Would You Let an AI Supercomputer Run Your Company's Entire Customer Service Operation?

Not the entire operation, and not on one box. A DGX Station GB300 has the capacity to run private AI models for a large share of customer conversations. Reviewers have measured thousands of tokens per second across dozens to over a hundred simultaneous requests. But customer service is a 24/7, customer-facing system, and a single Station has one GPU, one 1,600 W power platform and no built-in failover. Singapore's agentic AI guidance also expects humans to stay accountable for what an AI agent does. The sensible design is AI that drafts, triages and answers routine questions, with human approval for anything consequential, a fallback when the box is down, and a clear hand-off to a person.

How much customer traffic could one DGX Station handle?

Per-request figures are our division of each reviewer's total by its stated concurrency. Results depend on model, quantisation, prompt length and serving software. The B200/GB200 row is from different NVIDIA hardware and is shown only as a proxy for a 70B-class model; no DGX Station or B300 figure for that test has been published, and there is no MLPerf result for the DGX Station. Sources: StorageReview reviews of the MSI WS300 (24 Aug 2026) and ASUS ET900N G3 (23 Sep 2026), ServeTheHome's MSI WS300 review (23 Oct 2026) and NVIDIA's TensorRT-LLM performance overview (14 Sep 2026). SourceIT has not benchmarked these systems.

The takeaway: capacity is rarely the bottleneck for a typical company's customer chat volume when using a mid-sized model. Reliability, accuracy and control are.

Why is one Station risky as the whole operation?

A customer-service platform is judged on the hours it is down, not the hours it is up. Every DGX Station OEM version runs one GB300 superchip on a single 1,600 W power platform. A failed drive is covered by the RAID 1 OS mirror, but a failed board, pump or power supply takes everything offline until the OEM repairs it. Warranty terms differ: SourceIT's HP and Supermicro listings state 3-year onsite cover, while MSI, ASUS and GIGABYTE terms are confirmed at quotation. Even next-business-day onsite service means a day without the AI. Patches, driver updates and MIG layout changes also need maintenance windows. For the full picture, read Your DGX Station Could Be a Single Point of Failure for Your Entire AI Team.

Let the AI answer the easy questions at 3am. Do not let it be the only thing standing between your customers and your company.

What does a safer design look like?

• Human hand-off built in. Platforms already support this: Microsoft documents Copilot Studio handing a conversation to a live agent in Dynamics 365 Customer Service with "the full history of the conversation and all variables".
• Approval for consequential actions. Refunds, account changes, cancellations and anything involving money or personal data should need a human click. OWASP's LLM06:2025 "Excessive Agency" lists human approval, least privilege and authorisation enforced in downstream systems, not by the model, as core mitigations.
• Fallback when the box is down. Route to human agents, a second Station (NVIDIA documents a switchless two-node link), or a cloud model approved for that data.
• Retrieval, not memory. Answer from your own policies and product data through a RAG pipeline so responses can cite sources and be corrected without retraining.
• Logs and monitoring. CSA's "Securing Agentic AI" addendum (17 June 2026) asks for central logging of agent actions and tool calls to a SIEM, plus guardrails against prompt injection. Customer messages are untrusted input, so assume someone will try.
• Power protection. A 24/7 service on a 1,600 W system needs an appropriately sized UPS and a licensed electrician's check of the circuit.

What do Singapore's rules expect?

IMDA's Model AI Governance Framework for Agentic AI (22 January 2026) is guidance rather than law, but its message is direct: "humans are ultimately accountable", with checkpoints for human approval and limits on an agent's autonomy, tools and data access. Customer conversations contain personal data, so the PDPA's protection, retention and accountability obligations apply wherever the model runs. Running the model on your own DGX Station keeps conversations on premises, which simplifies questions about overseas transfer, but it puts security and retention squarely on your team. Related reading: Would You Trust an AI Agent With Your Company's Credit Card? and What Happens When Your AI Agent Can Control Excel, Outlook and Your Entire Desktop?

So what should the Station actually do in customer service?

Good fits are agent-assist (drafting replies for human agents to send), triage and routing, summarising long email threads, searching internal knowledge for the agent, and answering routine questions such as opening hours, order status or warranty terms from approved sources, with an easy route to a person. Poor fits are unsupervised refunds, complaints handling and anything where a wrong answer has legal or financial consequences. Splitting the GPU with MIG also lets the customer-service model run in its own isolated slice, so an internal team's experiment cannot slow customer replies; see Can Seven AI Teams Share One DGX Station?. For agents running around the clock on smaller hardware, compare Can a GB10 Run AI Agents 24/7 Without Anyone Watching?

Buying in Singapore

SourceIT supplies five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. For a 24/7 service, pair it with an online UPS such as the APC Smart-UPS On-Line 3kVA Rackmount at S$4,480.00 or the APC Smart-UPS SRT 5000VA tower at S$6,741.25 (both listed 4 Nov 2026), sized against your full load. Browse the NVIDIA DGX Spark and DGX Station collection. For a quotation, 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

How many customer chats can a DGX Station handle at once? It depends on the model. StorageReview measured 22,161 tokens per second across 128 concurrent requests with GPT-OSS-20B on an MSI XpertStation WS300, about 173 tokens per second each. Larger models handle fewer simultaneous conversations.

Is there a DGX Station benchmark for a 70B model? We found none published. NVIDIA's TensorRT-LLM figures for Llama 3.3 70B are for B200 and GB200 GPUs, so treat them only as a rough proxy, not a DGX Station result.

What happens if the DGX Station fails? Any AI service running only on it stops until it is repaired. Plan a fallback such as routing to human agents, a second Station or an approved cloud model.

Should an AI agent be allowed to issue refunds? Not without human approval. OWASP's Excessive Agency guidance and Singapore's IMDA agentic AI framework both call for human checkpoints on high-impact actions.

Does running customer-service AI on premises help with the PDPA? It keeps conversations under your control and avoids overseas transfer questions, but your organisation remains responsible for protecting, retaining and disposing of that personal data properly.