20VC: Nebius Co-Founder on AI Infrastructure Bubbles | The Real Impact of Open Source on OpenAI & Anthropic | How Price Elastic is Demand for Compute | Could Nebius Sell 10x More Compute If They Had It & more with Roman Chernin

Current AI Infrastructure Investment is Just the Beginning, Not a Bubble

Roman Chernin posits that the substantial investment currently directed towards AI infrastructure is not indicative of a bubble, but rather the foundational stage of a much larger technological transformation. He sees it as the initial phase of building out what he refers to as "useful AI."

He emphasizes the very early stage of widespread AI adoption, noting that practical enterprise use cases are just starting to emerge. For instance, he highlights that AI-powered coding, a significant working use case, has only been effective at scale for a few months.

Chernin predicts a massive future demand for AI infrastructure, suggesting that the industry will eventually require tens or hundreds of times more resources than are currently available. This perspective reinforces that the present investment is merely the initial step in a long-term development cycle.

We are just at the beginning of this amazing moment when Jensen calls it like useful AI.

Open-source models optimize costs for existing use cases, enabling frontier AI to focus on new challenges.

Most enterprises are still in the very early stages of AI adoption, typically utilizing it for only a small fraction of their overall volume and use cases. As companies begin to scale their proven AI applications, they actively seek methods to improve their economic efficiency and accelerate growth.

While frontier models from providers like OpenAI, Anthropic, and Google offer cutting-edge capabilities, they may not always be the optimal long-term solution for established use cases. Once a specific AI application has been validated, businesses often pivot to open-source models. These models can be fine-tuned and post-trained to create specialized solutions that are more cost-effective or even higher quality for their particular needs than a general-purpose frontier model.

This shift does not harm frontier AI providers. Instead, it allows them to continuously advance to the next set of challenges, focusing on complex, unsolved tasks that lack established budgets or solutions. The market for these advanced, unaddressed problems is vast, providing ample opportunity for frontier models to grow exponentially by pushing the boundaries of what AI can achieve.

Maybe you don't need the best in the world universal model, but you can create a specialized model that in your particular case will work even better.

Cheaper AI Models Drive Exponential Demand and Infrastructure Growth

The availability of more affordable AI models, like DeepSeek, initially triggered market anxieties among investors who predicted a decline in demand for AI infrastructure. The prevailing thought was that reduced AI costs would translate to less need for supporting hardware and services, potentially indicating a market downturn.

However, this expectation was swiftly contradicted. As AI inference became more economically viable, a significant number of businesses recognized they could successfully integrate AI into their production systems. This newfound accessibility unlocked a wide array of previously unfeasible applications, leading to a dramatic expansion of the total market for AI solutions.

This situation illustrates the Jevons Paradox, where increased efficiency or a reduction in the cost of a resource leads to an overall increase in its consumption. For AI, cheaper models do not diminish total compute demand; instead, they significantly amplify it, creating substantial benefits for infrastructure providers.

The AI market is still in its very early stages, with numerous unsolved problems yet to be tackled. This extensive potential ensures ample room for both advanced frontier models and highly specialized, open-source solutions to coexist and thrive, as the overall ecosystem of AI applications continues its rapid growth.

so many people figured out that they can run inference in their production workloads with DeepSeek and economics will work.

Nebius Organizes Business Growth Around Four Core Dimensions

When intelligence becomes cheaper, demand doesn't necessarily decrease; instead, consumption often increases because it allows for solving more complex tasks or making previously unviable tasks economically feasible. This shift in economics is a fascinating phenomenon.

Nebius approaches company building through four main dimensions. The first is capacity, focusing on deploying megawatts, gigawatts, and GPUs. As an infrastructure company, achieving a large scale is fundamental for its relevance and existence.

Expanding physical capacity, however, is fraught with real-world complications. Launching new data centers requires navigating intricate supply chains, regulatory processes, and overcoming various real-world obstacles, often preventing rapid deployment.

The second dimension is product development, where the goal is to quickly adapt to new types of workloads and customer demands. Early in the AI journey, customers like large labs and hyperscalers primarily needed raw compute infrastructure, leading to bare metal deals, but this is only the initial stage of evolving customer needs.

