In the early 2010s, dozens of cloud infrastructure providers were jockeying for position. Rackspace, Joyent, Terremark, Savvis, GoGrid, the telcos, the hosting companies, a long tail of regional players.
Today, Synergy Research tracks more than 120 cloud service providers. Three of them — AWS, Azure, and Google — hold 63% of the market.
Parts of the AI market are following the same arc. Other parts won't. It depends entirely on which layer of the stack you're looking at — and that distinction matters a lot.
The story everyone is telling about AI is a story about models. The real story is about which integrated stacks cohere.
Layer 1: The Model Layer Is (Mostly) Already Decided
Frontier models for enterprise have consolidated on Anthropic, OpenAI, and Google.
Open source has real roles — specialized workloads, sovereignty requirements, cost arbitrage on commodity tasks. But operational complexity, security surface, evaluation burden, and the speed at which the frontier moves push large enterprises toward the same three names.
This isn't a prediction. Walk through the AI roadmaps of the Fortune 100. It's already happened.
A meaningful rise in token pricing could change this longer term. But for now, the ROI on GenAI is high enough that I don't see price sensitivity becoming the deciding factor any time soon.
Layer 2: The Integrated AI Stack Is the Real Battle
The layer above the models is where the next consolidation will happen. And it's not a single category. It's six interlocking layers that have to work as one system: data, semantic, governance and entitlements, model routing, agent orchestration, and observability across all of it.
Thousands of vendors are competing here today, many selling pieces. In three years, the winners won't be the best individual components. They'll be the platforms that integrate all six into a coherent stack.
This is where I think the market is most misread. The next dominant players won't be standalone orchestration tools or standalone governance tools. They'll be integrated AI stacks — the way the hyperscalers became integrated cloud stacks.
Three plausible categories of winners:
Hyperscalers.Bedrock, Azure AI Foundry, Vertex — extending out from infrastructure.
Traditional enterprise software companies.Salesforce, ServiceNow, SAP, Oracle, Microsoft — extending out from the system of record.
A new entrant.A startup or emerging player building the integrated stack from scratch.
And while many are calling for the early demise of software development firms, I hold the opposite view. Firms are going to need more help, not less, in building an integrated AI platform. If the traditional software companies build genuine integrated AI platforms — six layers, working as one — they're positioned to win enormous share. If they build a portfolio of standalone AI tools and call it a strategy, they'll be disintermediated by whoever does the integration work properly.
That's the fork in the road for every legacy software company right now. Tools or platform. Most are still choosing tools.
The piece most observers underweight is observability. When agents start composing other agents, the diagnostic question — what did the system do, why, with whose permissions, against which data — becomes the load-bearing capability. Whoever owns observability across the stack ends up with the gravity.
Layer 3: The Application Layer Stays Wide Open
This layer doesn't consolidate. It can't.
Internal teams will vibe-code their own apps fast and cheap when the problem is uniquely theirs. Vertical specialists with proprietary content and deep domain expertise will own categories the platforms can't reach. The integrated stack winners will offer their own application suites — without necessarily displacing the other two.
Most enterprises will run all three at once. That's the natural state, and it's not a problem to solve. It's a feature of the market.
What Leadership Should Be Doing Now
By the time the integrated stack layer consolidates, firms that picked early will have a multi-year compounding advantage on data, evaluation, and integration. Waiting is the expensive choice.
Five moves to make now:
1. Pick a model strategy, not a model.
Your frontier provider may shift. The capability to evaluate, swap, and route across them is what you actually need to build.
2. Make the stack bet deliberately.
Stop letting individual teams pick competing orchestration, governance, and observability tools. The cost of consolidating later is bigger than the cost of choosing now.
3. Invest in what no platform can give you.
Proprietary data, institutional expertise, evaluation capability. These are the only durable advantages once the platform layer settles.
4. Treat integration as the strategy.
A coherent stack of "good enough" components beats a stack of best-of-breed components that don't talk to each other. Every time.
5. Reframe what your technology organization is for.
The job is shifting from building tools to building the infrastructure that lets tools be built at scale. Your platform team's output isn't applications anymore. It's the substrate — data, governance, orchestration, observability — that lets the rest of the business build applications safely and quickly. That's a different operating model, a different talent profile, and a different definition of success. The firms that make this shift early will out-build everyone else by an order of magnitude.
In the early 2010s, three providers ended up owning the cloud. The firms that figured out the stack early are the ones that won the last decade. The same window is open right now. It won't be open for long.
McMillanAI helps business leaders navigate AI with clarity and confidence.
Start the Conversation