Cloud Compute Myths: Why On-Demand Isn't the Enemy and Spot Isn't Free

Imagine you’re two weeks from launch, and your CFO just asked you to justify a $40,000 monthly AWS bill. You’ve read the blog posts: “Savings Plans slash costs by 66%!” “Spot Instances are 90% off!” But you also remember the horror story from your last gig—the one where a Spot fleet got interrupted mid-batch and the retry logic was a mess. So what do you actually do?

We’ve been there. And we’ve learned that the cloud compute market—now a $443.6 billion industry (MarketLine)—is full of half-truths and oversimplifications. Let’s bust the six myths we hear most often.

Myth 1: On-Demand Is Always a Waste of Money

Yes, On-Demand is the most expensive way to run a steady workload. But it’s also the only purchase option that gives you zero commitment and instant scalability. For spiky, unpredictable, or short-lived workloads—think a two-day marketing campaign or a quick data-science experiment—On-Demand can be the smartest choice. You’re paying for flexibility, and for many teams, that’s worth the premium.

The real waste is running steady-state production on On-Demand when you could commit to a Savings Plan or Reserved Instance. If you know you’ll have a database running 24/7 for the next year, leaving it On-Demand is like paying for a hotel room when you could buy the house. So, don’t blindly follow the “always use Savings Plans” advice—but don’t ignore it either. Match the purchase option to the workload’s predictability.

Myth 2: Spot Instances Are Too Risky for Anything Important

Spot Instances can save up to 90% off On-Demand (AWS EC2 Spot), and they’re perfect for stateless, fault-tolerant, or flexible workloads like big data, CI/CD, and web servers. The catch? They can be interrupted with a two-minute warning. That’s scary if you’re running a stateful application without checkpoints.

But here’s the thing: many modern workloads are already distributed and can tolerate interruption. If you design your jobs to be idempotent and use queue-based retries, Spot can be a game-changer. We’ve run batch processing on Spot for months without a single lost job—because we built for failure from day one. The risk isn’t the interruption; it’s the lack of planning for it.

Myth 3: Serverless Is Always Cheaper

Serverless compute like AWS Lambda and Azure Functions sounds perfect: you pay only for what you use, and the free tiers are generous. But that model breaks down for high-utilization, predictable workloads. If you’re running a service that receives millions of requests per hour, the per-request and per-GB-second charges add up fast.

For example, beyond the free tier, Lambda charges $0.20 per million requests and $0.0000166667 per GB-second (AWS Lambda Pricing). Run a 1 GB function for 10 seconds per request at 10 million requests a month, and you’re looking at a hefty bill. Meanwhile, a $0.096/hour m5.large instance (AWS EC2 Price List API) could handle the same load for a fraction of the cost.

Serverless shines for low and spiky traffic, but for steady, high-volume workloads, a properly sized VM or container is often cheaper. Don’t let the “pay-per-use” marketing fool you—it’s still a pricing model, and it can be more expensive than you think.

Myth 4: More vCPUs Always Means Better Performance

We’ve all been tempted to bump up the instance size to fix a performance issue. But that’s often a waste of money. For many workloads, the bottleneck isn’t CPU—it’s memory, network, or I/O. And sometimes, the architecture is the problem, not the hardware.

Instead of blindly scaling up, look at your metrics. If you’re running a memory-bound application, a memory-optimized instance like R5 or X2idn (AWS Documentation) will serve you better than a compute-optimized C5. And if your workload is bursty—like a development server that’s idle most of the day—consider a burstable T3 instance. T3 instances accumulate CPU credits when running below baseline, letting you burst above it when needed (AWS EC2 T3). That’s a cheap way to handle short spikes without paying for full-time CPU capacity.

Myth 5: The Cheapest Instance Type Is Always the Best Deal

Price per vCPU-hour is a tempting metric, but it doesn’t tell the whole story. A Graviton-based instance might cost up to 20% less than a comparable x86 instance (AWS Graviton), but if your software isn’t optimized for ARM, you could see worse performance. Similarly, a GPU instance like p4d.24xlarge costs $21.957642/hour (AWS EC2 Price List API), but if your workload can use it, it can cut training time by 4x compared to previous-gen GPUs (AWS EC2 P5)—saving you money in the long run.

We once migrated a batch job from a c5.2xlarge to a Graviton m7g.large (which costs $0.0816/hour versus $0.1008/hour for m7i.large) and saw a 15% performance improvement because the code was compiled for ARM. The cost per job dropped by more than the raw price difference. The lesson: benchmark, don’t assume.

Myth 6: The Free Tier Makes Cloud Compute Basically Free

Yes, AWS gives you up to $200 in credits and Azure gives you $200 for 30 days, and Google Cloud gives you $300 for 90 days. But these are trial offers, not long-term strategies. After the credits expire, you’re paying full price.

What’s more, the always-free tiers are limited. AWS’s free tier includes 1 million Lambda requests and 400,000 GB-seconds per month (AWS Lambda Pricing), which sounds generous until you calculate that it’s only enough for a tiny, low-traffic app. Azure Functions’ free grant is similar (Azure Functions Pricing), and Google Cloud Functions gives you 2 million invocations per month (Google Cloud Functions Pricing). But if you’re running a production workload, you’ll blow through these limits in a day.

The free tier is a hook, not a solution. Plan for the post-free-tier reality from day one, or you’ll be shocked when the bill arrives.

The Bottom Line

Cloud compute pricing is not a one-size-fits-all puzzle. On-Demand has its place, Spot is viable if you design for it, serverless is great for spiky traffic but dangerous for steady loads, and the cheapest instance type isn’t always the best value. Stop chasing the single “best” option and start matching the tool to the workload. That’s how we keep our bills under control—and our sanity intact.

Sources

• AWS Documentation - https://docs.aws.amazon.com/ec2/
• AWS EC2 Pricing - https://aws.amazon.com/ec2/pricing/
• AWS EC2 Spot - https://aws.amazon.com/ec2/spot/
• AWS Lambda Pricing - https://aws.amazon.com/lambda/pricing/
• AWS EC2 Price List API - https://pricing.us-east-1.amazonaws.com/offers/v1.0/aws/AmazonEC2/current/us-east-1/index.json
• MarketLine - https://www.marketresearch.com/MarketLine-v3883/Global-Cloud-Computing-43554509/