The AI industry’s cost problem is getting harder to ignore. Companies that rushed to build on the most powerful models from OpenAI, Anthropic, and Google DeepMind are now confronting bills that can spiral out of control — and a growing number of them are turning to open source AI models as the more sustainable path forward. That shift, once tentative, is now visible enough that Amazon’s top technologist is calling it out publicly.
Summary
Key takeaways
- Amazon CTO Werner Vogels said companies are increasingly switching to cheaper open source AI models to control mounting AI costs.
- Uber burned through its entire 2026 AI budget in just four months, reportedly spending half a billion dollars in a single month.
- Open source models are typically free to download; users pay only for cloud infrastructure, often making them cheaper than proprietary alternatives.
- Amazon launched a new open-source AI tool at the UN AI for Good summit to help researchers search over 1,100 scientific datasets using natural language.
- Transparency around AI training data is emerging as a non-negotiable requirement in healthcare, government, and humanitarian sectors.
Rising AI Costs Propel Shift to Open-Source Models
Speaking on the sidelines of the UN’s AI for Good summit on July 10, 2026, Werner Vogels, Amazon’s Chief Technology Officer, put it plainly: “We see a shift happening between the cheaper open source models and the bigger expensive models.” It was a frank acknowledgment that the AI gold rush has a price tag — and many companies are no longer willing to pay it unconditionally.
Expensive Proprietary Models from Industry Leaders
The flagship models from OpenAI, Anthropic, and Google DeepMind sit at the top of the performance rankings. But performance at scale comes with a cost structure that has blindsided more than a few organizations. These systems bill by the token, meaning costs compound rapidly as usage grows across teams and products.
The starkest illustration of this came from Uber. The company reportedly burned through its entire 2026 AI budget in just four months — and, according to reports, spent roughly half a billion dollars in a single month after failing to cap AI usage for employees. That kind of spending trajectory forces even large organizations to reassess their approach fast.
Vogels framed it as an architectural question rather than a purely financial one. “Cost is a very important part of your architecture, you need to take that into account,” he said. “Do you really need to have the biggest, highest-end model to solve this? The answer is no, you don’t.”
Cost Advantages of Open-Source AI Models
Open source models — sometimes called open-weight models — can typically be downloaded for free. The main cost comes from the cloud computing infrastructure needed to run them. That setup often works out significantly cheaper than paying ongoing token-based fees to proprietary providers, especially at scale.
The cost advantage is not marginal. According to data from OpenRouter, a developer platform that aggregates access to multiple AI models, Chinese open source models can be 60% to 90% cheaper than leading Anthropic and OpenAI alternatives. That gap helps explain why the share of tokens used by U.S. companies on Chinese AI models via OpenRouter climbed as high as 46% — up from an average of just 11% over the prior 12 months.
The trend is not just about cost-cutting. As Benchmark’s Peter Fenton, an investor in open-source AI developer tool Ollama, put it: every company with high inference expenses has a “vital existential project” pushing them toward open-weight models. Ollama itself, which helps developers run open-weight models locally, now counts nearly 9 million monthly active users and sits in 85% of the Fortune 500 — a sign of how mainstream this shift has become.
Amazon CTO Highlights Transparency and Pragmatism in AI Adoption
The cost argument alone does not capture the full picture. Beyond the economics, something more structural is changing in how organizations think about the AI they deploy.
Werner Vogels’ Remarks at the UN AI for Good Summit
Alongside the cost conversation, Vogels pointed to a second force reshaping AI procurement: the demand for transparency. “Transparency becomes extremely important,” he said at the summit. “People want to know what is the data that goes into it.”
This is not an abstract concern. Companies are now scrutinizing not just what an AI model can do, but how it was built — what data it was trained on, what biases might be embedded, and how its decisions can be explained. That scrutiny reflects a broader maturation in enterprise AI adoption, moving past the initial hype phase into a harder-headed evaluation of AI transparency and data governance.
