The Cloud's New Normal: Less for More

S3 Standard storage has been $0.023/GB to start since December 2016.

AWS announced 107 price reductions between 2006 and 2021. The early reductions were significant because the early prices were high. Cloud was new, scale was small, and the providers were pricing to cover infrastructure they were still building out.

But the appeal was real. And for good reason. The introduction of elasticity and near-instant provisioning was game-changing for businesses accustomed to ordering hardware months before needed.

As adoption grew and economies of scale kicked in, hardware costs dropped and the providers passed some of that through. The m3 to m5 generation of EC2 instances saw a 28% price reduction on general purpose compute.

It would be the last time prices went down and the era of "use the latest instance; they're faster and less expensive" ended.

Think About It

Storage may not have changed, but compute certainly has. A lot. Look at M instances (4 GB RAM per vCPU), for example. First generations cost $0.350 per hour, down to $0.192/hr for fifth and sixth. And steadily up since. The pattern is consistent across offerings.

When confronted with this observation, I often told customers: "of course they did". AWS couldn't have been expected to continue lowering prices forever. They'd eventually hit zero. Not that customers would complain.

From m6i onward, prices flattened out and then started climbing. Each generation brought faster processors, and AWS will point to improved price-performance per unit of work. That's fair if your workload needed the extra performance. And the price-performance claim is real; it's measurable on most workloads.

You may recall, however, the majority of customers don't use the majority of their machines. Moving to m7i or m8i doesn't make them run better, it makes them cost more.

It’s no secret there are still customers running on m1 instances. That’s lazy. They could run on m5 for almost half the price. AWS likes this. Staying on m5 or m6i, on the other hand, is often a conscious decision. This one, not so much.

Remember It

I opened with S3 (object storage) pricing being static since 2016. When S3 launched in 2006 it was at $0.15 per GB per month. 10 years later, AWS had reduced it by nearly 85% to the current $0.023/GB. You get quantity discounts with S3, like buying in bulk, so prices go down as you store more. Economies of scale came into play and AWS learned what it needed about customer willingness to pay a premium for this (at the time) revolutionary new model.

On the other hand, the needle never moves on block storage. gp2 Elastic Block Storage (solid-state drives) has been $0.10/GB-month for as long as it's existed. Same with st1 and sc1 (conventional hard drives) at $0.045/GB-month and $0.015/GB-month, respectively. Spinning hard drives have dropped in price over time. EBS pricing has not.

When gp3 launched at re:Invent 2020 at $0.08/GB-month, AWS positioned it as a price reduction. It was a new product at a new price point. gp2 didn't change. If you were already running gp2, your bill stayed the same unless you did the migration work yourself. S3 Standard, as mentioned, hasn't moved since 2016. The underlying storage media has continued getting cheaper at roughly 30% per year throughout.

EFS (AWS' version of NFS, network filesystems) launched in 2016 at $0.30/GB-month. Much like S3, less-expensive tiers were later introduced, where "infrequent access" and "cold" came at a discount.

One has to wonder. Is the lack of pricing reductions due to having reached the maximum economy of scale, an implicit request to move to the latest and greatest technology as it's released, or just... Wait. What's the term for it again? Oh yeah, milking a cash cow?

Talk About It

Network has gone the other direction. Egress to internet is still $0.09/GB for the first 10 TB. Unless you're migrating off AWS, which is something I'll never understand. Public perception vs vendor lock-in armor? Tough to say.

Cross-AZ data transfer is still $0.01/GB each direction for traffic that never leaves the provider's own network. This is a killer. Sure, each AZ is geographically different but customers unaware of this would build multi-tier without locking the AZ for a single request get absolutely murdered on this.

NAT Gateway costs $0.045/hr to exist (roughly $32/month before a single byte flows through it) plus $0.045/GB in processing charges on the same traffic you're already paying egress on. Every private subnet that needs to reach the internet pays that toll. This particular cost feels like a tax on discrete egress; a huge win for security (server sees client, client can't see server) but a Juniper firewall in a data center does the same for free. Add cross-AZ charges for extra happy fun sauce.

