Why AI is bringing the workstation to the datacentre
This is a guest post for the Computer Weekly Developer Network written by Benjamin Boxer in his capacity as CEO of Phaze.
Phaze aims to build the best enterprise remote desktop experience to access and remotely collaborate on workstations for 3D and CAD software.
Boxer writes in full as follows…
For most of the past two decades, computing followed a simple trajectory. Hardware became smaller, faster and cheaper, so organisations put more power on every desk. Everybody came to expect a better laptop or a more capable workstation every other year. But GPUs and hardware shortages are now pushing that trend in the opposite direction.
Companies still need more compute than ever, but they’re becoming less interested in colocating it with every employee. The reason has less to do with remote work or virtualisation than with the economics and power consumption of GPUs. Engineers, developers, designers, architects and VFX (visual effects) professionals rely on GPUs pushing graphics, simulations and models to their limits to create the products and entertainment loved by billions of people.
Their demands have shifted GPU compute from a specialised resource into something almost every technical team depends on. Once that happened, the workstation stopped looking like a personal device and started looking like infrastructure.
AI changed the workload
GPU workstations have never been mainstream machines. They existed for people whose work justified the expense: engineers running CAD software, artists rendering 3D scenes, researchers training models and scientists running simulations. Most software developers could write code all day on a laptop without thinking about GPU utilisation.
This trend is changing.
The software that these specialists rely on is adding AI features that require significant GPU resources. Today, most of the tokens generated in that software come from cloud machines. In the future, the inference may run locally or on-prem to save costs.
Developers spend far less time typing into a code editor.
AI coding assistants generate code continuously. Researchers run local language models to prototype features, evaluate prompts and test agent workflows. Teams increasingly use autonomous coding agents that spend hours building, testing, debugging and validating software with little human intervention.
The same pattern shows up outside software development. Adobe continues adding AI features across Creative Cloud. Autodesk has embedded generative capabilities throughout its design products. Video editing, image generation, transcription, search and content creation all lean on accelerated compute.
None of these applications would justify a dedicated GPU workstation on their own. When you combine them, however, workstations start to require GPUs, RAM and CPUs that push the boundaries on traditional power demand from a device. GPU acceleration is no longer reserved for a handful of specialised applications. It has become part of everyday software.
This matters because organisations are starting to buy GPU infrastructure for everyone at the same moment that GPUs, RAM and CPUs are increasing in price exponentially.
Open-weight models change economics
The next change is coming.
Enterprises are starting to think about how tokens are generated and when to use cheaper inference versus the state-of-the-art models. Calling a hosted model through an API remains the easiest way to add AI to an application, but every request carries a cost. Those costs stay manageable when AI remains an experiment. They become much harder to ignore when every internal tool, customer-facing product and engineering workflow starts generating inference requests all day long.
Open-weight models offer a different path. Companies can run models on their own infrastructure, fine-tune them for specific domains, keep proprietary data inside their own environments and avoid paying for every token sent to a third-party service.
That decision shifts spending instead of eliminating it. Instead of renting someone else’s GPUs, organisations buy their own.
The consequence is easy to miss. Companies that once thought about AI as software suddenly have to think about racks, cooling, power, networking, utilisation and lifecycle management. The hardware to run AI becomes part of infrastructure planning.
Individual workstations don’t make sense
A high-end workstation is one of the most expensive endpoint devices an organisation can purchase. It consumes significant power, generates heat and becomes obsolete long before most office hardware. People who spend their days designing in 3D, running simulations and working with CAD software push these devices to their maximum output. They rely on the hardware to deliver the highest performance. As the hardware improves, the software always uses more of it. Software publishers push GPUs to their max because end users demand it.
These demands are expanding to more use cases because more software products rely on GPU acceleration. It isn’t feasible to give every employee a workstation, especially if it may sit idle at times. Many companies have already started responding by moving workstations into racked environments.
Conversations with infrastructure teams increasingly include deployments where roughly one-third of workstation fleets now live in datacentres instead of beneath individual desks. That percentage will continue climbing as GPU demand spreads across more engineering teams.
The user still gets a workstation. The workstation simply isn’t in the same room anymore.
Remote desktop: The critical path
Centralising workstations creates a new requirement: access has to feel local.
Boxer: Once workstations become shared infrastructure, remote access stops being an optional convenience.
For years, remote desktop software filled a narrow role. It was important to highly regulated industries, like healthcare and finance, that need to ensure data security and access control. Outside of those industries, however, remote desktop was a tool for IT teams. They used it to troubleshoot machines. Contractors used it to reach office computers from home. It solved an access problem rather than an infrastructure problem.
Professionals increasingly work from the office, home, customer sites and while traveling. They don’t want to rebuild their development environment every time they change locations.
Once workstations become shared infrastructure, remote access stops being an optional convenience. It becomes the layer that connects people and AI agents to the compute they need.
Professionals move between the office, home, customer sites and travel without wanting to rebuild development environments every time they change locations. A workstation becomes part of their workflow, not a physical object tied to a desk. Computer use agents are going to be an essential tool to every developer and IT team. Computer use agents can QA software for engineers and troubleshoot broken devices for IT teams.
Deploying computer use agents to every client device, however, is a real challenge and security risk. To reduce complexity, monitor activity and secure computer use agents, the go-to deployment will be through a remote desktop.
Once workstations become shared infrastructure, remote access stops being an optional convenience. It becomes the layer that connects people and AI agents to the compute they need.
Centralisation solves more than utilisation
Keeping compute inside the datacentre keeps source code, proprietary models, credentials, customer data and development environments there as well. Sensitive assets stop spreading across hundreds or thousands of endpoint devices.
Operations become easier too. Infrastructure teams patch systems once instead of touching every workstation individually. Hardware upgrades happen in racks instead of under desks. Failed machines no longer interrupt a developer’s environment because another workstation can take its place.
Power management, monitoring, provisioning and lifecycle management all become centralised.
While regulated industries have valued these for years, the need hasn’t spread. While the benefits are there, the cost was an employee’s experience working on a computer with lag from remote desktop software. That has changed with a new breed of remote desktop software that prioritises latency and performance, so a user feels like the computer is sitting right in front of them.
