Oct 9, 2025 · 1h 24m · lennys-podcast
Scale AI CEO on Meta’s $14B deal, scaling Uber Eats to $80B, & what frontier labs are building next
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Scale AI CEO Jason Droege joins Lenny Rachitsky to discuss the governance and future of frontier AI data infrastructure, while sharing foundational lessons from scaling Uber Eats to an $80 billion enterprise. He provides an insider look into expert model evaluations, the realities of enterprise AI adoption, and durable business strategy.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Lenny holds 25.8% of the talking time here. How this is scored →
Lenny as informed peer 3.9 Guest teaching 5.1 Guest disagreement 1.4 Lenny pushing back 1.5
speaking balance: gold is Lenny, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 13:27 Calling competitor narratives bogus
Jason forcefully pushes back against competitor claims regarding data labeling, calling the market positioning completely bogus and divorced from reality.
Hardest push from Lenny ▶ 28:42 Questioning long-term human data dependencyLenny challenges Jason directly on whether Scale's thesis of perpetual human data demand is simply a narrative aligned with its own commercial growth.
Biggest teaching moment ▶ 26:30 Reframing enterprise data valueJason educates Lenny on why ingesting massive volumes of raw enterprise data fails, explaining that the true bottleneck is capturing and digitizing domain-specific human judgment.
Lenny holds their own ▶ 1:03:28 Lenny analyzes Costco's margin structureLenny demonstrates strong business expertise by citing an analysis of Costco, explaining how negative product gross margins function when offset by high-margin subscription membership revenue.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Lenny as informed peer | Guest teaching | Guest disagreement | Lenny pushing back | Why |
|---|---|---|---|---|---|---|
| Early Entrepreneurship Lessons from Scour | 3 | 5 | 1 | 1 | Lenny prompts Jason to share formative lessons from co-founding Scour with Travis Kalanick. Jason explains the brutal reality of startup negotiation and how surviving a massive lawsuit taught him that standard business rules are often made up. | |
| Scale AI Governance and the Meta Investment | 5 | 4 | 1 | 2 | Lenny clarifies the details surrounding Meta's massive investment in Scale and Alex Wang's new role. Jason lays out the governance structure, confirming Scale remains independent with no preferential data access for Meta. | |
| Evolution of Data Labeling and Expert Networks | 4 | 6 | 4 | 2 | Lenny asks about competitor narratives claiming Scale is stuck in low-cost labeling while labs need experts. Jason bluntly dismisses these claims as bogus marketing, presenting stats showing 80% of their network holds bachelor degrees or above. | |
| Reinforcement Learning Environments and Agentic Training | 3 | 6 | 1 | 1 | Lenny asks about reinforcement learning dominating future economic activity. Jason breaks down the technical reality of RL sandboxes, explaining the need for generalizable agentic data across complex software environments like Salesforce. | |
| Digitizing Enterprise Judgment and Healthcare AI | 4 | 7 | 1 | 1 | Jason walks through a real-world healthcare case study where AI detects subtle patient allergies across hundreds of pages. He explains to Lenny why massive enterprise petabyte ingestion is mostly useless without digitizing specific human judgment. | |
| Human Data Durability and Frontier Model Trajectories | 5 | 6 | 2 | 5 | Lenny challenges Jason on whether Scale's belief in perpetual human data demand is merely self-serving financial bias. Jason acknowledges the incentive but defends the durability of human judgment and explains why evals define what good looks like. | |
| Enterprise AI Reality vs. Hype in Production | 4 | 5 | 2 | 2 | Lenny cites recent studies showing enterprise AI pilot failures and developer productivity bottlenecks. Jason contextualizes the data, explaining that productionizing AI reliably takes months of operational chiseling rather than minutes. | |
| Sponsor Message: Mercury Banking for Startups | 3 | 6 | 1 | 1 | Following a sponsor read, Jason shares how his team deconstructed restaurant unit economics by literally weighing cold cuts from suppliers. He illustrates how aligning seller incentives unlocked Uber Eats' 30% take rate model. | |
