Cloud vendors still sell three layers of service, but their price tags have stopped respecting the borders. On October 12, 2026, AWS moved short-term memory in Amazon Bedrock AgentCore from a flat per-event rate to consumption-based pricing of $1.00 per GB ingested, $0.20 per GB retrieved and $0.10 per GB-month stored. That is a managed platform service (PaaS) billed with the kind of meter you would expect on raw storage (IaaS). Meanwhile Amazon Quick, a finished business application (SaaS), lists a flat $250 monthly infrastructure fee per account on top of its per-seat price.
So the old IaaS vs PaaS vs SaaS question, who manages what, now arrives with a second question attached: which meter runs on your invoice? This guide answers both. It maps the three cloud service models onto the current catalogs of AWS, Microsoft Azure and Google Cloud as of October 27, 2026, lays out the published prices, and ends with a verdict you can apply to your own stack. One honest limit up front: no neutral benchmark exists that pits an infrastructure layer against an application layer, so the numbers below are vendor-published prices, analyst placements and customer claims, each labeled for what it is.
The audience is the engineer or technical lead who has to pick where a workload lives. If you only need a refresher, the first two sections cover the definitions. If you are weighing a migration, skip to the pricing, lock-in and migration sections. Even consumer products belong in the picture, because the best cloud storage services are SaaS in its simplest form: you rent a result and never see a server.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
IaaS vs PaaS vs SaaS: The Three Layers in Plain Terms
Think of the three cloud service models as a ladder. At the bottom, infrastructure as a service (IaaS) hands you virtual machines, block and object storage, and networking. You choose the operating system, patch it, install the runtime, deploy the code and run the backups. The provider owns the building, the power, the physical servers and the hypervisor. Amazon EC2, Azure virtual machines and Google Compute Engine are the standard IaaS examples.
One rung up, platform as a service (PaaS) takes the operating system and runtime off your plate. You bring code or data, and the platform handles scaling, patching and often high availability. AWS Lambda, Amazon RDS and Aurora, Azure SQL, Azure Databricks and Google’s Model Garden are typical PaaS examples. You give up low-level control and gain speed, and you usually pay by request, by capacity unit or by token instead of by server-hour.
At the top, software as a service (SaaS) delivers a finished application. You manage users, permissions, configuration and your own data. The vendor runs everything underneath. Microsoft 365, Amazon Quick and the Gemini agent Google announced on October 8, 2026 all sit here. Billing is nearly always per seat, per month.
The practical way to compare the layers is a responsibility ladder. Count what you still own at each rung:
- IaaS: you own the guest operating system, middleware, runtime, application, data, identity and access, and the network configuration inside your virtual network.
- PaaS: you own the application, the data and the access rules. The provider owns the operating system, runtime and scaling machinery.
- SaaS: you own your data, your user list and your configuration choices. The provider owns everything else.
Each provider tilts differently. AWS says in its own materials that it remains primarily an IaaS and PaaS provider, with SaaS arriving mainly through managed applications and AWS Marketplace listings. Azure spans all three, with virtual machines and Azure Local for infrastructure, managed services such as Azure Databricks and Azure SQL for the platform tier, and Microsoft 365 for software. Google Cloud, according to provider-tracking coverage, now positions itself as one integrated IaaS, PaaS and SaaS and AI platform. Note that this mapping is an analyst’s classification, not an official taxonomy from any of the three.
The boundaries also move with time. A service that once looked like PaaS may now be sold as a flat-rate product, and a SaaS tool may expose an API that makes it a platform. Treat the three labels as a spectrum of control versus convenience, not as sealed categories, and you will make better decisions than a team that argues over which box a product belongs in.
IaaS vs PaaS vs SaaS Specs Table: 12 Dimensions Side by Side
The table below compares the three layers on twelve dimensions that decide real projects. The product names in the examples row come from the current AWS, Azure and Google Cloud catalogs. Everything else describes how the model works in general.
