Hong Kong–based Votee AI is betting that the next phase of enterprise AI will not be won by ever-larger global models, but by smaller, localized systems built for language, regulation and data control.
The AI company, led by former banking executive Pak-Sun Ting, has developed what it says is the world’s first open-source Cantonese large language model with seven billion parameters and a corresponding benchmark (HKCanto-Eval) − positioning itself at the center of a growing push for ‘sovereign AI’ across Asia and elsewhere.
In an interview with The AI Innovator, Ting said the gap between consumer use of AI and enterprise deployment remains pronounced in Hong Kong and much of Asia, even as awareness of the technology is high.
“Hong Kong has one of the highest per capita usage of AI,” he said. “But in terms of actual adoption by enterprise and governments, it’s actually well below the curve.”
That gap, he argues, starts with language.
Global AI models often claim multilingual capabilities, but Ting said that while they work well for consumer use, they fall short for business and government use where language must be much more precise, particularly for languages with complex local variations such as Cantonese, spoken by about 80 million people.
“Generic models hit a wall when they go to nuanced spoken language,” he said. “We basically build large language models for low-resource languages.”
The consequences are not trivial. While lower accuracy may be acceptable for casual use, Ting said it breaks down in high-stakes environments. “When it comes down to banks, corporates, governments, you need something that has an 85%, 90%, even 95% accuracy rate,” he said.
Below that threshold, models can misinterpret inputs or generate irrelevant outputs altogether — a more severe failure than typical “hallucinations” seen in English-language systems. So while English-language hallucinations may get facts wrong, in low-resource languages the AI spews out nonsensical responses.
The global language gap is big. Ting noted that Southeast Asia alone has about 2,000 different languages. Africa, another target continent for Votee, has about 1,000 different tongues. “How do you bridge those gaps? That’s why we see this as not just an economic opportunity, but also a very big social mission.”
Language specialization for AI
Votee’s response has been to build models from the ground up for specific languages rather than adapting existing ones. Its Cantonese model, including open-source releases, is paired with enterprise software and a training platform called “Magic,” which automates the creation of specialized models for both languages and corporate use cases.
The company’s broader strategy reflects a shift in how AI is being deployed in Asia. Rather than relying entirely on global cloud-based models, governments and enterprises are increasingly prioritizing control over data and infrastructure.
“You want to have your own large language model on premises,” Ting said. “So you’re not feeding data outside, and you’re also able to control the cost of using your own model.”
This emphasis on sovereignty — data staying within jurisdictional boundaries — has become a central issue, particularly in regulated sectors such as banking and government. Votee’s positioning aligns with that demand, offering on-premises deployments and compliance with multiple regulatory frameworks, including ISO standards and regional AI guidelines.
The approach appears to be gaining traction. The company says it has worked with more than 200 enterprises and public-sector organizations, including banks, regulators and telecommunications providers, with deployments ranging from compliance systems to voice-based customer service.
In one example, Ting cited an internal compliance checker built for financial giant DBS Bank’s Hong Kong operations. Other use cases include training simulations, chatbots, and voice agents handling bookings or customer interactions in Cantonese.
A ‘fast and slow’ Asia
Still, the company is operating in a market where adoption remains uneven.
Ting described a “fast and slow” dynamic across Asia. Governments are moving quickly to fund AI initiatives and articulate national strategies, but practical deployment is constrained by infrastructure, regulation and resource limitations.
“AI is not just a model plus applications,” he said. “You have to have a full stack layer” of energy, infrastructure, GPUs, models and applications.
Malaysia is emerging as a regional data center hub, Ting said. Vietnam is looking into it as well. A lot of facilities are being built close to coastal areas, he added, perhaps to get easier access to water for cooling.
As for AI replacing jobs in the region, Ting said he is not seeing it at the moment. While some jobs will be lost, he sees AI as enabling smaller nations to better compete with larger countries. “Before, you need a big workforce. Now you can survive and maybe even thrive with a small workforce,” he said.
New opportunities could arise as well. Developing nations such as the Philippines could become “digital embassies,” where countries with data centers host the data of other nations.
Big digital divide
But a key to unlocking this opportunity comes back to getting the language right. While countries such as China, Japan, South Korea and Australia are ahead, supported by strong industrial bases and government investment, emerging economies in Southeast Asia are progressing more slowly, in part because of linguistic fragmentation and limited data resources.
“Part of it is because the language is not there yet,” Ting said. “And so you have this big digital divide.”
That divide extends beyond Asia. Ting estimates that billions of people in emerging markets are underserved by current AI systems, a gap he attributes to both technical and commercial factors.
Large technology companies, he said, are focused on scaling global models rather than investing in localized ones with smaller economic returns. “They’re racing to build the next big thing,” he said. “It’s not part of their objective to build very precise languages that understand culture, that understand nuances.”
There are also geopolitical considerations. In some cases, foreign providers face restrictions or resistance when working with governments, further reducing their incentive to localize.
Votee’s model — building language-specific systems through local partnerships — is designed to address that gap. The company has worked with institutions such as Hong Kong’s public broadcaster and universities to gather data and refine its models.
Ting’s bet is that the next wave of AI adoption — particularly in enterprise and government — will depend less on scale alone and more on precision, compliance and cultural fit.
If that holds, the market may fragment along linguistic and geopolitical lines, reshaping an industry that has so far been dominated by a handful of global players.
Be First to Comment