The Docker moment for AI has arrived

Containers changed how we ship code. Now the same thing is happening to the AI that writes it. What happens when AI stops being a service you rent and becomes a tool you own.

In 2013, Docker solved a problem every developer knew but couldn’t articulate: “It works on my machine” wasn’t a joke—it was a daily nightmare. Containers virtualized environments and changed how we ship software forever.

Twelve years later, we’re watching the same pattern unfold with AI.

Docker launched Model Runner in April 2025. The command looks familiar: docker model run. Pull models from registries. Expose OpenAI-compatible APIs. Package and share AI like you package and share applications. The parallel isn’t accidental, it’s the point.

But this isn’t just Docker chasing a trend. The infrastructure for local AI has been quietly maturing for two years. Ollama made running models as simple as ollama run llama3. Llama.cpp brought inference to C++ with aggressive optimization. Quantization techniques shrunk 70-billion-parameter models to fit on consumer GPUs.

The economics have tilted. A dual RTX 5090 setup now matches H100 performance for 70B models—at 25% of the cost. An 8GB graphics card runs DeepSeek-R1 8B, a reasoning model that would have required datacenter hardware eighteen months ago.

This matters most for one use case: coding.

Why coding leads the local AI shift

AI assists with many tasks. For coding, it transforms the workflow entirely.

The tight feedback loop matters. You write code. You have questions. You need explanations. You want refactoring suggestions. Every interaction that requires a round-trip to distant servers introduces friction. Every friction point breaks flow state.

Local inference removes the latency. Responses feel immediate because they are. Your context stays loaded. Your conversation doesn’t time out. The AI becomes part of your development environment rather than a service you visit.

Privacy matters more for code than for most content. Proprietary business logic. Internal APIs. Security-sensitive implementations. Sending this to external servers requires trust—in the provider’s security, in their data handling, in their terms of service.

Local models require no trust. The code never leaves your machine. There’s nothing to exfiltrate because there’s no transmission.

One developer recently documented building an entire feature at 40,000 feet—seven hours, zero internet, fully functional AI assistance. Just Ollama running locally on a MacBook while the plane crossed the Atlantic. This scenario was impossible two years ago. It’s routine now.

The IDE landscape is fragmenting

The AI coding tool market is evolving fast. Cursor raised $2.3 billion in November 2025, validating the thesis that AI-native development environments represent a massive market. But the competitive landscape has shifted since their early dominance.

OpenCode launched as a desktop app across macOS, Windows, and Linux. It’s open source and free. It connects to 75+ model providers, including local models through Ollama and Docker Model Runner. A stress test showed 379 file updates in 22 seconds for a rename-symbol refactor. Full LSP support means the AI understands code structure, not just text patterns.

Claude Code took a different path, turning the terminal into an AI-native environment. Where traditional AI coding tools add assistance to an existing editor, Claude Code lets AI orchestrate the entire development workflow.

The philosophical split matters: Do you want AI as a feature within your tools, or AI as the foundation your tools build on?

What’s emerging is a spectrum of preferences:

Privacy-focused developers want local-first. No codebase leaving their machine. No terms of service parsing. No wondering which prompts train which models.

Cost-conscious teams want predictable economics. Hardware costs once, then runs indefinitely. No per-seat licensing. No token metering.

Flexibility-focused organizations want model choice. Run whatever performs best for their use case. Switch providers without switching tools.

The market is big enough for multiple approaches. Cloud-integrated tools like Cursor offer polish and seamless experience. Open alternatives like OpenCode offer freedom and local-first architecture. The interesting question isn’t which wins—it’s whether local-first becomes the default for certain developer profiles.

The virtualization thesis

Docker didn’t just containerize applications—it created a shared language for deployment. Dockerfiles became infrastructure documentation. Images became artifacts you could version, share, and trust.

Docker Model Runner follows the same playbook. Models as OCI artifacts. Registries as distribution. Compose files that spin up AI alongside your application stack.

The question isn’t whether we’ll virtualize AI. We already are. The question is whether this becomes as routine as container deployment—whether running a local model becomes as natural as running a local database.

I’m betting yes. Not because local models will match GPT-4 or Claude 3.5 on every benchmark, but because “good enough and private” serves most development workflows. The frontier models will keep advancing. Local models will keep following, six to twelve months behind, running on increasingly capable consumer hardware.

The gap that matters isn’t raw capability. It’s whether local models cross the threshold of usefulness for daily coding tasks. They have.

What I’m planning to test

I’m setting up a local development environment to see how this works in practice:

• Ollama running DeepSeek-R1 8B for reasoning tasks along with other models for coding

• Docker Model Runner for integration with existing container workflows

• OpenCode as the primary interface, configured to hit local endpoints

The experiment: Can a fully local AI stack handle a real project? Not a demo. Not a toy. An actual feature with tests, documentation, and production deployment.

If it works, the implications are significant. The average developer laptop becomes a capable AI development environment—no subscriptions, no cloud dependencies, no external trust required. That doesn’t kill cloud AI tools, but it creates a genuine alternative where none existed before.

Docker changed deployment by making environments portable. The same transformation is coming for AI. The tooling exists. The economics work. The only question left is adoption—and coding is where that adoption will happen first.

---