| Dimension | IaaS | PaaS | SaaS |
|---|---|---|---|
| What you rent | Compute, storage, networking | Runtime, managed database, model endpoint | A finished application |
| Who patches the OS | You | Provider | Provider |
| Who writes the code | You | You | Vendor |
| Typical billing unit | Instance-hour, GB-month, data egress | Request, capacity unit, token, GB of data | Seat per month, sometimes plus a platform fee |
| Time to first value | Days to weeks | Hours to days | Minutes to hours |
| Customization | Near total | High within the platform’s limits | Settings, plugins, APIs |
| Scaling work | You design and automate it | Mostly automatic | Invisible to you |
| Staff skill needed | Systems and network engineering | Application and data engineering | Administration and change management |
| Lock-in risk | Lowest (portable images and containers) | Medium to high (proprietary APIs) | Highest (your data and workflows live inside) |
| Compliance control | Most direct (you pick region, keys, hardware) | Shared (provider certifies the platform) | Mostly contractual |
| Example products | Amazon EC2, Azure virtual machines and Azure Local, Google Compute Engine | AWS Lambda, Amazon Bedrock AgentCore, Azure Databricks, Google Model Garden | Amazon Quick, Microsoft 365, Google’s Gemini agent |
| Best fit | Custom stacks, GPU work, strict residency | Products you build and run | Functions you do not differentiate on |
Two rows deserve a second look. Staff skill is the quiet cost driver, because an IaaS estate that looks cheap per server-hour needs people to patch, monitor and secure it, and those salaries rarely appear in the cloud bill. Lock-in runs the opposite direction from convenience, so the layer that saves the most effort also asks the most from you at exit time. Later sections put numbers on both.
IaaS in 2026: Raw Infrastructure From AWS, Azure and Google Cloud
Infrastructure remains the foundation, and the three providers still compete hardest here on commitment discounts and specialized hardware. AWS sells Savings Plans for eligible compute and machine-learning usage on one- or three-year commitments. Microsoft advertises that its savings plan can cut up to 65 percent from select compute services. Google’s pay-as-you-go model gives automatic savings based on monthly usage, plus discounted rates for prepaid resources, according to its network documentation.
New customers get a soft landing. AWS currently offers $100 in Free Tier credits at sign-up plus up to $100 more as you explore, usable for up to six months. Those credits matter for IaaS experiments, because the hardest part of renting servers is learning what they cost before you commit. AWS also states that data transfer into its network is free, while data transfer out of services such as S3 and EC2 can use tiered pricing. That asymmetry shapes every exit plan, and the lock-in section returns to it.
Azure’s infrastructure story in October 2026 centers on hybrid. The latest Azure Local release is version 12.2610.1004.41, which uses OS build 26100.33650 and adds public-preview confidential virtual machines based on AMD SEV-SNP. It superseded version 12.2609.1003.7 from September. On the silicon side, Microsoft announced its Maia 300 accelerator for production deployment in 2027. It is an announced product, not something you can rent today.
Google Cloud’s recent infrastructure news leans toward AI and sovereignty. The whats-new.cloud update tracker lists Confidential G4 GPU instances, built on NVIDIA Blackwell, as generally available on October 21, 2026. Gemini 3.5 Flash became generally available on Google Distributed Cloud air-gapped deployments on October 20, which gives regulated customers a model that runs without an internet connection. A third-party price tracker, computeprices.com, lists Google Cloud’s eight-GPU H100 SXM configuration at $4.86 per GPU-hour, but that is not an official Google price page, so treat it as a rough reference only.
IaaS still wins where you need control the platform layers cannot give. A custom kernel module, a licensed appliance, an unusual GPU topology or a hard data-residency rule all point here. The price is operational duty. Even basic resilience becomes your job, and teams that compare disaster recovery tooling often discover that the tooling is the easy part and the runbook is the hard part. Infrastructure also changes under your feet. The same tracker says that starting January 5, 2027, Cloud DNS will follow CNAME chains between public and private zones and will stop returning full responses to ANY queries in public zones. Teams that own their DNS configuration should audit it before that date.
PaaS in 2026: Managed Platforms and the AI Agent Layer
The platform tier is where the three providers spent the most energy this autumn, and almost every launch involves AI. The shared pattern is simple. Each cloud now sells managed model access, agent runtimes and data services as platform products, and bills them by consumption. Google also lets customers pull the cost against existing cloud commitments. Here is what each provider shipped.
AWS: Bedrock, AgentCore and Serverless Databases
Amazon Bedrock pricing stays model- and usage-dependent, with token-based and reserved-capacity options. AWS announced general availability of OpenAI’s GPT-6.1 Terra on Bedrock in its latest what’s-new posts, so the service now fronts models from several labs behind one API. The AgentCore Memory change described earlier is the clearest sign of where PaaS billing is heading: the new short-term-memory prices apply in every AWS Region where the service is offered, and AWS says long-term-memory pricing did not change.
