Quick answer: Verify seven hardware fields before you shortlist an AI kiosk: NPU sustained compute, memory and expansion headroom, thermal design under 24/7 duty, full-load power draw, camera and privacy handling, peripheral IO, and dual-source availability. Each field should be backed by a supplier test report, not a spec-sheet line. Peak numbers are marketing; sustained numbers are procurement.
Overview
Verify seven hardware fields before you shortlist an AI kiosk: NPU sustained compute, memory and expansion headroom, thermal design under 24/7 duty, full-load power draw, camera and privacy handling, peripheral IO, and dual-source availability. Each field should be backed by a supplier test report, not a spec-sheet line. Peak numbers are marketing; sustained numbers are procurement.
Most AI kiosk pages you will find are vendor product listings or component-maker blogs that stop at the feature list — what the hardware has, and why it is good. That is not enough to shortlist on. What follows is a buyer-side verification pass: each spec becomes an evidence request you can send to every supplier in your shortlist and compare answers side by side. That comparison is what protects your BOM, your field-failure exposure and your legal sign-off.
This is written for procurement and product managers at kiosk integrators and OEM buyers building or reselling edge-AI self-service units. It is not a consumer tablet comparison, and it is not a “what is edge AI” explainer.
| Spec field | What suppliers usually state | What to request as proof |
|---|---|---|
| NPU / compute | Peak TOPS, processor generation | Sustained inference rate at rated ambient and duty cycle, with test conditions |
| Memory & expansion | Max RAM, slot count | Max capacity in the shipped configuration and free lanes after peripherals |
| Thermal design | “24/7 rated”, fan or fanless | Burn-in report: full-load AI inference at rated ambient, logged throttling events |
| Power draw | Nameplate or idle watts | Measured watts at idle, typical and peak AI load; PSU headroom; input range |
| Camera & privacy | Resolution, frame rate, “AI camera” | On-device vs cloud inference, retention defaults, biometric disable by SKU, shutter option |
| Peripheral IO | Scanner, payment, audio modules | Decode latency, interface standard, operating temperature range, pilot-unit test plan |
| Dual-source BOM | Single processor/module part number | Named alternates per critical line item and sourcing status at quote stage |
1. NPU and compute: separate peak TOPS from sustained throughput
Verify the NPU’s sustained inference rate under thermal load, not the peak TOPS figure, and confirm processor generation, memory ceiling and expansion lanes so the platform can absorb model updates without a board redesign. Peak TOPS is measured in conditions your kiosk will never run in.
Peak TOPS is the most-abused number in this category because it is the easiest to publish and the hardest to falsify at the quotation stage. Ask instead for sustained throughput at the ambient temperature the unit will actually see, at the duty cycle it will actually run.
Vendor documentation is a useful format template even when the vendor is not your supplier. Premio’s published description of its CT-XAR01 industrial motherboard, for example, states Intel Core Ultra Series 2 processors with an integrated NPU, support for up to 96 GB DDR5 memory, multi-display and peripheral connectivity, and MCIO and PCIe Gen5 expansion ( premioinc.com ). That is the level of field-by-field specificity to demand — not because that board is right for your build, but because your shortlist should answer in the same format.
Scripted question to send: “What sustained inference rate do you guarantee at [X]°C ambient over a [X]-hour continuous run, and at which facility was it measured?”
2. Memory, storage and expansion headroom
Confirm RAM ceiling, storage endurance class and free expansion lanes, because AI kiosk deployments typically grow from a single pilot model to multi-model vision workloads within the unit’s service life. Headroom is a lifecycle decision, not a spec-sheet line.
Capture these fields for every shortlisted unit:
- DDR generation and maximum capacity as shipped , not as supported by the platform
- SSD endurance class (TBW) and the storage module’s temperature rating for the enclosure it sits in
- Expansion lanes still free after the standard peripheral set is fitted — M.2, PCIe, MCIO
The common trap is a spec sheet that lists expansion slots without stating how many are consumed by the configuration you are quoting. A platform that advertises three slots and ships with two occupied leaves you one upgrade path, not three.
The sourcing consequence is direct: locked-down expansion forces a mid-life board swap, which changes MOQ, tooling assumptions and lead-time math. If you are already managing lifecycle risk on the commercial side, the same contract pass should cover price and BOM exposure — see the dual-source BOM and tariff clause checklist .
3. Thermal design for 24/7 duty: request sustained-load evidence
Verify thermal design against sustained 24/7 duty at the highest ambient the unit will actually see, using supplier burn-in data or a witnessed test, because enclosure airflow paths — not the chipset alone — determine field failure. An AI workload changes the thermal profile of a kiosk substantially versus a passive transaction terminal.
Request the following in writing:
- The ambient temperature the thermal design is rated to, and whether that rating includes the AI inference workload
- The airflow path and dust/filter strategy, plus the filter service interval assumption
- Fan versus fanless, and the documented failure mode of the chosen cooling approach
What to require as evidence — the method, not a promised outcome: continuous full-load AI inference at rated ambient for an agreed duration, with logged failure points and throttling events. Specify the duration and ambient yourself, based on the deployment site, rather than accepting a supplier’s chosen test window. Any “24/7 rated” claim that arrives without a test report should be treated as unconfirmed, and priced as risk.
4. 24/7 full-load power draw: size circuits and PSU headroom correctly
Verify measured power draw at full AI load, not idle or nameplate-only figures, since site electrical planning, UPS sizing and PSU headroom all depend on the peak the unit actually pulls. Idle numbers understate the electrical work at every site.
