Singapore is stepping up its AI game. Recognising that graphics processing units (GPUs) are costly yet crucial for developing AI applications, the government is ready to spend up to S$500 million to ensure that local enterprises can access to tomorrow’s high-performance computing infrastructure.
It is taking a multi-faceted approach to secure and ensure access to GPUs each of which can cost as high as US$40,000. It may include purchasing the GPUs, leasing from external providers offering AI compute services. Expanding partnerships with chipmakers and cloud service giants to secure GPU access is another avenue.
The local compute resources will drive companies across sectors like financial services, transport and logistics and healthcare to create innovative AI solutions. It also hopes to catalyse capability building, attract AI industry players to set up shop locally and incentivise enterprises to discover proprietary datasets and apply their domain expertise.
This plan was outlined in Parliament yesterday by Josephine Teo, Minister for Information and Communications, during the Committee of Supply debate on her ministry.
The S$500 million is about half of the S$1 billion-plus investment for the National AI Strategy 2.0 unveiled last December. The refreshed strategy, which recognises AI as vital for business and society, goes beyond just an economic booster.
This investment in AI compute infrastructure is one of five AI initiatives in the Singapore roadmap for 2024. The rest include talent development, capability building and AI governance.
The compute infrastructure initiative signals that Singapore aims to excel in AI application development to benefit local companies. It is not competing with tech titans to build big foundation models which would cost much more.
Meta, for example, is expected to spend billions acquiring 350,000 GPUs to build massive compute infrastructure. Tech giants like OpenAI and Microsoft also require huge amounts of GPUs for AI training and inferencing.
With scant details on resource administration and allocation for Singapore’s plan, some concerns come to mind. Will companies have to apply for access? If so, criteria could include:
- Clear, fair eligibility rules
- Flexibility, as AI evolves rapidly, since outcomes may be hard to define
- Sustained access to enable robust solution development, instead of a one-off effort
- Tech-savvy administrators who understand the technology and innovations
One beneficiary could be Singapore-based KeyReply, which builds specialist AI-powered virtual assistants for enterprises to automate conversations and processes for customer support and marketing.
Founder Peiru Teo welcomed Singapore’s new initiative. Thus far, she has been cautious about AI training, focussing on using smaller foundation models which require less compute time.
With the availability of compute resources, she aims to train full vertical, broad-based healthcare chatbots, instead of specific use case using smaller datasets.
“Without such access, we will be priced out because GPUs are expensive,” she said. The criteria to get access, she added, must also be clear and not disadvantage smaller companies like startups and mid-sized companies.