Deccan AI Series A: $25M to Scale India AI Training Network

Deccan AI series a funding closed at $25 million this week, all equity, with A91 Partners leading. Deccan AI supplies the post-training work that makes AI models production-ready. Data generation, evaluation, reinforcement learning. The kind of work frontier labs outsource because it doesn’t scale in-house.

Susquehanna International Group and Prosus Ventures joined the round. Founder Rukesh Reddy said the company grew 10x over the past year and now runs at a double-digit million-dollar revenue run rate. He wouldn’t share specifics, but 80% of revenue comes from the top five customers. That’s standard when your buyers are frontier AI labs. There aren’t many of them.

Post-Training Is Where AI Companies Actually Spend

The Deccan AI series a round valued execution over geography. Here’s why: OpenAI and Anthropic build base models in-house. But post-training work scales differently. You need expert feedback, domain-specific evaluations, reinforcement learning environments. That requires thousands of contributors who can generate high-quality training data fast.

Deccan competes with Scale AI, Surge AI, Turing, and Mercor. All chasing the same market: AI labs that need post-training data yesterday. Typical engagement timelines run days, not weeks. Labs request large volumes of domain-specific data with near-zero error tolerance. One bad data point can degrade model performance in production.

Deccan works on coding improvements, agent capabilities, and API integrations. The company’s Helix evaluation suite and operations automation platform serve enterprise customers. Active projects number in the dozens at any time across about 10 customers, including Google DeepMind and Snowflake.

Why Deccan AI Series A Backers Bet on India

Most of Deccan’s contributor network sits in India. The company employs 125 people, split between the San Francisco Bay Area headquarters and a large operations team in Hyderabad. The contributor network exceeds 1 million people. Students, domain experts, PhDs. Around 5,000 to 10,000 contributors stay active in a typical month.

About 10% of the contributor base holds advanced degrees. That share climbs higher among active contributors depending on project requirements. Pay ranges from $10 to $700 per hour. Top contributors earn up to $7,000 monthly. The sector has faced criticism over gig worker conditions and compensation, but Deccan’s rates sit at the higher end of the market.

Competitors spread operations across 100-plus countries. Deccan concentrated in India for quality control. Reddy’s thesis: easier to maintain standards when operations run in one country instead of everywhere. The company recently started sourcing U.S.-based talent for niche expertise in geospatial data and semiconductor design, but India remains the core.

That concentration drove the Deccan AI series a thesis. India supplies AI training talent and data. It doesn’t build frontier models. Those stay concentrated in the U.S. and China. India’s position in the AI value chain: supplier, not developer. Deccan leverages that position instead of fighting it.

The Unit Economics Question Nobody Answers

Post-training work is what justified the Deccan AI series a check. Revenue concentration at 80% from five customers creates obvious risk. Lose one and you’re down 16% overnight. That’s the reality when your TAM is a dozen frontier labs and a few hundred enterprises willing to pay for custom AI work.

Reddy described Deccan as a “born GenAI” company, built for high-skill post-training work from day one. Traditional data labeling firms started with computer vision tasks and scaled up. Different DNA. Deccan skipped the low-skill work entirely.

The work itself is evolving. Models now move beyond text into world models that understand physical environments. Robotics, vision systems, spatial reasoning. That means new types of training data and evaluation methods. More complexity, higher skill requirements, tighter tolerances.

Quality control remains unsolved at scale. Reddy said tolerance for errors in post-training sits close to zero. One mistake affects production performance. That makes this harder than early-stage data labeling, which can absorb some noise. Post-training requires domain expertise and accuracy simultaneously. Hard to scale, harder to price.

What $25M Buys You

Series A funding typically covers 18-24 months of runway and early scaling. Deccan needs to convert contributor growth into margin expansion while maintaining quality. The tradeoff every services business faces: scale the team or scale the margin. You rarely get both.

Customer concentration won’t improve until the frontier AI market expands or Deccan moves downmarket. Neither happens fast. Frontier labs consolidate, not proliferate. Enterprise AI adoption scales slower than SaaS. So revenue concentration stays high until the market structure changes.

The India focus creates margin advantages. Lower labor costs, concentrated operations, easier quality control. The risk: competitors source globally and find talent Deccan can’t access. Niche expertise in semiconductor design or geospatial data doesn’t concentrate in one country. Deccan already sources some U.S. talent for that reason.

Execution comes down to speed and accuracy. Can Deccan deliver high-quality domain data in days while scaling contributor volume? Most startups can’t. Quality breaks at scale or speed kills accuracy. Founders who solve both build defensible businesses. Those who don’t become commoditized labor marketplaces.