As users grow, how do you scale?
Out with more units, up with denser GPUs
Sovereign GaiXer starts with one unit and grows as more people use it. There are two directions. Adding units (“out”) is a feature of the current product. Replacing one unit with a larger GPU server (“up”) is at the concept stage. This article explains how the two differ and how to decide when, and in which direction, to grow.
Key points
Scale-out in Sovereign GaiXer means placing a second or third unit on the same network, where the units find each other and form a cluster (a group that works as one), adding capacity as users grow. This is the “out” direction, and it is a feature of the current product. The first unit is a Lenovo ThinkStation PGX with NVIDIA GB10 (128GB unified memory), and the guideline is 10–20 users per unit. The other direction, “up,” means replacing one unit with a large GPU server that carries more GPUs. It is at the concept stage, and support will be decided after testing each hardware model. The goal is that even as the hardware grows, you keep using the same AI-OS, App Store and business apps.
- What you will learn, 1: there are two ways to grow, “out” (more units) and “up” (denser GPUs). How they differ, and how much is possible today.
- What you will learn, 2: what happens when you add a second unit, and how to go from one unit to three or four as users grow.
- What you will learn, 3: when to grow and in which direction, judged by usage data and AI value conversion figures.
How many ways can you grow? Out with more units, up with denser GPUs
As more people in your organization use generative AI, one unit will eventually not be enough. At that point, there are two ways to grow. “Out” means adding more of the same machine. “Up” means replacing one machine with a larger one that carries more GPUs. In technical terms, out is scale-out: adding capacity by adding units, and up is scale-up: adding capacity by making one unit bigger. In Sovereign GaiXer, scale-out is a feature of the current product, and scale-up is at the concept stage.
Solid boxes show what the current product can do (out, or scale-out). Dashed boxes show what is at the concept stage (up, or scale-up, and the combined goal of up and out). The idea is to start with “one unit” at the bottom left, grow to the right first, and in the future also grow upward.
GPU: the part that does generative AI’s computing
Generative AI writes text by running a huge number of calculations at once. The GPU does this work. The Sovereign GaiXer unit also contains an NVIDIA GPU.
GPU density: how many GPUs one machine carries
In this article, “GPU density” means the number and performance of GPUs in one machine. Raising density means growing up, which changes how big a job one machine can handle.
Cluster: several units working as one
A cluster is a group of machines run together as one system. Adding units means growing out, and you add capacity by forming this kind of cluster.
Think of the checkout counters in a supermarket. When more customers come, adding more checkout counters is growing out. Replacing one counter with a larger, faster one is growing up. Adding counters can be done right away, but installing a large one means rethinking floor space and wiring. Sovereign GaiXer is the same: out is available now, and up is a concept for the future.
How many units do you start with? One for now, with a guideline of 10–20 users per unit
You start with one unit. The current product’s hardware is the Lenovo ThinkStation PGX. It carries NVIDIA’s GB10 chip and 128GB of unified memory: a design in which the computing parts share one memory, like sharing one large desk. This lets generative AI run entirely inside the unit. The guideline is 10–20 users per unit.
NVIDIA GB10 (128GB unified memory)
This is the computing power that runs the LLM (large language model) at the heart of generative AI inside the unit. Answers are produced by one unit inside your organization, without sending data to the cloud.
Guideline: 10–20 users per unit
Of these, a few people generating answers at the same time is assumed. There is no limit on the number of registered users. Responsiveness depends on how the unit is used, so the number of units for your scale is decided through discussion.
About 150 × 150 × 50.5 mm, about 1.2 kg, one power outlet
It runs on the included 240W AC adapter and needs no server rack or dedicated air conditioning. A second unit is the same size.
JPY 3,806,400 before tax, paid once
Includes the hardware, the software license and a three-year hardware warranty. No additional charges as users or generations increase (excluding electricity and other operating expenses).
The first unit is like the first copier placed in a new department. While few people use it, one is enough. When queues start forming in the morning or before deadlines, it is time for a second one. There is no need to buy a large machine from the start.
What happens when you add a second unit? Place it on the same network and a cluster forms automatically
Growing out is a feature of the current product. Just place a second unit on the same network, and the two units find each other and form a cluster (a group that works as one). This is the AI-OS’s “automatic clustering from the second unit” introduced in Part 1.
