“From Netflix to StubHub: Building AI Foundations in 2026” | The Data Huddle by Yuki – Sundar Velayutham, Principal Data Engineer at StubHub

By Dana Bilder Morozova
September 28, 2026 | 5 min read

The Data Huddle by Yuki is a podcast about modern data leaders at global companies. The people who make data work every day, but rarely get the spotlight. Honest conversations about their careers, wins, mistakes, and how they see the future of AI and data.

What does nearly two decades of running data at scale teach you about the moment the industry is in right now? In this episode, Dana Bilder-Morozova sits down with Sundar Velayutham, a principal data engineer with nearly two decades in the field, for a conversation about what actually made Netflix a data company, what separates a great data team from an average one, and why AI has made infrastructure busier rather than cheaper.

About the Guest

Sundar Velayutham started in database engineering, working with traditional OLTP systems like Oracle, and moved into big data at Yahoo during the early Hadoop years – almost two decades ago, when the term “data engineering” was still being coined. That led to Netflix, where he spent roughly eight years across multiple domains, starting on the experimentation data engineering team. From there he went to Apple, where he moved from building systems himself to directing technical decisions across a large engineering group. Today he works on data and AI foundations.

Read that list again and the interesting part is not the logos. It is that at every point where the obvious next step was management, he did not take it.

Episode Highlights

“Data, Data, Data – Every Product Decision Was Based on How Users Would React to It”

When Sundar joined Netflix roughly twelve or thirteen years ago, it was not big tech.

“Netflix, when I started there, it was not big tech. We were less than 10 data engineers in the company.”

The DVD-by-mail business was still substantial and streaming had only just begun. What leadership had was a view of where media was going, and the discipline to test every step towards it against real user behaviour.

“They would run over 200 tests at any given time, on every product surface, to understand how would we increase streaming, how would we improve retention, how would we make sure that users stick with us for a longer time.”

“So data, data, data – every product decision, change this button, make this box art, everything was based on how users would react to it.”

Two hundred concurrent experiments is not a culture statement. It is an infrastructure commitment. Every one of those tests has to be instrumented, collected, stored, joined and computed, and the results have to arrive fast enough to act on. A company that calls itself data-driven and runs three tests a quarter is describing an aspiration. Netflix built the platform that made the claim true, and then used it to decide which shows to make.

“That’s How You Get Your Next Job”

On paper his career looks like a straight line built on brand names. Sundar’s own account is almost entirely about people.

“In hindsight, I’m more grateful to what I learned along the way, the relationships that I built. The leaders, the ICs, the people that mentored me, and the people that I mentored over the years at Netflix – that’s how you get your next job.”

He moved to Apple because a leader he had worked with at Netflix built a team there and asked him to come. The move after that happened the same way: people he had worked closely with years earlier, now somewhere new, telling him they were building something worth joining.

“That’s why you want people more than just the title and the company.”

Worth taking seriously if you are earlier in a data career. Senior data hiring at this level is not a job board, it is a small network of people who have already watched each other work. The version of that advice that is actually actionable is unglamorous: the colleagues you have now are the hiring pipeline you will have in five years.

The Difference Between a Great Data Team and an Average One

Asked what separates the two, Sundar does not mention a single tool.

“Most of them understand that the most important thing is to move the business forward. That’s the key. How does the business get better? How does the business engage more of their customers? How do they make more revenue? It’s a business at the end of the day.”

The structural version of that is vertical integration.

“Companies that get successful are often where the data is closely aligned with the product and integrated vertically into every product vertical. That’s where the biggest wins come from.”

Notice what is absent from the answer: architecture, stack, headcount, maturity model. A data team measured on pipeline count will optimise pipeline count. A data team embedded in a product vertical and measured on what that vertical earns will eventually kill its own pipelines when they stop paying for themselves – which is a different, healthier, and much rarer instinct.

“Building Is Back on the Table”

Sundar spent years doing deeply hands-on work: re-engineering petabyte-scale pipelines, designing data models, building systems meant to heal themselves. Then, in the last two to three years, the ceiling moved.

“You can build things that you were limited by your skill set. You could do the work of an SWE, you could do the work of an analyst, you could be a data engineer – you could wear multiple hats and do end-to-end engineering work today.”

“Every good IC who’s senior enough will say that being hands-on and building is back to being awesome. It’s not easy-easy, AI still has its quirks, but building is back, and being hands-on is back on the table.”

This is the most optimistic thing in the episode, and it is also the explanation for the least optimistic thing. The reason senior engineers are building again is that the cost of attempting something dropped. When the cost of attempting something drops, the number of attempts goes up. That has to land somewhere, and it lands on the platform.

What “AI Foundations” Actually Means

His current remit is data and AI foundations. Stripped of the buzzword, it means building the frameworks and primitives that stop every team in a company from solving the same problem privately.

“Right now everybody builds, because everybody is in the camp of – I can build that, I should build my own framework, I should just do this on my own thing.”

He expects the same consolidation the data world already went through once.

“Just like we aligned on really good data platforms, data catalogs, data pipeline technologies, we will also align on AI frameworks and AI foundations – such that the primitives that exist that empower other engineers become standard, and nobody would have to go build their own.”

Every team that builds its own version of the same context layer pays for it twice: once in engineering time, and again in the compute that a half-tuned framework burns in production. Standardisation is usually pitched as a governance argument. It is at least as strong as a cost argument.

The Rule He Never Compromises On at Scale

Two rules, actually. The first is to know what the data is feeding before designing anything at all.

