Miju Labs

← All companies

Handshake AI

A decade-old campus network that already owned the PhDs its competitors were recruiting — zero to $1B gross in fifteen months, on a valuation nobody has re-marked.

medium confidence5 minupdated 2026-08-29ai labs · data · supply
Vertical
Expert data for frontier labs
Founded
Parent 2013; Handshake AI launched January 2025
Headquarters
San Francisco
Raised
$435M (last priced round: $200M Series F, January 2022, Coatue + Valiant Peregrine)
Last valuation
$3.5B (2025 mark); $3.3B at the Series F — the mark predates all the AI revenue
Revenue
~$1.0B GROSS annualised for the AI arm (April 2026), ~$300M net after contractor payments; ~$1.10B gross group-wide, ~$450M net
Status
Active; ~100 US roles cut in the legacy recruiting business; contractors publicly allege withheld pay
Who runs it · 8 people in the index

Latest (Sep 2026): Handshake AI went from launch (Jan 2025) to about $1B gross annualised by April 2026 (Sacra: ~$1.10B group gross, ~$450M net), working with eight frontier labs (TIME, Aug 2026). It acqui-hired data-quality startup Cleanlab on 28 Jan 2026 (nine staff into its research org); contractors on OpenAI projects reported suspensions with pay withheld in Dec 2025-Jan 2026, and in Aug 2026 it began paying up to $30K for professionals' own work documents. The last priced round is still the $3.3B Series F (Jan 2022; ~$3.5B 2025 mark), with no new round found as of 2026-09-19. Note: X/LinkedIn handles are the parent Handshake's; no separate Handshake AI accounts were verified.

What Handshake AI is saying
Running long-horizon, agentic data projects is the best preparation for building robust enterprise AI apps bc it forces you to build thoughtful expert <> agent interactions and optimize LLM usage. Data companies have a massive advantage in serving the enterprise. We now spend more on LLMs than on headcount. That does not make human operators, and especially experts, less valuable. It actually makes them more productive and more valuable. Purely synthetic data without experts embedded in the authoring process doesn't work, and data companies that don't dig into their own data will ship slop. They won't produce the nuance, realism, or complexity the task demands, and the labs struggle to catch all these issues because they rely on agentic graders that have the same blind spots. But the token spend is significant, and it goes to three places: (1) hybrid synthetic + expert environment / task authoring, (2) rollouts to validate task complexity, and (3) QC on the data. You cannot produce a high-quality long-horizon agentic task without LLMs across all three buckets. We also use agents to automate internal processes, but that is a small fraction of the bill. Any data company telling you the bulk of its LLM spend is internal automation is lying. Producing this data at high quality and reasonable cost comes down to two problems. (1) the expert-agent interaction paradigm - how the expert creates, reviews, interprets, and audits the LLM's outputs (synthetic input files, prompts, rubrics, and so on). Solving this means knowing precisely where agents fail or underperform, designing the interface through which experts validate those outputs, and building the tools that let experts steer the agent going forward. (2) optimizing every QC check for the right tradeoff between cost, quality, and latency. That means constructing complex and diverse eval sets, configuring an agentic system with the right models (often chained prompts, guardrail LLM calls, and deterministic checks), running the eval, then iterating on the harness, prompts, chaining, and model choice until the mix is optimal. Both are now core skills of our SPLs. Enterprise deployments follow the exact same pattern. Understand the workflow and where models fail or underperform, design the user-agent interaction to mitigate those failures, build evals that test the agentic system across the workflow, and optimize the system against those evals. Running data pipelines forces us through this loop hundreds of times a month across a wide range of project types, and we have to build product and infrastructure to do it fast and repeatably. You cannot build an optimal enterprise deployment without these muscles, and the largest data companies have already done it hundreds of times.

Handshake is the fastest zero-to-$1B in the expert data vertical and the only large name in it that looks cheap. Both facts have the same cause: it already owned the supply, and nobody has re-priced it since.

Why it won so fast

Handshake spent a decade signing up universities. By the time it launched Handshake AI in January 2025 — as "a startup inside a startup," with separate teams and offices (Sacra; Lenny's Newsletter) — it had 17–20M students, 1,600+ institutions, roughly 500,000 PhDs and ~3M advanced-degree holders in network.

