Rising AI bills push Singapore companies to chase ROI over ‘tokenmaxxing’

Companies in Singapore are moving beyond the artificial intelligence experimentation phase and becoming increasingly selective about how they deploy the technology. This is as rising AI spending pushes them to demand clearer returns on investment (ROI).

While the cost of using AI models has fallen, overall AI bills continue to jump as businesses use the technology far more extensively.

Consultancy Bain & Co estimated that the price of tokens had halved between 2024 and 2025.

Yet, the consumption of tokens – the units that large language models use to process information – surged 4.5 times during the same period, driving overall spending.

“Not everything needs to be done using the latest frontier model, where you are maxing out your tokens,” said Singtel group CEO Yuen Kuan Moon in an interview earlier in July.

This shift comes even as Singapore’s AI integration into work is low, but adoption remains high.

Google has said Singapore is Gemini’s highest adoption per capita market globally. Separately, GIC has noted that the Republic leads the world in Claude use per capita.

Yet, a recent Salesforce-YouGov study found that only 6 per cent of Singaporean desk workers surveyed use AI as a core part of their work daily, compared with 11 per cent globally.

Some industry observers attribute the increase in AI implementation cost to the “tokenmaxxing” culture, where companies encourage employees to maximise their token usage to get the most out of their AI models.

The term gained traction earlier this year when reports emerged that companies such as JPMorgan Chase and Meta were tracking their engineers’ AI usage.

The increase in token consumption is further driven by the rise of agentic AI workflows – where AI agents are using more tokens to retrieve contexts, read files and carry out more work.

Jiang Tianyi, CEO of AvePoint : AVP +5.66%, said that the data protection firm’s largest cost pressure has come from its internal AI-accelerated development workloads.

“On top of this, AI agents that can run endlessly and interact with other AI agents while executing tasks or workflows can cause costs to rise exponentially if not managed,” he told The Business Times.

Choosing the right model

Rather than using the most powerful AI model for each task, companies are becoming more careful in matching models to workloads.

Mohan Jayaraman, expert partner at Bain & Co, noted that many companies still default to premium models during early pilots and broad employee-access programmes because “they are easy to procure”.

“The winning business case is not to use the best model everywhere, but to be clear on the cost and value per task and the right model for the workflow,” he told BT.

He added that as projects scale into higher volumes, leading companies will reserve their frontier AI models for complex reasoning, high-stakes synthesis and ambiguous tasks.

Open-weight models, on the other hand, can be used for structured tasks such as classification, extraction and summarisation, he said.

The cost of tokens for frontier AI models – such as Claude’s Fable 5 and OpenAI’s ChatGPT GPT-5.6 Sol – is higher than that of smaller, open-weight models.

Unlike frontier models, open-weight AI models allow companies to customise their AI according to their workload within their own computing infrastructure.

Singtel uses a hybrid approach to select the model needed, depending on the complexity of the task.

“Using an intelligent and dynamic selection process, we assess the requirements of each task and select the most appropriate frontier or open-weight model,” William Woo, Singtel’s group chief information officer and group chief digital officer, told BT.

He added that open-weight AI models can be deployed and customised within a company’s own environment, allowing for greater control over data and governance.

“Singtel runs tasks involving sensitive or highly confidential information on open-weight AI models hosted on RE:AI, our sovereign AI cloud,” he said.

As for AvePoint, Jiang said the firm uses AI to route workloads among different models. This allows it to control the types of tokens being used.

“The companies that win will not be those with the most advanced model access, but those that rightsize models, who measure cost per task and redesign operating models around AI,” said Jayaraman.

Measuring business value, not just cost

As companies continue “rightsizing” their models for their tasks, other considerations – such as cybersecurity and governance – are also included when it comes to scaling a project from pilot to production.

“While time savings often justify an AI pilot, they rarely determine whether it scales,” noted James Wilson, partner in technology consulting and advisory at KPMG in Singapore.

He said the projects that are scaled from the pilot stages are those that demonstrate measurable business value relative to deployment – particularly in areas such as revenue growth, cost reduction, improved customer outcomes and risk reduction.

“Scaled AI, in our view, needs a three-way model: IT governs the platform, finance governs spend and unit economics, and the business owns productivity, revenue or risk outcomes,” said Jayaraman.

The notion of productivity gains is further strengthened by Singapore’s high labour cost.

“Singapore’s high labour cost does strengthen the value case for AI, as each hour saved has a higher economic value than in many lower-wage Asean markets,” noted Jayaraman.

“Our view is that Singapore companies will first use AI to increase throughput, improve service levels, reduce turnaround time and redeploy scarce talent into higher-value work.”

Within Singtel, the telco measures the impact of AI by comparing the speed, quality and overall impact against a baseline.

In addition, the group also accounts for the total cost of implementing AI, including technology, governance and operational oversight.

“So far, we have not discontinued any of our major AI initiatives, as they have met our economic expectations,” said Woo.

As businesses continue scaling up their AI implementation, Wilson said they should not mistake AI activity for productivity.

“Real productivity should show up in outcomes such as faster cycle times, lower costs, fewer errors, higher quality, improved customer experience or increased revenue,” he added.

Source: The Business Times © SPH Media Limited. Permission required for reproduction.

ADV: Global Rule of Law Conference | Strengthening Societies and Supporting Economic Growth Lottery winnings and scam losses are shared between divorcing spouses