If we are not large enough, nobody needs us to exist.

Nebius Offers Two Core AI Infrastructure Layers: Bare Metal and Multi-Tenant Cloud

Nebius provides foundational AI infrastructure, starting with scaled physical infrastructure, or bare metal, designed for large hyperscalers like Meta or Microsoft. This layer is consumed in massive volumes, with deals often discussed in terms of megawatts of compute.

The second layer is a multi-tenant cloud offering, targeting research-heavy teams. These customers prefer managed infrastructure-as-a-service, including virtualized storage, compute, and networking, complete with APIs, observability, and security. Unlike the bare metal layer, this cloud service is typically measured and sold in GPU hours, representing the efficient time spent on compute tasks.

Nebius is also developing a third layer: managed inference, under a product called Nebula Stalking Factory. This service is aimed at vertical AI customers who want to avoid managing specific GPU types (e.g., B200s vs. H200s) or deploying optimization tools like vLLM or sGLM themselves, providing a fully managed platform for inference workloads.

If the first layer speaks in megawatts... Then when you speak about this managed cloud, people speak GPU hours because this is the key unit you sell, the efficient hours you can spend on compute with storage, with complementary services.

Nebius Expands AI Infrastructure to Managed Inference and Agentic Workflows

Nebius is introducing a new primitive for AI product builders: managed inference measured in tokens, rather than GPU hours. This service, often called a "Token Factory," enables companies to easily diversify their model usage or shift between specialized and open-source models, such as moving away from Anthropic, without needing to manage the underlying infrastructure.

The company sees the next evolution in AI infrastructure shifting towards agentic applications and workflows. For users building end-to-end agents, the focus moves beyond specific models or token counts. Instead, developers aim for efficient execution of an entire task, prioritizing the desired final outcome.

The platform aims to provide the intelligence to manage these agentic workflows. It can dynamically determine the most suitable model for a given call, considering factors like whether to use a single smarter model, combine two lighter models with a judging mechanism for optimal results, or adjust the context size, all to achieve an optimized end-to-end task execution.

This advanced layer, referred to as Layer 4, is being developed with an optimization engine, positioning it as a competitor to services like OpenRouter. The goal is to provide tools that help scale agentic workflows efficiently, addressing the economic considerations developers face when moving from development to production.

The magic that platform can make is actually think for you which model better to use in this particular call.

Nebius aims to meet compute demand by diversifying its product portfolio

Nebius recognizes a substantial market demand for its compute capacity. While selling ten times their current capacity overnight isn't realistic, the company confirms a definite existing demand for such a scale.

The key challenge for Nebius is not the absence of demand but strategically building a diversified portfolio of demand. This involves balancing various customers across their product offerings, which span multiple layers.

Their strategy involves selling across different product tiers, including bare metal, managed infrastructure, and inference services, with plans for new layers in the future. Nebius believes that moving higher up the stack allows them to create greater value for customers and serve a broader market, as bare metal caters to a relatively small segment.

Not overnight, but we definitely have demand for that.

Nebius cultivates a diversified customer portfolio and a full-stack platform to avoid over-reliance on mega-players.

Nebius intentionally builds a diversified customer portfolio to avoid over-reliance on a few large clients like Meta or Microsoft. While these mega-players offer substantial business, they often require only physical infrastructure, bringing their own entire software stack. This limits the additional value Nebius can provide beyond basic capacity.

To ensure long-term business protection, Nebius focuses on developing a comprehensive full-stack software offering. This strategy allows them to serve a broader range of customers who value a platform beyond just bare-metal infrastructure, attracting clients that appreciate a more advanced and integrated solution.

This approach enables Nebius to be selective, prioritizing customers who value their built platform rather than solely focusing on price. It helps them avoid becoming a mere capacity provider to mega-players, securing a more sustainable and less concentrated business model.

