Every week, someone declares that AI is going to destroy SaaS. That the subscription software model is dead. That by 2030, you'll build your own tools instead of buying them. That the great SaaS consolidation is coming, and most companies you pay monthly won't survive it.
That narrative is loud. It is also, in important ways, wrong. Or at least, it is asking the wrong question.
The right question is not whether AI kills SaaS. The right question is: what is AI actually built on?
When you follow that question all the way down the stack — past the chatbots, past the models, past the APIs, all the way to the physical infrastructure where AI actually lives — you find something that should change how every business leader thinks about the next decade.
AI is built on SaaS. Not metaphorically. Literally. The infrastructure that makes AI possible is cloud computing, which is SaaS. The tools used to build AI products are SaaS. The way AI gets distributed to your organization is SaaS. If anything, the AI economy does not threaten the SaaS model. It depends on it.
I'm an AI consultant. I help organizations adopt AI with intention. And I want to give you the clearest picture I can of how this technology actually works — because what you believe about the infrastructure shapes what decisions you make about your own adoption.
Let's go all the way down.
It Starts with Chips
AI does not exist in the cloud as some kind of weightless, frictionless intelligence. It runs on physical hardware. Specifically, it runs on chips — mostly specialized processors designed to handle the kind of math that machine learning requires at massive scale.
For the last several years, the dominant chip for AI work has been the GPU — Graphics Processing Unit. NVIDIA built a near-monopoly in this space. Their chips power most of the AI training and inference happening in data centers today. The company's stock became a proxy for the entire AI industry.
But GPUs are not the whole story, and they are increasingly not the future story. The major cloud providers — Amazon, Google, Microsoft, Meta — have been investing heavily in their own custom chips. Google has its TPUs. Amazon has Trainium. These custom chips are called ASICs: Application-Specific Integrated Circuits. They are built for one job and they do it more efficiently than a general-purpose GPU.
The economics here are significant. Google's TPUs routinely show two to three times better performance per watt than comparable GPUs. Amazon's Trainium is priced to undercut GPU-based instances by 30 to 40 percent for high-volume workloads. When you are processing billions of AI requests, that efficiency gap compounds into enormous cost differences.
At the very top of the power curve, NVIDIA's current Blackwell chips draw over 1,000 watts per chip. The next generation is heading toward data center racks that consume 300 to 600 kilowatts each. A single AI training cluster can require as much electricity as a small town. This is why the AI chip market is projected to exceed $400 billion by 2030.
None of this is abstract. Every time you send a prompt to an AI tool, that query lands on one of these chips somewhere in a data center. The chip does the work. The result comes back to you. What sits between those two moments is the infrastructure stack.
The Infrastructure Stack Is SaaS, All the Way Up
Start at the bottom: chips. Then move up.
Above the chips sits compute infrastructure — the data centers themselves. These are run by hyperscalers: Amazon Web Services, Microsoft Azure, Google Cloud. Hyperscalers spend billions building the physical capacity that AI requires. In just the first quarter of 2025, U.S. hyperscale companies spent $81 billion on capital expenditures — a 71 percent year-over-year jump — almost entirely for AI and cloud infrastructure.
That compute infrastructure is what you are paying for when you pay for cloud services. AWS, Azure, Google Cloud. These are SaaS businesses. You pay monthly. You scale up when you need more. You do not own the hardware. That is the SaaS model.
Above the compute layer sits the software frameworks — the tools developers use to actually build AI models. PyTorch, TensorFlow, CUDA. These frameworks abstract away the complexity of the hardware and let developers focus on building. Most of them are open-source, but they live and distribute through cloud platforms, which are SaaS.
Above the frameworks sit the model platforms — the services where companies like OpenAI, Anthropic, Google, and Meta host and distribute their AI models. You access these through APIs. You pay per use. That is consumption-based SaaS.
Above the model platforms sit the application layers — the AI-powered tools your organization actually uses. Every AI feature in your CRM, your HR software, your project management tool, your email client? It is calling an API that sits on top of a model that runs on cloud compute that runs on chips in a data center. The whole stack is SaaS, wrapped in SaaS, delivering to SaaS.
One recent analysis described this clearly: just as the modern SaaS ecosystem emerged from the cloud computing stack, the next generation of AI products is emerging from the same infrastructure layers currently being built. The structure is not new. The scale is new.
So Where Does the "SaaS Is Dead" Narrative Come From?
The narrative is real, even if it's incomplete. And it's worth understanding what is actually happening.
Microsoft CEO Satya Nadella declared that traditional business SaaS applications are dead. His company's leadership has laid out a timeline where AI agents replace conventional software workflows by 2030. HubSpot flagged "AI agent substitution" as a new competitive category on their Q4 2025 earnings call. Salesforce, ServiceNow, and Workday all reported slower net new revenue growth despite record AI investment.
