For the past two years, the dominant mental model for AI has been fairly simple: frontier model companies build the infrastructure, startups build the applications.
OpenAI, Anthropic, Google DeepMind, Meta and others train the large models. Developers and startups rent access to those models through APIs. In this view, the model layer becomes a kind of new cloud infrastructure, while the application layer is where founders create workflow-specific products, vertical software and new user experiences.
That distinction is starting to blur.
Something important is happening in AI, and it is still not talked about enough: the model companies are moving from renting infrastructure to selling solutions.
They no longer want to be only the intelligence provider in the background. Increasingly, they want to own the workflow, the user interface, the enterprise deployment, the data feedback loop and the end customer relationship.
Google I/O: models are becoming products
Google I/O 2026 was a good example of this shift. Google did not just announce new Gemini models. It announced a much broader push to put Gemini into Search, shopping, Workspace, creative tools, developer environments and new form factors. Its official I/O collection includes Gemini Omni, Gemini 3.5, updates across developer tools, Search and AI-powered products. Suggested link: Google I/O 2026 official collection.
This matters because Google is not merely saying “our model is better.” It is turning the model into workflows.
Search becomes more agentic. Shopping becomes more automated. Creative work moves into tools like Flow, Stitch and Pomelli. Developer work moves further into AI Studio and Antigravity. Gemini becomes a consumer interface, a work assistant and a developer platform at the same time. Independent coverage also highlighted Google’s Gemini redesign, Gemini 3.5 Flash, Gemini Spark and agentic updates to Antigravity.
That is the shift in a nutshell: the model is no longer just a capability. It is becoming a product surface.
OpenAI’s Deployment Company: from API calls to operational change
OpenAI is moving in the same direction, but from the enterprise deployment side.
On May 11, 2026, OpenAI announced the OpenAI Deployment Company, a majority-owned and controlled company designed to help organisations build and deploy AI systems across their most important work. As part of the launch, OpenAI agreed to acquire Tomoro, bringing roughly 150 Forward Deployed Engineers and Deployment Specialists into the new company from day one.
This is a major signal.
OpenAI is not just saying “use our API.” It is saying: we will help you identify priority workflows, connect AI to your internal systems, redesign processes and drive adoption inside the organisation.
That looks much less like cloud infrastructure and much more like a combination of software, consulting, systems integration and operating transformation.
It also acknowledges a reality many enterprise AI buyers already understand: the bottleneck is not always model quality. It is deployment.
Enterprises need permissions, audit trails, security, data access, integrations, human approval loops, process redesign and internal adoption. The hardest part is not running a prompt. The hardest part is turning model capability into durable business change.
Anthropic: professional agents, not just chat
Anthropic is also moving from model access into packaged professional work.
Earlier this month, Anthropic launched ten ready-to-run agent templates for financial services, covering tasks such as pitchbook creation, KYC screening and month-end close. These agents ship through Claude Cowork, Claude Code and Claude Managed Agents, allowing teams to put Claude on real financial work in days rather than months.
That is not “chat with a model.” It is “delegate a defined financial workflow.”
Anthropic also launched Claude for Small Business, which plugs Claude into tools small businesses already use, including QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace and Microsoft 365. The workflows include payroll planning, month-end close, sales campaigns and invoice chasing.
Again, the language is important. The product is not being sold as intelligence in the abstract. It is being sold as work performed inside existing tools.
Why the model companies are moving up the stack
There are several reasons this is happening.
First, the economics of the model layer are difficult. Training and serving frontier models is expensive. If monetisation depends only on API usage, the model provider risks becoming a high-capex utility in a brutally competitive market. Moving up the stack allows these companies to capture more margin and justify higher subscription or enterprise pricing.
Second, applications create better feedback loops. A model used inside a real workflow generates more useful signals than a model called through a generic API. If Claude is helping an analyst build a pitchbook, Gemini is helping a user shop, or OpenAI is embedding models inside an enterprise process, the model company learns which tasks matter, where the friction is and what users are willing to pay for.
Third, the biggest AI companies already have distribution. Google has Search, Android, Workspace, YouTube, Chrome and Cloud. OpenAI has ChatGPT, the API and a growing enterprise footprint. Anthropic has strong positioning in coding, enterprise and professional workflows. Once you have frontier models and distribution, the incentive to package more of the use case is obvious.
What this means for startups
For startups, this raises an uncomfortable question: where do you build if the platform beneath you keeps moving up?
The simplistic answer is that AI application startups are doomed. I do not believe that.
Every major platform shift creates this anxiety. Cloud platforms moved into databases, analytics and security. Apple moved into apps and services. Microsoft bundled products that could have been standalone companies. And yet, each wave still created large independent companies.
But the bar for AI application startups is going up.
A thin wrapper around a model is increasingly vulnerable. If the product is mainly “we call GPT, Claude or Gemini and put a UI on top,” the risk is obvious. The model company can copy the feature, bundle it, subsidise it, distribute it and improve it with every model release.
The more defensible startups will be the ones that own something deeper than model access:
- a specific workflow
- proprietary data
- a hard-to-reach customer segment
- unique distribution
- regulatory or domain expertise
- deep integration into systems of record
- trust, compliance and human-in-the-loop review
In other words, the opportunity is not to build “AI apps” in the abstract. It is to transform painful workflows that the general-purpose platforms will not understand deeply enough, prioritise quickly enough or customise specifically enough.
The shift from software to outcomes
This also connects to a broader theme: AI-native companies may not look exactly like SaaS companies.
If frontier models can perform larger chunks of white-collar work, startups do not always need to sell software seats. They can sell outcomes.
Tax filings completed. Claims processed. Leads qualified. Compliance reviews performed. Financial reports generated. Customer support tickets resolved. Creative assets produced.
Ironically, the model companies are validating this direction themselves. Anthropic is packaging agents around jobs to be done. OpenAI is building a deployment company around enterprise transformation. Google is embedding Gemini into the everyday surfaces where people search, shop, build, create and work.
The market is moving from “rent this intelligence” to “buy this result.”
The investor question
For investors, this changes the diligence questions.
It is no longer enough to ask which model a company uses, or whether the demo looks magical. We need to ask:
- What workflow does the company own?
- What proprietary data does it accumulate?
- How deeply is it embedded in the customer’s operations?
- What would make it hard for OpenAI, Anthropic, Google or Microsoft to absorb this use case into a broader product suite?
- Is the startup selling software, or is it selling a measurable business outcome?
At Remagine Ventures, this is exactly the type of question we are asking founders. Not “are you using AI?” but “what painful workflow are you transforming, why are you uniquely positioned to own it, and why will the platforms not simply make you a feature?”
The model companies are moving up the stack because the value is moving up the stack.
That does not mean the AI application layer is dead. It means it is becoming more competitive, more vertical and more outcome-driven.
And that is where the most interesting companies will be built.
He is a former General Partner at Google Ventures (GV) in Europe, former head of Google for Entrepreneurs in Europe, and founding head of Campus London, Google's first startup hub. Eze writes on Israeli tech, venture capital, artificial intelligence, and founder strategy.
He is also the founder of Techbikers, a nonprofit that brings together the startup ecosystem on cycling challenges in support of Room to Read.