| Independent Thinking and Evaluating New Business Models | 3 | 5 | 1 | 1 | Lenny explores Jason's reputation for independent thinking when vetting business opportunities. Jason emphasizes that true alpha comes from contrarian insights that founders are willing to obsess over for a decade rather than chasing consensus. | |
| Scaling Uber Eats and the McDonald's Partnership | 4 | 5 | 2 | 1 | Jason details failed early Uber experiments like convenience vans before finding product-market fit with Eats. He shares how initially rejecting McDonald's due to brand vision accidentally created immense leverage for an exclusive global rollout. | |
| Gross Margins as a Filter for Business Viability | 6 | 4 | 1 | 2 | Jason details how gross margins serve as a primary diagnostic for business defensibility. Lenny jumps in with an insightful comparison to Costco's membership subscription model to illustrate scale-driven low-margin defensibility. | |
| Survival as a Precursor to Winning in Tech | 3 | 6 | 1 | 1 | Jason argues that survival is the mandatory prerequisite to thriving, contrasting Silicon Valley's risk-happy narrative with asymmetrical decision-making. He recounts his humbling post-dot-com experience selling used golf clubs on eBay. | |
| Hiring Principles, Team Balance, and Core Attributes | 4 | 5 | 1 | 1 | Jason outlines his three-part hiring rubric: problem solving, cross-functional humility, and leadership. He emphasizes composing executive teams as complementary organisms where familiarity outranks individual resume pedigree. | |
| AI Corner: Daily Workflows and Knowledge Synthesis | 4 | 4 | 1 | 1 | Lenny and Jason compare notes on their daily AI habits. Jason reveals he uses conversational voice mode during his commute as a technical tutor to master cutting-edge research concepts and synthesize complex company documents. | |
| Lightning Round: Recommended Books, Media, and Mottos | 4 | 3 | 1 | 1 | In the rapid-fire lightning round, Jason shares his favorite books, reflections on cognitive biases, experimenting with Google Veo 3 on old screenplays, and his guiding life motto: the end is never the end. |
Statements from this episode (26)
MPAA and RIAA sued Scour for $250 billion, settled for $1 million
“Because the MPAA and the RAA were the ones who sued us, the entertainment industry sued us or the associations that represent the entertainment industry, and then they settled it for a million dollars.”
Meta paid $14B for a 49% non-voting stake in Scale AI
“The transaction was Meta invested a little bit over fourteen billion dollars to get 49% of the company non-voting stock. Didn't take a new board seat.”
Only 15 Scale AI employees moved to Meta in the $14B deal
“Only about 15 people went over in the transaction. So scale has about 1100 employees or so now, and we have two major businesses.”
Scale AI has two business units each generating hundreds of millions
“Each of those businesses has each of them has hundreds of millions of revenue. So we kind of have two unicorns inside the company today that sustains.”
Scale AI provides data labeling services for the Department of Defense
“We have a relationship with the department of defense where we do labeling for them.”
80% of Scale AI's expert labelers hold degrees; 15% have PhDs
“80% of the people that we have in our expert network have a bachelor's degree or grader, which is very contrary to some of the positioning that's out there and some of the understanding of this industry. About 15% have a PhD that's grader.”
Scale AI has built RL environments for AI agents for over a year
“There's these things called RL environments that effectively are sandboxes for AI agents to play in to accomplish a goal so that they can learn how to accomplish that goal. We've been doing this for over a year.”
Training data value scales with its generalizability across tasks and environments
“And the more generalizable it is, the more valuable the data is. So our job is to provide the most valuable data to model builders that accomplishes the goal of making the agents as useful as possible for their end users.”
Scale AI sells model data and custom enterprise applications
“So we have two sides of our business. One, we supply data to model builders. We sell the data. And then the other is, is we actually do solutions. We do, we sell applications and services to healthcare systems, insurance systems, et cetera.”
