Bengaluru, July 21 -- As companies scale artificial intelligence (AI) adoption, rising token consumption is pushing clients of India's largest technology services firms to rethink which AI models they deploy for different tasks. Three of India's five largest tech services firms-Tata Consultancy Services Ltd, HCL Technologies Ltd and Wipro Ltd-said during their latest earnings calls that customers are rethinking their AI model mix to contain rising token bills, shifting routine work to smaller models and reserving larger ones for more complex tasks. Tokens are the basic units of data that AI tools such as OpenAI's ChatGPT and Anthropic's Claude break language into. The more a company uses a model, the more tokens it consumes-and the larger the bill. "The unit cost is becoming cheaper while total AI consumption is growing exponentially, so enterprise AI bills are increasing even as the underlying technology becomes more efficient," said Phil Fersht, chief executive of HFS Research. The issue has emerged as clients tighten technology budgets, squeezed between rising AI costs and uncertainty from the conflict in West Asia. Token costs, one chief executive said, were barely a topic a few months ago. They are now firmly on the radar of enterprises scaling AI adoption. "If you see about 3 months ago or 6 months ago, token cost was not a topic. It's only when enterprises started scaling the adoption of AI, the whole various dimensions of token costs came into play. Now, as (for) the token consumption cost we believe, it really depends on the exact models that you're implementing, and customers are looking at creative ways to reduce their overall token costs," said C Vijayakumar, chief executive of HCLTech, during the company's post-earnings analyst call on 13 July. HCLTech is using a tiered structure that sells AI solutions through small and large models based on client requirements. The approach echoes a broader shift in how IT services firms are packaging AI work. Cognizant Technology Solutions Corp, for instance, charges clients based on the extent of AI involvement-from work done by AI and verified by humans to fully automated work done without human intervention. For now, Vijayakumar said, the approach would reduce costs, although rising token consumption would increase overall costs. HCLTech ended last quarter with advanced AI revenue of $171 million, up 10.3% on a quarterly basis in constant currency terms. The company classifies 'Advanced AI' as revenue from agentic AI, AI factories and physical AI. Constant currency does not take currency fluctuations into account. Wipro CEO Srinivas Pallia said clients are also becoming more focused on the link between AI spending and business outcomes. "As the tokenization landscape evolves, clients are focused on net productivity and require a tighter linkage between investment and outcomes," said Pallia, during the company's post-earnings analyst call on 16 July. He added that clients are spending big on AI tokens as adoption of automation tools increases. "The tokenization, token economics, whatever you call it, it's going -- it's actually skyrocketing right now. So the-especially the CFOs are saying, you know, "Hey, you know, what's the-what's my ROI? So do I use a high-end LLM for a particular process or a workflow? Do I use a, you know, open source model," Pallia added. The emerging response is not to pick one AI model and stick with it. It is to use different models for different jobs. "We believe, most enterprises will have multiple models, one LLM (large language model) plus many SLMs (small language models) or multiple LLMs within the same family, they probably will use an older model for certain queries and a newer model for certain workloads," said Tata Consultancy Services' (TCS) K. Krithivasan during the company's post-earnings conference call on 9 July....