Last month, Zhipu AI released GLM-5.2 and it outscored GPT-5.5 on SWE-bench Pro, one of the most rigorous AI coding benchmarks in use. It is open weight, MIT licensed, and free. Before that, Alibaba’s Qwen 3 pushed the benchmarks further. Before that, DeepSeek R1 matched GPT-4 class performance in January 2025.
The performance gap between open source models and closed frontier models has narrowed significantly, and on some of the most demanding benchmarks, open source models are now winning.
For businesses deciding how to adopt AI, that changes the strategic question entirely. It is no longer “which frontier model subscription do we buy?” It is “which open source model do we run, and where do we run it?”
The Problem With the API
The open weight models coming out of Chinese labs are genuinely capable. They are also free to download and run yourself. Most businesses are not doing that. They are using the API or the web interface, which means the inference is happening on servers in China.
China’s National Intelligence Law, passed in 2017, requires organizations operating in China to cooperate with state intelligence work. The Chinese government can compel access to data stored on those servers. DeepSeek, Qwen, and GLM all carry this exposure when accessed through their hosted APIs. GLM-5.2 ships under an MIT license that lets anyone download the weights, but running 744 billion parameters on your own hardware is not a realistic option for most organizations without a deliberate infrastructure strategy.
The answer is not to avoid these models. The answer is to not use their API.
Running Open Source Models on Infrastructure You Control
When the model weights are open, you decide where the inference happens. The options range from managed cloud services to dedicated hardware, and the right fit depends on what your organization already operates and how much control you need over the stack.
For organizations in the Microsoft ecosystem, Azure AI Foundry provides access to thousands of open source models from Hugging Face deployed directly within your Azure tenant. Inference runs inside your cloud boundary under your existing data governance policies, with no data leaving your Azure environment. AWS Bedrock offers a comparable approach for organizations already invested in AWS infrastructure. Both are managed paths that remove the complexity of building and maintaining the inference stack yourself.
Beyond the major hyperscalers, dedicated GPU cloud providers, on-premises hardware, and hybrid approaches all represent viable paths depending on data residency requirements, latency needs, and existing infrastructure investments. The common thread across all of them is that inference happens in an environment your organization controls, under terms you negotiated, in a jurisdiction you chose.
The Co-Location Transition Nobody Is Talking About Yet
Traditional co-location providers have been losing ground to Azure and AWS for the better part of a decade. The hyperscaler value proposition of managed infrastructure, global scale, and consumption-based pricing has steadily eroded the case for maintaining your own hardware in a third-party facility.
AI inference is going to reverse that pressure, but not in the way most people assume.
The transition is not about co-lo facilities hosting customer-owned GPU rigs. It is about co-lo providers owning the hardware themselves and renting inference capacity directly. That is a more significant business model shift, and it is a natural one. These facilities already have what inference workloads require: dense power capacity, precision cooling, physical security, and high-bandwidth connectivity. The missing piece has been the economics of the GPU hardware itself.
That is changing. As next-generation inference chips reach the market and fabrication capacity expands, GPU hardware costs are coming down and supply is increasing. The two go directly hand in hand. When the economics reach an inflection point, co-lo providers can invest in GPU hardware at competitive prices and offer rented inference capacity at rates that undercut hyperscaler pricing, without the global redundancy overhead and margin requirements that make cloud inference expensive.
For co-location providers, this is the transition that stops the erosion. They already have the physical plant, the enterprise customer relationships, and the operational expertise. Adding rentable inference hardware turns them from passive real estate into active infrastructure providers competing directly in the AI compute market.
For businesses, it opens a path to frontier-grade inference capacity without purchasing and maintaining the hardware, without routing data through a hyperscaler, and with contractual jurisdiction guarantees a cloud API cannot provide.
This Is an Infrastructure Decision, Not a Software Purchase
The businesses that get this right over the next two years are going to treat AI adoption as an infrastructure strategy, not a subscription decision.
The open source models are already frontier-quality. Azure AI Foundry and AWS Bedrock make self-hosted inference accessible today without building the stack from scratch. The co-location path is coming as hardware economics shift. Organizations that build the internal capability to evaluate, deploy, and govern open source models on infrastructure they control will have cost flexibility, data sovereignty, and architectural independence as the model landscape continues to evolve.
Organizations that sign multi-year API contracts with frontier labs are going to find themselves paying for a performance advantage that no longer exists, under data handling terms they did not fully evaluate, at a moment when a capable and self-hostable alternative was already available.
The capability is free. Where you run it is the decision.
How is your organization approaching AI infrastructure? Are you running models in your own environment or relying on frontier API providers?
Discussion
Comments are powered by GitHub Issues via utterances. A GitHub account is required to comment.