Startups rethink cloud use as AI bills mount

Own GPUs, open models and tighter workload management gain traction.

Rising AI and compute-intensive workloads are pushing up cloud infrastructure costs for startups, with growth-stage firms now spending more than twice as much on cloud as they did two years ago, prompting companies to adopt tighter cost controls and explore alternatives to conventional cloud services.

A growth-stage startup with a reasonably scaled product that spent around Rs 8-20 lakh a month on cloud infrastructure two years ago now typically spends Rs 18-45 lakh, according to industry experts. For late-stage and enterprise-scale companies, monthly cloud spending has risen from Rs 50-80 lakh to Rs 80-150 lakh over the same period. The increase has been driven by more complex workloads, AI experimentation and the addition of GPU computing to infrastructure stacks.

The sharp rise is also changing how startups look at cloud expenditure. With companies already under pressure to control customer acquisition, hiring and marketing costs, cloud infrastructure is increasingly being treated as a major P&L (profit and loss) item rather than simply an engineering expense.

Proptech unicorn NoBroker, for instance, has cut its AI usage costs by 80% in less than 24 months by combining frontier models with newer open-weight models deployed on both cloud and local infrastructure. “We constantly optimise the breakeven point between utilising our own provisioned GPUs for self-hosted models versus falling back to provider-hosted environments,” Akhil Gupta, co-founder and CPTO of NoBroker and founder of ConvoZenAI, told FE. 

Gupta said business workloads typically peak during the day and fall at night, leaving provisioned GPUs underutilised even though the cost remains unchanged. NoBroker has therefore shifted heavy and predictable workloads to its own GPUs, improving utilisation and reducing its dependence on provider-hosted computing. 

Other startups are focusing on eliminating inefficiencies within their existing cloud infrastructure. Quick-commerce logistics firm Zippee has reduced cloud spend per shipment by 30-35% over the past year by right-sizing overprovisioned instances, moving batch workloads to spot capacity and tightening data-retention policies.

The company also identified idle development environments, old snapshots, unnecessary inter-service data transfers and rising observability costs as areas where cloud expenditure can build up without directly contributing to business growth. “Workload-level ownership, right-sizing and storage discipline” helped bring down costs, said Mohit Singh, head of Product, Zippee.

As cloud bills rise, a growing number of startups are turning to specialised providers for optimisation. Firms such as Zepto, Lenskart and MobiKwik have used third-party vendors to manage cloud costs. MobiKwik, for instance, has saved 12% on its overall AWS costs over the past year with the help of a third-party provider, according to sources.

Cloud optimisation firms including CloudKeeper and Utho Cloud are seeing growing demand from startups and enterprises. CloudKeeper said its customers have achieved an average reduction of around 20% in cloud costs, while Utho Cloud said its clients typically see reductions of around 60%, with some cases reporting savings of 75-80%.

The financial impact can be significant for startups. Utho Cloud founder and CEO Manoj Dhanda said cloud spending can rise from 8-10% of operating expenditure at an early stage to 18%, 25% or even 35% as companies scale without adequate FinOps discipline. For AI-first and GPU-heavy companies, the share can be higher.

With cloud expenditure increasingly competing with product development, hiring and customer acquisition for scarce capital, startups are now being forced to treat infrastructure efficiency as a core financial discipline rather than a back-end technology exercise.

The article was originally published in Financial Express