What does Markov (YC S26) actually sell?

Markov builds datasets and environments for training computer-use AI agents, capturing human screen recordings of software tasks like Chrome and Blender use. Its flagship computer-use-large dataset, released March 16, 2026, contains 48,478 recordings totaling about 12,300 hours under a CC-BY-4.0 open license.

Is Markov profitable or well funded?

No financials are public. Markov is a YC S26 batch company listed with funding stage unknown and about 2 employees as of July 8, 2026, per Seedtable. No revenue, customer count, or investor names beyond YC's standard batch investment have been disclosed.

Could OpenAI or Anthropic make Markov obsolete?

Possibly. OpenAI's Computer-Using Agent (Operator) and Anthropic's macOS Claude Computer Use agent, expanded around March 25, 2026, both self-generate GUI interaction data, meaning Markov's main buyers are also building the substitute for its human-captured trajectories.

Who is Markov's biggest competitive threat?

Well-capitalized data-supply incumbents like Mercor (valued at $10B after an October 2025 $350M round, later acquiring RL-environments startup Deeptune in July 2026), Scale AI (backed by a Meta stake around $15B, June 2025), and Surge AI (~$1B reported revenue) can replicate Markov's approach at far greater scale.

Does Markov have a defensible moat?

Its data is released under CC-BY-4.0, meaning anyone can copy it freely; a competing open dataset from Paradigm Shift AI already covers similar ground with 3,167 tasks. No exclusive lab partnerships or proprietary lock-in have been disclosed as of mid-2026.

Has anyone tried Markov's approach before and failed?

Yes. YC-backed Pig.dev attempted a similar Windows-desktop computer-use agent but pivoted away from the idea in mid-2025, according to TechCrunch reporting from July 18, 2025, signaling real technical difficulty in the computer-use category.

YC Summer 2026 · United States

Markov

Markov builds open-source screen-recording datasets to train computer-use AI agents.

Little protection left

Assessed 27 July 2026 · Evidence as of 27 July 2026

Not much stands between this business and its pressures.

On the recordSince 27 July 2026 · Checkable from H2 2026 · Every call on the record Checkpoint: YC S26 Demo Day funding announcement

The call

Markov gives away its main product under an open license while the companies buying it are learning to make their own.

Markov collects videos of people using computers, then sells this data to train AI agents to do computer tasks automatically. The problem is that richer companies like Scale AI and Mercor can copy this idea without losing money on their main business, and the AI labs that buy from Markov (like OpenAI and Anthropic) are also building their own agent systems that learn from synthetic data instead of human videos, which shrinks Markov's potential customers. The company open-sourced its technology under a free license, which means anyone can use it. The one thing that could change this would be if Markov moved into a specialized area (like software for regulated industries) or made an exclusive deal with a major AI lab that locked out competitors.

Why

AI pressure: OpenAI Operator/CUA and Anthropic's macOS Claude computer-use agent. The labs buying human-captured trajectories now generate their own via self-play, shrinking demand for Markov's data. Could AI do the core job? Yes, likely.

Bigger players: Mercor, Scale AI, Surge AI. Well-funded data suppliers already sell competing trajectory and labeling data and can copy Markov's open datasets freely. Could a giant copy it? Yes, easily.

What protects it. Being early to open-source a large OSWorld-style dataset gives visibility, but CC-BY-4.0 licensing means nothing stops copying.

What would change this call

Mercor, Scale AI, or Surge AI launches or acquires a computer-use trajectory data product. Shows well-funded incumbents are directly absorbing Markov's niche.
OpenAI or Anthropic announces a synthetic or self-play trajectory generation milestone. Shows buyers reducing need for human-captured data like Markov's.
Markov announces an exclusive licensed data deal or pivots into a regulated vertical. Would indicate Markov is building defensibility instead of staying fully open.

