AI-Native

Amber Yang built an AI model in high school that predicted space debris collisions, won the International Science Fair, and turned it into a startup while studying at Stanford. Eventually these lessons led her to a role in truly AI-Native investing as a Partner at Lightspeed.

In this episode, Anu and Markiesha sit down with Amber to unpack how a technical investor builds conviction in AI infrastructure – why the founder archetype has gotten younger, how some infra deals close in 30 minutes, and which founder types usually succeed at different layers of the AI stack.

Amber also talks through the buzzwords she hears most often and how she guides them to something more concrete, why compute constraints rarely make her pass at the early stage, and where she thinks the capital-intensive parts of the stack end up.

What You’ll Learn:

• Why the AI-native founder archetype keeps getting younger

• Why raising seed capital can be easier than landing a research job at top labs like OpenAI and Anthropic

• How infra deals close in 30 minutes, and the years of relationship-building behind them

• How founders can find success in the infra layer and the application layer

• What a world model means to her, and the buzzwords she’s forced to challenge in meetings

• Why compute constraints rarely stop her interest at the early stage

Learn more: www.nebius.com/podcast

Ryan Wang started his PhD at MIT in 2018, right as the transformer architecture was reshaping AI, and spent late nights restarting GPU racks in a campus basement. Years later, the startup he co-founded, Eigen AI, would rank first among inference providers serving the same open model, earning a shoutout in Jensen Huang’s 2026 GTC keynote. Along the way he learned that at scale, the model is the easy part. The real challenge is in everything it takes to serve it.

In this episode, Anu and Markiesha sit down with Ryan, now the SVP of Applied AI at Nebius, to unpack why inference optimization has become one of the biggest levers in AI economics including how optimization is now squeezing four to five times more tokens out of the same GPUs, and why a fine-tuned 70B open model can beat frontier closed models on specific tasks.

Ryan also shares the prompt-design choice that keeps your cache intact, why cached context costs a tenth of recomputing it, and when founders should buy inference instead of building a team and delves into why cheaper tokens will grow the inference market rather than shrink it.

What You’ll Learn:

• Why everything about inference changes at ten times the scale

• The prompt-design choice that keeps your cache intact

• How open models can outperform frontier closed models

• When founders should buy inference as a service and when to build an in-house team

• Why Ryan thinks the gap between open and closed models has shrunk from a year to a few months

• The story behind Eigen AI’s monumental GTC shoutout

• Why he sees dramatic expansion of the inference market

• What changes he sees coming to inference infra, including agentic AI infrastructure

Learn more: www.nebius.com/podcast

GPU scarcity has forced AI teams into a world their infrastructure was never built for — reserving capacity across multiple providers, geographies, and SKUs and watching 30 to 40 percent of expensive allocations sit idle. In this episode, Zongheng Yang, co-founder and CEO of SkyPilot, joins the hosts to unpack why being multi-cloud stopped being optional for AI teams and what it actually takes to manage compute efficiently once you have it.

Before founding SkyPilot, Zongheng was a speech researcher at Google Brain and a co-author on Tacotron, a PhD student at UC Berkeley's Sky Computing Lab (the birthplace of Ray, Spark, and VLM), and a contributor to the early open-weight model Vicuna. He and his co-founders built SkyPilot to solve their own problem as AI researchers: compute scattered across multiple clouds being impossible to fully utilize. Four years and 14 million downloads later, SkyPilot has become an open source standard for orchestrating AI compute across providers, with more than 280 contributors and customers including Nubank, Abridge, and Hippocratic AI running production workloads on it.

Zongheng breaks down why decades-old solutions like Kubernetes and Slurm quickly buckle under the combinatorial complexity of multiple clouds, multiple GPU SKUs, and multiple research teams competing for the same fleet. He explains why idle GPUs, not just scarce ones, are the bigger hidden cost — a team spending $100 million a year can lose $15 to $30 million to poor utilization alone — why SkyPilot bet on open source and the Apache 2.0 license from day one, and why he thinks the winning bet right now is to start building for a compute-abundant world rather than hoarding capacity for a scarce one.