Regarding price elasticity, Nebius recently increased its prices by 30% a couple of months ago and still experiences significant pipeline pressure on supply. This indicates that demand for their specialized infrastructure, particularly for training workloads, remains strong despite the price hike, suggesting a degree of price inelasticity in the current market.

we actually raised prices like just thirty percent. yeah, just couple months ago, and we still, still have, pipeline pressure, let's say, on supply.

Total Cost of Ownership Matters More Than Nominal GPU Price

The real cost of AI compute extends far beyond the nominal price per GPU hour. Many mistakenly focus on raw capacity and its sticker price, but this overlooks the broader picture of Total Cost of Ownership (TCO). A platform's quality and its optimizations significantly impact the actual economic outcome for a customer.

Factors like effective uninterrupted runtime for tasks, or the number of tokens that can be extracted during inference, are crucial. For example, specific use cases might lead to vastly different real costs, even with the same GPU nominal price, depending on how efficiently the platform allows you to operate.

A well-built software platform can implement optimizations that drastically change the price of tokens or the overall operational cost, sometimes by an order of magnitude. This system-level approach, providing a high level of service beyond just raw infrastructure, unlocks greater economic value for customers and enables their growth.

Nebius addresses this by focusing on its software platform to optimize TCO. This means customers can achieve better results and more sustainable economics, ensuring that their product can scale effectively without being bottlenecked by an overly simplistic view of compute costs.

It's not only GPU hour cost, it's all the optimizations you do, all the real, we call it TCO, total cost of ownership.

Nebius Adapts Product Strategy to Evolving AI Customer Needs with Full-Stack Differentiation

Nebius continuously adapts its product strategy by closely observing evolving customer needs and market demand. Key shifts include a transition from AI model training to inference, from merely using models to developing agents, and a significant move from AI labs to enterprise customers. This transition brings new requirements for inference platforms, assistance with model fine-tuning, and data collection to enable a "flywheel" for continuous model improvement.

A core part of Nebius's differentiation is its full-stack integration, encompassing "full-stack down" (deep control over physical infrastructure like data centers, racks, and servers for speed and cost efficiency) and "full-stack up" (following customer needs to serve diverse segments beyond raw infrastructure). This approach enables the company to lower the barrier for builders who may not be AI researchers or inference engineers, abstracting infrastructural and AI complexities.

By allowing customers to focus on their core use cases, Nebius aims for a more diversified customer portfolio. The company sees long-term demand coming from enterprises that require platforms and tools compatible with their complex, legacy environments, an area Nebius is strategically targeting for future growth.

I think the one, the most fascinating mo-moment for me is that I think what we see is that barrier to build is going down.

Nebius's Token Factory offers managed inference for open-source models.

Companies often begin with OpenAI for its ease of use and production-ready service, achieving significant growth. However, they soon encounter limitations like high costs or the inability to customize models with proprietary data, prompting a search for more flexible alternatives.

Many turn to open-source models, expecting cost savings and greater control. While open-source benchmarks might appear competitive, deploying and scaling these models for production is complex. It requires significant effort in optimization, orchestration for large GPU clusters, caching, and observability, mimicking the full infrastructure management provided by services like OpenAI.

Nebius addresses this challenge with Token Factory, a managed inference platform. It allows businesses to utilize open-source or specialized models, including those with custom-tuned weights, without the burden of infrastructure management. Token Factory handles all necessary optimizations, scaling, and cost economics, providing a reliable, production-grade service akin to OpenAI, but with the flexibility of open-source.

For example, if a product built on OpenAI grows to run on hundreds of thousands of GPUs, moving to open-source models would require building out all the orchestration, caching, and observability from scratch. Token Factory removes this complexity by managing the underlying infrastructure, allowing users to focus solely on their model and application. It supports both existing open-source vanilla models and custom-tuned models.

How Token Factory reduces AI inference costs through model optimization and managed updates

Token Factory reduces AI inference costs by up to 70% by optimizing baseline open-source models for specific customer use cases. This is not a magical process but involves targeted improvements to the models themselves to make token generation more economical.

The optimization process incorporates several techniques, including model distillation, which creates smaller models that maintain the same quality, speculative decoding, and advanced caching. By applying these methods, Token Factory constructs a system that meets unique customer requirements with optimized economic efficiency.

