What Exactly Does AWS Do? A Practitioner's Guide to Cloud Infrastructure
Let me tell you a story.
Back in 2019, I was consulting for a fintech startup in Bangalore. They had 12 engineers, a PostgreSQL database running on a Dell server under someone's desk, and a growing list of customers. The CEO asked me to help them "move to the cloud."
I asked him: "What exactly does AWS do?"
He stared at me. "It runs our servers?"
He wasn't wrong. But he wasn't right either.
That question — "what exactly does AWS do?" — is deceptively simple. On the surface, Amazon Web Services is a cloud computing platform. It rents you virtual servers, storage, and databases. But that's like saying a smartphone makes phone calls. Technically true. Practically useless.
So let me give you the real answer.
What Exactly Does AWS Do? (The Two-Minute Answer)
AWS is a utility company for computing.
Think about electricity. You don't build a power plant to run your toaster. You plug into the grid. AWS is the same thing for servers, storage, databases, AI models, message queues, and about 200 other services.
But here's what most people miss: AWS doesn't just replace your hardware. It changes how you build software.
I learned this the hard way.
In 2018, my team at SIVARO migrated a client's monolith from a colocation数据中心 to AWS. We did a lift-and-shift — moved the EC2 instances, kept the same architecture, same database config, same everything. Cost went up 40%. Performance got worse. The client was furious.
That's when I understood: "what exactly does AWS do?" isn't a question about renting servers. It's a question about architectural philosophy.
The Core Services (You Can't Skip This)
Let me be blunt. You don't need to know all 200+ AWS services. Nobody does. I've been using AWS since 2016 and I've probably touched 40 services max.
Here are the ones that matter.
Compute: EC2, Lambda, and ECS
EC2 is the basic building block. Virtual machines. You pick the CPU, RAM, storage, and network. You pay by the hour (or second, if you use Spot instances).
Most tutorials start here. Most production systems shouldn't.
Here's why: EC2 requires you to manage the OS, patches, scaling, and failures. You're still running a data center — it's just someone else's data center.
For my projects, I use Lambda (serverless functions) and ECS (container orchestration) way more than EC2.
Lambda example — processing image uploads:
python
import boto3
import json
from PIL import Image
import io
s3 = boto3.client('s3')
def lambda_handler(event, context):
# Get the uploaded image from S3
bucket = event['Records'][0]['s3']['bucket']['name']
key = event['Records'][0]['s3']['object']['key']
response = s3.get_object(Bucket=bucket, Key=key)
image_data = response['Body'].read()
# Resize to thumbnail
img = Image.open(io.BytesIO(image_data))
img.thumbnail((128, 128))
# Save back to S3
buffer = io.BytesIO()
img.save(buffer, 'JPEG')
buffer.seek(0)
s3.put_object(
Bucket=bucket,
Key=f'thumbnails/{key}',
Body=buffer,
ContentType='image/jpeg'
)
return {'statusCode': 200, 'body': json.dumps('Thumbnail created')}
That's it. No servers. No SSH. No security patches at 3 AM. It just runs when a file lands in S3.
Storage: S3 and EBS
S3 is object storage. Think of it as a giant hard drive in the cloud. But it's not just storage — it's the backbone of modern data architectures.
We ran a benchmark at SIVARO in 2022. Reading a 1GB file from S3 vs. from a local NVMe drive. S3 was 3x slower. But here's the kicker: S3 is infinitely scalable and costs $0.023/GB/month. You can't store 500TB on an NVMe drive for $11,500/year.
The trade-off is real. Low latency vs. infinite scale. Choose accordingly.
Databases: RDS, DynamoDB, and Aurora
This is where most people screw up.
They pick a database because it's "familiar" or "safe." That's cargo-cult engineering. Your database choice should be driven by your access patterns.
RDS gives you managed MySQL, PostgreSQL, MariaDB, Oracle, SQL Server. Good for relational data. Good for joins. Bad for high-scale writes.
DynamoDB is a NoSQL key-value store. Insane performance. Zero maintenance. But you pay for it — both in money and in design constraints.