Serverless data services keep getting cheaper at the margin. AWS’s database page advertises a 50 percent reduction in DynamoDB on-demand pricing, up to 67 percent lower DynamoDB global-table pricing and ElastiCache Serverless for Valkey from $6 per month. Those figures come from a single page, so confirm them against your own region before you plan around them. The same update tracker also shows Aurora Serverless raising its capacity ceiling to 384 Aurora Capacity Units for the PostgreSQL-compatible edition, a 50 percent jump from 256. For containers, AWS states that Amazon ECS carries no separate management fee when you run on EC2, Fargate or Outposts, so you pay only for the compute underneath. If you want a hands-on example of renting a managed capability instead of assembling servers, our walkthrough on how to build a game streaming service on AWS shows the pattern in practice.
Azure: Databricks, SQL Managed Instance and Foundry
Azure Databricks added DeepSeek V4.1 Flash in October with two commercial options: priority pay-per-token for latency-sensitive traffic and reserved provisioned throughput for steady loads. The same release notes list detection for 50 new sensitive-data classes, support for 16 more countries and regions, and general availability of Feature Views for batch, streaming and request-time scenarios. Platform churn runs both ways. The notes also mark Gemini 2.5 Pro and Gemini 2.5 Flash as retired in Azure Databricks, and flag Zhipu’s GLM 5.2 as a legacy model.
Azure SQL Managed Instance gained flexible memory and automatic index compaction for Business Critical instances in September, and automatic backup immutability reached general availability during the August to September update cycle. Microsoft is also pushing migration with discounts. Its October partner announcements, as summarized by Microsoft Learn, list up to 30 percent off eligible Azure database pay-as-you-go usage for qualifying SQL Server migrations, up to 20 percent off eligible Microsoft Fabric pay-as-you-go capacity, and up to 15 percent off eligible App PaaS and data-service usage for Microsoft Foundry customers. These are promotions with eligibility rules, not list prices.
Google Cloud: Model Garden and Agent Compute Engine
Google launched Gemini 4 Argon (model ID gemini-4-argon) on October 13, 2026 at an introductory $2 per million input tokens and $10 per million output tokens through November 30. From December 1 the listed standard price doubles to $4 and $20, and cached input tokens carry a 95 percent discount. On October 21 the company began offering OpenAI models in Model Garden through Google Cloud Marketplace. Google’s regional blog names GPT Astra, GPT 6.1 Sol and GPT 6.0 Luna, and says they are in private preview for selected enterprise customers, so they are not generally available. Usage appears on a single Google Cloud invoice and can draw down eligible cloud commitments.
Anthropic’s Claude models arrived with a choice of Anthropic-managed or Google-managed deployment, both at standard pricing and both billed through Google Cloud. The tracker also lists Agent Compute Engine, announced October 22 as a purpose-built compute service for AI agents. A public preview is expected in the coming months, so it is not generally available, and no price has been published. Google also says its agentic Quick Estimator for VMware and infrastructure inputs is generally available and produces Compute Engine total-cost-of-ownership projections. In short, the platform tier now includes tools that help you leave the infrastructure tier.
SaaS in 2026: Amazon Quick, Microsoft 365 and the Gemini Agent
Software as a service used to mean a clean per-seat price. The agent era complicates that. Amazon Quick is the best current case study because AWS publishes the full price structure. The Professional plan costs $20 per user per month and the Enterprise plan costs $40 per user per month. Both add a flat $250 per account per month infrastructure fee, which AWS says covers the AI infrastructure behind the product. Extra agent hours cost $3 each, metered to the second, and index storage beyond the pooled allocation costs $5 per GB per month.
That structure reveals what SaaS vendors now pass through. A flat platform fee means small teams pay a heavy per-head premium. Amazon Quick lists Professional at $20 per user per month and Enterprise at $40 per user per month, and both plans carry the same flat $250 per account per month infrastructure fee. Spread across ten users, that fee weighs far more per head than across one hundred. Seat price is the headline, but the fixed fee decides whether the product suits your headcount.
Microsoft and Google are fighting the same battle with different tactics. Microsoft announced a 30 percent promotional offer on Microsoft 365 E3 starting October 27, 2026, limited to eligible new Cloud Solution Provider customers who buy at least 300 licenses. Google announced the Gemini agent on October 8 as a universal work agent integrated with Workspace, able to handle knowledge work, content creation, coding and tasks across connected systems. It also announced real-time spend caps and other spending controls, which tells you that agent usage bills are a real worry for buyers.
SaaS also arrives through marketplaces. Mintlify launched its Enterprise offering in the AWS Marketplace on October 27, 2026, in the AI Agents and Tools category, and Temenos said in May that it added Digital Banking and Payments to its Core Banking SaaS offering on AWS, according to that company’s announcement. Google makes the same marketplace pitch for third-party models, with purchases landing on one Google Cloud invoice and drawing down existing commitments. The extreme SaaS case is entertainment: a streaming subscription or a cloud gaming plan hides every layer below it. Readers comparing the best cloud gaming services are comparing SaaS products whose only visible spec is latency.
