March 25, 2026: Deccan AI Lands $25M to Turn india Into the World’s AI Refinery

On March 25, Deccan AI closed a $25 million Series A round. A91 Partners led the deal. Susquehanna International Group and Prosus ventures also invested. The startup is 18 months old. It is based in San Francisco and runs its main office in Hyderabad. The new money lets Deccan grow its India contributor network. The network already has more than one million users. These users are students, PhDs, and experts. They create, label, and test data. This data helps make large language models safe for customers.

The raise is the biggest this year for post‑training services. Post‑training is work that starts after a foundation model is built. OpenAI, Anthropic, Google DeepMind, and Snowflake outsource parts of this work. Deccan has become a preferred vendor for all of them. It runs many projects at once. Projects include coding tests and reinforcement‑learning environments. These environments teach models to call external APIs.

Inside the Factory: 1 Million Indians Quietly Teaching AI to Behave

Deccan’s edge comes from scale and specialization. About 125 full‑time staff design workflows. Daily annotation is done by 5,000 to 10,000 active contributors each month. These contributors come from a registry of more than one million Indians. Founder Rukesh Reddy said that about 10 percent of the pool has advanced degrees. The share rises to over 30 percent on projects that need medical, legal, or advanced math skills.

Contributors earn per task. Rates rise for domain expertise. A chemistry PhD can earn about $25 an hour. An undergraduate can earn $6 to $8 per task. That rate is still double India’s median urban internship wage. Reddy says the network changes quickly by design. He wants people who just finished a semester of convex optimization. He does not want people who memorized the textbook three years ago.

Quality checks are strict. Every submission is spot‑checked by a second rater. If agreement falls below 92 percent on a coding snippet, or below 96 percent on a medical Q&A, the batch is re‑worked. Workers do not get paid for re‑work. Reddy says the policy keeps client rejection rates below 0.3 percent. Google DeepMind confirmed this figure for its internal benchmark sets.

From Text to Robots: World Models Create a New Data Hunger

Until last year, post‑training mainly cleaned conversational text. Now, world models are the new frontier. World models ingest video, lidar, and haptic data. They let robots or drones predict physical outcomes. This shift increases data volume. A single 30‑second warehouse clip creates 1,800 annotated frames. Each frame needs depth, segmentation, and affordance tags.

Reddy says the network changes quickly by design.

You may also like

Entrepreneurship & Business

Failing forward reshapes entrepreneurial growth pathways

Forecasts from the World Economic Forum suggest that firms embedding iterative failure analysis will outpace peers in revenue growth by up to 12% annually.

Read More →

Deccan’s solution is Helix. Helix is an evaluation suite. It creates 3‑D reinforcement‑learning environments in minutes. Customers upload a Blender file of a factory floor or a Unity mock‑up of a drone race. Helix then sends collision‑boundary tasks to contributors. Contributors tag occlusions, object permanence, and Newtonian violations. Snowflake used Helix in February. It retrained an in‑house manipulation model. Helix cut simulation error rates by 18 percent in two weeks. This came from an internal slide shown to TechCrunch.

The company now pilots live feedback loops. Instead of annotating static logs, contributors watch robot video streams in near‑real time. They flag anomalies. Early trials with an unnamed logistics partner showed a 27 percent drop in pick‑and‑place failures on a conveyor belt. The workflow costs $1.40 per minute of video. Reddy says margins are not yet good.

Red‑Hot Market: Scale AI, Mercor, and a $2B Scramble for Data

Demand for post‑training labor has made the sector a land‑grab. Scale AI is worth $7.3 billion. It has the biggest U.S. footprint. It relies on subcontractors in Kenya and the Philippines. Mercor is Deccan’s closest rival. Mercor recruits globally through a talent marketplace. It claims faster matching. Mercor raised $8 million in January. Surge AI, Turing, and Labelbox together raised more than $500 million since 2023. Reddy says none of them match Deccan’s depth in India technical talent.

The differentiator is error tolerance. Reddy says frontier labs will stop a 50‑billion‑parameter run if post‑training noise lowers accuracy below 99.9 percent. That means every data row must be courtroom‑grade. Scale AI disputes the claim. It cites its Human Preference Evaluator benchmark, which tops 99.5 percent. Scale AI declined to share contract audits.

That means every data row must be courtroom‑grade.

Investors bet the market will keep growing. A91 Partners partner Kaushik Anand said his firm modeled a $2 billion annual spend on post‑training by 2028. Robotics and agentic systems drive that spend. They need continuous re‑calibration. Anand said the sector is rare. In that sector, supply is the bottleneck, not demand.

You may also like

AI & Technology

AMD commits up to $5 billion to Anthropic

AMD's commitment of up to $5 billion to Anthropic marks a significant investment in AI infrastructure, promising to enhance AI capabilities and create numerous job…

Read More →

Cash Burn Versus Quality: The Tension Inside Zero-Error SLAs

The $25 million round gives Deccan two years of runway if headcount stays flat. That is unlikely. Reddy plans to double Hyderabad staff to 250 by December. He will raise contributor payments 40 percent. This will help keep top PhDs from rivals. Some early investors worry about burn rate. A Prosus Ventures memo in February warned that per‑task costs could rise 25 percent before Helix automation helps.

Geopolitical risk also exists. Indian data‑labeling companies like Wipro and TCS have faced client audits over worker conditions. Deccan avoids the issue by calling contributors gig workers, not employees. This model skirts India’s 2020 labor‑code amendments. It could be re‑litigated if unions push for health‑care coverage. Reddy says he watches the Karnataka gig‑workers case closely. He set aside $1.2 million for compliance retrofits.

The Long‑Term View Deccan’s edge depends on how fast world‑model workloads grow faster than India’s graduate pipeline.

Customers keep coming. Since January, Deccan signed three new U.S. robotics startups. It also signed one European telecom testing LLM‑based customer support. Reddy will not disclose revenue. He said March receipts were triple November’s. That puts annual recurring revenue on pace to cross $35 million by October. If sustained, Deccan could get a Series B at a valuation above $250 million within a year.

The Long‑Term View

Deccan’s edge depends on how fast world‑model workloads grow faster than India’s graduate pipeline. IITs and NITs produce only 35,000 engineers a year. Reddy estimates a 150,000‑person shortfall in specialized raters by 2029. The company pilots micro‑courses that give certificates in physical‑scene reasoning. It hopes to turn final‑year undergrads into qualified annotators in six weeks. If up‑skilling fails, the $25 million cushion lets Deccan recruit in Vietnam or Eastern Europe. Those moves would erode the cost advantage that underpins its model.

For now, Deccan’s cash lets it bid aggressively on multi‑year robotics contracts. Competitors shy away from those contracts. The next 18 months will decide if that scale builds durable market power. Or if quality becomes a commodity anyone can rent by the hour.

You may also like

AI & Technology

AI and Hybrid Work Redefine Career Capital in 2026

LinkedIn’s 2026 labor market report shows a surge in AI‑related postings, and IMD flags flexible work as a core trend.

Read More →