What Data Do Frontier Labs Buy? A Deep Dive into How Handshake AI Reached $1 Billion in One Year

Handshake AI reached a $1B valuation in one year by supplying expert-level data, revealing the trend toward data supply chain specialization.
Handshake transformed from a campus recruiting platform into a data supplier for frontier AI labs, reaching a $1 billion valuation within a year. Its core business is organizing experts in finance, biology, alignment, and other domains to produce structured, verifiable, high-quality training data. Financial data focuses on analysts' reasoning processes rather than public reports, biology data requires PhD-level judgment, and alignment data emphasizes ethical boundaries and consistent judgment criteria. The article notes that general data has been commoditized while expert-level data remains a seller's market; competition among frontier labs is extending from compute to the data supply chain, with the ability to organize expert networks, design high-quality tasks, and validate data effectiveness becoming key.

As the large model race enters its second half, an overlooked question has surfaced: what data are frontier labs actually buying? Handshake AI went from zero to a $1 billion valuation in one year, precisely by supplying highly specialized data to frontier labs. This article is based on the first in-depth interview with Garrett Lord and the team at the data lab, breaking down specific cases in three major domains—finance, biology, and alignment—to reveal the truth behind the data supply chain.
From Recruiting Platform to Data Lab: Handshake's Unexpected Pivot
Handshake was originally known as a campus recruiting platform connecting college students with employers. But the explosive growth in demand for high-quality, domain-specialized data driven by generative AI led the company to discover a new frontier. According to Garrett Lord, Handshake reached a $1 billion valuation within a year, with its core business becoming the supply of training data to frontier AI labs.
This transformation was no accident. General internet data has been largely "consumed" by large models, and the remaining incremental value lies in expert-level, structured, verifiable domain data. Leveraging its campus network and academic resources, Handshake can reach professionals in finance, biology, medicine, and other fields, building a moat in data collection.
"We're not selling data—we're selling the judgment of domain experts." — Garrett Lord's positioning in the interview.
Financial Data: Beyond Financial Reports, It's Decision Logic
In finance, frontier labs' needs go far beyond public financial reports and news. The cases Handshake provides show that labs need analysts' reasoning processes—how to derive stock-level judgments from macro indicators, how to identify anomaly signals in financial reports.
Specifically, Handshake organizes finance professionals to complete structured tasks: given a company's financial data, write an investment memo including valuation assumptions, risk points, and catalysts. This data is used to train models to mimic professional analysts' chains of thought, rather than simply memorizing numbers.
The value of this data lies in its scarcity. High-quality records of the full investment analysis process are rare on the public internet, and Handshake incentivizes professionals to produce them through paid tasks. According to the team, the unit price of financial data is significantly higher than general text because supply is limited and verification costs are high.
Chain of Thought (CoT) refers to the technique where a model explicitly generates a series of intermediate reasoning steps before giving a final answer. In training data, chain of thought appears as structured text in the form of "question → step-by-step reasoning → conclusion," such as the complete logic of a financial analyst deriving a stock-level judgment from macro indicators. Compared with data that only provides answers, samples containing chain of thought enable models to learn "how to think" rather than "what to remember," thereby improving generalization on complex tasks. The investment memos collected by Handshake are typical chain-of-thought data, whose value lies in making implicit expert reasoning explicit and trainable.
Biology and Alignment: Domain Barriers Determine Data Premiums
Demand in the biology domain also has distinct professional characteristics. Handshake's cases cover experimental design, literature interpretation, and clinical data analysis. For example, biomedical PhDs are asked to read a paper and design follow-up experiments, or to extract biologically meaningful conclusions from omics data. Such tasks require domain experts and cannot be completed by ordinary annotators.
Alignment data is another rapidly growing category. Handshake collaborates with AI safety researchers to build datasets used to train models to refuse harmful requests and follow complex instructions. Unlike biology data, alignment data places more emphasis on judgment criteria and edge cases, requiring annotators to understand the ethical and safety principles behind model behavior.