Cloud Service Model Pricing: Metered, Committed and Per Seat
Every list price in the cloud falls into one of three shapes. Metered prices charge for what you use, by the GB, the token or the second. Committed prices trade a one- or three-year promise for a discount. Per-seat prices charge for people, with or without a fixed fee on top. Most real bills mix all three. The table below collects the current published prices we could verify. Microsoft’s pages did not give us a single Azure VM or database price we could quote with confidence, so none appear here.
| Layer | Provider and product | Published price or term | Billing shape |
|---|---|---|---|
| PaaS | AWS, Bedrock AgentCore Memory (short term) | $1.00 per GB ingested, $0.20 per GB retrieved, $0.10 per GB-month stored | Metered since October 12, 2026 |
| PaaS | Google Cloud, Gemini 4 Argon (introductory) | $2 per million input tokens, $10 per million output tokens, through November 30, 2026 | Metered by token |
| PaaS | Google Cloud, Gemini 4 Argon (standard) | $4 per million input tokens, $20 per million output tokens from December 1, 2026, cached input 95% cheaper | Metered by token |
| PaaS | Azure Databricks, DeepSeek V4.1 Flash | Priority pay-per-token or reserved provisioned throughput (no figures published in the notes) | Metered or reserved |
| PaaS | Google Cloud, OpenAI and Claude models | Billed on one Google Cloud invoice, draws down eligible commitments, Claude at standard pricing | Marketplace plus commitment |
| IaaS | AWS, Savings Plans | One- or three-year commitment on eligible compute and machine-learning usage | Committed |
| IaaS | Azure, savings plan | Up to 65% off select compute services | Committed |
| IaaS | AWS, Free Tier credits | $100 at sign-up plus up to $100 more, valid up to 6 months | Credit |
| SaaS | AWS, Amazon Quick Professional | $20 per user per month plus $250 per account per month | Seat plus platform fee |
| SaaS | AWS, Amazon Quick Enterprise | $40 per user per month plus $250 per account per month | Seat plus platform fee |
| SaaS | AWS, Amazon Quick overages | $3 per extra agent hour, $5 per GB per month of extra index storage | Metered |
| SaaS | Microsoft 365 E3 promotion | 30% off for eligible new CSP customers buying 300+ licenses, starting October 27, 2026 | Promotional seat discount |
The trend line matters more than any single row. PaaS is converging on pure consumption, as AgentCore Memory and Argon show. SaaS is acquiring infrastructure-style fees, as Quick’s $250 line shows. IaaS keeps its commitment discounts but now competes with credits and marketplace drawdowns. One more date to circle: Argon’s standard price is double the introductory price, so a workload budgeted in October will cost twice as much to run in December unless caching or a cheaper model absorbs the change.
Putting those prices into a small cost model makes the differences concrete. The script below uses only the figures in the table. The sample inputs are ours, chosen to be round: 500 GB ingested, 200 GB retrieved and 100 GB stored for agent memory, fifty Enterprise seats for Quick, and 100 million input plus 20 million output tokens for Argon.
def agentcore_memory_monthly(ingest_gb, retrieve_gb, stored_gb):
return ingest_gb * 1.00 + retrieve_gb * 0.20 + stored_gb * 0.10
def quick_monthly(users, per_user, infra_fee=250):
return users * per_user + infra_fee
def argon_monthly(m_in, m_out, standard=False):
p_in, p_out = (4, 20) if standard else (2, 10)
return m_in * p_in + m_out * p_out
print(agentcore_memory_monthly(500, 200, 100)) # 550.0
print(quick_monthly(50, 40)) # 2250
print(argon_monthly(100, 20)) # 400 (introductory)
print(argon_monthly(100, 20, standard=True)) # 800 (from December 1)The results are small on their own, but they come from three different meters: per-GB charges for the agent memory workload, per-seat fees for Quick Enterprise, and a per-token model bill that changes on December 1. That is exactly why finance teams struggle to forecast cloud spend. A useful habit is to model each service in its own unit first, then convert to a monthly figure, and only then add them up.
