When to run AI on your own servers

A year ago, if you wanted the best AI, you used an API from OpenAI, Anthropic or Google, and your data went to their servers. Open models existed, but they were clearly a step behind.

That gap has narrowed a lot this spring. Google released Gemma 4 on 2 April under the Apache 2.0 licence, in sizes from small models that run on a phone to a 31 billion parameter version for a single server. DeepSeek published a preview of V4 on 24 April, with a context window of a million tokens. A smaller Qwen 3.6 model from Alibaba, Moonshot's Kimi K2.6 and, in March, Sarvam's 30 and 105 billion parameter models for Indian languages all came with downloadable weights. In the Gulf, TII's Falcon-H1 Arabic and Jais 2, from G42's Inception and MBZUAI, are open too.

"Open weights" means the trained model itself is published. You can run it on hardware you choose, and nothing you send it goes back to the company that made it. For businesses that handle sensitive data, that matters.

Why businesses here are asking

Three reasons come up often.

Data protection is the first. India's DPDP Rules, notified last November, phase in obligations on how personal data is collected, used and kept, with most of them applying from May 2027. The UAE has its own data protection law, and sectors such as healthcare and finance have stricter rules again. Many organisations simply prefer that customer data never leaves systems they control.

Cost at volume is the second. If you process tens of thousands of documents or messages a day with a simple, repeatable task, paying per token adds up, and a dedicated model can work out cheaper.

Control is the third. A model you host does not change behaviour overnight because a provider updated it, and it does not disappear when a provider retires a version.

When self-hosting makes sense

• The data cannot leave your infrastructure, by law, contract or policy.
• The task is narrow and high-volume: classifying, extracting fields, routing, summarising, where a mid-sized model is good enough.
• You need it to work offline or at the edge, such as in a clinic, a warehouse or on a device.
• You want to fine-tune a model on your own data for a specialised job.
• You have, or can hire, people to run it.

When an API is still the better deal

• Your volume is low or uneven. A server sitting idle most of the day costs money; an API charges only when you use it.
• You need the very best model for complex reasoning, agents or coding. The top closed models still lead there.
• You have no one to look after GPUs, updates, monitoring and security.
• You need it working next week.

Count the running costs. A large model like DeepSeek V4-Pro needs a cluster of high-end GPUs. Even a 31 billion parameter model needs a serious graphics card, plus someone to keep it patched, monitored and backed up.

The middle path

For most teams, the best first step sits between the two.

AWS, Microsoft Azure and Oracle run cloud regions in the UAE, and all the major providers have regions in India. Running an open model on rented cloud servers in those regions, or through a managed service where one is offered there, gives you open weights and data that stays in the country, without buying hardware. Check which models are available in which region before you plan around one. Separately, the paid APIs of the big providers do not train on your data by default, and enterprise agreements can add stronger guarantees. For many workloads that is enough, and it is far simpler.

A sensible pattern is to sort your AI tasks by sensitivity. Use a top API model for general work that does not involve personal data. Send tasks with sensitive personal or financial data to an open model you host or that runs in-country. Keep the code that talks to models behind a thin layer of your own, so moving a task from one to the other is a configuration change.

A word on where models come from

Several of the strongest open models come from Chinese labs. When you run the weights yourself, your data does not go to the developer, which removes the most common worry. It is still worth testing how a model behaves on topics that matter to your business, reading its licence carefully, since some open licences have conditions, and checking whether your sector or your clients have rules about model origin.

Where to start

Pick one task where data sensitivity or volume is the real issue, and run a small test with two or three open models against the API model you would otherwise use, on your own examples. If an open model passes, you have a cheaper, more private option for that job. If it does not, you have learnt that for a few days of work, and with open models improving every few months, you can run the same test again later.