“You always have to understand what your data is feeding. Is it information, metrics, reporting – or is it real-time use cases that feed into the product, where the product is actually using the data to give a user experience?”

Everything downstream follows from that answer: instrumentation, collection, transport, logging, telemetry, the data model, the technology choice. In his words, “not one size fits all.” The second rule is standardisation, and the reasoning is very practical.

“In big data engineering there are like 20 ways to do the same thing. Hundreds of companies building products like Snowflake, Databricks, and tons of execution engines like Trino, ClickHouse, Druid. You have to always standardize. You have to understand what works well for what use case.”

Both rules point the same direction. Most expensive platforms are not expensive because someone chose the wrong engine. They are expensive because nobody decided, so the organisation ended up running four of them.

“A Lot of Work in the Past Two Years Has Been Thrown Away”

We asked whether he worries about infrastructure becoming more expensive and less predictable. His answer is the most quietly important thing in the conversation.

“There has been an increased load on our tech stack and infra from a compute perspective – from how we write code, how many PRs we generate, how much more data that we try, build and throw away.”

“A lot of work in the past two years has been thrown away. You build something, the technology changes, AI changes, some things become part of foundation harnesses, and then you throw away your work.”

This is the line item almost nobody has modelled. Productivity gains at the engineer level do not reduce platform spend, they multiply the work arriving at it: more code, more pull requests, more pipelines, more experiments, more data written and then discarded. The discarded part still ran. It still scanned, still shuffled, still billed.

And it is not a phase to wait out. Sundar is explicit that the experimentation itself is now the normal state of engineering – which means the throwaway compute is a permanent line in the budget rather than a one-off cost of adopting AI.

There is a second thread here worth pulling. He is blunt about what AI should retire, and it is the thing most cost overruns are hiding behind:

“Now people just talk to AI and get the data they need when they need it, instead of a dashboard refreshing petabytes of data underneath every day and running reports that nobody ever uses.”

How He Evaluates an AI Tool Before It Touches the Stack

Sundar receives hundreds of vendor emails. His filter is engineering, not marketing.

“You have your concept of unit tests. You have a set of inputs, an expected set of outputs. You make sure that you run those and see if you’re getting the output you expect – and within a reasonable cost. You don’t want a token max.”

Then it gets instrumented like anything else that runs in production.

“You build data pipelines that output observability data from these AI tools to make sure that it’s doing the right thing, and you have metrics on top. Accuracy metrics, correctness metrics, conciseness, verbosity. Did it pick the right data set? Did it do the right query? Do they churn – do they do 10 things trying to figure out what to do?”

That last question is the expensive one. A tool that arrives at the right answer after ten unnecessary queries passes a correctness test and fails an economics test, and only one of those two shows up in a demo. Note also how naturally cost sits inside his definition of quality: “within a reasonable cost” is not a procurement note, it is one of the unit tests.

On build versus buy, he is honest about his own bias – “I’m an engineer at my heart, so I think it’s always better to build” – and clear-eyed about what that means for the market. The buyer either already built something similar, or expects their platform vendor to ship it next quarter. Which is precisely why anything sold into this stack now has to prove it in production numbers rather than in a deck.

“I Can Confidently Say I Don’t Really Know”

Asked to name the biggest trend of 2027 and 2028, he gives the most credible answer we have had on this show.

“It’s one of those questions that I can confidently say I don’t really know.”

What he will commit to is shape rather than timing: frontier models keep their place, alongside a proliferation of smaller open-source and open-weight models, and the stack itself becomes agents.

“It will be agents talking to agents, learning from each other, learning from you, or already have learnt everything from you. A lot of it will get automated. Anybody who thinks that that’s not coming is wrong.”

On whether that replaces the people doing the work today, he is measured rather than reassuring.

“Every transformative technology will replace a lot of what people do today as data practitioners. But people are very flexible and adapt, and the industry adapts really fast. Roles will change, titles will change, we will do other things, and we will let AI do things that we do today, because that becomes a commodity. I don’t think that you necessarily get replaced. You do something else.”

He applies the same discipline to layoffs, separating two causes that usually get merged into one headline: organisations that hired faster than they could justify to shareholders, and genuine productivity gains where one engineer now covers the work of two. Both are real. Neither, in his reading, is AI taking the jobs.

The Bottom Line

The through-line of this episode is that AI did not make data teams cheaper. It made them faster, and speed shows up somewhere. It shows up in more code, more pipelines, more experiments, and a growing share of work that gets built and then thrown away – all of it consuming compute that someone is still paying for. Sundar is one of the few people saying this plainly, probably because he never left the IC track and can still see the load arriving.

The rest is discipline he has been practising for two decades: know what your data is feeding before you design it, standardise instead of collecting four ways to do the same thing, tie the team to what the business earns rather than what it ships, and test an AI tool the way you would test a system – fixed inputs, expected outputs, a cost ceiling, and observability on whether it churns. None of that is new. That is rather the point.

By Dana Bilder Morozova
Marketing & GTM strategist with a passion for brand positioning, creative formats, and bold product storytelling. At Yuki, I build narratives not around fluff, but around real people – the engineers, data experts, and decision-makers who make the product matter. Drawn to clarity, driven by curiosity.

Table of Contents

Free cost analysis

Take 5 minutes to learn how much money you can save on your Snowflake account.

By clicking Submit you’re confirming that you agree with our Terms and Conditions.

Follow us on LinkedIn

Related posts

Free cost analysis

Take 5 minutes to learn how much money you can save on your Snowflake account.

By clicking Submit you’re confirming that you agree with our Terms and Conditions.

Skip to content