Mozilla's September update to "The State of Open Source AI" contains a chart that places two curves side by side. One is the unit price of AI inference, which has fallen roughly 60 times over four years. The other is the spot price of memory, which has risen about 700% over the past year.
Those two lines capture exactly what enterprises have been weighing this year. Models keep getting cheaper, usage keeps climbing, and nobody wants their data sitting in someone else's building, so should AI move back into your own data center? Yet the moment you start pricing it, the servers turn out noticeably more expensive than last year.
Moving AI on-premise was never something a single phrase like "data stays in-house" could settle. It is an equation, and the largest term in that equation usually does not appear on the hardware quote.
## Open models are no longer the backup option
Start with where open models now stand.
According to Mozilla, eight of the top ten models by token volume on the model routing platform OpenRouter in August were open-weight models, and seven of those eight came from Chinese teams. In early August, DeepSeek became the first open model to reach number one by weekly requests on that platform, ending Google's 51-week run at the top.
Enterprise numbers are moving the same way. An enterprise survey released in September by the US research firm ETR found that open-weight models now account for 34% of respondents' AI token usage, up from 23% a year ago, with respondents expecting 41% within a year. On deployment, self-managed cloud accounts for 46% and on-premise infrastructure for 42%, now close to parity.
In other words, open models are no longer a stand-in for when budget runs short. They are among the main options enterprises evaluate seriously.
## Twelve points behind in production, and the bigger the company the wider the gap
The same Mozilla report also offers an honest number: open models reach production 12 percentage points less often than closed ones.
What deserves more attention is how that gap is distributed. Broken down by company size, the larger the organization, the wider the gap between open and closed production rates, reaching about 16 points at companies with more than a thousand employees. The report's reading is blunt: scale closes the gap for closed models but leaves the open-model gap where it was, which means the problem is tooling, not budget.
That line is worth sitting with. Large enterprises have money, people, and data centers, so on paper they are best placed to self-host, yet they are the ones stuck deepest. The reason is that the bigger the organization, the more systems to connect, the more scattered the data, the more complex the permissions, and the stricter the audit requirements. The model itself runs as soon as you download it. Getting it to read the right data safely, plug into existing processes, and have someone accountable when it goes wrong is where the real work lies.
ETR's survey echoes this. Asked for the biggest barrier to putting open-weight models into production, enterprises ranked incomplete security and compliance review first at 62%, up 9 points since July. Self-hosting cost came second, rising from 40% to 50%.
## Taiwan: lots of use, little scaling
The Taiwanese numbers sketch a very similar outline.
At its MIC FORUM Fall in September, the Market Intelligence and Consulting Institute (MIC) of the Institute for Information Industry released a survey showing that 53% of enterprises expect generative AI to become widespread within three years and 39% of organizations have begun using AI agents, but only 23% have actually expanded that use across internal operations.
MIC's deputy director noted that enterprises are combining open models, internal databases, and retrieval-augmented generation (RAG) to build dedicated AI systems on-premise. He cited financial institutions tuning models into credit specialists, hospitals using on-premise architecture to handle sensitive medical records, and petrochemical firms building models for industrial safety and raw-material forecasting. He added that future model selection will weigh five things more heavily: inference cost, deployment autonomy, data governance, reliability, and application integration.
Of those five, only inference cost depends mainly on the model itself. The other four, deployment autonomy, data governance, reliability, and application integration, are all questions of system architecture and integration. Most of what holds back the 16 points between 39% and 23% sits in those four.
Small and mid-sized businesses are at an even earlier stage. A survey for the Ministry of Economic Affairs' SME white paper found an AI adoption rate of 7.4% among Taiwanese SMEs, against an average of 39% among SMEs in OECD countries. More than 90% of SMEs already have basic digital infrastructure, but only 15.3% can actually visualize, forecast from, or add value to their operational data.
A survey of Taiwan's top thousand SMEs released on 19 September by Dun & Bradstreet and Taiwan's AI Foundation (AIF) found that 63.1% remain in the first two stages of AI adoption and 86.7% of AI investment decisions have yet to be formally executed. Meanwhile, roughly eight in ten employees are already using AI tools on their own at work. The company is still holding meetings to evaluate, while staff have long since started using it, just somewhere nobody is managing.
## What gets stuck is the data
String these numbers together and the answer is remarkably consistent.
A survey by Cloudera and Harvard Business Review Analytic Services this March found that only 7% of enterprises consider their data fully ready for AI. The leading obstacle respondents named was data silos and difficulty integrating across systems, at 56%.
That is why moving AI on-premise so often becomes a far larger project than expected. You can install a model in an afternoon, but for it to answer "why is this order late," it has to read the shipment record in the ERP, the production schedule in the MES, the inbound material status in the procurement system, and the customer email sitting in the support system. That data lives in different systems, different formats, and different permission scopes, and much of it has never been organized into anything queryable.
Where on-premise AI really costs money is not the machines. It is getting the data into a state the model can understand while only the right people can see it.
## Four questions to answer before moving on-premise
If you are evaluating self-hosting or on-premise deployment, we suggest getting clear answers to these four questions first.
- Is the data you want AI to handle already queryable by software, or is it scattered across spreadsheets, paper, and each department's own systems?
- Who defines access to that data? When the model answers a question, could a salesperson end up seeing numbers only finance should see?
- When an on-premise model errs or slows down, who investigates? Does anyone on your team handle model versioning and performance monitoring?
- Have you calculated total cost of ownership? Beyond servers and GPUs, there is power, facilities, redundancy, licensing, and at least one person's operating hours.
The first two are data governance, the last two are operational capability. If you cannot yet answer the first two, the best investment right now is usually not servers but data integration. Connect the data and clarify the permissions first. Whether you end up in the cloud, on-premise, or in a hybrid architecture, that investment will not be wasted.
## Models will keep getting cheaper. Integration will not.
Inference prices will keep falling, open models will keep closing in on closed ones, and the choices will only multiply. But whichever model you pick, the work of connecting it to your data and processes and keeping it running securely and reliably does not disappear because the model got cheaper.
That is where our work as a systems integration and digital transformation partner is centered. What we help enterprises do is not pick the strongest model. It is first mapping where the data lives, how systems connect, how permissions are divided, and who is accountable when something breaks, then placing AI inside an architecture that can be operated for the long term, in line with governance standards such as ISO 27001.
Because what decides how far AI can go was never the model's ranking. It is whether the system behind it can hold up.