Happy Birthday, Cloud: Twenty Years of Turning an Industry Upside Down

Twenty years ago today, on March 14, 2006 (Pi Day), Amazon quietly launched S3. Jeff Barr wrote the announcement blog post before catching a flight. The What's New entry was a single paragraph. No code examples. No demo. No keynote. Just a storage service, a flat price of $0.15 per gigabyte, and a REST API.

Nobody in the room knew they were watching an industry being born.

I've been thinking about this all week. If cloud were a person, it just turned twenty. Old enough to vote. Old enough to be called an adult. Still young enough to be figuring things out. And now it's carrying AI on its shoulders. A responsibility that would have been unimaginable when S3 was running on 400 storage nodes across 15 racks.

Here's the thing. The cloud story isn't really a technology story. It's an economics story. A story about how radically simple pricing became radically complex, how a side project inside a bookstore became a $128.7 billion business, and how the infrastructure built to serve web pages ended up powering the AI revolution.

Let me walk you through it the way I see it.

The beginning: Malloc for the Internet

S3 was internally called "Malloc for the Internet," a reference to the C memory allocation function. If you've written C, you know what malloc does: it gives you a chunk of memory, no questions asked. That was the design philosophy. Storage should be as invisible and fundamental as allocating memory in code.

The original pricing was absurdly simple:

Storage: $0.15 per GB per month
Data transfer: $0.20 per GB
Minimum fee: Zero

That was the entire pricing page. No tiers. No commitments. No reserved pricing. No 47-page whitepaper explaining discount structures. You stored data, you paid fifteen cents a gig. Done.

The entire service ran on about 1 petabyte of capacity and 15 Gbps of bandwidth. S3 was built with 8 microservices. Today it runs on over 300.

Five months later, EC2 showed up with a single instance type: the m1.small. One virtual CPU, 1.75 GB of RAM, 160 GB of local disk. Ten cents an hour. Bezos personally intervened to reduce the proposed rate from $0.15 to $0.10 because he wanted adoption over margins. There was no GUI. The AWS Management Console wouldn't arrive until 2010. You wanted a server, you called an API.

And here's the part that's easy to forget now: before this, running a web application meant buying physical servers, forecasting capacity months in advance, and investing tens of thousands upfront. Hardware costs ate 40–50% of total IT spend. An estimated 80% of IT budgets went to keeping the lights on rather than building anything new. By a server's fifth year, support costs were up 148%. By year seven, 300%.

SmugMug, the photo hosting service, was spending $80,000 per month on new storage hardware before S3. They switched in April 2006, just weeks after launch, and saved nearly $1 million in the first year. Their CEO called S3 "The Holy Grail."

That's the thing about revolutions. They don't announce themselves. They just show up on Pi Day with a one-paragraph blog post and quietly make everything that came before obsolete.

The price war years: a race that changed everything

For most of its first decade, cloud had one dominant story: prices go down.

AWS has cut prices at least 134 times as of late 2023. Probably over 140 by now. Bezos laid out the philosophy in his 2005 shareholder letter with remarkable clarity: "When we lower prices, we go against the math that we can do, which always says that the smart move is to raise prices." He kept lowering them anyway. Not because competitors forced his hand. 51 of those cuts happened before there was any competitive pressure.

S3 went from $0.15/GB to $0.023/GB. An 85% drop. EC2 workloads in 2017 cost up to 73% less than the same workloads in 2014. The flywheel was textbook: lower prices, more adoption, greater scale, lower costs, lower prices again.

Then came March 25, 2014. The day the price war went nuclear.

Google fired first: Compute Engine prices cut 32%, Cloud Storage cut 68%, BigQuery cut 85%. Google's SVP Urs Hölzle argued cloud pricing should follow Moore's Law.

Within 24 hours, AWS responded: S3 cut 51%, EC2 cut 30–40%, RDS cut 28%. Microsoft matched within days. Blog commenters thanked Google for "reducing our bills."

I love that detail. Users didn't care who won the war. They just wanted cheaper cloud.

But here's what nobody talks about: the price cuts largely stopped after 2016. S3 Standard hasn't budged from $0.023/GB in over nine years. Google Cloud actually raised storage prices in 2024, up to 25% for multi-region. The "race to zero" narrative was a 2006–2016 story. What followed is a different story entirely.

A race to complexity.

From a napkin to 300,000 SKUs

This is where it gets uncomfortable.

AWS went from 3 services in 2006 to over 300 today. From 1 EC2 instance type to 1,167. From a pricing page that fit on a napkin to approximately 300,000 different pricing SKUs. That's 7.5 times more variants than a typical grocery store's product catalog. AWS Cost and Usage Reports routinely exceed Excel's row limit of 1 million rows. Amazon literally has to split the monthly file because it's too large to open.