A major challenge in the rapidly evolving AI landscape is the constant release of new models, sometimes every few weeks. Each new model may perform differently across benchmarks, making it difficult for users to keep up with experimentation and adoption.

Token Factory addresses this by acting as a managed platform that abstracts all the work involved in benchmarking and transitioning between models. It ensures customers can stay at the technological frontier, seamlessly switching to the best-performing models for their specific use case without manual effort.

the platforms like ours Abstracts from you all the work that you need to do to actually change from one model to another, to benchmark all of them, and so on.

AI Model Development Accelerates with Specialized Niche Applications

The AI landscape continues to see rapid iteration and improvement, with many niche models constantly emerging. The belief is that significant model advancements will continue, and the industry is far from hitting a development wall.

Beyond general frontier analytics, there's a growing world of specialized models across new modalities, including life sciences, robotics, video, and image processing. These models are highly optimized for particular use cases.

For example, a team in Israel is developing a cyber defense foundational model. They start with open-source foundational models and then intensely train them for specific cyber defense agents, optimizing for the unique quality and latency requirements of such critical applications.

This trend necessitates robust, optimized inference infrastructure to allow customers to effectively deploy and use these specialized, highly-trained models. The growth trajectory in AI, seen in companies like Anthropic, Cursor, and Cognition, and expanding into healthcare and financial verticals, is quite amazing.

We spoke about this frontier analytics, but there is entire world of life science models, robotics, world models, video models, image models. They all have their own use cases as well, and we see more and more small specialized models for particular use cases coming, but like very much optimized.

Enterprises unlock exponential AI adoption by addressing the 'cold start problem' with foundational investments.

Many non-AI native enterprises initially struggle with AI adoption, often relying on closed models like OpenAI. They face economic challenges when these models don't meet specific use cases or effectively enhance human efforts. This "cold start problem" hinders rapid progress, as seen with companies like Revolut, which initially allocated 99% of its inference budget to closed models.

To overcome this, companies must make foundational investments internally. This involves establishing robust evaluation mechanisms to define what constitutes success for their specific context, creating clear metrics to maintain quality, and implementing a CI/CD process for AI development. Securely integrating models into existing production processes is also critical.

Once these foundational problems are solved, enterprises experience exponential growth in AI adoption. They gain the ability to "ship fast," make informed decisions about model evolution, and leverage various models to build more products. This rapid growth trajectory mirrors that of AI-native companies, demonstrating how initial slow adoption can transform into accelerated expansion once the right systems are built.

When they solved these foundational problems, they start growing exponentially.

The Future of AI: Diverse Models and Agentic Orchestration

The future of AI will see a diverse ecosystem of models emerge, catering to various needs. This includes highly intelligent frontier models for complex problems, extremely fast models for rapid iterations, and 'in-between' models that balance smartness with affordability. This variety allows users to select the most suitable model for each specific task.

A significant development will be the rise of an 'agentic layer' or intelligent engine. This engine will abstract the complexity of model selection from users and developers. Instead of manually choosing which model to call or how to orchestrate tasks, users will provide a high-level goal, and the engine will automatically determine the optimal sequence and combination of underlying AI models and tools.

This agentic orchestration will simplify development for enterprise use cases. Developers can focus on core customer needs rather than the intricate plumbing of various AI models. The engine will handle the reasoning and execution, deciding when to leverage different models or tools, such as an LLM or a search function, to achieve the desired outcome efficiently and effectively.

you give a task, there is an engine The reasoning engine that decides how to run this task and you get the result.

Cultivating AI Builders and Demand for European Sovereignty

Europe lags behind the US and China in AI model development. While there's a push for sovereign AI, the conversation in Europe often concentrates too heavily on physical infrastructure, like megawatts and computing power.

An alternative perspective argues that focusing solely on infrastructure misses the core driver. Companies that provide infrastructure, such as Neboos, are motivated to build out capacity when there's actual demand for it.

This demand primarily originates from local builders: the researchers, developers, and product companies actively creating and deploying AI models. These are the entities that will drive the need for computational resources.

