内容来源：https://www.livemint.com/ai/artificial-intelligence/india-ai-mission-bharatgen-meity-indian-ai-model-large-language-models-nvidia-gpus-ai-kosh-sarvam-ai-11758193038836.html

内容总结：

印度电子和信息技术部近日宣布，在总额12亿美元的"印度AI使命"计划第三轮投资中，非营利联合组织BharatGen获得1.12亿美元资金支持，成为印度获得最高资助的人工智能实体。该组织由印度理工学院孟买分校孵化成立，计划构建参数规模达万亿级的AI大模型，重点开发涵盖22种印度本地语言的多语言数据集。

本轮共有8家初创企业和非营利机构获得总额1.7亿美元的资助，其中近60%资金将用于GPU芯片资源支持。BharatGen已获准使用13,642块英伟达GPU进行为期12个月的模型训练。电子和信息技术部长瓦什瑙表示，政府将继续追加1万块GPU以推动AI产业发展，并计划在9月28日前发布侧重于安全规范的非强制性AI治理框架。

尽管专家指出这种以本土语言为核心的开发模式可能限制企业的全球化商业潜力，但BharatGen已与HDFC银行、IBM等企业展开应用合作。项目负责人表示，将优先构建面向公共服务的AI模型，通过政府机构采购实现自给自足，暂不引入风险投资。目前印度政府已在18个月内投入该计划超五分之一的资金，用于构建本土AI模型和类似Hugging Face的AI应用市场。

中文翻译：

巴拉特生成（BharatGen）在印度人工智能计划中获得1.12亿美元资助，成为印度获资最高的人工智能实体。
印度电子和信息技术部于本周四宣布了其12亿美元印度人工智能计划的第三轮投资，巴拉特生成获得最大份额资助。

新德里：由印度孟买理工学院孵化、十个月前正式成立的非营利联盟组织巴拉特生成，现已获得印度中央政府98.8亿卢比（1.12亿美元）资助，成为印度获资最高的人工智能实体。
印度电子和信息技术部本周四宣布第三轮投资计划，在12亿美元总预算中筛选出八家初创企业和非营利组织，为其从零构建大型人工智能模型及多语种数据集提供支持。

巴拉特生成因提议构建训练参数高达万亿级的人工智能模型，成为本轮17亿美元（150亿卢比）净支出中获资最高的机构。这些资助大部分以图形处理器（GPU）芯片使用权形式发放，这类芯片占人工智能模型构建企业基础设施成本的90%。
该机构获准使用13,642块英伟达GPU进行为期12个月的模型训练。两名高级官员透露，仅巴拉特生成的净支出预估就达1亿美元（86亿卢比），占人工智能模型总支出的近60%。

其他获选机构包括Avataar AI、Fractal Analytics、Tech Mahindra旗下Makers Lab、Zeintech、Genloop、NeuroDX和Shodh AI。其中Avataar、Fractal和Zeintech计划构建700-800亿训练参数模型，成为印度人工智能计划中规模较大的项目。总体而言，在印度全力构建自主人工智能体系的背景下，巴拉特生成已成为电子和信息技术部最大资助对象。

累计十二家机构入选
2024年3月启动的印度人工智能计划在2025财年首年支出17.3亿卢比（2000万美元），本财年预算分配为200亿卢比（2.3亿美元）。4月26日，Sarvam获准以24.7亿卢比（2800万美元）计费成本优惠使用4,096块英伟达GPU。5月30日，另有三家初创企业Gan.ai、Gnani和Soket入选本土自主人工智能模型建设资助名单。

电子和信息技术部秘书克里希南在发布会后表示，本财年结束前将"再批准一批初创企业资助名单"，并称该部"将根据实际需要"提高人工智能计划的预算上限。完成本轮分配后，已有12家企业获得构建本地语言大型人工智能模型的资助，旨在为企业及公共服务机构提供应用服务。

印度联邦信息技术部长瓦什瑙指出："正如巴拉特生成项目所展现的，人工智能发展需要大量资源。长期来看还需追加激励措施——我们将增配1万块GPU以支持更多项目。"他补充说电子和信息技术部将在9月28日前发布人工智能治理框架，该框架并非强制性规范，而是着重安全性地为人工智能开发部署提供建议。

