Beyond US & China: Asia's Quiet AI Revolution

Asia has been quietly building its AI capabilities for quite a while. But no one was bothered before DeepSeek happened. 

By Asia, we mean nations beyond China. Amid the fierce battle between the US and China

for AI supremacy, no other Asian country was visible. 

China’s DeepSeek has proven it is possible to make AI models cheaply and affordably. You don’t need billions and billions of dollars (millions will do) to create foundational large language models (LLM) that power generative AI. What you need is a new approach to look at old problems.

India is rising to the challenge now. So is Southeast Asia. The same is true for Korea and Japan.

As such Asia has the largest number of people using AI applications. The demand is so much that heavy-weight tech companies—read Google, Microsoft, Nvidia, Oracle, and Amazon—have committed to investing billions of dollars into the region to set up AI data centers. 

So it shouldn’t come as a surprise if there are more and more Asian people and companies putting resources into making large language models to compete with top AI models like ChatGPT, Gemini, Claude, Minstral, and Llama.

In this edition of Hedwig by The Content House, we took a look at how different Asian nations are creating AI that speaks their language and understands cultural nuances.

India

The Indian government has decided to become the biggest cheerleader of entrepreneurs and researchers who have the capability and ambition to create India’s own large language models. It wants to solve the biggest bottleneck for the foundation AI companies—the costly, and often inaccessible, computing power.

As part of the INR 10,370 crore (US$1.1 billion) IndiaAI Mission, launched in March 2024, India aims to help as many as six firms develop foundational AI models. Almost half of the budget outlay for IndiaAI Mission will go into making 18,693 virtual graphics processing units (GPUs) accessible to Indian startups and researchers to build the foundational models.

Globally, GPU access costs US$2.5-$3 per hour. Under the IndiaAI mission, the AI model computation will cost less than 100 rupees (US$1) per hour after the 40% government subsidy to end users on the total price.

Moreover, the focus will be on companies that can achieve algorithmic efficiency to deliver a model at a much lower cost and in much less time (say, in the next eight to 10 months).

These homegrown models will compete with the best models coming out of the U.S. and China, like OpenAI’s o1 and DeepSeek-R1. 

The Indian government has called for proposals from AI companies to develop foundational models that keep India’s cultural context and her languages in mind.

Let’s look at the major AI companies working on foundational models.

LossFunk

Earlier this month, Paras Chopra announced building a new startup called Lossfunk—to build a highly efficient reasoning model from India.

The founder of successful Saas startup Wingify —who recently exited his company for around US$200 million—has chosen to jump into the fiercely evolving AI space.

He believes building indigenous AI capabilities is just as important for India as it was to build its own nuclear technology. To realize this ambition of homegrown AI power, he is setting up a foundational AI lab in India that produces state-of-the-art models and algorithms for the world.

Lossfunk will build highly efficient AI models using advanced techniques such as reinforcement learning, model distillation, and pruning to achieve state-of-the-art performance with limited computational resources. 

With the ambition to drive “superintelligence” from India, Paras plans to apply for government-sponsored computational resources under the IndiaAI mission. 

Currently, he is recruiting technical experts to launch a next-generation AI lab in India.

Krutrim

Krutrim, a part of Ola Group headed by Bhavish Aggarwal, is another AI company that launched an AI research lab in February 2025. 

The new Krutrim AI Lab aims to create the complete AI computing stack—AI computing infrastructure, AI Cloud, foundational models, and AI-powered end applications for the Indian market. This computing platform will help organizations and developers worldwide build custom AI applications and models.

To fuel this ambition, Krutrim has secured an investment of US$230 million in equity and debt, with a committed future investment of US$1.2 billion by next year.

Notably, Bhavish founded Krutrim AI in 2023 to develop open-source AI models tailored for Indic languages and support the AI research and developer community in the country. 

The newly launched Krutrim AI lab aims to create high-quality data sets and build AI that understands and processes text, speech, and visuals across multiple languages to enable cost-efficient AI innovation in resource-constrained environments.