What the Published Numbers Say: Three Vendors, One Analyst, No Neutral Benchmark
Readers often ask which layer is faster or cheaper, as if IaaS, PaaS and SaaS were competing products. They are not, so nobody publishes a head-to-head benchmark across them. What exists is a set of vendor statistics, analyst placements and customer outcome claims. The table sorts what we found by source and labels each figure honestly, so you can weigh a marketing number as a marketing number.
| Source | Metric | Value | Type |
|---|---|---|---|
| Google Cloud (Gemini at Work 2026 keynote) | Customers using Google Cloud AI products | Nearly 80% | Vendor claim |
| Google Cloud (same keynote) | Fortune 100 companies using Gemini Enterprise | Nearly 90% | Vendor claim |
| Google Cloud (same keynote) | Customers that each processed over 1 trillion tokens in a year | Nearly 500 | Vendor claim |
| Google Cloud (same keynote) | TPU 8i price-performance versus the previous generation | 80% better | Vendor claim |
| Microsoft Azure | Savings plan discount on select compute | Up to 65% | Advertised maximum |
| Microsoft Azure | Western Sydney University modernization saving | About 30% | Customer case study |
| Gartner via Microsoft | 2026 Magic Quadrant Leader placements | Server Virtualization Platforms, Global Industrial AIoT Platforms, Cloud-Native Application Platforms | Analyst ranking |
| Gartner via AWS | 2026 Magic Quadrant Leader placements | SCPS, Contact Center as a Service | Analyst ranking |
| AWS press release | Mintlify agent-traffic uptime | 99.999% | Headline claim |
| AWS homepage banner | Agent pilots that stall | 88% | Vendor claim, no method given |
Three things stand out. First, every figure here comes from a vendor or from an analyst the vendor is quoting, so none is independent. Second, the numbers measure different things, from adoption to price-performance to uptime, which means you cannot rank the clouds from this table. Third, the Gartner placements say something about breadth and execution, not about whether a given service fits your workload. A Leader badge in server virtualization tells you little about how Azure’s platform tier handles a small team shipping an API.
The AWS banner statistic about stalled agent pilots deserves a sentence of skepticism. AWS gives no method on the page we checked, so the 88 percent figure works as a marketing hook rather than a research finding. It still captures a real pattern. Teams that jump straight to a platform product without defining the business outcome tend to stall after the demo, regardless of which cloud they use.
Because no neutral benchmark exists, build your own. Measure three things on a pilot of each candidate layer. Track cost per unit of business output, such as cost per thousand support tickets handled or per thousand API calls served. Track hours to first production deploy, counting the integration work. And track engineer-hours per month spent on upkeep after launch. Those three numbers answer the IaaS vs PaaS vs SaaS question for your team far better than any vendor chart.
Real-World Examples: Nine Companies and the Layer They Picked
The best way to see the trade-offs is to watch what real organizations chose in 2026. The examples below span all three layers and all three clouds. Where a claim rests on a single announcement, the text says so.
- Mintlify (SaaS vendor on AWS): According to an AWS press release dated October 27, 2026, Mintlify runs on AWS to serve billions of agent visits at 99.999 percent uptime and launched its Enterprise offering in AWS Marketplace. The lesson: a SaaS company is itself a heavy IaaS and PaaS customer, and it sells through the same marketplace its buyers already use.
- WorkWave (cloud modernization on AWS): The company announced on September 22, 2026 a cloud modernization and an AI engine called WAIve, targeting up to a 50 percent reduction in cloud costs, according to its announcement. Treat the target as a goal, not a result.
- Oracle AI Database@AWS (managed database): On August 13, 2026, Oracle and AWS said the service had expanded to 22 AWS Regions and that Oracle Exadata Database Service on Exascale Infrastructure was generally available on it. For shops with large Oracle estates, this is PaaS that avoids a rewrite.
- Western Sydney University (Azure): Microsoft says the university took control of its modernization with Azure and saved approximately 30 percent in costs. It is a vendor case study, so read the saving as a claim, not an audit.
- Airwallex (Gemini Enterprise, SaaS and PaaS): Google says the global financial platform deployed Gemini Enterprise across its infrastructure. Google credits it with resolving customer problems ten times faster, cutting dispute-handling time by 40 percent and speeding software time to market by more than 95 percent.
- Hitachi (agents on Google Cloud): Google says Hitachi’s field technicians built agents for more than 200 use cases through its HMAX portfolio, including image-recognition agents that check photos of tools and wiring, with productivity gains of up to 400 percent. This is a customer-built layer on top of a platform.
- Temenos (SaaS on AWS): According to the company’s announcement, Temenos added Digital Banking and Payments to its Core Banking SaaS offering on AWS in May 2026. Banks rent the whole stack and manage only configuration.