Start with one unit
Place one unit and start using chat, internal document search, transcription and more. The guideline is 10–20 users per unit.
A cluster forms automatically
Place a second unit on the same network, and the units find each other and form a cluster. You do not need to start with work such as drawing a dedicated wiring diagram.
Add units as users grow
As more departments use it, you can add a third unit, then a fourth. You can check each unit’s status on the monitoring dashboard.
It is similar to pairing wireless earbuds with a smartphone. Put them near each other, and they find each other and connect. Likewise, when you connect a second Sovereign GaiXer unit to the same network, the two units set themselves up. There is no need to replace the first unit; it keeps working as before.
| Stage | What you do | What happens |
|---|---|---|
| Now: one unit | Connect one unit to the company LAN | Start using it, with a guideline of 10–20 users |
| Second unit | Place it on the same network | The units find each other and form a cluster automatically |
| Third, fourth unit… | Add units as users grow | Check every unit’s status in one place on the monitoring dashboard |
What does growing up mean? A concept for running on a larger GPU server
Growing up is at the concept stage. The idea is to replace the unit with a large GPU server: a business machine that carries multiple GPUs, and run Sovereign GaiXer on it. Whether a server is supported will be decided after testing each hardware model. It is not a feature available today.
On a GPU server
One example is a GPU server with up to eight RTX PRO 6000 Blackwell GPUs, each with 96GB of memory. One machine could handle bigger jobs than the current unit.
On a large GPU server
One example is a large platform such as NVIDIA HGX B300 (eight GPUs). This stage assumes running large models or company-wide use.
Some work is too big for more units alone
For example, when you want to use a large model that does not fit on the current unit (PGX), or when many more people across the company use it at the same time. For such needs, one larger machine may suit better than more units (concept).
When a family grows, building an extension is growing out, and moving to a bigger house is growing up. Moving gives you more space at once, but if you do not inspect the new house carefully, your furniture may not fit. That inspection is why “testing each hardware model” is essential for growing up.
What is the goal? A company-wide AI platform, up and out, with the same AI-OS on any hardware
Combining out and up leads to a cluster of large GPU servers supporting a company-wide AI platform. This is the goal of Sovereign GaiXer (concept). There are three key points.
- Out: more users mean more units, and more GPUs per company. You can add computing power step by step to match the people using it.
- Up: handle large models that do not fit on PGX and company-wide concurrent use (concept). One larger machine takes on big jobs that more units cannot reach.
- Even when the hardware changes, the AI-OS, App Store and business apps stay the same. The aim is to grow without changing how employees use it or how IT manages it.
SmartphoneThe device you use every day
Sovereign GaiXerA “generative AI smartphone” for your company
When you switch to a new smartphone, your apps, store and controls work just as they did yesterday. Sovereign GaiXer aims for the same: even as the hardware grows, you keep using the three layers above as they are.
| Stage | Direction | Hardware (example) | Status |
|---|---|---|---|
| Now: one unit | Starting point | Lenovo ThinkStation PGX (NVIDIA GB10, 128GB unified memory) | Available |
| From the second unit | Out | Multiple PGX units clustered | Product feature |
| GPU server | Up | Up to eight RTX PRO 6000 Blackwell GPUs (96GB) | Concept (decided after testing) |
| Large GPU server | Up | NVIDIA HGX B300 (eight GPUs) | Concept (decided after testing) |
| Goal | Up and out | A cluster of large GPU servers | Concept |
When, and in which direction, should you grow? Decide with usage data and AI value conversion figures
Decide on adding units with numbers, not hunches. The AI-OS includes, from the start, the data you need to decide whether to add units. Who is using it and how much, and what that usage would cost on cloud AI. These two figures are the basis for adding a second unit.
Usage visibility
See usage by department and by person. Knowing which departments have growing numbers of users shows when to add a second unit and which department to expand to next.
AI value conversion
AI usage processed in the unit is shown as what the same work would have cost on cloud AI, in Japanese yen. “How much work this unit has done, in money” becomes material for the internal approval to add units.
