From Model Wars to Platform Wars: The Evolution of the AI Value Chain

---

When ChatGPT was released in late 2022, the AI industry appeared deceptively simple. The winners would be the companies that built the most intelligent models.

OpenAI, Anthropic, Google, Meta, and others entered an intense race to push the frontier of artificial intelligence. Model quality became the primary differentiator. Every benchmark mattered. Every improvement in reasoning, coding, multimodal understanding, and context length was celebrated as a competitive advantage.

Around these model providers, an entire startup ecosystem emerged.

Thousands of companies were founded on the assumption that frontier models would become commodities, while value would be created through applications, workflows, and user experiences. Investors poured billions into startups building AI writing assistants, agent platforms, visual workflow builders, prompt management systems, copilots, retrieval engines, and industry-specific AI solutions.

For a brief period, the AI value chain appeared neatly divided.

• Model companies would provide intelligence.

• Startups would provide applications.

• Cloud providers would provide infrastructure.

But that assumption is rapidly collapsing.

The AI industry is entering a new phase where the largest model providers are expanding aggressively across the entire value chain, transforming themselves from model companies into platform companies.

The Great Expansion of OpenAI and Anthropic

OpenAI’s original mission was clear: build the most capable AI models.

Yet over the past two years, OpenAI has steadily expanded beyond the model layer.

The launch of Custom GPTs allowed users to build specialized AI assistants directly within ChatGPT. The Assistants API introduced capabilities such as retrieval, code execution, function calling, and orchestration that previously required significant engineering effort. More recently, OpenAI introduced agent frameworks, application ecosystems, and increasingly sophisticated workflow capabilities.

Each expansion moved OpenAI further into territory previously occupied by startups.

Anthropic has followed a similar trajectory.

Initially known primarily for Claude, Anthropic has evolved into a broader developer and enterprise platform. Claude Code, the Model Context Protocol (MCP), Agent SDKs, and enterprise workflow capabilities position Anthropic not merely as a model provider but as an operating layer for AI-powered work.

The strategic logic is obvious:

• Building the best model is extraordinarily expensive.

• Training costs continue to rise.

• Inference costs remain significant.

• Competitive advantages can be short-lived.

• A model company that only sells tokens risks becoming a supplier in a highly competitive market.

To create durable value, these companies need to move closer to users, own workflows, control developer ecosystems, and build switching costs.

In other words, they need to occupy more of the value chain.

The Startup Squeeze

This evolution has created a brutal reality for many AI startups.

Historically, startups worried about competing against current products.

Today, they must compete against future roadmaps.

A startup may spend two years building a workflow orchestration platform, agent framework, or AI productivity tool, only to discover that a frontier model company has incorporated similar functionality directly into its ecosystem.

The result can be devastating.

We have already seen examples across the industry.

Jasper’s rapid rise was followed by a difficult repositioning after ChatGPT made general-purpose AI writing widely accessible.

Chegg publicly attributed significant business challenges to the rise of ChatGPT.

Adept and Inflection ultimately became examples of another trend: valuable AI companies being absorbed into larger ecosystems through acqui-hires and strategic talent acquisitions.

The lesson is becoming clear: thin wrappers around frontier models are increasingly vulnerable.

If a startup’s primary value proposition can be replicated by the platform provider, its long-term defensibility becomes questionable.

The safest positions are likely to be businesses that own proprietary data, deeply embedded workflows, regulatory expertise, industry-specific execution capabilities, or unique distribution channels.

Realization: Models Alone Do Not Win

Perhaps the most fascinating development is that even the frontier AI leaders appear to have recognized a fundamental truth:

The best model does not necessarily win the market.

This became particularly evident when observing enterprise AI adoption.

The early launch of Copilot was a disaster for Microsoft. Users are deeply unsatisfied by the performance despite its close integration with its own ecosystems.

After unlocking itself from OpenAI, Microsoft has repositioned itself as AI “Platform” company that is model agnostic.

The introduction of Copilot Cowork, incorporating Anthropic’s capabilities, reflects a broader recognition that enterprises care more about outcomes than model purity.

In my own testing, Cowork feels significantly more capable than earlier Copilot experiences.

More importantly, it demonstrates a shift in Microsoft’s strategic thinking.

Microsoft increasingly appears less interested in winning the model war itself and more interested in becoming the platform through which enterprises consume AI.

That may prove to be the smarter strategy.

The Rise of the AI Platform Companies

This shift mirrors a pattern we have seen before in technology.

The most valuable companies are often not the inventors of every component.

They are the owners of the platform.

Cloud computing provides a useful analogy.

Customers rarely choose AWS, Azure, or Google Cloud because of a single database service or individual feature.

They choose the platform because of the surrounding ecosystem:

• Identity management

• Security controls

• Networking

• Data storage

• Compliance frameworks

• Monitoring

• Governance

• Integration capabilities

• Operational tooling

AI is increasingly following the same path.

An enterprise can absolutely purchase APIs directly from OpenAI or Anthropic.

Yet most organizations still prefer consuming AI through their cloud providers.

The reason is simple.

A model is only one piece of the solution.

Organizations still need identity integration, data pipelines, governance controls, observability, security frameworks, deployment tooling, compliance processes, and enterprise-grade operational capabilities.

The cloud service provider already owns these layers.

This places Microsoft, AWS, and Google Cloud in a remarkably strong position.

Rather than participating directly in the brutal economics of frontier model development, they can offer customers access to whichever models prove most effective.

Whether the winning model comes from OpenAI, Anthropic, Google, Mistral, DeepSeek, xAI, or a future entrant becomes less important.

The platform still captures value.

The Emerging AI Value Chain

The industry increasingly appears to be organizing itself into three layers.

Layer 1: Intelligence Providers

These companies build the frontier models.

Examples include OpenAI, Anthropic, Google, Meta, xAI, Mistral, and others.

This layer is characterized by enormous capital expenditure, rapid innovation cycles, and intense competition.

While technically impressive, it may also become the most commoditized layer over time.

In addition, with agentic pipelines, we do not require the most advanced models in majority of agentic stages. In my own practice, the open-weighted China models can more than satisfy 90% agentic workload.

Layer 2: AI Platforms

These companies provide access, governance, integration, and enterprise consumption mechanisms.

Microsoft, AWS, and Google Cloud are emerging as the dominant players.

This layer creates significant switching costs and recurring revenue streams.

Layer 3: Business Workflows

This layer owns actual business processes and user behavior.

Examples include Microsoft 365, Salesforce, ServiceNow, SAP, Workday, and industry-specific operational platforms.

Ultimately, this may be where the deepest enterprise lock-in exists.

Users do not buy AI.

They buy outcomes.

The workflow layer delivers those outcomes.

The Future: From Model Wars to Platform Wars

The first phase of AI was defined by a race for intelligence.

The next phase may be defined by a race for control of the customer relationship.

History suggests that technology creators do not always capture the largest share of economic value.

Operating systems captured value from hardware manufacturers.

App stores captured value from developers.

Cloud platforms captured value from software vendors.

The AI industry may follow a similar pattern.

OpenAI and Anthropic have already recognized that models alone are insufficient and are rapidly expanding into applications, agents, ecosystems, and workflows.

At the same time, Microsoft, AWS, and Google Cloud are positioning themselves as the universal distribution platforms through which enterprises consume AI.

This creates a fascinating possibility.

The ultimate winners of the AI era may not be the companies that build the smartest models.

They may be the companies that own the platforms through which those models are delivered.

The first decade of AI was about who could create intelligence.

The next decade may be about who controls access to it.