Krutrim has already open-sourced a few new Indic models across text, voice, and vision over the past year to enhance AI adoption in India. In January 2024, it released its first LLM, Krutim-1, with 7B (billion) parameters. Built on this foundational model, Krutrim-2 was launched in Feb 2025. Optimised for 22 Indic languages along with English, Krutim-2 is 12B parameters advanced LLM, built on the Mistral-NeMo architecture with a context window of 128K tokens.

Parallelly, it is also developing AI training and inference platforms that enable AI research and development across industry domains.

Interestingly, it is setting up its own data center facility in partnership with NVIDIA to set up India’s first GB200 cluster by March 2025. This feeds into Krutrim’s ambition to build

India’s largest supercomputer in India by the end of the year. Bhavish plans to scale up its data center capacity to a massive 1 GW by 2028.

Additionally, Krutrim has developed ‘BharatBench,’ a benchmark for Indic performance to test how well AI models understand and perform in Indian languages and cultural contexts. 

Sarvam AI

Sarvam AI is a low-profile pioneer in generative AI, founded in 2023 by Dr. Vivek Raghavan and Dr. Pratyush Kumar, and backed by prominent VCs like Lightspeed, Peak XV Partners, and Khosla Ventures.

What makes Sarvam AI stand out is its open-source multilingual foundational models that support voice-first interfaces, aiming to make AI accessible and practical for India's diverse linguistic landscape. 

The company launched two LLMs in 2024: Sarvam-2B and Sarvam 1, trained from scratch on 2T (trillion) and 4T tokens, respectively, with an equal split between English and Indic languages. 

Designed to cater to India's diverse linguistic landscape, these open source models support 10 Indic languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu, breaking language barriers and making AI more inclusive.

Developers can use the base models to build their own voice-first, multilingual, and action-driven AI solutions.

Sarvam AI aims to develop the “full-stack” for Generative AI ranging from research-led innovations in training custom AI models to an enterprise-grade platform for authoring and deployment. Besides, it works with Indian enterprises to co-build domain-specific AI models on their data. Interestingly, the company plans to use generative AI to improve public services in India on a large scale, leveraging the existing India stack digital infrastructure.

Southeast Asia

Southeast Asia has started its own quiet journey toward developing in-house AI capabilities.

As such, the demand for AI services, led by generative AI applications, is explosive. This is why global tech giants have committed to investing over US$30 billion in the region over the next few years. Southeast Asia's young and digitally native population is the key to accelerating AI adoption.

SEA-LION

When it comes to building foundational models, Singapore has taken the lead.

In December 2023, Singapore launched a S$70 million initiative, funded by the National Research Foundation, to develop the SEA-LION project—Southeast Asia's first LLM. 

AI Singapore, the government-backed non-profit firm is the one developing SEA-LION (Southeast Asian Languages in One Network), a family of open-source LLMs that better understand Southeast Asia’s diverse contexts, languages, and cultures. 

Related to this effort is project SEALD (Southeast Asian Languages in One Network Data), a collaborative research initiative between AI Singapore and Google Research, with the aim to enhance diverse datasets (in languages spoken in Southeast Asia) for training LLMs in Southeast Asian languages.

SEA-LION models are trained on regional languages including Thai, Vietnamese, Bahasa Indonesia, Malay, Filipino, Tamil, Burmese, Khmer, Lao, Chinese, and English. This extensive linguistic coverage enables the models to handle languages—traditionally underrepresented in major language models—and understand cultural context awareness.

That’s crucial because existing LLMs—mostly developed in the US and China and heavily on English data—display strong bias in terms of cultural values, political beliefs, and social attitudes. This can lead to biased responses, especially on topics requiring local context.

Last year, Singapore earmarked S$1 billion (US$744 million) to invest in AI computing, talent development, and industry growth over the next five years. Notably, about half of this funding is designated for securing high-performance computing resources for AI innovation and capability building across sectors such as transport, healthcare, and financial services.

In initial tests, SEA-LION showed promising results in terms of speed, accuracy, and conciseness. As per the benchmark proposed by AI Singapore specifically for evaluating LLMs in Southeast Asian languages, SEA-LION was assessed to be better than other LLMs like the GPT-4 in understanding regional sentiments, while in reasoning tasks, SEA-LION ranked second only to GPT-4.