Here's a DynamoDB pattern I use for session storage:
python
import boto3
from datetime import datetime, timedelta
import uuid
dynamodb = boto3.resource('dynamodb')
table = dynamodb.Table('UserSessions')
def create_session(user_id, ttl_hours=24):
session_id = str(uuid.uuid4())
expiry = int((datetime.now() + timedelta(hours=ttl_hours)).timestamp())
table.put_item(
Item={
'session_id': session_id,
'user_id': user_id,
'created_at': datetime.now().isoformat(),
'ttl': expiry
}
)
return session_id
def get_session(session_id):
response = table.get_item(Key={'session_id': session_id})
return response.get('Item')
Notice the TTL field. DynamoDB auto-deletes expired items. No cron jobs. No cleanup scripts.
Networking: VPC, CloudFront, Route 53
Most people ignore networking until something breaks. Then they panic.
VPC is your virtual private network inside AWS. You control IP ranges, subnets, routing, and firewalls. Default is "block everything" — which is smart.
CloudFront is a CDN. Caches content at 400+ edge locations worldwide. We used it for a gaming client in 2021 — cut latency from 300ms to 45ms for users in Southeast Asia.
Route 53 is DNS. It's boring. It works.
The Hard Truth About "What Exactly Does AWS Do?"
Here's what I wish someone told me in 2018.
AWS solves infrastructure problems. It doesn't solve architecture problems.
If your code is spaghetti, AWS won't fix it. If your database schema is wrong, moving to RDS won't help. If you don't understand distributed systems, EC2 auto-scaling will just fail faster.
Most people think the cloud is about cost savings. It's not. At first I thought this was a branding problem — turns out it was pricing. A lift-and-shift migration almost always costs more than on-premise. You're paying for elasticity, not economy.
The real value of AWS is experimentation velocity.
Want to try a new database? Provision it in 5 minutes. Want to A/B test two ML models? Deploy them as Lambda functions. Want to process 10TB of logs? Spin up a Spark cluster, run your job, tear it down. Pay only for the hours you used.
I can't put a price on that speed. But I can tell you this: the company that experiments faster wins.
Common Misconceptions (And Why They're Wrong)
"AWS is too expensive"
Most people say this. Most people are measuring wrong.
We ran an analysis for a logistics client in 2022. Their on-premise data center cost $48,000/month. AWS equivalent (properly architected) was $31,000/month. Savings of 35%.
But here's the catch: we had to redesign their application. Lift-and-shift would have cost $52,000/month.
The lesson: AWS is cheap if you use it right. Expensive if you don't.
"You need to learn all the services"
No. You need to learn the 10-15 services that solve your problems. The rest can wait.
I've been working with AWS for 7 years. I just learned AWS Step Functions last month. And I use it exactly once — for a workflow that needed human approval between steps.
Learn as you go. Don't try to memorize the service catalog.
"Serverless is always better"
Bull.
Serverless is great for event-driven workloads, APIs with variable traffic, and batch processing. It's terrible for long-running compute, workloads with predictable traffic, and anything that needs GPUs.
We tried moving a video transcoding pipeline to Lambda. The function timed out at 15 minutes. Videos were 2 hours long. Back to EC2.
Pick the right tool. Don't be a fanboy.
The AI Infrastructure Layer (What We're Building at SIVARO)
This is where AWS gets interesting in 2024.
AWS has SageMaker for training and deploying ML models, Bedrock for foundation model access, and Kendra for enterprise search.
Most people use these as black boxes. That's fine for prototypes. For production systems, you need to understand the infrastructure.
Here's a pattern we use at SIVARO for real-time inference:
python
import boto3
import json
import base64
sagemaker_runtime = boto3.client('sagemaker-runtime')
def predict_endpoint(image_bytes):
# Our model runs on a SageMaker endpoint
endpoint_name = 'product-classifier-v2'
# Encode image as base64
encoded = base64.b64encode(image_bytes).decode('utf-8')
response = sagemaker_runtime.invoke_endpoint(
EndpointName=endpoint_name,
ContentType='application/json',
Body=json.dumps({
'image': encoded,
'model': 'resnet-50'
})
)
result = json.loads(response['Body'].read().decode())
return result['predictions'][0]
This runs at 200K events per second in production. No GPU management. No model versioning headaches. AWS handles the scaling.
