Hey builders,

I recently had the chance to sit down with Roman Chernin, Co-founder of Nebius, a company building the cloud foundation for AI inference.

Nebius recently closed a $17 billion deal with Microsoft and a $3 billion deal with Meta. It is one of the world’s fastest-growing AI infastructure companies.

Full Interview here:

What we discuss:
🔹 The Microsoft & Meta deals - why hyperscalers are coming to Nebius for dedicated GPU capacity
🔹 Full-stack AI infrastructure - from data centers to managed inference platforms
🔹 When to stick with closed-source models vs. switching to fine-tuned open-source models
🔹 Token Factory - their managed inference and post-training platform
🔹 Real results: How Prosus achieved 26x cost reduction vs. proprietary models
🔹 NVIDIA's $20B Groq deal - what it means for the inference market
🔹 Why enterprise AI adoption will happen faster than anyone expects

Today I'm breaking down:
→ What Nebius actually does (and why it matters for your AI stack)
→ The key takeaways from my conversation with Roman that are relevant for people building in AI
→ MCP Connect Day in Paris this Thursday (I'm speaking!)
→ How you can track ChatGPT App Store data as apps are now getting approved daily!

Let's dive in.

Why Should You Care About AI Infrastructure?

Here's the thing most people don't realise: the cost and complexity of running AI models, particularly inference, is where most companies hit a wall.

You've probably experienced this. You build an AI feature. It works great in development. Then you try to scale it and watch your costs explode. Latency becomes a nightmare. The whole system becomes unreliable.

The natural instinct is to reach for expensive proprietary models - but that crushes your margins at scale. The alternative? Open-source models. But here's what most people miss: unoptimized open-source models aren't actually that good out of the box.

You need both pieces working together - optimized models AND the infrastructure to run them efficiently.

Nebius is solving this by building a full-stack solution specifically designed for the real-world application of AI:
→ Silicon: Specialised hardware optimised for AI workloads
→ Software: A robust platform for deploying and managing AI models
→ Operational Support: The expertise to ensure seamless operation and scalability

They're enabling businesses to move beyond simply training AI models and actually using them effectively - with open-source models that are properly optimized to deliver real performance.

Key Takeaways from My Conversation with Roman

1. When Should You Switch from Closed Source to Open Source?

Most AI companies start with closed source models (OpenAI, Anthropic, Gemini).

Why? They don't want additional technological risk. They focus on the product.

But when they hit product-market fit and start scaling, closed models may not be sufficient.

Two reasons to switch:
- Cost performance ratio
- Quality tuning for your specific use case

You don't need the best model in the world.

You need a model sufficient for YOUR use case that can be tuned to YOUR data.

2. Open Source Can Slash Your AI Costs by 10x-100x

Prosus (Nebius customer) achieved 26% cost reduction vs proprietary models.

But that's not even the ceiling.

Roman says they regularly see "order of magnitude optimization" - tens or even hundreds of percent improvements.

The key: optimizing for YOUR specific use case, not the average.

If you have high-volume, repeatable tasks - open source + tuning is where the economics flip.

3. What the GROQ Deal Tells Us About Inference

NVIDIA just did a $20B deal to license GROQ’s inference technology.

Why? Roman: "Inference is quite diversified. Behind 'inference' there are quite different workloads."

- Super large reasoning models = one set of requirements
- Low latency models = completely different (latency & price are key)

We'll see more specialized inference offerings from all major players.

But Nebius plays a different game - the service layer.

Building infrastructure at scale, making it reliable and cost efficient, plus the software stack.

Better silicon underneath = more use cases they can serve on top.


4. Enterprise AI Adoption Will Happen Faster Than You Think

People debate how fast enterprises will adopt AI.

"Compliance issues, products not ready, models not secure..."

Roman's take: You'll be surprised how fast they move.

AI has consistently surprised us with scale and pace.

If you think enterprises will be slow, you'll probably be wrong.

The companies betting on slow adoption may get left behind.

5. The AI Adoption Curve: From Startups to Enterprise

Three waves of AI adoption:

- Wave 1: AI-native startups (vertical AI companies, model builders)

Move fastest, make decisions fast Can work with bare metal offerings


- Wave 2: Digital enterprises (e.g. Shopify, Booking .com)

Need higher level of abstraction

Require compliance, security, multi-cloud


- Wave 3: Classical enterprises (banking, manufacturing, transportation)

Mostly inference, not training Looking for complete platforms, not building foundations


Each wave needs different products.

The use cases shift from training → inference as you move up the curve.

Speaking at MCP Connect Day - Paris, February 5th

Just two days ago Anthropic released MCP Apps. Combined with ChatGPT's App Store, VS Code, Goose, and upcoming support from Block, Microsoft, JetBrains, AWS & Google DeepMind - we're seeing something big emerge.

They're all using the same MCP App standard. ONE app can now live inside MULTIPLE LLMs at once.

Which creates an entirely new problem: discoverability.

My talk: "If the LLM Can't Find You, You Don't Exist"

When millions of apps compete to be surfaced under the same prompts across multiple platforms, just having the best app doesn't cut it. You need to be the one that gets recommended - everywhere.

200+ AI builders will be there. Speakers from OpenAI, Mistral, Hugging Face, GitHub, Apify and more.

Link to tickets here (not sure how many are left).

Staying on top of ChatGPT App Store Data!

ChatGPT is finally approving non-launch partner apps into the app store!

They’ve been doing this more and more frequently over the last week. In fact,

We’re tracking all the App Store data and have it publicly available at appdiscoverability.com/track for anyone who wants to stay up to date on it!

As always, feel free to drop me a reply with feedback or questions. I read every reply!

Happy building,

Elliot

Learn how your business could grow with Apps on ChatGPT

If your company or brand is interested in getting an app inside ChatGPT, we'd love to talk to you.

Please book a strategy call with our team below.