至此，该部在18个月内为本土人工智能模型已投入超计划总资金五分之一，包括从云计算企业采购GPU基础设施、提供优惠使用权，以及构建类似Hugging Face的人工智能市场平台AI Kosh用于托管印度本土开发的人工智能模型与应用。

巴拉特生成执行副总裁巴勒表示，其目标是构建"训练数据集达数千亿参数的大型人工智能模型"。他补充说："电子和信息技术部的资助将以GPU使用权和现金形式发放，助力优化运营。我们已从国家跨学科信息物理系统使命计划获得23.5亿卢比（2700万美元）资金，去年在印度孟买理工学院完成孵化。过去十个月我们构建并发布了包括Param在内的多个模型，现正着手开发涵盖22种语言的印度语言数据库Bharat Data Sagar并开源共享。"

专家称商业化前景存疑
行业专家认为电子和信息技术部的方案未必能助力入选初创企业实现商业成功。高德纳咨询公司新兴技术总监分析师维尔玛指出："印度人工智能计划资助的初创企业多聚焦印度语言数据构建人工智能模型，因此可能主要服务于农业、教育、零售等特定领域及公共事业。"她表示这种模式难以形成可全球推广的人工智能框架，"专注本土市场将限制初创企业的全球业务潜力，使其仅能服务印度市场。企业客户并非政府关注重点——他们仍倾向于使用谷歌、OpenAI等机构的全球模型。"

印度孟买理工学院计算机科学讲席教授、巴拉特生成首席研究员拉马克里希南透露，这个目前以非营利模式运营的实体已获得实际商业关注："与HDFC银行、IBM和Zoho等机构的洽谈显示，它们都有兴趣在真正本土的人工智能模型上构建应用。使用全球模型处理印度语言成本极高，因为这些语言在训练集中占比不足2-3%，且存在诸如卢比符号渲染等细节缺失。企业正是看中我们模型在这些领域的优势，早期应用案例成果显著。"

巴勒与拉马克里希南确认，虽然电子和信息技术部不会因资助持有机构股权，但将委派一名代表进入董事会。巴勒表示："政府部门支持我们将模型转化为商业模式，但我们优先目标是构建能有效服务公共事业领域的模型——政府机构本身将是关键客户。这将实现资金自给自足，无需引入专注商业回报的风险投资。"

