Why Businesses Are Bringing AI Back Into Their Own Offices
The first wave of business AI was a browser tab pointed at someone else's data centre. The next wave, for a growing number of organisations, is a box under a desk or in the server room, because the questions that matter now are where the data goes, what the bill will look like next year and who controls the model. That does not make the cloud wrong; it makes "cloud only" a choice that has to be argued rather than assumed.
This post sets out why some businesses are moving AI workloads back on-premises, what the published survey evidence says (and who paid for it), the honest case for staying in the cloud, and where desk-side hardware such as the ASUS Ascent GX10, the ASUS ExpertCenter Pro ET900N G3 and NVIDIA RTX Spark laptops fits.
The reasons in one table
| Reason to bring AI in-house | What it means in practice | The cloud's counterpoint |
|---|---|---|
| Privacy and PDPA | Prompts containing personal or client data stay on hardware you control | Enterprise cloud contracts offer data-handling commitments; local hardware still has to be secured |
| Data residency | No question of which country the prompt is processed in | Some providers offer regional hosting options |
| Predictable cost | A one-off purchase instead of a usage bill that grows with adoption | No upfront spend; you pay only for what you use |
| Latency | Requests travel across the office network, not the internet | For most chat use, cloud response times are acceptable |
| Model choice | Run the open-weight model you choose, pinned to a version you control | The strongest frontier models are only available as cloud services |
| Operations | Your team decides when to update | Your team also has to do the updating |
Is AI repatriation a real trend or vendor marketing?
Both, and it is worth reading the evidence with that in mind. In August 2026 Cloudera published a survey fielded by Wakefield Research from 5 to 22 June 2026 among 1,500 enterprise architects, cloud infrastructure leads and data architects in nine markets, including 100 respondents in Singapore. Two-thirds (66%) said they had moved AI workloads from public cloud back to private cloud in the past year, and 84% reported increased infrastructure costs driven by AI workloads.
Earlier, in March 2026, Cloudian released a survey of 203 enterprise IT decision-makers run through the Centiment platform in February 2026. It found 93% had repatriated some AI workloads from public cloud, were doing so, or were evaluating it, and 91% would choose on-premises, private cloud or hybrid over public cloud for AI involving sensitive company data.
The caveat: both surveys were commissioned by companies that sell private-cloud or on-premises data infrastructure, and "evaluating" is a much lower bar than "moved". We read them as evidence that the question is being asked seriously, not as proof that the cloud is in retreat. In our own enquiries the pattern is similar in general terms: more of the AI conversations we have start with data handling and running cost rather than with which model is smartest.
Privacy and PDPA: the strongest argument
Singapore's Personal Data Protection Act sets out obligations that apply whatever tool you use. The PDPC's summary of the data protection obligations includes the Transfer Limitation Obligation (personal data may be transferred to another country only according to the requirements prescribed under the regulations) and the Protection Obligation (reasonable security arrangements must be made to protect personal data in the organisation's possession).
A local model removes one question entirely: if a staff member pastes a customer email into an assistant running on an office GX10, the prompt is processed on hardware in Singapore that the organisation controls. It does not remove the Protection Obligation. A desk-side AI box holding HR files or client documents needs access control, patching and backups like any server. On-premises moves the responsibility to you; it does not make it disappear. Our post on local AI agents on your desk covers permissions and human review.
Predictable cost, without redoing the maths
The cost case is real but narrower than the headlines suggest. A fixed purchase is easier to budget than a token bill that rises with every team that adopts the tool. But local hardware only pays back quickly when it replaces steady, high-volume use of an expensive per-token model. We have already worked through that break-even in when running AI locally beats the cloud, with the full numbers in the 12-month local AI vs cloud API cost model.
Latency and model choice
A local model answers over the office network, so there is no internet round trip and no dependence on a provider's queue at busy times. It matters most for agents that make many calls in a row.
Running your own open-weight model means you decide which version runs and when it changes, so a workflow that works on Monday is not altered by a provider's silent update on Tuesday. Memory decides which models you can run: the GX10's 128 GB of unified memory is what puts 70B-class models within reach, as explained in why 128GB of unified memory matters for local AI.