- Kodiak AI (IaaS and PaaS at scale): The company said on October 1, 2026 that it uses AWS as its primary cloud provider for AI development and safety simulations. Simulation workloads are the kind of bursty, hardware-hungry job that favors rented infrastructure.
- Recursive (IaaS commitment): Recursive announced a $410 million, multi-year agreement on July 28, 2026 to run its automated AI research system on AWS, according to the company. At that scale the deal is a commitment contract, not a pay-as-you-go account.
The pattern across nine stories is consistent. Companies that sell software buy infrastructure and platforms. Companies that run operations buy SaaS and build thin layers of agents on top. Organizations with large legacy estates pay for managed versions of the software they already own. Almost nobody sits purely in one layer. Consumers show the same pattern, because a player who follows a free cloud gaming setup guide is renting a finished service on top of someone else’s servers and still has to supply the network and the controller.
What Practitioners Say About the Three Layers
Few people explain these terms better than the engineers who work on cloud sustainability, because they trace resources from the data center to the invoice. On the podcast The Week in Green Software, host Chris Adams described the infrastructure layer this way: “IaaS for most people stands for infrastructure as a service, right? So you pay for things on a kind of monthly or hourly basis, almost like metered electricity.” That metering idea is the thread running through every price in this article.
Adams then framed the middle layer by what it spares you: “That’s platform as a service, which as I understand it, you might not want to be working with all these low level things yourself.” The phrase low level things covers patching, capacity planning and failover. It is also what you pay a premium to avoid, and the AgentCore pricing change shows that premium turning into a GB-month meter.
Max Schulze of the Sustainable Digital Infrastructure Alliance (SDIA) pointed to the overlap that confuses buyers: “To me it’s the most ironic one because past, like a database as a service is also a software as a service, but it’s tools that developers use to build more software.” In other words, the same product can be PaaS to its developer and SaaS to its seller. Azure Databricks and Amazon Bedrock fit that description.
Schulze also offered a blunt description of what SaaS often is underneath: “And SaaS is when you then glue a lot of different kind of open source components together into a product.” Whatever the exact recipe, SaaS products rest on infrastructure that someone pays for, which helps explain why Amazon Quick lists an infrastructure fee of its own.
Security, Compliance and Lock-In Across the Layers
Security duties shift with the layer, but they never vanish. On IaaS you secure the guest operating system, the network rules and the identities, and you also own recovery, so a plan built on disaster recovery services across AWS, Azure and Google Cloud belongs in the budget. On PaaS you secure the application, the data and the access policy. On SaaS you secure the people, meaning who gets an account, what they can export and which third-party apps they connect. Customer-side mistakes such as loose access rules are the failure mode a provider can least fix for you, so choose the layer whose duties your team can actually staff.
Sovereignty and residency requirements are pushing vendors to offer controlled versions of every layer. Azure Local, now at version 12.2610.1004.41, brings Azure to customer-controlled hardware and adds preview confidential virtual machines. Google’s Distributed Cloud air-gapped option now supports Gemini 3.5 Flash, and Gemini for government services entered private preview on October 20. These offerings suit regulated buyers, but they cost more in effort and often in dollars than the public cloud versions.
Lock-in is the cost nobody puts on a pricing page. The higher the layer, the more the vendor’s roadmap becomes your roadmap. October 2026 offers a handful of live examples from Microsoft alone. Azure Communication Services is being retired as a standalone offering, and Microsoft began limiting new onboarding on October 23, 2026. AI Builder capacity add-ons retire on November 1, 2026. The AMQP protocol for Azure SQL Managed Instance change-event streaming was deprecated on August 15, 2026, though existing stream groups keep working until April 2027. Per the October 2026 Azure Databricks release notes, Gemini 2.5 Pro and Gemini 2.5 Flash are retired there. Microsoft also renamed application landing zones to workload landing zones in its Cloud Adoption Framework guidance.
None of these changes is unusual, and several come with long lead times. The point is that PaaS and SaaS users cannot postpone a retirement the way an IaaS user can keep running an old virtual machine. Budget engineering time for forced migrations, and prefer open interfaces such as standard SQL, containers and OpenAI-compatible model APIs where you can.
Exit costs deserve a line in every business case. AWS states that data transfer into its network is free, while transfer out of services such as S3 and EC2 can use tiered pricing. Moving a large dataset away from any provider is therefore the expensive direction. Ask for egress terms before signing a multi-year commitment, and test an export early, when you still have leverage.