Key numbers

48,478 screen recordings, ~12,300 hours in dataset (2026-03-16)
737 downloads/month on Hugging Face (2026-04-26)
About 2 employees per directory listing (2026-07-08)

Named risks

Open CC-BY-4.0 licensing lets well-funded incumbents copy datasets freely with no legal barrier. Severity: high
Frontier labs (OpenAI, Anthropic) are simultaneously Markov's main buyers and its substitute via self-generated trajectories. Severity: high
Only about 2 employees and no disclosed funding amount as of July 2026 limits ability to scale or compete. Severity: high
Precedent of YC peer Pig.dev abandoning a similar computer-use approach in 2025 signals category-level technical difficulty. Severity: medium
No named enterprise or lab customers disclosed, leaving commercial traction unverified. Severity: medium

Rivals, as read in July 2026

Rival	Direction	Evidence
Mercor	losing to	Mercor at $10B valuation (Oct 2025) acquired RL-environments startup Deeptune, Jul 2026, expanding into Markov's niche
Scale AI	losing to	Meta-backed at ~$15B (Jun 2025), can replicate Markov's data supply at far greater scale
Surge AI	losing to	Reported ~$1B revenue, self-funded (Mar 2026), dominant in expert human-data supply
Paradigm Shift AI	contesting	Publishes competing open dataset computer-use-data-psai with 3,167 tasks, directly overlapping niche
OpenAI (Operator/CUA)	losing to	Buyer and competitor; ships self-generated computer-use trajectories reducing demand for purchased data
Anthropic (Claude Computer Use)	losing to	Expanded to macOS desktop agent, Mar 25 2026, self-generating trajectory data as both buyer and rival

Questions people ask

What does Markov (YC S26) actually sell?: Markov builds datasets and environments for training computer-use AI agents, capturing human screen recordings of software tasks like Chrome and Blender use. Its flagship computer-use-large dataset, released March 16, 2026, contains 48,478 recordings totaling about 12,300 hours under a CC-BY-4.0 open license.
Is Markov profitable or well funded?: No financials are public. Markov is a YC S26 batch company listed with funding stage unknown and about 2 employees as of July 8, 2026, per Seedtable. No revenue, customer count, or investor names beyond YC's standard batch investment have been disclosed.
Could OpenAI or Anthropic make Markov obsolete?: Possibly. OpenAI's Computer-Using Agent (Operator) and Anthropic's macOS Claude Computer Use agent, expanded around March 25, 2026, both self-generate GUI interaction data, meaning Markov's main buyers are also building the substitute for its human-captured trajectories.
Who is Markov's biggest competitive threat?: Well-capitalized data-supply incumbents like Mercor (valued at $10B after an October 2025 $350M round, later acquiring RL-environments startup Deeptune in July 2026), Scale AI (backed by a Meta stake around $15B, June 2025), and Surge AI (~$1B reported revenue) can replicate Markov's approach at far greater scale.
Does Markov have a defensible moat?: Its data is released under CC-BY-4.0, meaning anyone can copy it freely; a competing open dataset from Paradigm Shift AI already covers similar ground with 3,167 tasks. No exclusive lab partnerships or proprietary lock-in have been disclosed as of mid-2026.
Has anyone tried Markov's approach before and failed?: Yes. YC-backed Pig.dev attempted a similar Windows-desktop computer-use agent but pivoted away from the idea in mid-2025, according to TechCrunch reporting from July 18, 2025, signaling real technical difficulty in the computer-use category.

Timeline

2026-07-26Mercor reportedly acquires Deeptune, an RL/simulation environments startup
2026-07-08Seedtable lists Markov at about 2 employees, funding stage unknown
2026-04-26Hugging Face shows computer-use-large dataset at 856 GB, 737 monthly downloads
2026-03-28Hugging Face lists markov-ai/computer-use dataset (OSWorld-based) at 747 monthly downloads
2026-03-25Anthropic's Claude Computer Use expands to a production macOS desktop agent
2026-03-16Markov releases computer-use-large: 48,478 screen recordings, ~12,300 hours, CC-BY-4.0
2025-10-01Mercor closes $350M Series C at $10B valuation
2025-07-18YC peer Pig.dev pivots away from Windows computer-use agent approach
2025-06-01Meta takes strategic stake in Scale AI valuing it around $15B

Revision history

27 July 2026

First assessed. Little protection left. The text above is fixed as of this date and is not edited afterwards.

In the same band

Cerenovus Belvedir Hoplite