Let me give you one example that captures the absurdity. There are now five different ways to pay for a single EC2 instance:

On-Demand: pay by the second, no strings
Reserved Instances: 1 or 3-year commitment, up to 72% off, with three payment options, two types, and two scopes
Spot: bid on spare capacity, up to 90% off, but your instance can vanish with 2 minutes' notice
Savings Plans: commit to a dollar-per-hour spend, three sub-types, up to 72% off
Dedicated Hosts: physical server dedication for compliance or licensing

And here's the kicker: Savings Plans launched alongside Reserved Instances, not replacing them. AWS still recommends RIs in certain scenarios. So enterprises manage both simultaneously. I've seen teams spend weeks each quarter just on commitment planning. It's become its own job function.

Then there's the hidden cost layer that catches everyone eventually.

Egress: Sending data into AWS is free. Sending it out costs $0.09/GB. One analyst estimated that's an 8,000% markup over actual bandwidth costs. Gartner puts egress at 10–15% of a total cloud bill.

NAT Gateways: $0.045/hour even with zero traffic, plus $0.045/GB for data processing. A standard 3-availability-zone deployment costs $98.55/month before any data moves. I've seen a team hit $9,000/month in NAT Gateway charges from accumulated small data movements nobody was tracking.

IPv4 addresses: As of February 2024, AWS charges $0.005/hour for every public IPv4 address. Something that was previously free. That's $43.80/year per IP. Analysis showed customers saw an average 2.6% bill increase overnight. For startups with lots of microservices, it can hit a quarter of their monthly bill.

The bill shock stories are legendary. A $23/month site that cost $2,657 overnight when a file went viral. A $47,000 bill that should have been $8,000. A company that discovered 47 abandoned dev environments running 24/7 across feature branches.

It's like ordering what you think is a simple meal at a restaurant and discovering there's a cover charge, a service fee, a plate rental fee, a condiment surcharge, and a 9% tax for eating on the left side of the room. The food might be great. But the bill makes you question your life choices.

For most people, cloud meant AWS

Let me state something that's obvious but rarely said out loud: for the better part of a decade, AWS was cloud. Not a cloud provider. The cloud.

When someone said "move to the cloud" between 2006 and roughly 2015, they meant AWS. Microsoft Azure launched in 2010. Google Cloud Platform went GA in 2012. But AWS had a multi-year head start, the developer mindset, and the ecosystem. By the time competitors showed up in earnest, AWS had already built the playbook everyone else was copying.

The numbers tell the story. AWS still holds 28% of the cloud infrastructure market. Azure is at 21%. Google Cloud at 14%. The Big Three collectively control 63%. But AWS's share has gradually eroded from roughly 33% in 2020. Not because AWS got worse, but because Azure and Google Cloud got serious.

Google Cloud didn't turn profitable until Q1 2023, after years of operating losses that hit $5.6 billion as recently as 2020. They were literally paying billions for the privilege of competing with AWS. That's how dominant AWS's position was. And in many ways, still is.

Here's what's interesting about multi-cloud today: comparing costs across providers has become its own industry. AWS m5.xlarge is not the same as Azure StandardD4sv3 is not the same as GCP e2-standard-4, even though all three are roughly 4 vCPU and 16 GB of RAM. Billing granularity differs. Discount structures are incompatible. Regional pricing varies in ways that defy logic.

The FinOps Foundation had to create FOCUS (the FinOps Open Cost and Usage Specification) specifically because no standardized format existed for comparing cloud bills across providers. I wrote about the reality of FOCUS adoption recently, and let me just say: beautiful idea, painful execution.

Entire companies (Apptio, nOps, Cast.ai, Vantage) exist solely to help organizations navigate this maze. Think about that. An industry had to be built just to help people understand the bills from another industry.

Twenty and carrying AI on its shoulders

Now here's where the birthday metaphor gets real.

At twenty, cloud isn't just an adult. It's become the load-bearing wall of the entire AI economy. Without cloud infrastructure, AI as we know it simply doesn't reach everyone. Full stop. The idea that any startup, any researcher, any developer can access supercomputer-class compute with a credit card? That's cloud. That's the unlock.

A P5.48xlarge on AWS (8 NVIDIA H100 GPUs, 192 vCPUs, 2 TB of RAM) costs roughly $33/hour after AWS cut P5 prices 45% in June 2025. Anyone can rent a supercomputer for about $4 per GPU-hour. In 2006, you couldn't even get a virtual machine with 1.75 GB of RAM without calling an API that barely worked. The progression is staggering.