The larger picture behind this initiative is to establish Singapore as an AI hub, with a robust local AI talent pool.

Sailor

Sea AI Lab, the AI research arm of Singapore’s Sea Group, has developed Sailor, a suite of open language models in partnership with the Singapore University of Technology and Design, specifically tailored for Southeast Asian languages.

Built from Qwen 1.5, these models are continually pre-trained on 200 to 400 billion tokens across seven languages: Indonesian, Thai, Vietnamese, Malay, Lao, English, and Chinese. 

Sailor offers models of varying sizes, including 0.5B, 1.8B, 4B, 7B, and 14B parameters, to cater to different computational requirements and use cases. Benchmark evaluations demonstrate Sailor's strong performance in tasks such as question answering, commonsense reasoning, and reading comprehension in Southeast Asian languages.

Their latest model Sailor2 was released in December 2024, keeping in mind a strong demand for models in the 8B and 20B parameter range for production use, alongside a 1B model for specialized applications, such as speculative decoding and research purposes. 

Sailor2 builds upon the foundation of Qwen2.5 and is continuously pre-trained on 500B tokens to better support 15 regional languages including English, Chinese, Burmese, Cebuano, Ilocano, Indonesian, Javanese, Khmer, Lao, Malay, Sundanese, Tagalog, Thai, Vietnamese, and Waray.

Japan

Japan is enhancing its AI research capabilities through a mix of government, academic, and corporate efforts. 

Last year, the Land of the Rising Sun committed over US$65 billion to support its domestic semiconductor and AI industries. This funding, set to continue through 2030, aims to make Japan powerful again—by enhancing technological infrastructure and reducing dependency on foreign technology. 

The country has also launched an AI Safety Institute to bolster safe and ethical AI research.

Now the government is working closely with industry leaders to support R&D in critical areas such as robotics, semiconductors, and natural language processing. 

Meanwhile, Japanese tech titans are also keeping up with global AI giants. 

For example, SoftBank Group formed a 50/50 joint venture with OpenAI to develop and market AI solutions tailored for Japanese businesses. SoftBank plans to invest US$3 billion annually to deploy OpenAI's technologies across its group companies, which in turn will facilitate the integration of advanced AI into various sectors of the Japanese economy.

Another tech titan that is super interested and invested in the Japanese AI scene is NVIDIA. 

In November 2024, NVIDIA partnered with SoftBank to build Japan's most powerful AI supercomputer and develop the world's first live 5G AI-RAN (Radio Access Network). 

NVIDIA has also invested in Japanese startup Sakana AI, which builds cost-effective generative AI models using small datasets. 

Sakana AI

Sakana AI is an AI R&D company based in Tokyo with a focus on developing new kinds of foundation models based on nature-inspired intelligence. The name Sakana is derived from the Japanese word さかな (sa-ka-na) which means fish. 

The name is inspired by the idea of a school of fish coming together and forming a coherent entity from simple rules. That’s because the company wants to create advanced transformative AI models inspired by natural systems, particularly evolution and collective intelligence observed in nature, such as schools of fish and beehives. 

This approach involves creating multiple smaller AI models that work collaboratively, akin to a swarm, to achieve complex tasks. Unlike the traditional approach of building large, monolithic AI systems, offering potential advantages in flexibility and efficiency.

These cutting-edge foundation models aim to automate and speed up scientific discovery.

Founded in July 2023 by former Google researchers, Sakana has received an outsized welcome in Japan, which is eager to catch up to the US and China in artificial intelligence. 

In September 2024, the company raised US$200 million in a Series A funding round, at a valuation of US$1.5 billion, from New Enterprise Associates, Khosla Ventures, and Lux Capital, with participation from Translink Capital, 500 Global, and NVIDIA. Additionally, it has got NVIDIA’s backing for research, infrastructure, and AI community building in Japan.

