内容来源：https://techcrunch.com/2026/03/25/deccan-ai-raises-25m-as-ai-training-push-relies-on-india-based-workforce/

内容总结：

随着全球对AI模型训练与优化需求激增，专注于AI模型后期训练与评估服务的初创企业Deccan AI近日宣布完成2500万美元A轮融资。本轮融资由A91 Partners领投，Susquehanna International Group和Prosus Ventures跟投，公司估值未公开。

尽管OpenAI、Anthropic等前沿AI实验室自主开发核心模型，但为确保AI系统在实际应用中的可靠性，从数据生成、评估到强化学习等后期训练环节正日益外包。2024年10月成立的Deccan AI正是瞄准这一需求的新兴服务商之一。

公司总部位于旧金山湾区，在印度海得拉巴设有大型运营团队，目前拥有约125名正式员工，并依托超过100万贡献者网络（包括学生、领域专家和博士人才）开展业务。创始人Rukesh Reddy透露，其客户包括谷歌DeepMind和Snowflake，已签约约10家企业客户，同时并行推进数十个活跃项目。

Deccan的服务涵盖提升AI模型的编程能力、智能体功能，以及训练模型与API等外部工具交互。随着AI模型从文本向理解物理环境的"世界模型"演进，公司还提供机器人、视觉系统等领域的专业化训练服务。

Reddy指出，后期训练对数据质量要求极高，容错率"接近零"，且常需在数日内处理海量高质量数据，平衡速度与精度成为行业核心挑战。为此，Deccan将主要人才基地设在印度，通过集中化管理提升质量控制效率。平台贡献者时薪约10至700美元，顶尖人才月收入可达7000美元。

目前全球AI训练服务市场竞争激烈，Scale AI、Surge AI及Turing、Mercor等公司均在数据标注、评估与强化学习领域展开角逐。Reddy表示，与传统数据标注企业不同，Deccan自创立即专注于高技能要求的生成式AI后期训练，过去一年业务规模增长十倍，年化营收已达数千万美元级别，其中前五大客户贡献约80%收入。

尽管印度在全球AI价值链中仍以人才与训练数据供应为主，但Deccan已开始从美国等地吸纳地理空间数据、半导体设计等领域的尖端人才，逐步拓展全球化专业网络。

中文翻译：

随着人工智能模型训练和精调需求的增长，初创企业Deccan AI近日完成了2500万美元的首轮大规模融资。这家公司专注于提供训练后数据与评估服务，其大量工作由印度专家团队完成。

本轮纯股权A轮融资由A91 Partners领投，Susquehanna International Group和Prosus Ventures参与投资。尽管OpenAI和Anthropic等前沿AI实验室自主构建核心模型，但随着企业致力于提升系统在现实应用中的可靠性，从数据生成到评估与强化学习等训练后工作正日益外包。Deccan正成为满足这一需求的新兴初创企业之一。

成立于2024年10月的Deccan提供多元化服务，涵盖提升模型编码与智能体能力、训练系统与应用编程接口等外部工具交互等。该公司与前沿实验室合作开展专家反馈生成、评估测试及强化学习环境构建等任务，同时通过评估套件Helix和运营自动化平台等产品服务企业客户。随着模型从文本处理向理解机器人、视觉系统等物理环境的"世界模型"演进，其业务形态也在持续拓展。

据公司披露，其客户包括谷歌DeepMind和Snowflake。创始人鲁凯什·雷迪（上图）在接受采访时表示，公司已签约约10家客户，并同时运行约二十个活跃项目。这家总部位于旧金山湾区、在海得拉巴设有大型运营团队的初创企业，目前拥有约125名员工，并依托超过100万贡献者网络开展工作，其中包括学生、领域专家和博士人才。雷迪向TechCrunch透露，每月约有5000至10000名贡献者活跃在平台上。

雷迪表示，约10%的平台贡献者拥有硕士或博士等高等学位，而活跃贡献者中该比例会根据项目需求更高。随着大语言模型的兴起，AI训练服务市场迅速扩张，Meta旗下的Scale AI及其竞争对手Surge AI，以及初创企业Turing和Mercor等公司都在竞相提供数据标注、评估和强化学习服务。

"质量仍是待解难题，"雷迪指出，训练后阶段的容错率"近乎为零"，因为错误会直接影响模型在生产环境中的表现。这使得训练后工作比早期阶段更为复杂，需要高度精确、领域特定的数据，规模化难度更高。他还强调这项工作具有极强时效性，AI实验室有时需要在数天内获取海量高质量数据，如何在速度与精度间取得平衡成为挑战。

该行业曾因工作条件和薪酬问题受到批评，通常依赖大量零工生成训练数据。雷迪表示，Deccan平台上的时薪约在10至700美元之间，顶级贡献者月收入可达7000美元。

尽管客户主要为美国AI实验室，但Deccan绝大多数贡献者位于印度。Turing和Mercor等竞争对手也从印度招募合作者，但业务范围更广泛地覆盖新兴市场。雷迪解释，Deccan将主要团队集中在印度是为了更好地管控质量："许多竞争对手在百余个国家寻找专家，而将运营集中在单一国家能显著提升质量管控效率。"

