GPU cloudHow Aravolta brought GPU-level monitoring to KONST's bare metal cloud
KONST is a Taipei-based AI infrastructure company operating high-density GPU data centers and GPU cloud services across Asia and Europe, and its fleet had outgrown the monitoring its team built in-house. Aravolta brought KONST's GPU servers, switches, firewall, and UPS onto one platform, onboarded remotely, and read GPU telemetry straight from the exporters the servers already ran.
We had built our own monitoring stack, and it served us well in the early days. But as our clusters grew, the gaps became visible: every new site or vendor meant another collector to write, and simple questions about how a rack behaved last week took real effort to answer. Aravolta gave us GPU, network and power telemetry in one place within days, without our team writing a single integration. What stood out just as much is the product itself: the UI is clean and genuinely pleasant to work in, our engineers found their way around it without training, and the depth of features behind that simple surface kept surprising us. For a lean infrastructure team running multi-site GPU clusters, that was exactly the kind of leverage we needed.
David Wacker
Systems Architect, KONST
The challenge
A home-built stack the fleet outgrew
A bare metal GPU cloud lives or dies on more than its GPUs. Servers, power, cooling, and network all have to be visible. In the early days, KONST's engineers built their own monitoring from open-source tools: SNMP data flowed into custom dashboards, and alerts went to the team's chat. It did the job and got the company through its first clusters.
But what works for a handful of racks doesn't scale to multiple sites and growing cluster sizes. Each new vendor or site meant more mapping and more upkeep, and more of the team's time went into maintaining the tooling rather than the infrastructure it was meant to watch.
What KONST needed
GPUs and the facility, in one place
KONST's team knew what the next platform had to do.
GPUs next to the facility
GPU usage and temperature had to sit next to the switches and power behind each rack, without the team maintaining the tooling.
Any time range
Questions about how a rack behaved last week had to be quick to answer, over whatever window the team picked.
New sites without new collectors
KONST runs sites in more than one country, so adding a site or a vendor couldn't mean writing another collector.
Data the team could build on
KONST wanted its data in its own data lake, so the platform needed an open API.
GPU monitoring
Down to the individual card
Aravolta read the GPU exporters KONST's servers already ran, so nothing was installed on the servers and no collector was written on KONST's side. Aravolta's own collector, running on its node, pulled the readings on a schedule.
A GPU server's asset page lists its cards one by one, with power draw, temperature, utilization, memory used, and fan speed for each, in the same platform as the switches, firewall, and UPS that keep the rack running.
Power and temperature
Draw and heat for each card, so a hot or power-hungry GPU stands out.
Utilization and memory
How hard each card is working and how much of its memory is in use.
Clocks and fans
Clock speeds show when a card is throttling, and fan speeds show how hard it is being cooled.
Onboarding
Remote, and fast
Aravolta's engineers onboarded the site remotely and responded quickly at every step. KONST didn't write a collector or map a single point.
- 1
A node on the management network
The Aravolta node joined the site's management network, and the rest of the work was done remotely.
- 2
Network and power, discovered
The node found the switches, firewall, and UPS over SNMP and matched each one against Aravolta's device library, so their metrics were generated automatically, down to the status and traffic of each switch port.
- 3
Links mapped
Aravolta read LLDP from the switches and worked out how they connect, with no diagram to draw by hand.
- 4
Alerts from the start
Each device got a not-reporting alert as it came in, so anything that went quiet would be flagged without anyone writing a rule.
What KONST got
The whole stack in one place
Every vendor, one data model
Switches, a firewall, and a UPS from different vendors, matched against Aravolta's device library, plus GPU telemetry, all in one format.
GPUs in context
Each card's power, temperature, and utilization sits next to the switches and UPS behind the rack, in the same platform and the same dashboards.
Dashboards anyone can build
Any metric over any time range, built in minutes and shared with the team.
Built to build on
A REST API to every asset and metric, so the team could pull data into its own tools and its own data lake.
For GPU clouds
Why card-level data matters
For a bare metal GPU cloud, the GPUs are the product, and each one depends on the rack around it.
Spot a struggling card
Temperature, power, and clock speed side by side show which card is holding back, with the rack's power and network in the same view.
Answer customers with data
When a customer asks why a job ran slow, the team can pull that card's history for the exact window in question.
Grow without new tooling
Each new site comes in through a node and the device library, so a bigger footprint doesn't mean more collectors to write and maintain.
Running GPU servers?
See every card's power, temperature, and utilization next to the infrastructure behind it.
Book a demoAbout KONST
KONST is a Taipei-headquartered AI infrastructure company that delivers compute, data centers, and enterprise AI adoption as one connected stack. Through its Konstra AI, Glows.ai, and Horizon AI brands, KONST builds and operates high-density GPU data centers, offers bare metal and cloud GPU compute, and runs the governance layer that lets enterprises consume AI as a service.
- Industry
- AI infrastructure: GPU data centers, GPU cloud, and enterprise AI
- Headquarters
- Taipei, Taiwan
- Previous monitoring
- An in-house stack built on open-source tools
- Scope
- GPU servers, switches, firewall, and UPS
Aravolta products
- GPU monitoring
- Network monitoring
- Power monitoring
- Dashboards