Scale AI and Mercor had been recruiting PhD annotators off Handshake. Handshake disintermediated its own customers (Aakash Gupta on X) [WEAK — commentary rather than reporting, though consistent with the timeline]. This is the cleanest Which side you build first solution in the atlas: the expensive side of the marketplace was already assembled and already paid for.

Growth from there: $5–10M gross ARR at launch → $550M in January 2026 → ~$1B gross annualised by April 2026, with net revenue after contractor payments around $300M (Dealroom). Lord's own numbers are $50M in the first four months and $100M in eight.

The mispricing

Contractor costs run 60–70% of gross — the same take structure as Mercor. So the arithmetic is:

BasisFigureMultiple on the $3.5B mark
Group gross annualised (Apr 2026)~$1.10B~3.2x
Group net after contractor payouts~$450M~7.8x
AI arm alone, gross~$1.0B—
AI arm alone, net~$300M—

At 3.2x gross, Handshake is the cheapest large name in the sector — cheaper than Innodata's 6.2x on audited public numbers. Even restated to net at ~7.8x it is well below Mercor at roughly 29–38x net. The gap is not a judgement about business quality; it is that the $3.5B mark predates the AI revenue entirely and there has been no priced round since January 2022. See What the public market pays for labour and GMV is not revenue.

What the arithmetic also shows is a ~30%-margin labour business bolted onto an ~80%-gross-margin SaaS business. Blending them produces a group gross margin that means very little, which is exactly the accounting problem GMV is not revenue exists to name.

Pay, and how far "expert data" has drifted

Lord states contractors average $100–125/hour, with a range from $75/hr (software engineers, improv actors) to $175/hr (investment bankers) to $300+/hr (MDs and PhDs) (BI via AOL). Sacra corroborates $100–125 for maths, physics and CS. Entry for supply runs through the "MOVE" (Model Validation Expert) Fellowship.

Handshake has also hired improv actors at up to $74/hour to record unscripted scenes for an unnamed leading lab (AOL/BI). That is a useful marker of how far this category has moved from annotation.

Pay withholding

This is the sharpest supply-side problem documented anywhere in the vertical.

Contractors on OpenAI projects report accounts suspended between late December 2025 and January 2026 with pay withheld. Business Insider interviewed five contractors; four were unpaid. Dozens more appear on Reddit, there are at least two lawsuits, and one court ruled a contractor was owed $6,475. Handshake's stated grounds are credential discrepancies, task times 3–4x benchmark, and work performed outside the US. Its support line: "This decision is final. There is no appeal process, and any work associated with this violation is not eligible for payment" (AOL/BI).

A community tracker documents a May 2026 payment crisis on "Project HH" — workers receiving 20–50% of earned pay, effective rates falling to $8–25/hr — plus a 7.5% account-ban rate and 55% negative sentiment across 1,654 Reddit posts between June and August 2026 (Breaking Even) [WEAK — community data analysis, not press].

Note that Handshake's stated grounds are also a Who is actually on the other end story: the company is asserting that a meaningful share of its credentialled supply is faking credentials or using LLMs to complete work. Both readings can be true, and both are expensive.

Restructuring

Around 100 US roles cut, 15% of a 650-person staff, concentrated in the legacy recruiting business (Upstarts). Cleanlab acquired in January 2026 for data quality.

Gap in the record

The legacy college-recruiting business is given as ~$150M gross ARR in one source and $190M (2024) in another; the AI team's growth is reported as 15 → 150 people in one account and 3 → 150 in another. Neither discrepancy was reconciled.

Worse, neither figure survives the group arithmetic: $1.10B of group gross less the AI arm's ~$1.0B leaves about $100M for the legacy line, not $150M and not $190M. The group net figure of ~$450M is built by subtraction from these same numbers, so it inherits the error. Treat the 3.2x and 7.8x multiples as accurate to within about a tenth, not to the decimal.

Gap in the record

Handshake does not disclose a take rate or gross margin. The ~30% figure here is derived from the stated 60–70% contractor share, not disclosed.