The real competitive advantage in AI isn't owning the model anymore; it's owning the data, expertise, and processes that make the model work for your specific business. A new analysis from Anaconda reveals that as open-source AI models become freely available to everyone, companies need to focus on customization, evaluation, and organizational knowledge to stand out.

What Happened When Bridgewater Tried to Use Frontier Models?

Investment firm Bridgewater Associates discovered this lesson the hard way. In June 2026, their AIA Labs team tested frontier models, the most advanced AI systems available, on six real-world tasks that their investors face daily, such as determining whether a news article matters to their portfolio or spotting signals in central bank documents.

The frontier models scored below 50% accuracy out of the box. Even after Bridgewater's experts carefully rewrote instructions and reframed questions, the best frontier model only reached 78.2% accuracy. That fell short of the 80% threshold Bridgewater needed to trust the system with real investment decisions.

So Bridgewater tried something different. Instead of relying on a proprietary frontier model, they downloaded Qwen3-235B, an open-source model that anyone can access. Initially, it performed worse than the frontier models, scoring just 44.8%. But after Bridgewater trained it on their own expert-labeled investment data, the open-source model jumped to 84.7% accuracy and made roughly 30% fewer mistakes than the best frontier model, all while costing about one-fourteenth the price per task.

How Can Companies Build Better AI Systems Than Their Competitors?

The key insight from Bridgewater's success is that the model itself is only part of the equation. Qwen3-235B is public; anyone can download it. What Bridgewater had that nobody else possessed was their labeled dataset created by expert investors and a rigorous process for ensuring that data was trustworthy.

Bridgewater's first attempt at labeling data failed because they hired non-experts to label examples. When the team examined the model's reasoning, they discovered the labels themselves were frequently wrong. Expert investors could produce correct labels, but expert time is expensive, so labeling everything by hand wasn't practical.

The solution was a routing system that directed human attention where it mattered most. Bridgewater trained a model on the cheaper labels, then checked every example where the model disagreed with its own training data. If a model couldn't reproduce a label from its own training set, either the example was genuinely difficult or the label was wrong. Both cases warranted expert review.

Steps to Building a Competitive AI System

• Capture Expert Judgment: Have your domain experts define what a good answer looks like, using real tasks from your actual work, not generic examples.
• Build Evaluation First: Create a test dataset with clear metrics before you start training, and never let training data touch your test set.
• Find the Disagreements: Use the model's own errors to locate examples where your training data might be unreliable or mislabeled.
• Route Expert Time Strategically: Send only the contested examples to experts for review, leaving the rest of the data alone to save time and money.
• Test Each Component: Remove each piece of your training setup one at a time and re-measure to understand what every piece contributes.
• Feed Results Back: Use the cleaned dataset and improved model as inputs to the next iteration, creating a continuous improvement loop.

None of these steps depend on which specific base model you choose. Bridgewater could swap Qwen3-235B for a different open-source model released next quarter, and the entire system would still work.

Why Model Availability Creates New Risks for Enterprises

The shift toward open-source models is happening at a moment when model and provider availability are becoming increasingly fragile. Pricing changes constantly to keep pace with compute demand. When Moonshot released their Kimi K3 model, throughput dropped from 30 tokens per second to 13, forcing them to pause new subscriptions just to protect existing users.

Governments are also stepping in. Frontier models are being pulled from markets due to security and geopolitical concerns. This creates an all-around crunch on model and provider availability that makes relying on a single external model risky.

"AI doesn't have an intelligence problem. AI has a context problem. How do we feed it that context? And that context is in the data," explained Ali Ghodsi, CEO of Databricks.

Ali Ghodsi, CEO at Databricks

This observation points to a fundamental shift in how enterprises should think about AI. General intelligence becomes business value only when it's connected to organizational knowledge and a system to use it. That system is a loop, not a single product: route work to the right model, feed it organizational context, capture expert judgment, evaluate outcomes, govern the path to production, and feed learnings back.

What Does Owning Your AI System Actually Mean?

Bringing open-source models into your technology stack can create real strategic value. They enable private deployment, offline use, regional control, customization, and greater independence from hosted APIs.

However, two companies running the same open-source model do not achieve the same results if one has better context, labels, evaluations, engineering standards, and feedback loops. The competitive moat only forms under two conditions: the system must stay neutral enough that the organization can change models, providers, and infrastructure without replacing the workflow around them, and the system must be the organization's own, with context, evaluations, policies, and judgment that stay portable.

If you cannot leave a vendor, everything you built inside their tool belongs to them too. If you can leave any vendor but nothing accumulates while you're there, you have flexibility and no advantage, because your competitor can buy the same flexibility tomorrow. Bridgewater's labeled dataset is the thing worth having. It's theirs, and it works on whatever model they point it at next.

The lesson is clear: in an era when open-source AI models are freely available to everyone, the real competitive advantage belongs to organizations that can build trustworthy data, develop rigorous evaluation processes, and maintain independence from any single vendor. The model is just the engine; the data and expertise are the fuel that makes it run.