The capital being poured into cloud-AI infrastructure is hard to comprehend. Combined hyperscaler CapEx hit $443 billion in 2025. The projection for 2026: $602 billion. Amazon alone is spending over $100 billion. Microsoft: $80 billion. Google: $75 billion. Goldman Sachs projects $1.15 trillion in cumulative CapEx from 2025 to 2027. In 2025, tech CapEx as a percentage of GDP nearly matched the combined scale of the broadband buildout, the interstate highway system, and the Apollo program.

Cloud infrastructure investment is on par with the highways and the moon landing. Combined.

And AI has introduced entirely new pricing dimensions that didn't exist five years ago. Per-token billing: GPT-4o charges $2.50 per million input tokens, $10.00 per million output tokens. The spread between the cheapest and most expensive model is 420x. Reasoning tokens (the model "thinking") cost money you can't even see. Cached versus uncached inputs. Batch versus real-time. It's a new layer of complexity stacked on top of the existing 300,000 SKUs.

In the 2026 State of FinOps survey, 98% of respondents reported managing AI spend. In 2024, that number was 31%. In two years, AI cost management went from niche concern to universal reality.

Cloud didn't just enable AI. Cloud is the accelerator. The delivery mechanism. The reason a researcher in Bengaluru can access the same GPU cluster as a well-funded lab in San Francisco. Without cloud, AI would still exist, but it would be locked behind the doors of organizations rich enough to build their own data centers. Cloud democratized compute in 2006. It's democratizing intelligence in 2026.

The discipline that cloud's complexity created

Every action has an equal and opposite reaction. Cloud's complexity created FinOps. And I say that as someone who lives in this space every day.

The FinOps Foundation started in February 2019, born out of Cloudability's quarterly customer meetings where practitioners kept asking for a vendor-neutral community to discuss one specific problem: why is managing cloud spend so hard? From 1,500 members in mid-2020, it's grown to over 96,000 practitioners across 15,000 companies, including 93 of the Fortune 100.

The fundamental shift that created FinOps is often overlooked. Cloud moved IT spending from capital expenditure (buying servers on 3–5 year depreciation cycles, predictable, centrally controlled) to operating expenditure. Variable. Usage-based. Decentralized. Any engineer can spin up resources instantly. Finance teams accustomed to annual procurement cycles suddenly faced monthly invoices that fluctuated wildly.

And the waste numbers haven't improved much despite years of trying. Flexera's State of the Cloud survey has tracked self-estimated waste for 14 years now: 30% in 2021, 32% in 2022 (it actually went up), 27% in 2024 and 2025. The Harness FinOps report projected $44.5 billion in cloud waste for 2025 alone. Average CPU utilization sits at 15–20%.

It's like the entire industry collectively decided to rent apartments four times bigger than needed, then hired consultants to figure out which rooms to stop heating. The consultants built a $13.5 billion market doing exactly that.

Cost optimization has been the top cloud initiative for nine consecutive years in Flexera surveys. Organizations exceed cloud budgets by an average of 17%. Only 2% of CIOs (two percent!) report spending less on cloud than projected. And yet 84% say managing cloud spend is their number one challenge.

If that doesn't tell you something about the gap between cloud's promise and its operational reality, I don't know what does.

What I think about at twenty

I spend my days building a platform that helps organizations close the gap between finding cloud waste and actually eliminating it. So I have a particular vantage point on this birthday.

Twenty years ago, Bezos saw a 300-year-old brewery's electric generator in a Luxembourg museum and had an insight: "Everybody has their own data center, and that's not going to last. You're going to buy the compute off the grid." He was right. We stopped generating our own electricity. We moved to the grid.

But nobody warned us the electric bill would be 300,000 line items long.

Cloud storage costs have fallen 85%. The cognitive cost of understanding your cloud bill has increased by orders of magnitude. S3 stores over 500 trillion objects now. It handles 200 million requests per second. It holds hundreds of exabytes across 39 regions. From 8 microservices to 300. From one pricing tier to a labyrinth of Standard, Infrequent Access, Intelligent-Tiering, Glacier, Glacier Deep Archive, Glacier Instant Retrieval, Express One Zone, Tables, and Vectors.

That's the paradox of cloud at twenty. The technology has delivered on every promise. The economics have created a new category of problems. And now AI, the most capital-intensive technology wave in human history, is running on top of it all.

Cloud at twenty isn't a teenager anymore. It's a young adult carrying more weight than anyone expected. It's the foundation of a $419 billion market. It's the accelerator that's bringing AI to every corner of the economy. It's also a 300,000-SKU pricing puzzle that spawned an entire discipline just to make sense of the bills.

Happy birthday, cloud. You've come a long way from that one-paragraph blog post and $0.15 a gigabyte. The next twenty are going to be even wilder.

And honestly? I wouldn't bet against you.

Nishant Thorat is the founder of CloudYali, a FinOps Action Platform governing cloud spend. He's been working in cloud cost management long enough to remember when the pricing page actually fit on a single screen.