The team emphasized in the interview that data from different domains cannot be simply mixed. Financial data requires precision and traceability, biology data requires scientific rigor, and alignment data requires consistency in value judgments. Domain knowledge is the first filter for data quality.
Alignment in the AI safety field refers to the technical goal of making model behavior consistent with human intent, values, and safety norms. Training alignment data typically includes: teaching models to identify and refuse harmful requests, follow complex instructions in ambiguous situations, and understand ethical edge cases. Unlike general capability data, alignment data emphasizes annotators' consistency in judging "what constitutes acceptable behavior," thus requiring experts with safety and ethics backgrounds. Handshake's collaboration with AI safety researchers to build such datasets is precisely to provide labs with high-quality value judgment samples rather than simple question-answer pairs.
Lessons from the Data Supply Chain: Specialization Is the Next Battleground
Handshake's case reveals a trend: competition among frontier labs is extending from compute and model architecture to the data supply chain. General data has been commoditized, while expert-level data remains a seller's market.
For data suppliers, the key capability is no longer crawling scale but organizing expert networks, designing high-quality tasks, and validating data effectiveness. Handshake leveraged its existing campus recruiting resources to quickly build an expert pool spanning multiple domains—this is the underlying logic behind reaching $1 billion in one year.
For AI practitioners, this means a shift in focus: the ceiling of model capability increasingly depends on whether you can obtain data that others cannot. Professional data in vertical domains such as finance, biology, law, and medicine is becoming a scarce asset.
Data supply chain refers to the complete chain from data requirement definition, collection, annotation, and validation to delivery for model training use. As general internet data is gradually exhausted, frontier labs increasingly rely on specialized suppliers to organize domain experts to produce data on demand. Key links in this chain include: designing structured tasks that elicit expert reasoning, establishing multi-round validation mechanisms to ensure data accuracy and traceability, and managing incentives and quality control of expert networks. Handshake's competitiveness lies precisely in transforming its existing campus recruiting expert resources into a scalable data supply chain, rather than simply expanding crawling scale.
Conclusion: Competition in the Data Deep Waters Has Just Begun
The rise of Handshake AI is not an isolated case but a signal of data supply specialization. As model parameters and compute gradually converge, data quality—especially the depth of domain expert participation—will become the core of differentiation. The practice of Garrett Lord's team shows that only labs willing to pay for expert judgment can train truly professional models. This competition in the data deep waters has only just begun.
One-line Summary
Handshake AI reached a $1 billion valuation in one year by supplying expert-level data, revealing the trend toward data supply chain specialization.
Paragraph Summary
Handshake transformed from a campus recruiting platform into a data supplier for frontier AI labs, reaching a $1 billion valuation within a year. Its core business is organizing experts in finance, biology, alignment, and other domains to produce structured, verifiable, high-quality training data. Financial data focuses on analysts' reasoning processes rather than public financial reports, biology data requires PhD-level professional judgment, and alignment data emphasizes ethical boundaries and consistency in judgment criteria. The article points out that general data has been commoditized, while expert-level data remains a seller's market; competition among frontier labs is extending from compute to the data supply chain, and the ability to organize expert networks, design high-quality tasks, and validate data effectiveness has become key.
Related articles

Gemini 4 Hydrogen In-Depth Review: Real Value Beyond Cost-Effectiveness
Based on Artificial Analysis selection data and hands-on feedback, this review unpacks the Gemini 4 Hydrogen value paradox: high scores don't equal high value, performance splits across automation workflows and coding tasks, pricing and Google subscription bundles are separate ledgers, and SynthID output labeling could be a dealbreaker.

Gemini 4 Hydrogen Five Extreme Tests: 96.7% Accuracy and One-Prompt Website Building
A YouTube creator runs five extreme tests on Gemini 4 Hydrogen: single-prompt game generation, 3D solar system simulation, 30-minute video timestamp localization, converting a 287-page report into an interactive dashboard, and 96.7% accuracy in hallucination resistance.

Gemini 4 Hydrogen Released: A 2.5x Faster Everyday AI Model
Google releases Gemini 4 Hydrogen, claiming 2.5x the speed of Gemini 3.8 Flash, with four-tier thinking control, a million-token context, and Orchestrator multi-model collaboration. API pricing is $1.50 per million input tokens.