The Next AI Fight Is Over the Right to Leave

A real choice is coming into view

A meaningful choice among AI models is beginning to emerge.

Open-weight models are becoming more numerous, more capable and more varied. Chinese developers are releasing large models built for coding, reasoning, multimodal work and agentic tasks, including Moonshot AI’s Kimi K3 and DeepSeek V4. American firms are producing both large general-purpose open-weight systems and much smaller models designed to run locally. Europe’s Mistral and other developers continue to expand the open-weight field.

These models do not all match the strongest proprietary systems across every task. They do not need to.

The relevant development is that credible alternatives are multiplying. Some open-weight models are already highly effective for many practical uses. Others approach frontier performance in selected areas. They can be customized and self-hosted or, in smaller forms, operated on personal and organizational hardware.

For the first time, it is becoming plausible that many customers could have meaningful choice among a broad range of capable AI models rather than among a small number of dominant providers.

That is an opening for dispersion.

But it will matter only if customers can use it.

The model may be replaceable before the relationship is

An open-weight model can make the intelligence technically replaceable. A developer can download and modify it or deploy it through different hosting environments. Compatibility layers can also make substitution easier: DeepSeek V4, for example, supports both OpenAI and Anthropic API formats.

Most people, however, are not simply using a model.

They are developing a relationship with a larger agentic system.

The AI agent an end-user works with begins to know how the user writes. It learns recurring preferences. It retains past conversations, files and project history. It connects to email, calendars, documents and other applications. It may eventually manage recurring tasks, maintain workflows and act under standing permissions.

The model produces the agent's intelligence. The surrounding system accumulates the customer.

That distinction will become more important as AI becomes more useful. A person may have access to five capable alternative models and still be unable to leave the provider that holds years of accumulated context. A business may be able to purchase intelligence from another source but be unwilling to rebuild its integrations, permissions, evaluations and institutional knowledge.

A market can therefore contain many models while remaining highly concentrated at the customer level.

Open weights make the model more replaceable.

Portability would make the provider more replaceable.

We have solved a simpler version of this problem before

The early cellphone market offers a useful comparison.

Customers could change carriers, but changing carriers generally meant losing the telephone number through which family, friends, employers, clients and institutions knew how to reach them. The customer was formally free to leave. The practical cost of leaving was high.

Congress created a number-portability obligation in the Telecommunications Act of 1996. The Federal Communications Commission then developed the technical and operating rules, with wireless number portability reaching the 100 largest metropolitan statistical areas on November 24, 2003.

Number portability did not eliminate every switching cost. But it removed one of the incumbent carrier’s most powerful advantages: control over something the customer could not realistically leave behind.

AI presents the same problem at much greater complexity.

The equivalent of the telephone number is not one identifier. It is the accumulated working relationship between the customer and the system.

The principle remains the same.

Customers should be able to change AI providers without surrendering everything that made the previous service useful.

What would have to move?

We often use the phrase data portability. That is too narrow for AI.

Practical AI portability has at least three levels.

Data portability means that customers can retrieve their conversations, files and account information in a usable format.

Context portability means that a replacement assistant can reconstruct what the former system learned about the customer: preferences, continuing projects, important relationships, working style and relevant history.

Operational portability goes further. It allows the customer to restore the workflows, connected applications, permissions, recurring tasks, agent configurations and evaluation history that made the former system function.

The first level is already available from major providers. The second is beginning to appear. The third remains much less developed.

That third level will become decisive.

Imagine a small company that has spent three years building an AI system into its operations. The system knows which employees may approve particular actions. It understands the company’s document structure. It has access to selected databases. It follows established review procedures. It has been corrected repeatedly and tested against the company’s standards.

A file containing old conversations would not allow another provider to recreate that system.

The customer would possess the data but still lose the capacity to switch.

A partial switch is already possible

The technology required for portability is not purely hypothetical.

Google now allows eligible personal Gemini users to import both chat histories and a representation of memory from other AI platforms. The documented process for ChatGPT begins with OpenAI’s data export, whose ZIP file includes chat history and other relevant account data. Gemini also supplies a prompt that asks the former assistant to summarize the user’s preferences, remembered facts and general context. That summary can then be added to Gemini’s own memory system.

