The cloud bill isn't the problem. Not knowing what's inside it is.
OpenAI spent $8.7 billion on inference alone in the first three quarters of 2025, according to leaked Microsoft financial documents. That's more than double what they spent in all of 2024. They have hundreds of millions in revenue and still couldn't make the unit economics work.
Your startup is not OpenAI. But the pattern is the same: cloud and AI costs don't scale linearly with your product anymore. One new feature using a hosted model, one ephemeral environment left running over a long weekend, one GPU instance nobody remembered to shut down — and your bill for the month looks nothing like your forecast.
The Flexera State of the Cloud 2025 report found that 27% of cloud spend is wasted — and this figure has been consistent for five consecutive years despite "cloud cost optimization" being the #1 stated priority every single year. The waste isn't going down because knowing you have a problem isn't the same as being able to see it.
What changed — and why 2025 is different from 2022
Three years ago, a startup's cloud bill was mostly compute and storage. Predictable. Scaling slowly. Now it's managed services, LLM API calls, GPU workloads, preview environments that spin up automatically per pull request, data egress from multi-region setups, and Kubernetes clusters running at generous resource limits because nobody wanted to deal with OOMKilled pods during a demo.
The dangerous part: GPU instances cost 5–10x standard compute. An H100 instance sitting idle overnight is roughly $24 per GPU in wasted spend — per night. A team running model fine-tuning jobs that finish at 3am but don't terminate until someone checks in the morning is burning real money in their sleep. This isn't hypothetical. Midjourney moved its inference fleet from Nvidia A100/H100 GPUs to Google TPU v6e in Q2 2025 and cut monthly inference costs from $2.1M to under $700K — a 65% reduction. Six weeks of migration work, 11-day payback period.
"78% of organizations detect cloud cost anomalies late. Only 22% notice them quickly." — CloudZero, State of Cloud Cost 2024
The anomaly detection gap is the actual problem. By the time you see the bill, the damage is already done. A Harness study from February 2025 found that enterprises take an average of 31 days to identify idle or orphaned resources, and 25 days to detect overprovisioned workloads. For a startup burning $50k/month, that's a lot of runway to lose before anyone notices.
FinOps isn't a tool category. It's an ownership model.
The reason cloud waste sits at 27% despite everyone knowing about it: nobody owns the problem end-to-end. Finance sees the invoice. Engineering sees the infrastructure. Neither sees the full picture. FinOps — Financial Operations for cloud — is just the practice of giving one person or function the job of sitting between those two views and acting on what they see.
WPP, the advertising group, ran this experiment properly. Three months after deploying FinOps practices, they had saved $2 million. Over the following year, that scaled to a 30% annual reduction in cloud spend. The lever wasn't exotic tooling — it was autogenerated right-sizing recommendations on instances they'd never looked at.
Lyft did a version of the same thing with AWS Cost Management, cutting costs 40% in six months. The savings came from identifying idle resources and moving predictable workloads to reserved instances — resources that had been visible in their bill the whole time, just not attributed to anyone.
Three things worth doing before you buy any tooling
The first is tagging. Every cloud resource needs environment, team, and project tags. Without them, your bill is a single number with no internal structure. With them, you can ask "which feature is costing us $8k/month" and get an answer. The FinOps Foundation's 2025 framework puts cost allocation as the second-highest priority for mature FinOps teams — only workload optimization ranks higher. You can't optimize what you can't attribute.
The second is TTLs on non-production environments. Every staging, preview, and dev environment should have an automatic shutdown policy. A Lambda function or Cloud Scheduler job that terminates untagged environments after 72 hours is a two-hour engineering task that recovers 4–8% of monthly spend. Non-production environments left running over weekends are one of the most common and entirely avoidable sources of waste.
The third is a 30-day utilization review. Pull CPU and memory metrics for every running instance. Anything sitting at under 20% CPU utilization is a candidate for downsizing. This is not a performance risk — it's an instance that was provisioned based on peak estimates and never revisited. Right-sizing 100 overprovisioned instances typically saves $75k/month. Most teams have access to this data in their cloud provider's native console and have never looked at it.
The AI cost problem specifically
If you're calling hosted LLM APIs (OpenAI, Anthropic, Gemini), your inference costs are now a variable that can change dramatically with product usage patterns. Token consumption doesn't scale the way API calls used to. One new workflow that chains multiple model calls together can 10x your inference spend before it shows up in your budget review.
The teams managing this well are tracking token consumption per feature, setting per-environment spend limits, and building cost awareness into the deployment pipeline — not discovering overruns in a monthly finance review. Deloitte projects that companies implementing FinOps practices will save a combined $21 billion globally in 2025. That money isn't coming from cutting features or downgrading infrastructure. It's coming from looking at what's already running.
Further reading: Flexera State of the Cloud 2025 · Harness FinOps in Focus 2025 · Deloitte TMT Predictions 2025 — FinOps
I am Ayesha Siddiqua – I work at the crossroads of cloud strategy and startup growth. I've had hundreds of conversations with CTOs, Heads of Engineering, and founders trying to navigate the same hard questions — when to hire, what to automate, how much to spend on infrastructure, and when "good enough" is actually good enough. I don't write about DevOps from a purely technical lens. I write about it because I believe the infrastructure decisions that get made (or ignored) in a startup's first two years quietly determine whether the company scales or stalls. I'm associated with Frigga Cloud Labs, a DevOps consultancy built for growing startups.
This blog is my way of contributing to the conversation. If it made you think, I'd love to hear from you.
Let's connect on LinkedIn