The PodKnowledge|← All Episodes
Based on Lenny's Podcast data
Lenny's Knowledge Sketch

Scale AI's $14B Meta Deal
& What's Actually Next

Jason Droege
CEO, Scale AI
OCT 9 2025
The Company

Scale AI: The Data Layer
Under Every Frontier Model

TRAINING DATA MARKET
"Every major AI model you use was trained with Scale's help. We're the infrastructure that makes AI reliable."
  • Scale provides data labeling, RLHF, and AI evaluation services
  • The $14B Meta deal: landmark enterprise AI contract
  • Training data quality determines model quality more than architecture
  • The government bet: defense AI is Scale's fastest-growing vertical
Framework

The Scale AI Model

RAW DATASCALE LABELTRAINED MODEL
$14B
Meta multi-year contract
$13.8B
Scale valuation
10+
frontier model customers
  • Scale's core service: human-in-the-loop data annotation at massive scale
  • RLHF platform: the feedback loop that aligns models to human values
  • Eval-as-a-service: ongoing model evaluation for enterprise customers
  • Government/defense: separate unit building AI for US national security
Jason's thesisScale is not a services company. It's an AI infrastructure company that happens to need humans to deliver the infrastructure.
Enterprise AI at Scale

What Scale Learned

  • Insight 1: Data quality compounds, better training data produces dramatically better models
  • Insight 2: Enterprise AI needs continuous evaluation, not just pre-deployment testing
  • Insight 3: Government AI is 5 years behind commercial AI but catching up fast
  • Insight 4: The data moat is real, it's harder to replicate than the model architecture
The data moat

Scale's most defensible asset is not its technology, it's its relationships with expert annotators in specialized domains.

The government opportunity

US defense AI is a multi-decade market. Scale is positioning as the trusted data partner for national security AI.

Playbook

Think About AI Data

  • Your training data IS your model, invest in its quality before model selection
  • Evals are not a one-time task, build continuous evaluation into your AI roadmap
  • Domain-specific data is worth 10× general data for domain-specific tasks
  • The data flywheel: more users → more data → better model → more users
The Scale predictionJason believes the AI training data market will be larger than the cloud market within 5 years. The data that trains AI is the new oil.
Contrarian

AI Data Myths

✗More data is always betterINSTEAD →✓ More high-quality data is always better. Scale's value is quality, not just volume.
✗Open source data eliminates the need for ScaleINSTEAD →✓ Open source data produces open source quality. Frontier models require frontier data quality.
✗AI will soon generate its own training dataINSTEAD →✓ AI-generated training data causes model collapse. Human feedback remains essential for alignment.
✗Data labeling is a commodityINSTEAD →✓ Expert data labeling is never a commodity. Medical, legal, and defense labeling require rare expertise.
Based on Jason Droege's episode on Lenny's Podcast. All ideas on this page are from the episode.Watch on YouTubeFollow @jdroege on X
GO DEEPER IN THE EPISODE
06:01 Jason’s early career and lessons learned10:27 The current state of Scale AI12:37 The shift to expert data labeling17:02 Challenges and strategies in finding experts18:48 Reinforcement learning and AI environments28:18 The future of AI and human involvement31:21 The role of evals35:25 What AI models will look like in the next few years41:43 Building Uber Eats and understanding customer needs48:19 The importance of independent thinking50:45 Setting high standards for new businesses53:03 Exploring and selecting business ideas57:07 The McDonald’s story01:00:13 The role of gross margins in business feasibility01:04:49 Why Jason says, “Not losing is a precursor to winning”01:09:12 Hiring and building teams01:12:11 AI corner01:14:47 Lightning round and final thoughts
Every sketch here began as a transcript.

Send us yours and we’ll draw it. $1 a transcript.

Sketch mine →
𝕏︎ X / Twitterin LinkedIn📸 Instagram🔗 Copy link
0:00