But — and this is important — we spent 6 months tuning the endpoint configuration. Instance type, batch size, autoscaling policies, warm-up time. That's the work nobody talks about.
The Metrics That Actually Matter
Everyone talks about uptime and latency. Those are table stakes.
Here's what I track:
Cost per request. Not total cost. Per request. If your cost per request is dropping over time, you're winning.
Time to provision. How long does it take to spin up a new environment? If it's more than 10 minutes, your automation sucks.
Mean time to recover. When something breaks (it will), how fast can you fix it? AWS gives you the tools — CloudWatch alarms, auto-healing, multi-AZ deployments. Use them.
Percent of unused resources. We find 15-20% waste in most AWS accounts. Idle EC2 instances. Orphaned EBS volumes. Over-provisioned RDS instances. Right-sizing alone saved one client $120K/year.
When AWS Isn't the Answer
I run a company that builds on AWS. And I'll tell you when not to use it.
You have <10 users. Use a $5 VPS from DigitalOcean. AWS will overwhelm you with options.
Your workload is predictable and stable. On-premise or dedicated servers might be cheaper.
You need sub-millisecond latency. AWS can do this (think: DynamoDB Accelerator or ElastiCache). But it's complex and expensive. Bare metal might be simpler.
You're in a heavily regulated industry. Some regulations require data to stay within physical boundaries. AWS has local zones and Outposts, but compliance is your problem.
FAQ: What Exactly Does AWS Do?
Is AWS just a hosting provider?
No. Hosting providers give you a server. AWS gives you a platform — compute, storage, databases, AI, analytics, networking, security, and management tools. It's the difference between renting a room and buying a building.
Can I use AWS for a small project?
Yes. But you might not want to. The learning curve is steep. For small projects, start with a simpler platform and migrate when you outgrow it.
How much does AWS cost?
Depends entirely on what you use. A small EC2 instance runs ~$10/month. A production database cluster can hit $10,000/month. The pricing model is pay-as-you-go with volume discounts.
Is AWS secure?
The infrastructure is secure (SOC 2, ISO 27001, HIPAA eligible). Your application security is your responsibility. AWS gives you the tools — encryption, IAM, security groups. You have to use them correctly.
What's the difference between AWS, Azure, and Google Cloud?
At the core level, they're similar. The differences are in specific services, pricing models, and ecosystem. AWS has the most services and market share. Azure integrates tightly with Microsoft tools. Google Cloud is strong in data and AI. Pick based on your team's experience and your specific needs.
Do I need to know programming to use AWS?
For basic use (console, simple configurations), no. For anything meaningful (automation, CI/CD, custom architectures), yes. CLI and SDK access is essential.
How do I learn AWS?
Build something. Don't read tutorials. Start with a simple web app. Deploy it on EC2. Add a database. Set up a CI/CD pipeline. Break it. Fix it. That's how you learn.
What exactly does AWS do that a regular server can't?
Scale. Elasticity. Managed services. A regular server has fixed capacity. AWS can add 1000 servers in 5 minutes. A regular server requires manual patching. AWS handles OS patches for you (with managed services). A regular server is one failure away from downtime. AWS offers multi-AZ deployments with automatic failover.
The Bottom Line
"What exactly does AWS do?" is the wrong question.
The right question is: "What can I build now that I couldn't build before?"
For me, that answer is production AI systems processing 200K events per second. For you, it might be a CRUD app that never goes down. Or a data pipeline that processes 10TB daily. Or a mobile app backend that scales from zero to millions of users.
AWS is a tool. A powerful, complex, sometimes infuriating tool. But a tool nonetheless.
The magic isn't in the tool. It's in what you build with it.
Nishaant Dixit — Founder of SIVARO. Building data infrastructure and production AI systems since 2018. Built systems processing 200K events/sec.