By Thorsten Meyer
Every argument about who wins in AI is really an argument about a single question: is this a winner-take-all market or a winner-take-many one? And the honest answer is that we already ran this experiment, at enormous scale, one technology cycle ago. It was called cloud computing. I think it is the most useful map we have for the decade ahead — not because AI will repeat it, but because it rhymes, and the places where the rhyme breaks are exactly where the interesting money and risk live.
So let me lay out what cloud actually taught us, and then be honest about where the analogy holds and where it snaps.
The prediction everyone got wrong, twice
Here is the part that should make anyone humble about confident AI forecasts. The cloud era was mispredicted not once but twice, in opposite directions, by extremely smart people.
The cloud era was mispredicted in both directions by the sharpest investors alive. Both errors were the same mistake: dividing a fixed pie that was about to explode.
In 2007, when Amazon first described AWS at any length, the consensus reaction was a shrug. Reselling compute and storage looked like a low-margin commodity business — a scale game, cost-to-serve racing to zero, nothing durable. If you had polled the sharpest investors of that moment on whether AWS would become a high-margin, defensible, category-defining business, you'd have gotten a room full of no's.
By 2014, the error had flipped completely. Now the fear was that AWS would eat everything — not just infrastructure but the applications on top, offered at 8% gross margins that would crush the beautiful 85%-margin software businesses above it. "Your margin is my opportunity." The hyperscaler would consume the whole stack.
Both predictions were wrong, and they were wrong in the same underlying way: they treated the market as a fixed pie to be divided, when the pie was about to expand by more than an order of magnitude. The global cloud market reached roughly $400 billion in 2025 and is forecast near $778 billion by 2030. You cannot reason about a market growing like that as if the question is who gets the biggest slice of a fixed quantity. The slice math is the wrong math.
IoT and Edge Computing for Architects: Implementing edge and IoT systems from sensors to clouds with communication systems, analytics, and security
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Lesson one: it became an oligopoly, not a monopoly
The first durable lesson is about market structure. Cloud did not collapse to one winner, and it did not stay fragmented. It settled into a three-firm oligopoly with a long tail: as of 2026, AWS at roughly 30–31%, Azure around 24–25%, Google Cloud near 12–13% — the Big Three together holding about 67–68% of global infrastructure, a combined share that has been remarkably stable for two years even as the market itself exploded.
Notice the shape. Not a monopoly. Not a free-for-all. A small number of enormous scaled winners, differentiated from each other — AWS on breadth, Azure on enterprise integration, Google on data and growth — with everyone else fighting over the remaining third. That is the natural resting state of a capital-intensive platform market, and it is the single most likely shape for the foundation-model layer of AI. When you hear "one lab will win it all," the cloud precedent says: probably not. When you hear "it'll be commoditized across a dozen equal players," the precedent says that too: probably not. The base rate points at an oligopoly of a few, and I'd weight that heavily.
TrueNas Guide For Beginners: Build Secure Storage Systems, Homelabs, and Private Cloud Solutions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Lesson two: the biggest winners were built on top of the giants
This is the lesson people forget, and it's the most important one. The fear in 2014 was that AWS would eat the layers above it. Instead, some of the largest value creation of the entire era happened on top of the hyperscalers, often in direct competition with the hyperscalers' own products.
Cloud didn't collapse to one winner or stay fragmented. It settled into a stable three-firm oligopoly — and then the biggest value grew on top of the giants.
Not a monopoly, not a free-for-all. A few enormous differentiated winners, everyone else fighting over the last third. The likeliest shape for the AI model layer.
The 2014 fear was that AWS would eat the layers above. Instead, huge value grew there — often competing head-on with the hyperscaler's own products.
Snowflake is the canonical case: a data-warehouse company that runs on AWS, competes head-on with Amazon's own Redshift, and is worth roughly $102 billion as of August 2026. It "out-Amazoned Amazon on Amazon" — and it did so precisely by being the thing Amazon structurally could not be: cloud-neutral, running identically across AWS, Azure, and Google so that enterprises spanning multiple clouds didn't need a separate data stack for each. That neutrality is a moat the hyperscalers can't copy, because their whole model is to keep you inside their ecosystem. Snowflake's recent $6-billion AWS commitment shows the relationship for what it is: not war, but symbiosis — the platform underneath, the winner on top, both growing.
And it wasn't one company. Datadog (observability), Cloudflare (a different kind of cloud), Confluent, Databricks, MongoDB — a whole cohort of $10B-to-$100B businesses grew in the layers AWS supposedly owned. The takeaway for AI is direct and, I think, underpriced: the model labs are the new hyperscalers, and the equivalent of Snowflake has probably not been built yet. The durable winners of the AI era may not be the labs at all, but the companies that build on top of them — especially the ones offering the one thing a lab structurally cannot: neutrality across all of them.