这种模式凸显了印度当前在全球AI价值链中的定位——作为人才和训练数据供应方，而非前沿模型开发者，后者仍集中在少数美国公司和部分中国企业手中。不过雷迪透露，Deccan已开始从美国等其他市场招募地理空间数据和半导体设计等专业领域人才。

与传统从计算机视觉起家的数据标注公司不同，雷迪将Deccan定位为"原生生成式AI企业"，这意味着公司从创立就专注于高技能工作。过去一年Deccan实现了10倍增长，目前年化收入达数千万美元规模。雷迪补充说，约80%收入来自前五大客户，这反映出前沿AI市场高度集中的特性。

英文来源：

As demand grows for training and refining AI models, Deccan AI — a startup supplying post-training data and evaluation work — has raised $25 million in its first major funding round, with much of that work carried out by an India-based workforce of experts.
The all-equity Series A round was led by A91 Partners, with participation from Susquehanna International Group and Prosus Ventures.
While frontier AI labs including OpenAI and Anthropic build core models in-house, much of the post-training work — from data generation to evaluation and reinforcement learning — is increasingly being outsourced as companies push to make systems reliable in real-world use. Deccan is emerging as one of a new set of startups serving that demand.
Founded in October 2024, Deccan provides services ranging from helping models improve coding and agent capabilities to training systems to interact with external tools such as application programming interfaces (APIs), which connect AI models to software systems.
The startup works with frontier labs on tasks such as generating expert feedback, running evaluations and building reinforcement learning environments, while also serving enterprises through products including its evaluation suite, Helix, and an operations automation platform. The work is also evolving as models move beyond text into so-called “world models” that better understand physical environments, including robotics and vision systems.
Deccan’s customers include Google DeepMind and Snowflake, according to the company. It has onboarded about 10 customers and runs a couple of dozen active projects at any given time, founder Rukesh Reddy (pictured above) said in an interview.
The startup, headquartered in the San Francisco Bay Area with a large operations team in Hyderabad, employs about 125 people and relies on a network of more than 1 million contributors, including students, domain experts, and PhDs. Around 5,000 to 10,000 contributors are active in a typical month, Reddy told TechCrunch.
Disrupt 2026: The tech ecosystem, all in one room
Your next round. Your next hire. Your next breakout opportunity. Find it at TechCrunch Disrupt 2026, where 10,000+ founders, investors, and tech leaders gather for three days of 250+ tactical sessions, powerful introductions, and market-defining innovation. Register now to save up to $400.
Save up to $300 or 30% to TechCrunch Founder Summit
1,000+ founders and investors come together at TechCrunch Founder Summit 2026 for a full day focused on growth, execution, and real-world scaling. Learn from founders and investors who have shaped the industry. Connect with peers navigating similar growth stages. Walk away with tactics you can apply immediately
Offer ends March 13.
About 10% of Deccan’s contributor base has advanced degrees such as master’s and PhDs, though the share is higher among active contributors depending on project requirements, Reddy said.
The market for AI training services has expanded rapidly alongside the rise of large language models, with companies such as Meta-owned Scale AI and its rival Surge AI, as well as startups Turing and Mercor competing to provide data labeling, evaluation, and reinforcement learning services.
“Quality remains an unsolved problem,” Reddy said, adding that tolerance for errors in post-training is “close to zero” as mistakes can directly affect model performance in production. That makes post-training more complex than earlier stages, requiring highly accurate, domain-specific data that is harder to scale.
The work is also highly time-sensitive, he said, with AI labs sometimes requiring large volumes of high-quality data within days, making it difficult to balance speed with accuracy.
The sector has faced criticism over working conditions and pay, with large pools of gig workers often used to generate training data. Reddy said earnings on Deccan’s platform range from about $10 to $700 per hour, with top contributors earning up to $7,000 a month.
India emerges as a hub for AI training talent
Even as its customers are largely U.S.-based AI labs, most of Deccan’s contributors are based in India. Competitors such as Turing and Mercor also source contractors from the country, but operate across a broader set of emerging markets.
Deccan chose to concentrate much of its workforce in India to better manage quality, Reddy said. “Many of our competitors go to 100-plus countries to find the experts,” he said. “If you have operations in just one country, it becomes far easier to maintain quality.”
That approach highlights India’s current position in the global AI value chain — as a supplier of talent and training data rather than a developer of frontier models, which remain concentrated among a handful of U.S. companies and a few players in China.
However, Reddy said Deccan has begun sourcing talent from a few other markets, including the U.S., for niche expertise in geospatial data and semiconductor design.
Reddy said Deccan was built as a “born GenAI” company, in contrast to traditional data labeling firms that began with computer vision tasks. This means it has focused on higher-skill work from the outset.
Deccan grew 10x over the past year and is now at a double-digit million-dollar revenue run rate, Reddy said, declining to share specifics. About 80% of its revenue comes from its top five customers, reflecting the concentrated nature of the frontier AI market, he added.