly as unstructured data proliferates. Going forward, a lot of that data will be owned by individual companies in their local systems and in a data lake. We will see multi-modal interplay with structured data and different kinds of tools emerging to use that data effectively.
Another trend I’m seeing is around data generation and data curation as a service for newer ML workloads. As AI workloads come up, much of their data has to be annotated by vendors like Scale AI or Labelbox using human annotators. We’re increasingly seeing companies curating that data themselves rather than handing it over to a third party to do so. And, as many models are reaching their limits, there’s also a question about how to generate net new data in different categories. New companies will emerge to help businesses generate that data and effectively close the loop between data generation, data curation, and AI model feedback, allowing for the creation of better data sets for use in use case-specific models.
And what about the application layer?
On the AI side, interfaces are going to become more multi-modal. Right now, text is the go-to interface for language-based models. But as speech models become more accurate on device and latency improves with audio, voice-based interfaces will proliferate. Of course, not everyone will use voice. There will also be visual and image-based interfaces, among others.
Another trend is going to be the growing use of web agents to assist with everything we do online, whether that’s buying groceries, managing finances, or booking flights. Web agents will become much more mainstream and virtually everyone will have their own web agent that’s trained for their specific needs, knows their history and preferences, and is able to do things on their behalf.