Which Layer for Which Job: Six Use-Case Recommendations
1. A two-person startup shipping a product. Choose PaaS for the product and SaaS for everything else. Managed databases, serverless functions and hosted model APIs let two engineers run what once needed a platform team. Avoid IaaS until a specific constraint, such as a GPU shortage or a licensing rule, forces it.
2. A regulated enterprise with strict data residency. Start with IaaS or its controlled cousins: Azure Local, Google Distributed Cloud air-gapped or a dedicated region. Add PaaS only where the provider can certify the service in your jurisdiction. Keep the keys and the audit trail on your side.
3. A team building AI agents. Use PaaS: Bedrock with AgentCore, Azure’s Foundry and Databricks services, or Google’s Model Garden. Model the memory and token meters before launch, set spend caps, and watch price changes such as Argon’s December 1 step. Treat any preview service, including Google’s Agent Compute Engine, as a pilot, not a production dependency.
4. A data and analytics group. Managed PaaS usually wins. Azure Databricks, Aurora and DynamoDB spare you cluster operations, and the recent discounts on eligible Azure database usage lower the entry cost for SQL Server migrations. Reserve IaaS for the rare workload that needs a custom engine.
5. Back-office collaboration and productivity. Buy SaaS. Amazon Quick, Microsoft 365 and Google’s Gemini agent all compete here, and the decision turns on which ecosystem your staff already live in. Model the fixed fees first. A $250 monthly platform charge changes the math for a team of ten far more than for a team of one hundred.
6. GPU-heavy training and custom kernels. Pick IaaS. Confidential G4 GPUs on Google Cloud exist for teams that need direct hardware access. Pair them with commitment discounts only after utilization is steady, because idle committed GPUs are an expensive mistake.
Migration Guide: Moving Between IaaS, PaaS and SaaS in Eight Steps
Migrations rarely cross all three layers in one jump. The common paths are IaaS to PaaS (replatforming a database or a function), PaaS to SaaS (retiring a custom tool for a bought one) and, less often, SaaS back to PaaS when seat costs or limits start to bite. The same eight steps apply to each.
- Inventory what you run. List every workload with its owner, data volume, dependencies and monthly cost. Google says its agentic Quick Estimator, now generally available, can turn VMware and infrastructure inputs into Compute Engine cost projections, which helps if you are leaving on-premises hardware.
- Mark which duties you own today. For each workload, write down who patches, who scales and who restores from backup. The gaps tell you which layer would remove the most toil.
- Pick targets by pain, not by fashion. Move the workloads where upkeep eats the most hours. Leave stable, low-touch systems alone.
- Model the new meter before you move. Convert the target service’s price into your usage units, as the cost script earlier did. Include any steps in the schedule, such as Gemini 4 Argon’s December 1 change, and any one-time promotions, such as Microsoft’s SQL Server migration discounts, which carry eligibility limits.
- Plan the data path and the exit. Inbound transfer to AWS is free, but outbound transfer from services like S3 and EC2 can be tiered. If you may return later, test an export now.
- Run a pilot with a rollback. Move one low-risk workload, run both versions in parallel, and compare the three pilot metrics: cost per unit of output, hours to deploy and monthly upkeep hours.
- Cut over in stages. Shift traffic gradually, keep the old system warm until the next billing cycle closes, and verify backups and access rules on the new layer.
- Decommission and tag. Turn off the old resources, remove stale credentials and tag the new ones by owner and cost center so finance can follow the new meters.
Three pitfalls recur. Teams forget the people cost, so a PaaS move succeeds technically while the staff who ran the old servers lose their purpose and leave. Teams also underestimate integration work, since a SaaS tool is only useful once it is wired into identity, billing and reporting. Finally, teams ignore lifecycle, so they adopt a preview service that changes before it reaches general availability. Treat preview offerings as pilots until the provider publishes pricing and a support commitment.
Pros and Cons of IaaS, PaaS and SaaS
The summary table condenses the trade-offs. No row is a veto. Each one is a cost you accept in exchange for a benefit.
| Layer | Main advantages | Main drawbacks |
|---|---|---|
| IaaS | Full control of the stack, flexible hardware choice, portable workloads, direct control over residency and keys | You staff patching, scaling and recovery, costs rise with idle capacity, slower time to value |
| PaaS | Fast delivery, automatic scaling, less operations work, consumption pricing that tracks usage | Proprietary APIs, meters that change (as with AgentCore and Argon), forced migrations when services retire |
| SaaS | Minutes to start, no infrastructure to manage, predictable per-seat budgets, vendor handles upgrades | Least control, highest lock-in, fixed platform fees that punish small teams, data portability depends on the vendor |
Read the table by asking which drawback you can live with. A fast-moving product team can tolerate PaaS lock-in because speed pays for it. A bank with a ten-year horizon may reasonably refuse the same trade. A twelve-person company that buys SaaS for payroll, chat and ticketing is rarely worse off for the lack of control, because none of those tools gives it a competitive edge.