Off-the-shelf RAG limitations force enterprises to do internal data labeling
“We're starting to see the labeling move into enterprises and into governments because you can only get so far with off the shelf plus rag, plus some fine tuning based on recorded data.”
Mission-critical enterprise accuracy remains extremely difficult for multi-agent systems
“The reality is, is that it's very hard to get mission critical use cases in agentic systems where agents are talking to agents to a level of accuracy that is necessary to accomplish a goal.”
Autonomous vehicles require less data labeling than they did historically
“Autonomous vehicles do not need as much data labeling as they did in the past.”
Jason Droege: No AI white-collar labor apocalypse in the next two years
“It might happen. The space is moving super fast, but I don't think it's going to happen. Like it's definitely not gonna happen in the next year. The idea that it happens in the next two years, I think is like very far-fetched, but nothing's impossible here.”
Scale AI's enterprise and government work primarily consists of model evaluations
“A lot of it's evals and within enterprise customers and government customers, it's mostly evals because somebody has got to establish the benchmark for like what good looks like.”
AI will force massive organizational change management within two to three years
“We're not there from the technology standpoint, but I do think in the next two to three years, if I, you know, take the bait and like have to make a guess is the technology will get to a point where it will push the change management and policy makers to say l…”
Moving AI from 70% POC to production is like adding uptime nines
“Basically what happens is the POCs Get to, like, 60 or 70% of the way there, and the human mind goes, oh, the rest is no big deal. But it's kind of like uptime in data centers where, like, every nine is, like, you know, an order of magnitude investment, you kn…”
Automating important enterprise workflows with AI takes six to twelve months
“If you have engineers who've worked with models before and they put in the time and I'm talking about like months, not like minutes, like you see on these videos to actually get legal approval, policy approval, regulatory approval change management, like an ac…”
Uber Eats launched with 30% take rates before settling at 25%
“And so we came in and we said, we're going to charge you 30% of the bill. And they were like, oh my God, is this Groupon all over again? This is way too high. Oh my gosh. And we explained the economics to them. And they were like, okay, we'll give it a try, bu…”
Incremental restaurant food delivery generates 70% to 80% gross margins
“If you were to take a restaurant location and triple demand based on the same labor, the same ingredients excuse me, the same labor, but you're just scaling ingredients. You've got a 70 to 80% incremental gross margin product.”
Uber's 10-truck mobile convenience store pilot in DC failed completely
“And we launched that in DC. We put like 10 of these trucks on the road. We put like 250 skews in them. And I mean, crickets is an understatement of how bad it was. I mean, we did, we couldn't get an order to save our lives.”
Uber Eats hit $20 billion in gross bookings within 4.5 years
“Well, we launched it in December of 2015 in Toronto, and within, like, two hours, we had done, like, 20,000 dollars for the sales. It was crazy how quickly we saw that, that it was the right idea and the unit economy for good. And then four and a half years la…”
Jason Droege rejected a McDonald's delivery partnership for months
“And so McDonald's actually approached us and they said, Hey, we'd love to do food delivery with you. And I'm, and I said, no. And they're like, hold on a second. Like we have like eighty million consumers a day. Like you don't, we don't want to do this togethe…”
Uber Eats kept mostly the same management team from $0 to $20B
“That management team for the most part, outside of some of the operations side, but like for the most part, that management team was the same management team from day one, when we had nothing to twenty billion dollars.”
Perplexity employees must ask AI before asking colleagues a question
“I did an interview with the founders of perplexity a few years ago, asking about how they work at perplexity. And the founder said that before they were ruled before they ask a question of anyone on the team, they have to ask AI first.”
AI tools that animate still photos into video will be emotionally life-changing
“I think those are going to be really like emotionally life-changing for people because you, because just like a little bit of movement in an image from a grandparent or a relative or whatever you haven't seen in a while, it really does make a big emotional imp…”