Here is what is actually happening. The pain is concentrated in a specific category: lightweight, horizontal SaaS tools that exist to automate one relatively simple, rules-based workflow. Scheduling tools. Basic analytics dashboards. Simple CRMs for small teams. Integration platforms that just connect other apps. Standalone AI writing assistants.
These tools are genuinely vulnerable. Not because AI is replacing software, but because AI is removing the bottleneck that made these tools necessary in the first place. If an agent can draft your outreach sequence, schedule your follow-ups, and synthesize your pipeline data without a human clicking through five different dashboards — you need fewer dashboards.
But the ERP systems? The payroll platforms? The healthcare record systems? The compliance infrastructure? The industry-specific software built around proprietary data, regulatory requirements, and a decade of integrations? Those are not going anywhere. Global SaaS spending is projected to rise from $318 billion in 2025 to over $500 billion by 2028. The enterprise core is not collapsing. The peripheral layer is repricing.
The SaaStr analysis put it well: "SaaS is being starved, not killed." The money that used to flow into seat licenses for lightweight tools is being redirected to AI infrastructure. The category is not dying. The budget is moving.
AI Makes SaaS Easier to Build, Not Obsolete
Here is the thing that gets lost in the death-of-SaaS conversation. AI is not a replacement for SaaS. AI is a way to build SaaS faster, with fewer engineers, at lower cost.
For the last two years, a movement called "vibe coding" has emerged — using AI tools to generate working software through natural language prompts. What used to take a team six months now takes a skilled builder a week. What took a week now takes a day.
This does not mean the end of software. It means the beginning of more software. Jevons paradox is at work here. When something becomes cheaper and easier to produce, people produce more of it. Cheaper compute did not reduce the demand for cloud services. It expanded the market. Cheaper model intelligence is not reducing the demand for software. It is enabling an explosion of new, specialized, AI-native software products.
The companies that will benefit most from this moment are not the ones asking, "Will AI kill us?" They are the ones asking, "What can we build now that we could not have afforded to build three years ago?"
The per-seat pricing model for software is evolving. That part of the narrative is accurate. IDC projects that by 2028, pure seat-based pricing will be obsolete for 70 percent of software vendors, replaced by consumption-based and outcome-based models. But "pricing model is changing" and "category is dying" are not the same sentence.
What This Means for the Leader Making Decisions Right Now
You are not buying chips. You are not managing data centers. But you are making decisions every week about which tools to pay for, which vendors to trust, and how to think about AI adoption inside your organization.
Here is what the infrastructure reality means for those decisions.
First: the AI tools you adopt are only as reliable as the infrastructure they sit on. When you evaluate an AI product, ask where it runs, which model it uses, and which cloud provider hosts it. These are not technical questions. They are vendor risk questions.
Second: the lightweight horizontal tools in your SaaS stack deserve a real audit. Not because AI is going to destroy them — but because AI may have already made the problem they solved irrelevant. The goal is not to cut tools for the sake of cutting. The goal is to stop paying for infrastructure that no longer serves you.
Third: the "SaaS is dead" panic is largely a story about investor sentiment and pricing multiples, not about whether your organization needs software. You still need software. Your operations depend on it. The question is which software still earns its place.
Fourth: if you are a leader in a SaaS company, or you have clients who are, the move is not to defend the old model. It is to build the new one. AI enables tighter, faster, more specialized products than any prior era of software development. The window to build something that matters is open right now.
And fifth: pay attention to the infrastructure layer, even if you never touch it directly. The AI chip market alone is projected to reach $453 billion by 2030. The companies winning that market — NVIDIA, Google, Amazon, and a growing field of challengers — are building the physical substrate that every AI product you use depends on. Understanding that dependency clarifies your risk.
The Headline Is Wrong. The Underlying Shift Is Real.
AI is not killing SaaS. That headline is wrong, and acting as if it is true will lead you to the wrong decisions.
The underlying shift — that the economics of software are changing, that per-seat pricing is under pressure, that lightweight tools doing simple jobs are being displaced by agents, and that the next wave of valuable software will be AI-native — is real.
But this shift is happening inside SaaS, not against it. AI runs on cloud infrastructure. Cloud infrastructure is SaaS. The tools used to build AI are SaaS. The products that deliver AI to organizations are SaaS. The model is not ending. It is expanding into territory it never could have reached before.
The question worth asking right now is not whether to believe in SaaS. It is whether the specific tools you are currently paying for are serving the work that actually matters — and whether the AI adoption decisions you are making are built on a clear understanding of what is actually happening under the surface.
Informed decisions beat reactive ones. That has always been true. AI makes it more urgent than ever.