The case for staying in the cloud
Scale and bursts. A cloud API absorbs a sudden spike in demand; a single desk unit serves a limited number of people at once. If usage is spiky or unknown, the cloud is the safer starting point.
Frontier quality. The most capable models are cloud services. For hard reasoning or specialist tasks, a local open model may simply not be good enough.
Maintenance. Someone has to update the operating system, model server and models, and restore the box if it fails. If no one owns that, local AI becomes a liability. The hidden costs of owning an AI workstation lists what goes beyond the purchase price.
Speed to start. A cloud pilot can begin today. Hardware needs a purchase order, delivery and setup. We often suggest proving the use case in the cloud with non-sensitive data first, as argued in the biggest mistake companies make when adopting AI.
Our verdict: bring the right work home
The question is not cloud or on-premises; it is which of your AI workloads you would be comfortable explaining to a client whose data went into the prompt.
Our view is that most organisations will end up hybrid. Steady, sensitive, high-volume work (document search over internal files, summarisation, classification, coding help, overnight agents) moves in-house. Occasional hard problems go to a frontier cloud model, using data you are comfortable sending. Light, non-sensitive usage should stay in the cloud for now. If you are unsure whether you need hardware at all, start with do you need an AI supercomputer?
What on-premises AI looks like on a desk
A shared office AI server. The ASUS Ascent GX10 1TB is S$7,200 as listed on 23 Oct 2026, with the 2TB model at S$8,050 and the 4TB model at S$9,850; all three use the NVIDIA GB10 with 128 GB of unified memory and showed stock on our store. The GB10 and DGX Spark buyer's guide compares the options.
A department-scale system. For larger models or more simultaneous users, the ASUS ExpertCenter Pro ET900N G3, a DGX Station GB300 desk-side system, is price on request and quoted per configuration.
AI that travels. For one person who needs local models away from the office, the MSI Prestige N16 Flip AI+ with NVIDIA RTX Spark comes as the C1XMT-224SG with 32GB (S$4,974.08), the C1XMT-223SG with 64GB (S$6,331.28) and the C1XMT-222SG with 128GB (S$10,854.12), as listed on 23 Oct 2026. Our RTX Spark laptop guide explains what each memory size can run. Note that an AI laptop and an AI PC are not the same thing; what AI PCs can actually do draws the line, and whether the office PC is dying looks at where desk hardware is heading.
Buying on-premises AI hardware in Singapore
SourceIT supplies ASUS and MSI with local warranty. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page. Setup and commissioning can be quoted alongside the hardware. Government buyers can purchase through GeBIZ and Vendors@Gov, where we are registered; see buying ASUS Agent Computers via GeBIZ. Before you ask for a quote, what IT vendors wish customers knew will save a round of emails.
Frequently asked questions
What does AI repatriation mean? Moving AI workloads that run on a public cloud service back to infrastructure the organisation controls, such as on-premises servers, desk-side AI systems or a private cloud.
Are businesses really moving AI back on-premises? Some are. A Cloudera survey published in August 2026 found 66% of 1,500 respondents had moved AI workloads from public cloud back to private cloud in the past year. It was vendor-commissioned, so treat it as a signal rather than proof.
Does running AI locally make us PDPA compliant? No. It keeps prompts on hardware you control, which addresses cross-border transfer questions, but the PDPA's Protection Obligation still requires reasonable security arrangements for the personal data on that hardware.
Is on-premises AI cheaper than the cloud? Only for steady, high-volume use that would otherwise run on an expensive per-token model. For light or occasional use the cloud is usually cheaper.
What hardware do we need to run a 70B model in the office? A system with enough memory to hold the model, such as the ASUS Ascent GX10 with 128 GB of unified memory. Larger models or many simultaneous users point to a GB300 system such as the ExpertCenter Pro ET900N G3.
Should we drop cloud AI entirely? Rarely. Most organisations are best served by a hybrid: sensitive, steady work locally and occasional frontier-model tasks in the cloud.