That is meaningful progress.

A customer moving from ChatGPT to Gemini would not necessarily arrive as a complete stranger. Past conversations could be retained. Important personal and professional context could be summarized and reconstructed.

But this is still a translation, not a transfer of the functioning relationship.

The new system receives the former system’s description of what it learned. It does not inherit the exact way the former system weighted that information, connected it across conversations or learned from repeated corrections.

Google’s documented import process covers chats and memory. It does not provide a process for migrating connected applications, standing permissions, scheduled tasks, project structures or custom workflows. Under the documented migration process, those elements would still need to be reconnected or rebuilt.

The result proves two things at once.

A significant degree of context portability is technically possible today.

The documented tools do not yet provide complete operational portability.

Access is not the same as exit

Regulation is also beginning to address pieces of the problem.

On July 16, 2026, the European Commission adopted binding specification measures requiring Google to provide competing AI services effective interoperability with 11 Android features relevant to AI services. The objective is to prevent Google’s own AI services, such as Gemini, from receiving operating-system access that competing assistants cannot match.

This is an important intervention.

It could allow another assistant to reach the customer and perform many of the same functions as Google’s own assistant. That addresses entry.

It does not necessarily address exit.

A rival assistant might receive equivalent technical access to Android while the customer’s accumulated memory, history and workflows remain embedded in Gemini. The competitor could enter the platform without being able to bring the customer along.

That is the gap current policy has not yet closed.

Interoperability allows another provider to connect.

Data portability allows information to leave.

Operational portability exists only when the customer can leave without destroying the capacity already built.

The standard should be practical, not symbolic

A meaningful portability standard should be judged by what happens after the customer presses “export.”

Can another system understand the material?

Can it reconnect applications, permissions and recurring work?

Can the customer resume useful operations without months of rebuilding?

An archive that no competing system can interpret is not effective portability. Nor is a thousand-page transcript that requires the customer to identify manually which facts, decisions and preferences still matter.

The standard should require structured, machine-readable exports of the customer-controlled parts of the system. That would include memory, instructions, project relationships, workflow definitions, tool connections, permission structures and relevant evaluation records.

It should also require a tested migration process. Providers should be able to demonstrate that another compliant system can use the exported material to restore a meaningful portion of the customer’s prior capacity.

This would not require every AI model to behave identically. Models will reason differently. Providers will offer different tools. Some proprietary capabilities will not transfer.

The objective is not perfect replication.

It is a credible exit.

Build portability before lock-in hardens

Portability will be easier to establish now than later.

Memory systems, agent frameworks and tool connections are still developing. No single architecture has yet clearly become the unavoidable default. Providers are already experimenting with imports, compatible API formats and open protocols for connecting AI systems to tools and data.

The technical standard should emerge through an open process with vendor participation, but no single provider should control it. Whether procurement, market pressure or regulation ultimately enforces that standard can be left for a separate discussion.

The immediate task is simpler: define practical portability before millions of people and organizations build years of irreplaceable context inside incompatible systems.

Standards adopted after lock-in forms become fights over how much incumbent advantage must be surrendered.

Standards adopted before lock-in forms can shape the market itself.

Portability turns technical choice into customer power

Open weights, miniaturization and common interfaces can create alternatives. But alternatives do not create customer power if accumulated context cannot move.

The layer that controls the customer’s accumulated context may become more powerful than the company that built the original model. The customer may be surrounded by alternatives and still remain unable to move.

Portability is therefore not a minor consumer convenience. It is part of the infrastructure of a competitive AI market.

It would allow people to adopt AI without treating their first choice as permanent. It would allow organizations to move when a better, safer or less expensive model appears. It would give new providers a chance to compete for existing customers rather than only for people who have not yet built anything.

Open models can expand the supply of intelligence.

Portability will determine whether that expansion produces freedom.

Without it, we may create a market with many models—and millions of customers who still cannot leave.