Notably, Khosla Ventures is also an investor in India’s Sarvam AI, which is developing voice-first, indic language foundational models for enterprises. The founder of Khosla Ventures, Vinod Khosla, sees Sakana differently though. As per him, most companies are building and training foundation models by using the same techniques as everyone else, but Japan’s Sakana AI is showing the path to innovation, referring to its endeavor to develop models that emulate the adaptability and efficiency of natural systems.

Last August, Sakana AI introduced "The AI Scientist," a system designed for fully automated scientific discovery. This system covers the entire research lifecycle—from generating novel research ideas and writing code to executing experiments and presenting findings in a complete scientific manuscript.

South Korea

South Korea is positioning itself as one of the world’s AI powerhouses through a combination of government initiatives, robust private sector R&D, and international collaboration.

The Korean government has been heavily investing in AI. The National AI Strategy (2019) set the foundation, emphasizing trustworthy AI and positioning AI as a key driver of economic and social transformation. The Digital New Deal (2020) further integrated AI into national development, launching initiatives like the AI Open Data Project to boost research and collaboration. 

A notable example is the AI Innovation Hub launched in 2024 in Seoul—a mega-lab with up to 35 petaFLOPS processing capacity and an initial investment of US$35 million—which supports over 630 AI researchers from more than 200 institutions involved in state-sponsored AI projects that include hyper-scale deep learning, spatial-temporal reasoning, and new, large language speech synthesis.

This facility is a core part of Korea’s plan to strengthen its domestic AI research ecosystem. Additionally, Korea’s semiconductor dominance (19% of total exports in 2022) complements its AI growth, as it remains one of the few nations producing AI-specific chips.

At the 2024 AI Seoul Summit, South Korea reinforced its commitment to AI safety by announcing plans to open an AI Safety Institute later this year, further integrating its domestic research capabilities with international safety standards.

Meanwhile, South Korea’s chaebols, a.k.a., large conglomerates such as Hyundai, LG, and Samsung are not only investing in domestic AI research but are also setting up research centers abroad to tap into diverse talent pools and research expertise.

Consequently, the competition among Korean companies to create LLMs that understand the Korean language and culture is intensifying. With their deep pockets and global reach, South Korean tech giants aren’t the ones to be left behind in building their own foundational AI models. Notable examples include:

LG AI Research

Five-year-old LG AI Research, the AI think tank of the LG Group, has developed a series of advanced bilingual AI models proficient in Korean and English under the Exaone brand.

LG AI Research released South Korea's first large-scale bilingual (Korean-English) LLM,

Exaone 1.0 in 2021. It was a 300B parameter model, which was largely compared to LLMs such as GPT-3 (2020), XLNet (2019), and Wu Dao 2.0.

Exaone 1.0 was one of the earliest multimodal AI models that could process both text and images.

Since then, LG’s AI research team has been optimizing Exaone models for practical industrial applications to improve efficiency and reduce costs. The latest in the series, Exaone 3.5 with 32B parameters, was released in December 2024. This high-performance model for specialized applications was developed with an investment of about US$4.8 million (7 billion won)—lower than the estimated US$6 million that DeepSeek reportedly spent on developing its V3 model.

Exaone 3.5 has two more variants featuring 2.4 billion parameters (for resource-constrained devices) and 7.8 billion parameters (general-purpose model).

These models support long-context processing of up to 32,768 tokens, which enhances their applicability in various real-world scenarios. Hence, they excel in real-world usability, long text processing, coding, and mathematics.

KT Corporation

KT Corporation, a leading South Korean telecommunications company, introduced Mi:dm in October 2023. Mi:dm is a customizable large language model aimed at corporate clients looking to develop their own advanced AI services.

Mi:dm had four different models, with options ranging from a "basic" version with around 7B parameters to an "expert" version, designed for high-performance applications,  featuring 200B parameters.

The name "Mi:dm" is derived from the Korean word "mideum," meaning "belief," reflecting the model's emphasis on reliability and trustworthiness. 

Mi:dm was trained on a dataset exceeding 1T tokens, enabling it to understand and generate human-like text across various contexts. Enterprises can fine-tune Mi:dm using their specialized datasets to develop tailored AI services that align with specific industry requirements.

