Frontier Systems

Physician-investor decoding frontier systems: Immunology & Machine Intelligence. Evidence-based Medicine, first-principles investing, independent thinking

[Frontier Systems @Rocky27fly]

@StephenMoore Finally someone is using their brain

[Frontier Systems @Rocky27fly]

@SMASIMHO I am now wondering if this is for some kind of “infrastructure partnership” program here? based on the description of the role. could be wrong will find out soon. Regardless a good outcome for Nebius either way- owned data center vs combination of owned + partnership program.

[Frontier Systems @Rocky27fly]

@SMASIMHO @nebiusai @romanchernin @naralokesh Your point on “national priority” for India is very accurate. We can see how fast google got their 1GW capacity/approvals with incentives. Only certain states can provide this in India currently and Nebius should move fast in those places.

[Frontier Systems @Rocky27fly]

India is behind in the model/LLM building comapred to US and China but when it comes to AI usage/token usage/AI applications in the future, India is the market every hyperscaler will compete to capture a good share of. The AI cloud build out is in the very early stages here. So if Nebius enters early, it can scale very well alongside the big3 clouds

[Frontier Systems @Rocky27fly]

Google is already building a 1 GW data center in Vizag, Andhra Pradesh, India (google’s largest AI hub outside US) with potential extension to 5GW later. The state of Andhra Pradesh, India has a favorable policy for AI and tech investments. @nebiusai @romanchernin you may want to check with them. @naralokesh is the concerned cabinet minister for this.

[Frontier Systems @Rocky27fly]

@alexmashrabov @togethercompute @nebiusai not giving you enough compute capacity?

[Frontier Systems @Rocky27fly]

@grok @YannickBuccella So what is the role of AI here in this current achievement?

[Frontier Systems @Rocky27fly]

@RayT168 @Burle84 Exactly why these clowns like to talk about Nebius -engagement

[Frontier Systems @Rocky27fly]

@nanalyzetweets The CEO’s past and his current activities on social media does raise red flags.

[dylan ツ @demian_ai]

A decade ago, Mäntsälä wasn’t one node in the @nebiusai infrastructure network, it was the entire network.

It was our first, and, for a time, our only, data center.

It’s where the company learned to design the whole stack: servers, racks, cooling, software, and operations.

Today, that original site has capacity of up to 75 MW.
Now we’re building a second data center in Mäntsälä, adding up to 70 MW from 2027.

Together with our 310 MW AI factory under construction in Lappeenranta, that will take Nebius to 455 MW of planned capacity across three Finnish sites 🇪🇺🇫🇮

Finland has the ingredients AI infrastructure needs: reliable, low carbon power, a climate made for efficient cooling, serious eng talent, and communities that understand the value these facilities can return locally.

From one data center to one of Europe’s largest purpose built AI infrastructure footprints.

We started in Finland, we’re still only getting started. 🇫🇮

[Nebius Academy @Nebius_Academy]

☀️ Your summer reading list could probably use more speculative decoding.

Meet 𝐓𝐡𝐞 𝐏𝐚𝐩𝐞𝐫 𝐏𝐢𝐜𝐤𝐬 – a weekly newsletter from @nebius_science, covering the papers on our R&D team’s radar.

👉 Subscribe now for a new selection every week: https://t.co/CEieCCA6KA

[Frontier Systems @Rocky27fly]

@AlexisWrem1747 “Foresight and vertical integration since 2019” while current quarterly AI cloud revenue is around 30 million-make it make sense😂😂

[Roman Chernin @romanchernin]

Thank you, @Enigma_AI, for the trust. All our growth is driven by great builders and your sleepless nights pushing the frontier of what is possible.

[Jay @jiwanpark1014]

$NBIS
Tavily overhauled its search engine specifically for AI agents (focusing on better snippet ranking, removing duplicate/conflicting info, and fresher indexing) and hit #1 on SealQA-Hard, SealQA-0, and SimpleQA, outperforming Exa, Perplexity, Brave, and You com.

The Upgrade:

Fixed 3 core search issues:

Reranking: Prioritizes exact, high-quality answer snippets over just popular website links.
Contradiction Handling: Filters out redundant text and conflicting data so LLMs don't get confused or hedge.
Freshness: Faster index updates for time-sensitive queries.

Why SealQA Matters: SimpleQA is too easy/basic, while multi-step benchmarks (like BrowseComp) mask pure search performance behind agent prompting. SealQA tests complex queries in a single search step.

As LLM reasoning improves, search quality is the main bottleneck. The live updates are already rolled out across Tavily’s API, SDKs, and MCP server.

[Frontier Systems @Rocky27fly]

@OakenshieldVGX Keep dreaming

[Jensen Huang @JensenHuang]

The mighty A100 fleet are mission-capable from 2020 through 2029. NVIDIA computing is more than chips. CUDA gives developers and NVIDIA engineers a common platform to continually upgrade Ampere, Hopper and Blackwell throughout their useful lives.
CUDA makes NVIDIA computing versatile. Versatility makes it fungible. Fungibility drives utilization and extends durability, making NVIDIA compute a productive asset: rentable, durable and financeable.