In this episode, you will learn:

• Why multi-cloud went from a nice-to-have to a necessity for AI teams almost overnight

• What breaks when a team moves from one cluster to managing dozens across providers

• Why idle GPUs can cost frontier teams tens of millions of dollars a year

• The difference between getting compute and knowing how to use it well

• Why SkyPilot's founders bet on open source and the Apache 2.0 license from the start

• What changes for AI teams once compute stops being the constraint

If you're building or scaling an AI team's infrastructure, this conversation will change how you think about the true cost of underutilized compute.

Alex Mashrabov spent years at Snap building face filters, probably the most widely used media AI product ever made.

Learn more at Nebius.com

When he left to start Higgsfield, he thought the hard part would be the models. It turned out to be something else entirely.

In this episode, Joshua Liss, Head of Media and Entertainment at Nebius, sits down with Alex, CEO & Founder at Higgsfield AI, to go deep on what it actually takes to build a platform at the center of a creator economy being remade by AI. Today, over 50% of creators on Higgsfield now produce content end to end with AI, a number that was unimaginable just 18 months ago.

Alex shares the secret behind Higgsfield's growth that he has never talked about publicly before, why the company's 50-person creative team is the real competitive moat, how Higgsfield trained Soul 2.0 on Nebius B300s to deliver a model that is 20 times cheaper and 20 times faster than the leading alternative, and what he sees coming for the entertainment industry that most people aren't ready for.

0:00 - Intro teaser

0:50 - Higgsfield's thesis: why most social content will be AI-generated

2:22 - The rise of AI-native creators going end-to-end

4:07 - The secret: Higgsfield's 50-person creative feedback loop

5:45 - Democratizing high-end content and the Higgsfield contest

8:52 - From Snap face filters to Higgsfield

13:35 - 20x faster, 20x cheaper

16:55 - Prediction: AI-native IP and an AI-first Netflix

What You'll Learn

• Why Higgsfield's real competitive moat isn't the research team but a 50-person creative feedback loop.

• The three stages of AI in entertainment and why stage three is closer than most think.

• How Higgsfield trained Soul 2.0 on Nebius B300s and cut inference costs 20x.

• Why the largest AI video competition in the world was won by names nobody had heard of.

• What it means to build like a seismologist, always hunting the next wave of creators before it breaks.

• Why any creator can now compete with what used to require a million-dollar production budget.

• Will the next Netflix will be AI first?

Lukas Ziegler, Robotics Evangelist, Investor, and Advisor, has spent the past 10 years working across robotics software, commercial roles, content, and investing. Now, he believes the industry is entering a new era.

Learn more at: ⁠nebius.com

In this episode, Evan Helda, Head of Physical AI at Nebius, sits down with Lukas to explore the shift from Robotics 1.0 to Robotics 2.0: a world where you don't program a robot, you just tell it what to do.

They discuss why the humanoid debate has been asking the wrong question, why the data gap between LLMs and robotics is vastly misunderstood, and why the companies that will win are already co-building with end users rather than shipping to them.

They also get into storytelling: why public exposure builds trust in robots faster than any marketing campaign ever could.

0:00 - Intro teaser

0:40 - Lukas Ziegler’s journey from engineer to robotics evangelist

2:40 - Robotics 2.0: the shift from programmed to general-purpose robots

4:45 - The biggest investment trends in physical AI

7:45 - Why robotics has a massive data problem

11:00 - Synthetic data, AI moats, and the future of RobotOps

17:00 - Why storytelling will determine robotics adoption

19:15 - Preparing for a future where robots free up our time

20:00 - Outro

What You'll Learn

• What Robotics 2.0 means and why it's a fundamentally different paradigm from Robotics 1.0.

• Why specialized humanoids may have a clearer path to deployment than general-purpose systems.

• The massive data gap between LLMs and robotics, and why physical AI will scale gradually.

• The three investment theses: verticalized humanoids, RobotOps infrastructure, and data collection.

• Why robotics companies should involve end users from the earliest stages of development.

• Why exposure beats education when it comes to public trust in robots.

What does it actually look like when enterprise AI moves from experimentation to real work?

Learn more at: nebius.com

In this episode, Markiesha Patrice, Head of Startup Platform & Community at Nebius, sits down with Debanjan Saha, CEO of DataRobot. They explore what the transition from AI pilots to production-ready agentic systems looks like from the inside.

Debanjan has been at the forefront of three technology revolutions: the internet, the cloud, and now AI, and he brings a clear-eyed, optimistic framework to a moment he believes is an order of magnitude larger than anything that came before it.