To address the issue of AI-generated inaccuracies, known as "hallucinations," KT integrated three core technologies—Document AI, to enhance comprehension of complex documents; Search AI, to utilize deep learning for real-time information retrieval; and FactGuard AI to ensure responses are grounded in source materials.

Collectively, these features reduce misinformation by up to 70% compared to other generative AI models.

KT has opened up the foundation model of Mi:dm to other companies. Upstage is one of the AI startups that has joined hands with KT to develop different AI business models. Beyond domestic applications, KT aims to extend Mi:dm's reach internationally, as evident by its partnership with Thailand's Jasmine Group to develop a Thai-based large language model.

Upstage

Founded in October 2020, Upstage builds custom large language models and works with businesses to tailor AI models for specific domains. Its flagship Solar LLM model is powering use cases in industrial environments. In April 2024, Upstage AI secured US$72 million (100 billion won) in a Series B round, which brought its total funding to over US$115 million and positioned it as one of the most-funded AI startups in South Korea.

Its flagship Solar Pro was released in December 2024. It is a high-performance LLM that delivers the capabilities of a 70B plus parameter model while operating efficiently on a single GPU. Solar Pro supports multiple tasks, languages, and domains, making it versatile for diverse business applications.

Upstage’s Solar LLMs are powering use cases across industries like insurance, legal, finance, and healthcare. Upstage has even partnered with Chosun Ilbo, a prominent Korean media outlet, to integrate AI throughout its newspaper and digital platforms, a big leap towards AI-powered journalism. Essentially, Upstage is focused on the development of purpose-trained LLMs for enterprises globally.

Trillion Labs

Trillion Labs is a South Korean AI startup founded in 2024 by Jaemin Shin, a former Naver research scientist. The company focuses on developing foundational LLMs tailored to the Korean language, culture, and customs. By utilizing extensive Korean datasets for pre-training, the company aims to develop AI that leverages Korea-specific information, addresses biases inherent in English-centric LLMs, and reduces reliance on English-centric models. 

Trillion Labs believes in “AI sovereignty for Korea” and wants to develop independent AI capabilities to prevent reliance on foreign-developed models.

As per Trillion Labs’ CEO Jaemin Shin, without AI sovereignty, South Korea risks relying on subpar AI performance compared to other nations. He aims to establish Korea as a leading AI nation and demonstrate that Northeast Asia, often overlooked in the generative AI market, can also take the lead.

In its pre-seed funding round in late 2024, Trillion Labs secured US$4.2 million (approximately 5.7 billion won) from investors including Strong Ventures, Kakao Ventures, Bass Investment, The Ventures, Goodwater Capital, and BAM Ventures. 

The company plans to leverage this funding to acquire high-quality language data and advance its technology to complete an accurate Korean LLM foundation model by the end of the year. 

Looking ahead, Trillion Labs intends to expand into regions with similar linguistic characteristics, such as Japan and Southeast Asia, positioning itself as a leading AI hub for the Asian market.

There are many more companies developing Korean-focused, specialized AI models for enterprises and real-world use cases. 

Other notable AI models include NCSoft’s Varco Vision, which processes both text and images with a superior understanding of Korean culture compared to similar foreign models; Moreh’s Motif, which generates more natural Korean responses than OpenAI’s GPT-4; and Naver's HyperCLOVA X, which is trained using 6,500 times more Korean data than GPT-4 and can grasp the intricacies of Korean culture, social norms, and linguistic nuances to generate responses that resonate deeply with Korean users.

All in all,  Asia is stepping up its AI game. Countries across the region are building their own AI models that understand local languages and cultures. 

From India's push to make AI affordable to Singapore's focus on Southeast Asian languages, to Japan's creative approach with Sakana AI, and South Korea's tech giants joining the race—Asia is no longer just using AI, it's creating it.

Like Hedwig? Subscribe to it here to get insights on the Asian startup ecosystem and building a sustainable business.

At The Content House, we offer research-based, analytical content that has a strong narrative quality to it. What makes us different in the crowded content market is our ability to convert institutional knowledge and expertise locked inside organizations into content that companies can use to increase brand reach.