Twelve months ago, the enterprise AI market was dominated by a straightforward narrative: if you wanted the best AI models, you went to OpenAI, Anthropic, or Google. These proprietary models were meaningfully better than anything available in the open-source ecosystem, and enterprise buyers accepted the vendor lock-in as the price of performance.

That narrative is unravelling quickly. Open-source models — particularly Meta’s Llama series, Mistral’s releases, and a growing collection of community-developed models — have closed the performance gap to the point where choosing between open-source and proprietary is no longer a question of quality but of trade-offs.

For enterprise technology leaders, this shift has significant implications for procurement strategy, infrastructure decisions, and long-term vendor relationships.

How the Gap Closed

The performance gap between proprietary and open-source models has narrowed through several mechanisms.

Architecture publication: Despite keeping model weights proprietary, companies like OpenAI and Google have published enough architectural details in research papers that the open-source community has been able to replicate many of the key innovations. Techniques like mixture-of-experts, reinforcement learning from human feedback, and various attention mechanisms have been rapidly adopted by open-source projects.

Compute accessibility: Training large models requires enormous compute resources, which was historically a barrier for open-source development. But Meta’s decision to train and release the Llama model family effectively democratised access to models trained with hundreds of millions of dollars of compute. Other well-resourced companies and research institutions have followed.

Community fine-tuning: The open-source ecosystem generates thousands of fine-tuned model variants optimised for specific tasks. While no individual fine-tune matches a top proprietary model across all benchmarks, fine-tuned open-source models often outperform proprietary models on specific domain tasks because they’ve been optimised for exactly that domain.

Hugging Face now hosts over a million model variants, and the rate of new model releases continues to accelerate. The sheer volume of experimentation happening in the open-source community means that innovations appear and propagate faster than any single company can match.

What This Means for Enterprise Procurement

Reduced Lock-In Risk

The most significant strategic implication is reduced vendor lock-in. When enterprises build applications on top of proprietary AI APIs, they become dependent on that vendor’s pricing, terms of service, and continued operation. Switching costs are high because applications need to be re-engineered for different APIs, and model behaviour differences can break downstream logic.

Open-source models deployed on enterprise infrastructure eliminate this dependency. The organisation controls the model, the data, the infrastructure, and the upgrade timeline. If a better model becomes available, switching involves swapping a model file rather than re-engineering an integration.

This doesn’t mean proprietary models are dead. For many use cases, the convenience of a managed API with guaranteed availability and continuous improvement justifies the lock-in. But enterprises now have a credible alternative that they didn’t have two years ago.

Data Sovereignty

For organisations in regulated industries — finance, healthcare, government, defence — data sovereignty requirements often make cloud-based proprietary AI APIs problematic. Sending sensitive data to a third-party API, even one with contractual data protection guarantees, may not satisfy regulatory requirements or internal risk policies.

Open-source models can be deployed entirely within the organisation’s own infrastructure, whether that’s on-premises data centres, private cloud, or sovereign cloud environments. The data never leaves the organisation’s control.

The Australian Signals Directorate has published guidance on AI security that explicitly addresses the risks of sending sensitive data to external AI services. For organisations subject to these guidelines, self-hosted open-source models provide a compliance-friendly path to AI adoption.

Total Cost Considerations

The cost comparison between proprietary APIs and self-hosted open-source models is more nuanced than it appears.

Proprietary APIs charge per token, which makes costs predictable and directly proportional to usage. For low-to-moderate usage, this is often the cheapest option because you’re not paying for idle infrastructure.

Self-hosted open-source models require GPU infrastructure, which involves significant fixed costs regardless of usage. But at high usage volumes, the cost per inference drops dramatically because the infrastructure cost is amortised across more requests. Organisations processing millions of AI requests daily often find self-hosted models significantly cheaper than API pricing.

The break-even point depends on the specific models, infrastructure choices, and usage patterns. A team experienced in business AI solutions can help model these trade-offs for your specific situation, but the general rule is that proprietary APIs suit experimentation and moderate-scale production, while self-hosted models suit high-volume production workloads.

The Technical Considerations

Choosing to run open-source models isn’t just a procurement decision — it’s an infrastructure and engineering decision.

Model Serving Infrastructure

Running AI models at production scale requires specialised serving infrastructure. Frameworks like vLLM, TensorRT-LLM, and Triton Inference Server optimise model execution for throughput and latency. Setting up and maintaining this infrastructure requires engineering expertise that many organisations don’t currently have.

Model Selection and Evaluation

The volume of available open-source models creates its own challenge: choosing the right one. For any given use case, there might be dozens of candidate models with different strengths. Systematic evaluation against your specific tasks, using your own data, is essential.

Standardised benchmarks help but don’t tell the whole story. A model that scores well on academic benchmarks may not perform well on your domain-specific data. Build evaluation pipelines that test candidate models against representative samples of your actual workload.

Ongoing Maintenance

Proprietary AI APIs are maintained by the vendor. Model updates, security patches, and performance improvements happen automatically. With self-hosted models, the organisation is responsible for all of this.

This includes monitoring model performance over time, evaluating new model releases for potential upgrades, managing the GPU infrastructure, and ensuring security of the model serving environment. It’s not trivial, and organisations should budget for ongoing engineering time.

A Pragmatic Approach

The most sensible enterprise approach right now isn’t all-proprietary or all-open-source — it’s a portfolio strategy.

Use proprietary APIs for prototyping, experimentation, and use cases where the convenience and managed service model justify the cost and lock-in. Use self-hosted open-source models for high-volume production workloads, sensitive data scenarios, and use cases where you need full control over model behaviour and updates.

Invest in the engineering capability to evaluate, deploy, and maintain open-source models even if you’re primarily using proprietary APIs today. The competitive dynamics of the AI model market are shifting fast, and organisations that can move between providers — or run their own models — will have significant strategic advantage over those locked into a single vendor.

The open-source AI ecosystem isn’t just an alternative to proprietary models. It’s a competitive force that’s driving down prices, improving transparency, and giving enterprise buyers options they didn’t have before. That’s good for everyone except the vendors who were counting on lock-in as a business model.