Some are, but nobody has measured how many. The surveys show a wide gap between using AI and getting value from it. McKinsey's 2026 survey found nearly nine in ten organisations use AI regularly, yet only 37% attribute any EBIT impact to it, unchanged from a year earlier. MIT NANDA reported in August 2025 that about 95% of enterprise generative-AI pilots it studied delivered no measurable profit-and-loss impact. There is no published study of SMEs buying AI servers that then sit idle, so the title is a fair question rather than a proven trend. The lesson holds anyway: define the process, data, users and success measure first, then choose between an API, cloud GPUs or your own hardware.
What should you answer before buying an AI server?
| Question to answer first | Why it matters (evidence) |
|---|---|
| Which process, and what measure of success? | Only 37% see any EBIT impact from AI (McKinsey 2026) |
| Buy or partner, or build in-house? | MIT NANDA: vendor and partner projects succeeded about 67% of the time; internal builds about one-third as often |
| How many people will use it at once? | Concurrency drives GPU count and whether you need one box or several |
| Is the data sensitive or residency-bound? | 74% rate data residency or in-region compute important (Deloitte 2026); PDPA transfer rules |
| Will you track cost per unit of inference? | Only 44% of organisations do (Deloitte 2026) |
| Can an API or cloud trial prove value first? | Cloud GPU rental starts at a few US dollars an hour; Singapore's Enterprise Compute Initiative funds cloud, not hardware |
| Who will run and patch it? | The Enterprise Compute Initiative requires at least a two-person tech team |
Source: McKinsey, The state of AI in 2026 (25 Aug 2026, 1,719 respondents); MIT NANDA, The GenAI Divide, as reported by Fortune (18 Aug 2025); Deloitte, State of AI in the Enterprise 2026 (3,500 leaders); DISG Enterprise Compute Initiative page (updated 17 Jul 2026). Cloud rental figure: RunPod RTX PRO 6000 pricing from about US$2.49 an hour (2026).
What do the surveys actually show?
Adoption is near-universal; value is not. McKinsey's August 2026 survey found 44% of organisations scaling AI across the enterprise, up from 38%, but the share seeing any EBIT impact stayed at 37%. About two in ten were scaling AI agents.
Pilots often stall. MIT NANDA's report, based on 150 leader interviews, a 350-employee survey and 300 public deployments, found about 95% of enterprise generative-AI pilots produced no measurable P&L impact. Its method has been criticised as thin, and "fail" here means no measurable financial impact, not that nothing worked. It also found that more than half of budgets went to sales and marketing, while the best returns came from back-office automation.
Few track the economics. Deloitte's 2026 survey found 71% track total inference or API spending, but only 44% track the cost per unit of inference. Without that figure, comparing an owned server against an API is guesswork.
These are mostly large-enterprise surveys about generative-AI projects, not SMEs buying hardware. We found no independent study of on-premise GPU utilisation at SMEs, so we won't claim most such servers sit idle. Even at hyperscale, utilisation is far from full: Meta reported 38–43% model FLOPs utilisation training Llama 3 405B, and a ByteDance study (OSDI 2025) found slow jobs wasting about 10% of allocated GPU-hours.
Why do companies buy the hardware first?
Some reasons are sound. Memory and SSD prices rose sharply through 2026, and lead times lengthened, so buying early has saved some firms money. Local hardware also lets teams experiment with sensitive data that cloud policies would block. Others are weaker: wanting to be seen doing AI, a budget that must be spent this year, or a vendor demo that worked on clean data. We cover the demo-to-production gap in Why AI Projects Look Amazing in Demos but Fail in Real Offices, and the broader adoption mistakes in The Biggest Mistake Companies Make When Adopting AI.
The costly mistake is skipping straight to a rack GPU server or a deskside AI supercomputer, with power, cooling and support commitments, before a use case has proved its value.
A S$10,000 AI box can be part of finding your strategy. A server costing ten times that should come after you have one.
What sequence works better?
- Pick one process and one measure. For example, hours spent answering supplier queries, or time to draft a tender response.
- Prove it with an API or off-the-shelf tool. A Singapore Chinese Chamber of Commerce survey reported 45% of SMEs started with off-the-shelf tools (September 2026).
- Check data and governance. The PDPC finalised advisory guidelines on personal data in generative AI on 20 July 2026. Decide what data the system may touch before choosing where it runs.
- Measure usage and cost per task during the pilot. This is the figure that decides between API, cloud GPUs and owned hardware.
- Buy hardware when the numbers support it: steady daily use, sensitive data, or API costs that exceed ownership costs. Our DGX Station vs cloud GPUs three-year ROI post shows one way to run that comparison.
- Name who will run it. Patching, model updates and user support need an owner. See who manages an AI supercomputer.
What about grants?
Be careful here. The Enterprise Compute Initiative, run by DISG with up to S$150 million, provides cloud credits from Google Cloud, Microsoft and AWS plus consultancy support; it does not cover on-premise hardware. Its eligibility rules, including at least ten Singapore staff, a tech team of two or more and CEO sponsorship, effectively reward having a strategy and team first. Hardware eligibility under newer schemes such as EDGE has not been stated. Don't assume an AI server is claimable; check with Enterprise Singapore or your adviser.
Which hardware fits which stage?
For exploring use cases with private data, a compact GB10 computer with 128 GB of memory runs models of up to about 120 billion parameters for a small team; see who should buy a S$10,000 AI computer. For production serving to many users, GPU servers or a DGX Station make sense once usage is known. Our post on networking versus GPU power explains why concurrency, not raw GPU count, usually decides the step up, and smaller AI workstations vs data centres covers where each belongs.
Buying in Singapore
For the discovery stage, the ASUS Ascent GX10 64 GB / 1 TB is S$6,670 and the NVIDIA DGX Spark 4 TB is S$9,300 (listed 8 Nov 2026). For production, GPU servers such as the Lenovo ThinkSystem SR675 V3 and HPE ProLiant DL380a Gen11, and DGX Station systems such as the MSI XpertStation WS300, are price on request. Browse AI workstations and AI servers. Prices change; ask for a current quotation. We can help size hardware against a measured pilot: use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Do most AI projects fail? MIT NANDA reported that about 95% of enterprise generative-AI pilots it studied had no measurable P&L impact. McKinsey found 37% of organisations see some EBIT impact from AI. Both mainly describe large companies.
Should we try cloud before buying an AI server? Usually, yes. A cloud or API pilot shows real usage and cost per task, which are the figures you need to justify owned hardware.
When does owning AI hardware make sense? When use is steady and measured, data must stay under your control, or API and cloud costs exceed the cost of owning and running the hardware.
Can an AI server be funded by a Singapore grant? Don't assume so. The Enterprise Compute Initiative funds cloud credits, not on-premise hardware, and other schemes haven't stated hardware eligibility. Check with Enterprise Singapore or your adviser.
Is a small AI computer a good first step? Often. A GB10 machine with 128 GB lets a small team test models on private data before committing to a server and its running costs.
