Private AI infrastructure is being sold by workload, not by GPU count. Lenovo AI Express shows the shift
Lenovo and NVIDIA now package hybrid AI capacity in small, medium and large tiers defined by users, tokens per second and model size, shipping in 15 to 25 business days. Agents change the sizing math.
Listen to this article · 6 min
AI-generated narration of the full article.
Buying private AI infrastructure used to look like an HPC project: architects choosing GPUs, interconnects and storage, months of design, and a cluster that was either too big for the first year or too small for the second. Lenovo’s newest offer reads more like buying a validated capacity tier.
What Lenovo announced
Lenovo AI Express, developed with NVIDIA as a quick-start path into Lenovo’s Hybrid AI Factory, comes in three predefined configurations. Lenovo describes each one by the workload it is meant to serve, not by its parts list first:
| Tier | Lenovo’s workload description | Hardware | Order to ship |
|---|---|---|---|
| Small | Focused inference for tens of users at 30+ tokens per second, models of roughly 7B to 70B parameters | ThinkSystem SR650a V4 with two NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs | From 15 business days |
| Medium | Higher-throughput inference and agentic AI for hundreds of users at 30+ tokens per second, models of roughly 70B to 400B | ThinkSystem SR675 V3 with eight RTX PRO 6000 Blackwell Server Edition GPUs | From 20 business days |
| Large | Full-scale generative AI for thousands of users, models up to a trillion parameters | ThinkSystem SR680a V4 with NVIDIA HGX B300 | From 25 business days |
The software options are NVIDIA AI Enterprise or Red Hat AI Factory with NVIDIA, with Veeam Kasten for data resilience. Services cover deployment, ROI validation and GPU optimization, plus a new Premier Support Plus tier for servers. The shipping times apply to eligible configurations after validation, and Lenovo says availability varies by region.
Lenovo also cites its own CIO survey on expected returns from AI. That is vendor research and should not be read as independent evidence.
Why the unit of purchase matters
Describing capacity as “users, tokens per second and model size” is the right move. It is the language of the workload, and it forces the buyer to answer the questions that actually determine whether a system is fast enough: how many people, how quickly they need responses, and how large a model the task needs.
There is a catch, and it is getting bigger: agents break the “users” number.
- Use case and its latency target
- Concurrent users at peak
- Model calls per task
- Tokens per call, in and out
- Model size the task needs
- Capacity tier, with growth
Business inputsWhere agents multiply the load
A person chatting with an assistant generates one model call per question. An agent completing a task can plan, call tools, retrieve, reflect and retry, which can mean dozens of calls per task, each carrying a long context. A tier sized for “hundreds of users” of chat may serve far fewer users of an agentic workflow. Background agents, which run without a person waiting, add load that no “user” count captures.
So the sizing exercise has to start one level lower:
- Calls per task for each agent, measured in a pilot, not guessed.
- Tokens per call, input and output, including retrieved context and tool results.
- Concurrency of agents, including scheduled and event-driven ones.
- Latency targets per step, because an overnight batch and a customer-facing reply have very different needs.
- Model mix. If most steps can use a smaller model and only some need a large one, as the routing-by-task argument suggests, the tier you need may be smaller than the largest model implies.
When a packaged tier makes sense
A validated tier with short lead times fits organizations that have already decided some inference must run on infrastructure they control, for the custody and sovereignty reasons we covered in our analysis of the sovereign AI stack, and want to avoid a bespoke design project. It sits one step above the desk-side option NVIDIA just announced with the 64GB DGX Spark: shared, data-center grade, operated by IT.
It does not remove the operating work. Someone still patches the stack, manages models, monitors utilization and plans the upgrade. A packaged tier makes buying faster; it does not make running cheaper.
What to do now
- Measure before you buy. Run the target agents in a pilot and record calls per task, tokens per call and latency per step.
- Translate agents into load, not into users, and include background agents.
- Size for the model mix, not for the largest model you might ever want.
- Plan the second year: utilization targets, the growth path between tiers and who owns capacity decisions.
- Check regional eligibility and lead times with Lenovo or its partners; the 15-to-25-day figures apply to eligible configurations.
The bottom line
Lenovo AI Express is a sign that private AI infrastructure is becoming a product bought by workload. That is progress. The buyers who get it right will be the ones who know their workload in agent terms, calls, tokens and concurrency, before they pick a tier.