Therefore, achieving sovereign AI in Europe means prioritizing the creation and growth of a robust ecosystem of local AI companies and innovators, like Lavables, Black Forest Labs, and Mithril. By fostering this builder layer, the necessary physical infrastructure will naturally emerge.

Megawatts will come, I think that it's what, what we in Neboos always told is, we will build infrastructure, the companies like us will build infrastructure if we have demand, and demand is coming from the builders.

End-User Products Drive AI Growth and Engineering Earns Nvidia's Respect

The most compelling investment area in the AI industry lies in end-user products. While building infrastructure is complex, creating applications that people genuinely need involves significant risk but ultimately drives the most growth. These innovators are considered the "heroes" of the AI journey.

Focusing on end-user products is crucial because they generate demand, which in turn fuels the development of better AI models and creates a powerful industry flywheel. This demand is essential for the continuous evolution and adoption of AI technologies.

When addressing the considerable power dynamics with Nvidia, companies like Nebius take a direct approach: focus on exceptional engineering and product development. By building strong products and effectively communicating their story, they believe other factors will align.

Nvidia is primarily an engineering-driven company. Therefore, demonstrating technical prowess and delivering high-quality engineering work is the most effective way to gain their respect and navigate the relationship.

I think the most amazing people in this industry are those who take a risk to go and build end user products, in my view.

The Four Hardest Dimensions of Building an AI Infrastructure Company

Roman Chernin emphasizes that "just doing your job" in the AI infrastructure space is complex, involving four critical dimensions for companies like Nebius. Building strong relationships with partners such as Nvidia is foundational, earned through consistently demonstrating strong engineering capabilities and mutual respect at multiple technical layers.

The first two dimensions involve simultaneously scaling compute capacity and developing innovative products. Companies must continuously expand their infrastructure to meet demand while also pushing the boundaries of what their platform offers to remain competitive in a rapidly evolving market.

The third dimension focuses on customer relationships, particularly in the cloud business, which is inherently a post-sales environment. It requires a strong customer-facing engineering team to ensure customer satisfaction and build lasting partnerships, as selling involves promising future value that must then be delivered.

Finally, the fourth and often most challenging dimension is securing capital. Operating in a highly capital-intensive industry means competing against the world's most capitalized companies, making substantial and continuous investment crucial for survival and growth.

Do your fucking job at the end of the day.

Time is a significant bottleneck in scaling AI infrastructure, even with unlimited budget

An unlimited budget would enable faster construction of data centers and quicker fulfillment with GPUs, allowing the company to serve more customers. The current capital expenditure is $20-25 billion, which is considerably less than larger hyperscalers.

However, the primary bottleneck depends on the time horizon. In the short term, specifically the next six months, additional capital provides little benefit as existing infrastructure is already set.

Over a 12-month period, capital can help accelerate some capacity constraints, but its real impact becomes evident over 24 months.

To overcome these time constraints, the company builds a portfolio of capacity by executing tasks in parallel. This involves securing power and land, then constructing data centers, and finally filling them with GPUs, ensuring that each subsequent stage has necessary resources already in place.

In the next six months, the capital cannot help. Six months is too short time. You have what you have, you need to deliver.

Navigating Public Resentment in Data Center Development

Public resentment and regulatory pushback are increasingly challenging data center construction. Around 40% of planned data centers are not being built due, in part, to public opposition, as exemplified by figures like Eric Schmidt facing criticism for AI advancements. This environment necessitates new strategies for development.

Companies like Nebius adopt a pragmatic business approach by treating data center projects as a portfolio. They aim to oversubscribe projects, ensuring that even if one data center faces delays, sufficient capacity can still be delivered to customers. Cloud workloads offer flexibility, allowing companies to build in various locations and shift demand to available capacity.

Beyond business pragmatism, companies must actively engage with local communities and authorities. This involves transparently explaining their operations, addressing local concerns, and working collaboratively to mitigate pushback. This engagement is now considered a fundamental duty, akin to how new industries like Uber had to explain their model to overcome initial public resistance.

It's just a part of your duty to engage and work with new communities.