[dylan ツ @demian_ai]

I’ve been at @nebiusai for two years.
Today feels like one of those days we’ll look back on.

When people look at AI infrastructure, they often reduce the story to one thing: GPUs.

But what we’re building at Nebius is better understood through an equation:

$/MW × MW activated × speed-to-revenue × funding efficiency + software leverage

And in Q2, every part of that equation moved.

1. First: $/MW.

Our 2026 base was around $12M in annual contract value per MW.
- The deals signed in Q2 came in above $20M/MW.
- Short-term capacity opportunities are now reaching $40–50M/MW.
- Pricing on older gen GPUs increased by more than 30% vs Q1.
- Our first Blackwell capacity auction cleared 15% above the highest price we had achieved before, and 20% above the existing pipeline.

That’s what real pricing power looks like in a market where certainty, scale and speed matter.

2. Second: MW activated.

We raised our 2026 contracted-power target from more than 4GW to 5GW, and expect to deploy more than 1GW of new capacity annually starting in 2027.

But contracted power isn’t live compute.

Between the two sit permitting, construction, energy, cooling, networking, GPUs, orchestration, testing and customer onboarding.

The unglamorous (and brutally difficult lol) work required to turn land and electrons into reliable production infrastructure.

After two years at Nebius, I’ve come to believe that this conversion engine is one of the real moats.

3. Third: speed-to-revenue.

-> Q2 revenue reached $582M: +454% YoY and +46% QoQ.

the ai cloud rev reached $575 M. ARR reached 3B$ (adjusted EBITDA margin reached 50%)

And this was before most of our planned 2026 capacity comes online in the second half.

Demand is no longer only a pipeline story.

- Total contract value won in Q2 grew nearly 4× QoQ. TCV from new customers more than 9×'d
- four landmark agreements averaged more than $1B each (with Reflection, Cohere, a US AI neolab and a major US quant firm)

These weren’t passive inbound wins. They took competitive processes, technical POCs and trust earned across multiple engagement cycles.

Customers chose scale, performance, reliability, technical support and the ability to grow with us.

This is becoming a strategic-partner business, not a GPU-rental business.

4. Fourth: funding efficiency.

Roughly 70% of Q2 deals included customer prepayments.

Across the four landmark agreements, those prepayments cover 50–60% of the associated capex.

Expected payback fell to 1 year and 10 months, compared with 2–3 years historically.

Add more than $9B of expected customer prepayments in 2026 and our first ~$775M asset-backed facility at SOFR +250bps, and the flywheel becomes:

Signed demand → prepayments → infrastructure → contracted cash flows → cheaper financing → more infrastructure.

That is a long way from “build it and hope.”

But let’s be precise: this remains infrastructure.

Q2 capex was approximately 5.7B$. Converting GW into production compute requires enormous capital and near flawless execution. The breakthrough isn’t that the business suddenly became capital-light. It’s that more of the build is becoming contracted, prepaid and financeable.

5. And then comes the final part: software leverage.

@nebiustf production inference workloads more than tripled in Q2.

Aether 3.6 strengthened the core cloud platform. @tavilyai brings the real time information agents need. Echo creates a natural language interface for operating infrastructure. @Eigen_AI_Labs and @clarifai deepen our inference capabilities.

This part is particularly personal for me because much of my own work sits close to Token Factory.

The opportunity isn’t simply selling GPU hours, it’s serving the complete AI lifecycle:

Training → post-training → inference → grounding → agents.

Training creates inference demand. Inference creates demand for tuning, evaluation and optimisation. Agents multiply model calls and require reliable access to real-time information.

A platform that serves that entire loop can improve customer outcomes, infrastructure utilization and economics simultaneously.

So the complete equation is:

Higher $/MW
× more MW activated
× faster conversion into revenue
× smarter funding
+ more software per unit of compute.

Each term reinforces the others.
And that’s why today’s announcement is so meaningful.

Two years ago, much of this existed as ambition, architecture and an extraordinary group of people willing to build from first principles.

Today:
- $3B ARR. 50% AI-cloud adjusted EBITDA margin
- More than $40B in customer commitments.
- A 5GW contracted-power pipeline.

Still early, but very real.

I’m incredibly proud of what this team has built, and even more excited by what’s ahead.

The Nebius story isn’t “more GPUs.”
It’s turning power into compute, compute into intelligence, and intelligence into products people use every day.

Strategy → execution → compounding.

We’re just getting started. $NBIS

Disclosure: I work at Nebius. This post is based exclusively on today's public earnings materials and public earnings call. The interpretations are my own and do not represent official company guidance or investment advice. Personal views, not investment advice.

Public sources:
https://t.co/DsFTpShaK2

https://t.co/fOqXoWDB8Z
https://t.co/bYmCk6du4k