Monitoring dashboard
From the second unit on, you can check the status of each clustered unit on the monitoring dashboard. Seeing which unit is busy helps you decide whether to add a third.
| Your current situation | What to do | What to base it on |
|---|---|---|
| One unit is coping and usage is stable | Keep using one unit | Watch the trend with usage visibility |
| People are waiting longer for answers | Out: place a second unit on the same network | Peak hours and the number of departments using it |
| You plan to have more departments use it | Out: add a third and fourth unit to match headcount | The 10–20 users per unit guideline and the AI value conversion amount |
| You want several hundred users | Out: discuss multi-unit or two-unit linked configurations | Headcount and usage (proposed in a 15-minute online consultation) |
| You want large models that do not fit on PGX | Up: at the concept stage, so please share it with us as a request | The models you want to use and how many people will use them at once |
Railway companies look at congestion rates to decide whether to run more trains or add cars to each train. “It feels crowded” will not get a budget for more trains approved. Usage visibility and AI value conversion are the numbers that show your internal AI’s “congestion rate” and “earnings.”
Common misconceptions and the actual arrangement
People often assume that AI kept on-premises is hard to expand later. In fact, you can start with one unit and add units to match the people using it.
| Misconception | Actual arrangement |
|---|---|
| When users grow, the only option is to replace it with a bigger machine | The current product grows out. Place a second unit on the same network and a cluster forms automatically; you keep using the first unit too |
| Adding a second unit needs complex setup by specialist engineers | Just place it on the same network, and the units find each other and form a cluster. You can check each unit’s status on the monitoring dashboard |
| More users mean extra charges per user | You pay the unit price for each unit you add (excluding electricity and other operating expenses and paid options such as GaiXer Care). There are no extra charges based on users or generations |
| It can already run on large GPU servers | Growing up is at the concept stage. Support will be decided after testing each hardware model |
| New hardware means relearning the apps and how to use them | In the goal, you keep using the same AI-OS, App Store and business apps even when the hardware changes (growing up is at the concept stage) |
| The timing for adding units can only be decided by gut feeling | Usage visibility and AI value conversion let you decide based on numbers |
Mini glossary: terms used in this article
Scale-out in Sovereign GaiXer
Scale-out in Sovereign GaiXer means placing a second or third unit on the same network, where the units find each other and form a cluster (a group that works as one), adding capacity as users grow.
Scale-up
Adding capacity by replacing one unit with a larger machine that carries more GPUs. In Sovereign GaiXer it is at the concept stage, and support will be decided after testing each hardware model.
Cluster
A configuration in which several machines are grouped and run as one system.
GPU
The part that runs the huge number of calculations generative AI needs, all at once. The Sovereign GaiXer unit contains an NVIDIA GPU.
GPU density
The number and performance of GPUs in one machine. This article uses it to describe growing up.
NVIDIA GB10 and unified memory
GB10 is the NVIDIA chip in the current hardware, the Lenovo ThinkStation PGX. Unified memory is a design in which the computing parts share the same memory. The current hardware carries NVIDIA GB10 and 128GB of unified memory.
GPU server
A business machine that carries multiple GPUs. One example configuration has up to eight RTX PRO 6000 Blackwell GPUs with 96GB each.
Large GPU server (such as NVIDIA HGX B300)
A large platform that carries eight GPUs together. Suited to large models and company-wide use.
Monitoring dashboard
A screen where you can check the status of every clustered unit in one place.
Usage visibility
An AI-OS feature that shows AI usage by department and by person. It helps decide when to add units and which departments to expand to.
AI value conversion
A feature that shows AI usage processed in the unit as what the same work would have cost on cloud AI, in Japanese yen.
Adding units and growing: frequently asked questions
What is scale-out in Sovereign GaiXer?
Scale-out in Sovereign GaiXer means placing a second or third unit on the same network, where the units find each other and form a cluster (a group that works as one), adding capacity as users grow. It is the “out” direction of adding units, and it is a feature of the current product. You can check each unit’s status on the monitoring dashboard.
How many people can use one unit?
The guideline is 10–20 users per unit. Of these, a few people generating answers at the same time is assumed. There is no limit on the number of registered users. Responsiveness depends on usage, for example when many people ask for summaries of long documents at the same time, so we will propose a number of units for your scale in a 15-minute online consultation.
Does adding a second unit need complicated setup?
No. Just place a second unit on the same network, and the two units find each other and form a cluster. There is no need to replace the first unit; you keep using it. Part 4 of the product guide explains how to connect to the company LAN.
How many units can you add?