Roman Chernin Foresees Space Data Centers Becoming Reality

Roman Chernin acknowledges the seemingly outlandish nature of building data centers in space, describing the idea as 'f*cking nuts.' Despite this, he holds a strong optimistic outlook on its feasibility and eventual reality.

His belief stems from the collective efforts of numerous intelligent individuals currently working to solve the challenges associated with bringing compute infrastructure into space. He argues that with so many smart people dedicated to the task, it is only a matter of time before it becomes a reality.

Roman draws a parallel to the rapid evolution of terrestrial data centers. He points out that just three years ago, the concept of multi-gigawatt data centers seemed far-fetched. Yet, today, such large, interconnected compute clusters are routine, demonstrating how quickly once-impossible technological feats can become standard practice.

This perspective underscores a broader confidence in human ingenuity and the accelerated pace of innovation, suggesting that what appears to be science fiction today can quickly become an engineering solution tomorrow, particularly when significant intellectual capital is invested.

My view is very simple. So many smart people now working to make it happen, that why wouldn't I believe it will happen?

Democratizing Building Creates New Jobs and Redefines Education

The advent of AI is democratizing the ability for anyone to become a "builder" or "developer," meaning individuals can now readily convert their ideas into digital assets. This accessibility is expected to unlock a vast number of new opportunities, empowering millions to create and innovate without needing specialized technical skills. This shift will generate an abundance of novel businesses and entirely new categories of work that are currently unimaginable.

However, this transformation also poses a significant challenge to the education system. With universal access to intelligence, the traditional model of learning facts becomes obsolete. There's no longer a need for rote memorization when information is instantly available.

Education must therefore evolve to prioritize critical thinking, adaptability, and continuous learning. As many professions will no longer be stable, the focus needs to be on training people to think, to learn new concepts constantly, and to navigate ever-changing environments effectively. This paradigm shift presents both substantial opportunities and inherent risks for society and the workforce.

The future demands a workforce that can fluidly adapt to new roles and learn on the fly, rather than relying on a static set of learned facts. Preparing individuals for this reality requires a fundamental re-evaluation of educational methodologies and goals.

We will see a lot of new businesses and a lot of new ideas just coming in life, and they will create a lot of new works that we don't even think exist.

Soft Skills for the AI Era: Empathy and Creativity

Roman Chernin advises focusing on soft skills for future generations, shifting from his previous belief that hard skills like math and engineering were paramount. He suggests that parents should prioritize helping their children develop human-centric abilities.

The first critical skill highlighted is empathic communication, which involves understanding humans, communicating effectively with them, and demonstrating empathy. This focuses on the unique human capacity for connection and interpersonal interaction.

The second essential skill is creativity, encompassing artistic expression and the ability to try new things. Chernin believes that children who cultivate these soft skills—empathy, communication, and creativity—will be highly sought after in the job market within the next ten years.

Being able to communicate with humans, understand the humans, be empathic to the humans, and have this creativity idea, like being able to try new things and like be creative, I think these two, if you can help your kids to develop those, I think they will, in ten years, they will be in demand.

Emphasizing relentless execution amidst industry consolidation and investor scrutiny

The company perceives industry consolidation into dominant 'super-empires' as a significant threat, contrasting it with a natural desire for market diversification. There's optimism that many individuals still aim to build independently, which inherently fosters a more varied and competitive landscape.

When a prominent investor like Leo Ashenberg discloses a substantial position, the company internally acknowledges it as a justification of their efforts. However, this is quickly reinterpreted as a 'credit and opportunity to deliver' on expectations, reinforcing the need to remain grounded and focused on work.

Roman Chernin emphasizes that in a "post-sale business" and an emotional market, all growth and customer credit are seen as opportunities to perform. The company's culture, influenced by its CEO and founder, stresses that nothing is guaranteed, and consistent delivery is paramount to staying relevant.

The team operates with an intense focus, often forgoing celebrations despite significant achievements. This continuous dedication is crucial in a fast-moving market where the ability to keep pace is vital for survival, much like a shark needing to move to stay alive.

Like it's like a shark, you're alive when you move, right? So this famous thing, so we, we have to move.