From navigating three major technology revolutions to leading one of AI’s pioneering companies, Debanjan shares his perspective on leadership, disruption, and the opportunities emerging as AI moves from experimentation to real-world impact.

0:00 - Intro teaser

0:50 - Hype or real? Why this revolution is different

4:20 - The two-timescale strategy for staying ahead

5:20 - What DataRobot builds: the agentic workforce

8:25 - Inside the platform: building and running agents

14:30 - Frontier agents: Chevron's oilfield safety

18:20 - The power of partnership: Nebius and NVIDIA

20:15 - Advice for builders: create 10X value

25:20 - The vision: building AI factories

26:50 - Outro

What You'll Learn:

-Why the AI revolution is an order of magnitude larger than the internet and cloud revolutions that preceded it.

-How the enterprise AI market is evolving from chatbot experimentation to high-value agentic transformation.

-The full build-to-run framework: testing, validation, identity management, governance, and real-time intervention.

-The one-way door vs. two-way door decision making framework he learned from his time at Amazon.

-Why startups must create 10x value to win enterprise customers.

What if AI’s next breakthrough isn’t a larger model, but a different way of thinking about intelligence?

Learn more at: ⁠nebius.com

In this episode, Markiesha Patrice, Head of Startup Platform & Community at Nebius, sits down with Sudip Roy, co-founder and CTO of Adaption. After helping build AI infrastructure during the scaling era at Google and Cohere, Sudip now argues that the industry’s future won’t be defined by ever-larger models alone.

He explains why leading foundation models are converging, why enterprises should think in terms of AI systems instead of individual models, and how continuous adaptation could become the next major shift in artificial intelligence.

0:00 - Intro teaser: The days of monolithic AI are over

0:47 - Challenging scaling laws: how Adaption Labs started

3:18 - The three pillars: data, continual learning, interfaces

6:52 - Adaptation is all you need: the future of AI models

8:40 - Proportional compute and the AI energy gap

11:00 - Solving the 5% reliability gap for enterprises

11:57 - Advice for builders: take contrarian positions

14:36 - Adaptive Data: synthetic data in 240+ languages

15:19 - Outro

What You'll Learn:

-Why leading foundation models are beginning to converge and what that means for the future of AI.

-Why thinking about tasks instead of models changes how enterprises should build AI systems.

-Why enterprise data remains AI’s biggest untapped resource.-How continuous learning closes the 5% enterprise reliability gap caused by domain-specific organizational data.

-Why challenging industry consensus can create breakthrough startup opportunities.

Henry Wang built a consumer app that 10 million people use for 150 minutes a day. For most of that time, the assumption in the industry was that you needed a frontier model to deliver a great AI experience.

Learn more at: nebius.com⁠

Henry found a different answer.In this episode, Joshua Liss, Head of Media and Entertainment at Nebius, sits down with Henry (Lifan) Wang, Co-founder of Kaon Labs, to get into how you actually scale a consumer AI product. Not just find product-market fit, but build the infrastructure that lets you serve tens of millions of users, test 100 models a week, and generate a trillion tokens a day without it costing you everything.

Henry explains why personalization at consumer scale isn't a recommendation system anymore, how Kaon Labs replaced two-week A/B tests with Elo ranking that resolves in two hours, why entertainment is the right beachhead for consumer AI right now, and what in-experience advertising could look like.

He also shares his contrarian take on what actually separates the consumer AI companies that break out from the ones that spike and die.

0:00 - Intro teaser: 10M users, 115 minutes a day & 1 trillion tokens

0:20 - How Kaon Labs built AI infrastructure for 10 million users

3:02 - Why AI roleplay is changing entertainment forever

5:10 - Hyper-personalization vs. recommendation algorithms

8:18 - Why scaling consumer AI is harder than enterprise AI

11:33 - Building and evaluating 100 AI models every week

16:08 - Monetizing AI with subscriptions & immersive advertising

19:43 - Rapid fire: The AI-native leaders every startup needs

22:08 - Lessons from building a 10 million user consumer AI company

25:22 - Outro

What You'll Learn:

-Why testing 100 models a week against real users is the core of Kaon Labs' competitive advantage.

-How Elo ranking replaced A/B testing: evaluating models in hours, not weeks.

-Why AI replaces recommendation systems with true personalization.