You can add a third, fourth unit and more as users grow. The scope of multi-unit clustering varies by configuration. For large organizations, such as those with several hundred people, there is also a configuration that uses an expansion port (ConnectX-7) to link two units. We will propose a configuration for your scale in a 15-minute online consultation.
How does pricing change when you add units?
You pay the unit price for each unit you add. One unit is a one-time purchase of JPY 3,806,400 before tax, including the hardware, the software license and a three-year hardware warranty. There are no additional charges as users or generations increase (excluding electricity and other operating expenses). Overseas terms require confirmation.
When will growing “up” onto large GPU servers be available?
Growing up (running on GPU servers or large GPU servers) is at the concept stage. Supported models, availability dates and prices have not been decided. Support will be decided after testing each hardware model. The plan is to first build a track record with growing out (multi-unit clusters), and to test growing up jointly with GPU makers and other partners.
If the unit is replaced with a large GPU server, will the apps or the way we use them change?
In the goal, you keep using the same AI-OS, App Store and business apps even when the hardware changes. It is the same idea as switching to a new smartphone and using the same apps the same way. However, support for growing up is at the concept stage.
Should we grow out or up first?
With the current product, the option is growing out. When more people use it, place a second unit on the same network. Growing up becomes necessary when you want large models that do not fit on the current unit, or when many more people across the company use it at the same time. Because growing up is at the concept stage, please share it with us as a request.
How do we decide when to add units?
Use the AI-OS’s usage visibility and AI value conversion. Usage by department and by person shows where the number of users is growing, and AI value conversion shows the unit’s work as what the same work would have cost on cloud AI. When people start waiting longer for answers, that is a guideline for adding a second unit.
Can organizations with several hundred people use it?
For large organizations, such as those with several hundred people, we will propose, in a 15-minute online consultation, configurations that grow out with more units or that link two units through the ConnectX-7 port. We decide the number of units and configuration based on headcount and usage.
Summary: three points on the two ways to grow
- Today, you can start with one unit. On a PGX with NVIDIA GB10 (128GB unified memory), the guideline is 10–20 users per unit (with a few people generating answers at the same time).
- Growing out (more units) is a feature of the current product. Place a second unit on the same network and a cluster forms automatically; you can then add a third and fourth. Decide on adding units with usage data and AI value conversion figures.
- Growing up (denser GPUs) is at the concept stage. The aim is to run on larger GPU servers while keeping the same AI-OS, App Store and business apps when the hardware changes. Support will be decided after testing each hardware model.
A small shop adds one checkout counter at a time as customers increase. That is how Sovereign GaiXer grows today. Even with more counters, staff operate them the same way, and customers line up the same way. And if the day comes to move to a large store, the goal of growing both up and out is to welcome many more customers while keeping the same way of using the counters.
Related reading
Turn NVIDIA GPUs into business AI from day one
The software that turns GPU computing power into AI you can use at work. The four things Sovereign GaiXer handles.
Product guide, Part 9Put idle GPUs to work
Filling the gap after buying a GPU server with software, and the path to company-wide use.
Product guide, Part 4What happens when you connect to the company LAN?
Three ways to connect, and what to agree with IT before placing a second unit on the same network.
Information is as of October 2026. All listed prices describe the Japanese offer; overseas pricing and commercial conditions require confirmation. Growing up (running on GPU servers or large GPU servers) is at the concept stage, and support will be decided after testing each hardware model. Supported models, availability dates and prices have not been decided. The scope of multi-unit clustering varies by configuration. “10–20 users per unit” is a guideline, and responsiveness varies with usage. GPU specifications are figures published by NVIDIA. All prices are before tax. “No additional charges” refers to the unit price excluding electricity and other operating expenses; paid options such as GaiXer Care are separate. When you add units, you pay the unit price for each unit added. “AI-OS” is a plain-language name for the role of the software foundation that supports secure business use of AI. The smartphone, checkout counter, moving house and railway comparisons are analogies explaining the product structure, not claims of compatibility or equivalence with specific products. NVIDIA, GB10, RTX, Blackwell and HGX are trademarks or registered trademarks of NVIDIA Corporation in the United States and other countries. ThinkStation is a Lenovo trademark. A trademark application has been filed for “Sovereign GaiXer.”
Company, product, and service names mentioned are trademarks or registered trademarks of their respective owners.