In February 2024, AWS introduced a charge for public IPv4 addresses: $0.005/hr per address, active or idle. That's $43.80 per address per year that didn't exist before. The limited supply of IPv4 addresses is a real constraint, and charging for them is a reasonable way to push adoption of IPv6. The trick here is that the majority of customer workloads through browsers using web tools and software reaching APIs can run on IPv6. Dependencies in the stack, like essential vendor software, may not. It's a surgical decision. And most companies are so busy with other priorities they just pay.

Summarizing Essentials

Storage stayed flat. Network got more expensive. Compute: both.

Hands-Free for a Fee

RDS (managed databases) prices track EC2 generational pricing with a management premium on top, so the same upward drift in compute flows straight through. Fair, considering that beneath every managed service lies actual compute.

Lambda per-request and per-GB-second rates have not changed since the service launched. That's actually in contrast to RDS considering the underlying hardware is getting more expensive. Props there.

None of this is dramatic on its own. The pattern is consistent: the infrastructure underneath these services has gotten cheaper for the vendors and none of the savings have surfaced in the pricing.

The Bottom Line

The cloud business that Amazon retail originally funded to get off the ground is now the most profitable segment of the entire company, by a wide margin. Credit where credit is due: Amazon invented cloud compute (hyperscaling, whatever you want to call it). They built EC2 to solve a problem, realized the value, and transformed the internet into what it is today.

And now, the relationship is reversed.

AWS operating margins went from 28.5% in 2022 to 37.0% in 2024. In 2025, AWS posted $45.6 billion in operating income. Amazon North America retail posted $29.6 billion at a 7.5% margin. That's worth sitting with for a moment. AWS generated 54% more operating profit than the entire North America retail operation despite being less than a third of its revenue.

Google Cloud went from negative 12% operating margin to roughly positive 15% over the same period. Input costs rose across the board: electricity, memory, silicon. Margins expanded anyway because the providers stopped passing cost improvements through.

Combined capex across Amazon, Microsoft, Google, and Meta was $251 billion in 2024. Amazon alone is targeting approximately $200 billion for 2026, directed primarily at AI infrastructure. Jassy was direct about it: "We're not investing approximately $200 billion in capex in 2026 on a hunch." The AI demand is real. The money funding that investment comes from cloud operating profits on existing customer workloads, not from AI revenue.

The economics are either sobering or make you want to jump off the wagon.

The Land Grab Became a Money Grab

The early pricing was never meant to last. AWS was funded by Amazon retail margins. Google Cloud was funded by advertising revenue. Microsoft Azure was funded by Office and Windows, the same playbook Microsoft ran with Xbox for years, selling hardware at a loss until the install base justified a change in pricing.

It worked. Enterprise cloud adoption is at 94%. More than half of corporate IT spend is in cloud. Let's not forget those budgets include software licenses, service contracts, data centers and every desktop, laptop, phone, and electronic device at every company.

The competitive pressure that produced 60-plus EC2 price reductions doesn't exist anymore because the race that required it is over.

The hyperscalers have won the market. GG

With the market captured, the behavior shifted. Google raised every Workspace tier about 20% in January 2025 by making Gemini mandatory in the base price. Disabling Gemini in the admin console does not lower the bill. Microsoft is doing the same with M365 and Copilot, with increases between 5% and 33% effective July 2026. Neither company called these price increases. And yes, of course AI is a factor now.

Real cost pressures exist. Electricity in key data center markets rose 267% wholesale over five years. DRAM surged over 150% to 300% (depending on type) as manufacturers shifted production toward HBM for AI accelerators. Nobody should dismiss that. But margins expanded 8 to 22 points across AWS and GCP during the same period. The providers absorbed every cost increase, grew their margins on top of it, and passed nothing back.

Tech for on-prem has improved. Companies are retreating back to data centers. Specific workloads at first, and entire footprints in some cases. They can recognize the reductions in costs and services by returning to CapEx and amortizing. They can't amortize OpEx, and the discounts aren't there as incentives.

S3 is still $0.023/GB. I don't expect it to change any time soon.