英文来源：

BharatGen is India’s highest-funded AI entity after receiving $112 million under IndiaAI Mission
On Thursday, the ministry of electronics and information technology announced the third tranche of investments under its $1.2-billion IndiaAI Mission. BharatGen walked away with the lion's share.
New Delhi: BharatGen, a non-profit consortium incubated in IIT Bombay and formally launched 10 months ago, has become the highest-funded Indian AI entity with more than ₹988 crore ($112 million) from the union government.
On Thursday, the ministry of electronics and information technology (Meity) announced the third tranche of investments under its $1.2-billion IndiaAI Mission, selecting eight startups and non-profit organisations to incentivise built-from-scratch large AI models, as well as datasets in multiple Indian languages.
Of the eight, BharatGen, was the largest recipient of the ministry’s net outlay of around $170 million ( ₹1,500 crore), having proposed to build AI models with up to one trillion data parameters in training. Most of these incentives will be in the form of access to graphics processing unit (GPU) chips, which account for up to 90% of the cost of infrastructure for companies that build AI models.
BharatGen received sanction from Meity to use 13,642 Nvidia GPUs for 12 months to train its models. Two senior officials said the estimated net outlay for BharatGen alone was close to $100 million ( ₹860 crore), nearly 60% of its overall spending on AI models.
Other companies selected for incentives include Avataar AI, Fractal Analytics, Makers Lab by Tech Mahindra, Zeintech, Genloop, NeuroDX and Shodh AI. Of these, Avataar, Fractal and Zeintech have proposed to build models with 70-80 billion training data parameters, putting them among the larger projects under the IndiaAI Mission. All said, BharatGen is now the biggest recipient of Meity’s AI incentives as India looks to build a sovereign AI model from the ground up.
A dozen and counting
The IndiaAI Mission, announced in March 2024, spent ₹173 crore ($20 million) in its first fiscal year, FY25, and was allocated $2,000 crore ($230 million) for the current fiscal year. On 26 April, Sarvam received approval from Meity for subsidised access to 4,096 Nvidia GPUs at a total billable cost of ₹247 crore ($28 million). On 30 May three more startups—Gan.ai, Gnani and Soket—were selected for incentives to build indigenous, sovereign AI models.
At a press conference after the announcement, S Krishnan, secretary at Meity, said the ministry would make “one more tranche of approvals for startups before the end of this fiscal year". He added that the ministry was open to increasing the fiscal limit approved under the AI Mission, “as required".
With this allocation completed, the ministry now has 12 companies incentivised to build large local-language AI models and offer services and applications to corporates and public service entities.
Ashwini Vaishnaw, union IT minister, said at the announcement, “The development of AI requires sweeping resources, as BharatGen’s project showed. This will require additional incentives in the long run—we will add 10,000 more GPUs to boost more missions."
Vaishnaw added that Meity would publish an AI governance framework by 28 September. The framework would not be regulatory or prescriptive, but seek to offer recommendations on development and deployment of AI, with an emphasis on safety, he said.
With this, Meity will have used more than a fifth of the mission’s funds in 18 months just for indigenous AI models—including sourcing GPU infrastructure from cloud computing companies, offering subsidised access, and building AI Kosh, an AI marketplace similar to Hugging Face, to host models and AI applications built in India.
Rishi Bal, executive vice-president at BharatGen, said the company’s goal is to build a large AI model “whose training dataset is in the hundreds of billions of parameters". He added, “The incentives from Meity will come in the form of GPU access as well as additional cash that will help in streamlining our operations. We had already raised ₹235 crore ($27 million) from the National Mission on Interdisciplinary Cyber-physical Systems (NM-ICS), and incubated our operations at the Indian Institute of Technology, Bombay, last year.
“In the past 10 months we’ve built and published multiple models including Param. Now, our goal is to tap a vast trove of data, for which we’re also building and open-sourcing an Indic language database in more than 22 languages, called Bharat Data Sagar."
Commercial success may elude startups, expert say
However, industry experts said Meity’s approach may not facilitate a commercial success for the chosen startups. “Most of the startups being incentivised by the IndiaAI Mission are building AI models focused on Indic-language data. To that end, these startups will likely cater to specific sectors such as agriculture, education or maybe even retail, alongside public utilities," said Anushree Verma, director analyst - emerging technologies at Gartner.
She added that this approach may not allow India to build an AI model framework that can be replicated around the world. “This focus will likely limit the potential of global business for the startups and have them cater to India only. Enterprises are not a part of the government’s focus here—they are still largely attuned to using global models made by the likes of Google, OpenAI and others," Verma said.
Ganesh Ramakrishnan, chair professor for computer science at IIT Bombay and principal investigator at BharatGen, added that the entity, which has been built as a non-profit firm for now, has already seen real-world business interest “through conversations with HDFC Bank, IBM and Zoho, among others, who are all interested in using a truly indigenous AI model to build applications on top of".
“The cost of running global AI models in Indian languages, in which the latter make for hardly 2-3% of the training dataset, is also very high. Plus, myriad nuances—such as rendering the rupee symbol—are missed out. These are the areas for which corporates are interested in using our model, and the early use cases and results have been very promising," he added.
Bal and Ramakrishnan affirmed that while Meity will not directly hold equity in the body in exchange for funding and incentives, it will appoint a member to BharatGen’s board of directors. “The ministries have been supportive of BharatGen taking its models to a commercial business model. However, we want to ensure that we first build a model that can be effectively used for public service utilities—where government bodies themselves will be key customers. This will help us be self-sufficient in our own billables, and not require us to have venture capital investors whose focus will be on commercial returns," Bal said.