Infrastructure as Code: Designing and Delivering Dynamic Systems for the Cloud Age
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Lesson three: "commodity" is a trap word
The cloud era should have permanently retired the lazy use of the word "commodity," and it's worth carrying that skepticism into AI. From the outside, running open-source models on standard hardware looks like pure pass-through resale — the same commodity framing that made people dismiss AWS in 2007. Up close, it isn't. Specialist inference providers extract multiples more speed and throughput from the same models on the same silicon, because running these systems efficiently is genuinely, deeply hard expertise, not a scale game.
The pattern is identical to cloud: the thing that looks like an undifferentiated commodity from a distance turns out, on inspection, to hide scarce and defensible expertise. Every time someone tells you a layer of the AI stack is "just a commodity" — inference, fine-tuning, orchestration, evaluation — the cloud precedent says look harder, because that's exactly where a durable business is often hiding in plain sight.
Microsoft Foundry in Action: A practical guide to building, monitoring, and governing AI applications using Microsoft's unified AI development platform
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Lesson four: enterprise adoption lags, then breaks
Cloud also teaches the shape of adoption, which matters for anyone timing this. For the first several years, blue-chip enterprises were deeply skeptical of cloud. The early heavy users were digital natives and startups — famously, one messaging app was at one point 40% of Google Cloud. The parallel today is exact: a single AI coding tool being a huge fraction of some provider's load, while conservative enterprises hang back.
Then, somewhere around 2014–2016, enterprise skepticism broke, and cloud became not just acceptable but mandatory — to the point where banks couldn't recruit good engineers because the talent wanted to work on modern platforms. That inflection is the one to watch for in AI. Enterprise AI adoption today is genuinely poor and poorly absorbed; the bears are right about that snapshot. But the cloud precedent says the snapshot is not the trajectory. The lag is normal. The break, when it comes, is fast, and it reprices everything.
Where the rhyme breaks — the honest bear case
I don't want to hand you the analogy as destiny, because the places it fails are where the real risk sits, and an honest builder names them.
Use cloud as a source of base rates, not a script. Four things it teaches about AI — and the one place the analogy snaps, which is where the real risk lives.
Speed and capital intensity. Cloud scaled fast, but AI is scaling faster — and it requires a staggering physical build-out: energy, power, land, chips, memory. That raises a genuine question cloud never faced at this intensity: whether capital itself becomes the binding constraint, and whether the build-out gets ahead of the revenue. The people warning about an "unparalleled capital implosion" in AI are not cranks; the analogy's own logic — enormous fixed cost, uncertain demand timing — is what makes the risk real.
Where value accrues is genuinely unsettled. In cloud, the value split between infrastructure and applications resolved in a fairly legible way. In AI it's an open fight: how much accrues to the chip makers, how much to the labs, how much to the application layer on top. "It all works" is a thesis, not a fact, and the honest version is that most companies in each layer will not work even if every layer thrives. The market being huge is cold comfort to the majority that become roadkill.
The neutrality play may be harder. Snowflake's cross-cloud neutrality worked partly because the clouds were interchangeable enough to abstract over. Models are more differentiated from each other than compute instances were, and switching between them is not friction-free. The "Snowflake of AI" is a compelling idea, but it's not a guaranteed slot waiting to be filled.
Where I land
Cloud is the right map for AI as long as you use it correctly: as a source of base rates and cautions, not a script. Its strongest lessons are that these markets resolve into oligopolies rather than monopolies; that the largest winners are often built on top of the giants rather than being the giants; that "commodity" is usually a failure of inspection; and that enterprise adoption lags and then breaks. Its strongest caution is that AI's capital intensity and speed introduce a failure mode — the build-out outrunning the demand — that cloud never faced so acutely.
If I had to compress it to one sentence: stop asking who wins the whole thing, and start asking who builds the neutral layer on top of whoever wins. That's the question cloud would tell you to ask, and it's the one I'd be spending my time on. The next seven pieces in this series pull apart the specific frontiers where that question gets answered — starting, next, with what SaaS becomes when the software can think.
Analysis and opinion from a builder, founder, and post-labor economist running a local-first inference operation. Figures verified at time of writing against Synergy Research Group, IDC, and public market data (cloud market shares and size, Snowflake market capitalization, AI's share of cloud spend); market structure claims reflect the author's interpretation of the cloud precedent, not a forecast. This is analysis, not investment advice. Part 1 of an 8-part series. Point-in-time as of 11 August 2026.