Verdict: How to Choose Between IaaS, PaaS and SaaS in 2026
Our verdict is a default order with one gate on IaaS. Buy SaaS for any function that does not set your business apart, such as collaboration, ticketing and finance. Build on PaaS for the product you ship, because the time saved on operations outweighs the lock-in for most teams. Reach for IaaS only when you can name the specific constraint, whether that is a GPU configuration, a licensing rule or a residency law. If you cannot name it, you are paying for control you will not use.
The data from this article supports that order. The AgentCore sample workload was billed at consumption prices, a small bill that would be hard to beat by building the same memory store on raw servers once staff time counts. Amazon Quick’s flat $250 account fee weighs far more on a small team than on a large one, so SaaS rewards larger teams. Gemini 4 Argon’s scheduled move from $2 and $10 to $4 and $20 per million tokens shows that the cheapest-looking PaaS meter can double on a published date. In all three cases, the lesson is to model the meter, not the brand.
On provider fit, our reading of the current catalogs is cautious. AWS tilts toward metered platform services and marketplace-delivered SaaS. Azure’s strength is hybrid infrastructure with Azure Local and the Microsoft 365 ecosystem, though it also carries the most visible lifecycle notices this month. Google Cloud’s pitch is model choice, with Gemini, Claude and (in private preview) OpenAI models behind one invoice, plus air-gapped options for regulated buyers. Each of these is a tilt, not a ranking, and none of them beats the others for every workload. For wider context on how the providers stack up, browse our cloud computing coverage.
The takeaway: decide by duty, not by label. Write down which responsibilities your team can carry, price the service in its own unit, and check the retirement and price-change dates before you sign. The IaaS vs PaaS vs SaaS choice is no longer a one-time architecture decision. It is a subscription to someone else’s roadmap, and you should read that roadmap as carefully as the price list.
Frequently Asked Questions About IaaS vs PaaS vs SaaS
What is the difference between IaaS, PaaS and SaaS? IaaS rents you infrastructure such as virtual machines, storage and networking, and you run everything above it. PaaS adds a managed runtime or data service, so you supply only code and data. SaaS delivers a finished application, and you manage users and configuration. Control decreases and convenience increases as you move up.
Which AWS, Azure and Google Cloud products are IaaS, PaaS and SaaS? Typical IaaS examples are Amazon EC2, Azure virtual machines and Azure Local, and Google Compute Engine. Typical PaaS examples are AWS Lambda, Amazon Bedrock, Azure Databricks and Azure SQL, and Google’s Model Garden. Typical SaaS examples are Amazon Quick, Microsoft 365 and Google’s Gemini agent. This mapping is an analytical classification, not an official taxonomy from the providers.
Is serverless PaaS or IaaS? Serverless functions such as AWS Lambda are usually classed as PaaS, because the provider manages the servers, the operating system and scaling, and you pay by requests and execution time. Some vendors blur the line, so judge by who patches the runtime rather than by the product name.
How much does Amazon Quick cost? AWS lists the Professional plan at $20 per user per month and the Enterprise plan at $40 per user per month. Both add a $250 per account per month infrastructure fee. Extra agent hours cost $3 each, and index storage beyond the pooled allocation costs $5 per GB per month.
How is Amazon Bedrock AgentCore short-term memory billed now? Since October 12, 2026, AWS bills it by consumption: $1.00 per GB for ingestion, $0.20 per GB for retrieval and $0.10 per GB-month for storage. AWS says long-term-memory pricing did not change.
When does Gemini 4 Argon’s introductory pricing end? The introductory price of $2 per million input tokens and $10 per million output tokens runs through November 30, 2026. From December 1 the listed standard price is $4 and $20, and cached input tokens carry a 95 percent discount.
Are OpenAI models generally available on Google Cloud? Not yet. As of October 27, 2026, the OpenAI models in Google Cloud’s Model Garden were in private preview for selected enterprise customers, so check availability before you plan a production launch around them.
Can I mix IaaS, PaaS and SaaS in one architecture? Yes, and most organizations do. A typical stack runs a custom product on PaaS, a few specialist workloads on IaaS and the company’s daily tools on SaaS. The key is to track each layer’s meter and lifecycle separately so that one price change or retirement does not surprise the whole budget.