-Why entertainment is the ideal proving ground for consumer AI.

-Product Market Fit matters, but scale is what separates breakouts from blips.

-What in-experience advertising looks like when it lives inside the story rather than interrupting it.

-The three things to look for in every key hire.

Brian Zhan doesn't want your business plan. In fact, if you hand him one at the seed stage, that might raise some red flags.

Learn more at: ⁠nebius.com⁠⁠

In this episode, Anu Maheshwari, Head of Venture Ecosystem at Nebius, sits down with Brian Zhan, partner at Striker Venture Partners. They unpack what it actually means to back people, not products, at the frontier of AI.

From reinforcement learning's generalization advantages, to recursive self-improvement loops in chip design, Brian makes the case that the VC consensus is missing the most important early bets by waiting for revenue.

Brian has been on the boards of Reflection, Skilled, and Periodic Labs. With a personal history in robotics research dating back to 2014, he maps out where the next trillion-dollar companies will emerge, why most VCs are too focused on revenue timelines to see them, and how Striker approaches seed investing with 10% of fund concentration per bet.

0:00 - Intro teaser: Why the best founders don't need a business plan

0:27 - Why Striker invests before there's a business plan

5:04 - Is AI in a bubble? Separating hype from opportunity

7:45 - How Brian identifies world-changing founders

9:24 - Reinforcement learning and the path to superintelligence

14:05 - Robotics, world models, and the future of physical AI

17:55 - Why most VCs think too small

20:18 - What Brian looks for in a founder pitch

23:37 - The OpenAI playbook & why compute partnerships matter

26:05 - Outro

What You'll Learn:

-Why a detailed business plan at seed is unnecessary.

-The precision behind frontier AI investing and why it's fundamentally different from SaaS.

-Why reinforcement learning underpins recursive self-improvement and why most investors don't actually understand it.

-How world models and simulation will accelerate robotics toward its ChatGPT moment.

-Solving energy supply by unlocking insight for everything else.

-How Striker evaluates founder-market fit in practice.

Tavily hit one million users without a single outbound cold email.

Learn more at: ⁠nebius.com⁠⁠

Not because of a viral moment, but because the founder understood earlier than most that the unit economics of agentic search are fundamentally different.

When you’re returning context to an AI agent rather than links to a human, token waste, latency, and information density become the variables that determine whether your product survives in production. Tavily built its architecture around that constraint from day one, and developers noticed. After being mentioned in 200+ research papers and 1,000+ YouTube videos, Fortune 500 enterprises were already running Tavily internally before a sales conversation ever happened.

In this episode, Markiesha Patrice (Head of Startup Platform & Community at Nebius) sits down with Rotem Weiss, Founder and CEO of Tavily, to get into the technical and strategic decisions behind building the search layer for the agentic web.

Rotem breaks down why optimizing for snippet retrieval over link retrieval changes the entire architecture of a search product, how Tavily's prediction engine pre-processes likely queries to cut latency without sacrificing accuracy, and why he believes the most important GTM decision an AI infrastructure company can make in 2026 is getting integrated into the coding assistants autonomously selecting tools on behalf of engineers.

0:00 - Intro teaser: the buyer isn't a human anymore

0:25 - Guest intro & what Tavily does

0:47 - Rotem's origin story & the first check

1:40 - The real-time data gap & launching GPT Researcher

4:30 - What is agentic search? Snippets, not links

6:55 - The prediction engine that kills latency

7:55 - Why 2026 is the year of token optimization

10:05 - Inside the Nebius acquisition

11:25 - The team & the "second CEO"

12:45 - "The buyer is the AI assistant" — winning coding tools

14:30 - GPUs as rocket fuel

16:30 - One million users, zero cold emails

19:20 - "I'm the worst at everything in the company"

20:30 - Israel vs. the US founder mindset

What You'll Learn

-Why returning the right snippet rather than the right link changes the entire architecture of a search product.

-The prediction engine Tavily uses to pre-process likely queries at the website level, reducing processing overhead without sacrificing accuracy.

-Why coding assistants autonomously selecting tools in a stack are becoming the real GTM lever for AI infrastructure companies.

-The open-source to enterprise conversion motion Rotem actually used.

-How the migration from large closed-source models to fine-tuned open-source models changes the token economics for AI companies trying to get to production.

Learn more at: https://nebius.com/startups/podcast