Bengaluru: For once, the problem is not the cheque. Venture-backed AI startups get cloud credits along with their funding, and established firms pay for compute from cash flow. What neither can do is get the chips quickly.
Deliveries of GPUs — the processors that train and run AI models — take three to four months, founders and investors said. An entire engineering team can lose a quarter waiting for one card.
"Early-stage startups get credit, so compute affordability is not an issue in the first year. However, the challenge with AI compute is not just affordability, but access," said Rahul Agarwalla, managing partner at SenseAI Ventures. He said demand is outstripping supply and companies are trying to book capacity years ahead, which is hard for a startup that cannot forecast its own needs.
Payments platform Juspay can afford to front-load the spending. It still waits. "I wish I could spend more on GPUs, and I say that because it is hard to get GPUs allocated. Right now, I don't have enough GPUs, and I am bottlenecked. I have been given commitments that it will take three to six months to come," co-founder and chief operating officer Sheetal Lalwani said at a media briefing this month.
The Bengaluru company swung to a net loss of ₹90 crore in 2025-26 after posting its first annual net profit of ₹62 crore, as it stepped up spending on AI infrastructure and overseas expansion.
Ashish Taneja, founding partner and chief executive of growX ventures, put the cost in plain terms: "When a GPU card takes three or four months to arrive, a whole engineering team can lose a quarter, and that hurts far more than a chip being a generation old."
There are two ways to pay for AI. Frontier labs charge by the token, so the bill rises with usage. The alternative is running open-source models on hardware you control, where the cost stays flat. Lalwani said Juspay runs about 95% of its AI-assisted coding the second way. "Renting an 8-node GPU costs ₹25-35 lakh a month, while buying the same work as tokens from frontier labs could cost even 5x to 10x that."
The bottleneck has layers. Advanced AI chips need high-bandwidth memory, and only SK hynix, Samsung and Micron make it at scale. Micron said in December 2025 that its entire 2026 supply was committed. Power and data-centre space come next.
The government's IndiaAI Mission has onboarded more than 38,000 GPUs that startups and researchers can rent at subsidised rates. Taneja said the pool has scaled faster than expected and works for training language models in a data centre. It does not work for everyone. Robotics company CynLr must keep compute next to the robot, and a defence firm cannot use shared servers.
Gokul N A, founder of CynLr, offered the number that decides the argument: running a robot the way large language models run today could cost $2-3 million per robot per year, against $230,000 a year for a worker across three shifts.