Data points to request per SKU:
- Measured watts at idle, at typical AI load and at peak AI load
- PSU rating and the headroom between rated output and measured peak
- Input voltage range and connector standard for the target market
This is a rollout-cost item, not an engineering footnote. An under-sized circuit or UPS adds site work to every deployment, not just the first unit, and it is far cheaper to resolve before the enclosure and power module are locked. Treat all electrical figures as requiring supplier-confirmed test data before they enter a site plan or an RFQ response.
5. Camera, biometrics and privacy handling: confirm, don’t assume
Verify where inference runs, what is stored, for how long, and what the kiosk displays as notice, because camera and biometric handling drives legal review in the US far more than compute choice does; specific obligations vary by state and must be confirmed with counsel.
Ask the supplier for written answers on:
- On-device versus cloud inference, and what leaves the unit if cloud is used
- Data retention and deletion defaults, including whether footage or embeddings persist after a session
- Whether biometric features can be disabled at SKU level, not just switched off in software
- Physical shutter or recording-indicator options for public-facing deployments
Separate two very different hardware roles. A camera used for people-counting and queue analytics is a vision-system component. A camera used for face matching is an identity system. They trigger different review paths inside your customer’s organization, and often different timelines for legal sign-off. If the kiosk also takes card payments, your privacy review sits alongside payment scope — see the PCI DSS scope checklist for branch self-service kiosks .
To be explicit: nothing here is a legal claim about any jurisdiction. The compliance determination belongs to your counsel, and the supplier’s role is to give you accurate hardware and data-flow facts to hand them.
6. Peripheral and IO specs that AI kiosks actually depend on
Verify scanner decode latency, camera resolution and frame rate, microphone array behaviour and payment module interface, because AI kiosks fail at the peripheral handoff more often than at the model. The model can be correct and the experience still fail if the scan is slow.
| Peripheral | Published spec to capture | What to test on the pilot |
|---|---|---|
| Barcode / document scanner | Decode latency, decode rate, field of view, operating temperature | Decode time on damaged and low-contrast media at expected volume |
| Camera / vision module | Resolution, frame rate, field of view, low-light behaviour | Inference accuracy at the site’s actual lighting |
| Audio array | Microphone count, noise handling, speaker output | Wake-word and conversational performance in ambient noise |
| Payment module | Interface standard, certified configurations | Transaction flow, timeout and error recovery |
Published spec formats give you a benchmark for the specificity to demand. Aila Tech’s self-service kiosk scanning listings, for instance, state decode time, decode rate, field of view, resolution and an operating temperature range rather than a generic “high-speed scanning” claim ( ailatech.com ). Ask every shortlisted supplier to match that granularity.
Peripheral choice also constrains enclosure design, which means it touches tooling and MOQ. If your cash or payment module selection is still open, the comparison in cash module selection and ROI is the adjacent decision. The broader risk review sits in the cash handling kiosk buying checklist .
7. Dual-source BOM and spec verification pack
Ask for a dual-source BOM and a verification pack — test reports, spec-by-spec sourcing and named alternates — at the quote stage, because single-source compute and memory are the AI kiosk BOM’s largest continuity risk. Waiting until after tooling is committed removes your leverage.
What the pack should contain, per critical line item:
- The specified part and at least one named alternate, with current sourcing status
- The test report supporting each performance claim you intend to rely on contractually
- Confirmation of which configuration was tested — board, memory, storage and enclosure together
- A written statement of which claims the supplier will and will not stand behind
Review it against your own spec acceptance criteria before you shortlist, not after. Market research is useful here only as a checklist of what to look for: Fact.MR’s AI kiosk market breakdown segments hardware into interactive displays, cameras and vision systems, processors and edge AI chips, sensors, printers and scanners, payment and POS modules, audio systems and connectivity modules ( factmr.com ) — a reasonable structure for auditing whether your BOM pack covers every subsystem you are depending on. It is not a substitute for your own test data.
FAQ: AI kiosk hardware spec verification
What NPU spec should I require in an AI kiosk, and is peak TOPS enough?
No. Peak TOPS describes a best-case condition the kiosk will not run in. Require sustained inference rate at a stated ambient temperature over a stated run duration, plus the test conditions used to derive it.
How do I verify a kiosk is genuinely rated for 24/7 full-load operation?
Request a burn-in report covering continuous full-load AI inference at rated ambient, with logged throttling and failure points. If the workload was not included in the thermal test, the rating does not cover your use case. Treat unbacked “24/7” claims as unconfirmed.
What power draw figures should a supplier provide for site electrical planning?
Measured watts at idle, typical AI load and peak AI load, plus PSU rating, headroom and input voltage range — per SKU, not per product family. Confirm them in writing before they enter a site plan.
Does an AI kiosk with cameras automatically create US biometric-privacy obligations?
It depends on what the camera does with the data, where inference runs and what is retained. A people-counting camera and a face-matching camera are different risk profiles. Obligations vary by state, and the determination should come from your counsel — this page makes no legal claim.
Turn the checklist into a quote
Seven fields, one artifact: NPU sustained compute, memory and expansion headroom, thermal burn-in evidence, full-load power draw, camera and privacy handling, peripheral IO specificity, and a dual-source BOM with a spec verification pack.
The checklist is only worth the paper it is written on if the supplier will sign to it. Request a 2026 quote plus a dual-source BOM and spec verification checklist — send us your target ambient, duty cycle and peripheral set, and we will return the quote with the test and sourcing documentation attached to each line. Start with our custom kiosk OEM/ODM manufacturing path to see how tooling, IP and exit terms are structured before you commit.