Importance of Trust in Sensitive Sectors
The transparency imperative is especially acute in sectors where the stakes of a wrong or unexplainable AI output are high. In healthcare, government, and humanitarian work, understanding how a system was trained can matter as much as its raw performance. “If these people serve vulnerable communities. If they don’t trust the system, they won’t use it,” Vogels said.
Open source models offer a structural advantage here. Because developers can inspect and modify the code, and more easily fine-tune models on their own proprietary data, they tend to align better with the transparency expectations of regulated or sensitive environments. The caveat is real, though: even most open-weight providers do not fully disclose all the data on which the model was initially trained. Openness is a spectrum, not a binary.
What makes this moment analytically interesting is the convergence of two separate pressures — cost containment and trust requirements — both pointing in the same direction. Organizations that might have justified high proprietary model costs during the experimentation phase are now facing a different calculation: can they maintain board-level confidence in their AI investments when costs are unpredictable and training data provenance is opaque? For many, the answer is reshaping their entire AI stack.
New Amazon Open-Source Tool Aims to Empower Scientific Researchers
Amazon’s Vogels did not just diagnose the industry’s direction at the summit — he also announced a concrete step in it. A new open-source AI tool from Amazon is designed to make scientific data meaningfully more accessible, with a particular focus on institutions that lack the technical resources of large research universities or well-funded labs.
Integration with AWS Registry of Open Data
The tool connects to the AWS Registry of Open Data, which houses more than 1,100 datasets from major scientific organizations including NASA, NOAA, and the NIH. Rather than navigating complex data catalogs — a process that could previously consume hours — researchers can now query the registry using plain natural language. A scientist could ask for satellite imagery with specific licensing terms, or request genomics datasets for a particular population, and receive relevant results without needing to understand the underlying data architecture.
Facilitating Access for Under-Resourced Institutions
The practical implication for research is significant. Under-resourced institutions — smaller universities, NGOs, public health agencies in developing regions — have long faced a structural disadvantage when it comes to data discovery. The technical overhead of working with large scientific registries favors institutions with dedicated data engineers. By lowering that barrier through natural language search, the tool opens access to datasets in fields like climate science and public health that were previously harder to reach for non-specialist teams.
It also positions Amazon squarely within the open-source AI ecosystem at a moment when that ecosystem is attracting serious capital and talent. Ollama’s recent $65 million Series B, bringing its total funding to $88 million, signals that the venture community sees open-source AI tooling as a durable business — not just a transitional phase before proprietary models reassert dominance. The open-source trajectory, in other words, has institutional momentum behind it now, not just cost logic.
The harder question hovering over all of this is what happens to the performance ceiling. Open source models are closing the gap with frontier proprietary systems, but the most complex tasks — the ones that justify the highest model costs — still tend to favor closed, heavily resourced systems. Companies may end up running hybrid stacks: open-source models for the bulk of their inference load, with proprietary models reserved for specific high-stakes tasks. That architecture, not a clean switch from one to the other, may be where the industry actually lands.
FAQ
Why are companies shifting to open-source AI models?
Companies are shifting to open source AI models primarily to reduce mounting AI costs. Proprietary models from providers like OpenAI and Anthropic bill by the token, which can generate unpredictable and very large expenses at scale. Open source models are generally free to download, with users paying only for the cloud infrastructure needed to run them — an arrangement that often works out significantly cheaper, especially for high-volume deployments.
What concerns does Amazon’s CTO highlight about AI transparency?
Amazon CTO Werner Vogels emphasized that transparency about AI training data is increasingly important for enterprise adoption. Organizations want to know what data was used to train the models they deploy. This is particularly critical in sectors like healthcare, government, and humanitarian work, where trust in the system is a prerequisite for actual use — especially when those systems serve vulnerable communities.
What is the purpose of Amazon’s new open-source AI tool?
The tool allows researchers to search the AWS Registry of Open Data — which contains more than 1,100 datasets from organizations including NASA, NOAA, and the NIH — using natural language queries instead of navigating complex technical catalogs. The goal is to reduce the time and technical expertise required to find relevant scientific datasets, with a particular focus on making that access more equitable for under-resourced research institutions.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.
