Enterprises shift to cheaper open-source AI models as costs soar, Amazon CTO says

Enterprises are increasingly turning to open-source AI models for enterprises as a way to control mounting AI costs, according to Amazon CTO Werner Vogels. Speaking at the UN AI for Good summit, Vogels noted a clear shift from expensive proprietary models from OpenAI, Anthropic, and Google DeepMind toward cheaper open-source alternatives. The move comes after high-profile cost blowouts, including Uber burning through its entire 2026 AI budget in four months, have made executives rethink deployment strategies.

What happened

Vogels said in an interview at the summit: "We see a shift happening between the cheaper open source models and the bigger expensive models." Stories of runaway AI bills have made executives skittish about building systems on the most advanced models that bill by the token. Uber reportedly burned through its entire 2026 AI budget in four months, while another company burned through half a billion dollars in a single month after failing to cap AI usage for employees.

Open-source models (also called open-weight models) can usually be downloaded for free, but users then pay for their own cloud computing infrastructure. Still, this often works out to be cheaper than using the most advanced proprietary models. Vogels emphasized: "Cost is a very important part of your architecture, you need to take that into account. Do you really need to have the biggest, highest-end model to solve this? The answer is no, you don't."

Alongside the cost shift, Vogels highlighted a growing demand for transparency in how models are trained and deployed. "People want to know what is the data that goes into it," he said. This is especially acute in healthcare, government, and humanitarian work, where understanding training data and decision processes can be as important as performance.

At the same summit, Amazon launched a new open-source AI tool that connects the AWS Registry of Open Data to AI assistants. The registry contains more than 1,100 datasets from NASA, NOAA, and the NIH, and the tool allows researchers to search using natural language instead of navigating complex data catalogs.

Why AI builders should care

For AI builders, founders, and product teams, this shift has direct implications for architecture and budget planning. The era of "tokenmaxxing" running every task on the most powerful frontier model is giving way to "tokenminimizing" where each task is matched to the right model. This practice, called model routing, helps control spend without sacrificing performance.

Open-source models offer lower per-token costs, but they require you to manage your own infrastructure. That tradeoff can be worthwhile for high-volume or predictable workloads. For regulated industries like healthcare and government, the ability to inspect and fine-tune open-source models on proprietary data provides a transparency advantage that proprietary APIs cannot match.

Chinese open-source models are also entering the picture, with some costing 10 to 50 times less than leading proprietary alternatives, according to JPMorgan. This further pressures the pricing of frontier models and gives builders more options.

Practical implications

If you are building AI products today, consider these steps:

• Audit your model usage. Identify tasks that don't require frontier-level reasoning and route them to smaller open-source models.
• Evaluate total cost of ownership. Open-source models shift costs from per-token fees to compute infrastructure. Run the numbers for your expected volume.
• Prioritize transparency. If your product serves healthcare, government, or vulnerable communities, open-source models may build trust that proprietary APIs cannot.
• Leverage open data. The AWS Registry of Open Data tool can accelerate research and data discovery for teams working with scientific datasets.

Caveats

Open-source models are not a free lunch. You pay for cloud compute, and for very complex tasks, frontier models may still be necessary. Even open-weight model providers rarely reveal all training data, so transparency is not absolute. Cost savings depend heavily on usage patterns and infrastructure efficiency. Model routing adds engineering complexity and requires careful monitoring to avoid regressions.

Despite these caveats, the direction is clear: enterprises are moving toward a more cost-conscious, transparent approach to AI. Builders who adapt their architecture now will be better positioned as budgets tighten.

FAQs

What are open-source AI models and how do they differ from proprietary models?

Open-source AI models (often called open-weight models) can be downloaded for free, inspected, modified, and fine-tuned on your own data. Proprietary models from companies like OpenAI, Anthropic, and Google are typically hosted by the provider and billed per token. Open-source models give you more control and transparency, but you must pay for your own cloud compute infrastructure to run them.

How can enterprises reduce AI costs with open-source options?

Enterprises can reduce costs by using cheaper open-source models instead of the most expensive frontier models for every task. Open-source models require paying for cloud compute, but total costs are often lower than per-token billing on high-end proprietary APIs. Architecture decisions like model size selection and use-case-specific routing also influence total spend.

What is model routing and why is it important for controlling AI spend?

Model routing is the practice of directing each task to the most appropriate AI model rather than running everything on the most powerful one. This avoids paying for expensive frontier models when a smaller, cheaper model can handle the task. It is a key technique for controlling AI spend while maintaining performance.

How do data provenance and transparency affect AI usage in healthcare and government?

In healthcare, government, and humanitarian work, understanding what data was used to train a model and how it makes decisions is critical for trust and compliance. Open-source models allow developers to inspect training data and fine-tune on their own data, which can help meet regulatory requirements and build confidence with vulnerable communities.