Worldwide spending on artificial intelligence infrastructure reached $318 billion in 2025, more than double the prior year, and IDC projects another $487 billion in 2026. Yet many deployments of enterprise graphics processing units (GPU) remain poorly utilized. Some industry estimates put average utilization around 5%.
Compare the GPU in a top-tier gaming rig against the GPU performance of most engineering workstations companies have in active use today, and the gaming hardware wins on the specification that matters most for AI.
I’ve spent much of the last three years working with my team to build virtualized cloud infrastructure focused on the heavy data- and graphics-processing demands typical of gamers, which means I’ve spent a lot of time analyzing how compute gets bought, deployed, and used. Conversations with engineering leaders and the buildouts I was working on pointed back consistently to the same structural problem: Enterprise information technology procures hardware for the workloads that exist at the moment of purchase, then operates it for three to five years. When the workloads change faster than the cycle, the hardware becomes a bottleneck before it becomes a write-off.
WHAT THE GAMING MARKET DID TO GPU HARDWARE
For years, consumer gaming was one of the main consumer markets that pressured GPU manufacturers to develop more video random-access memory (VRAM) and faster memory bandwidth. Games required it. They needed higher resolutions, more complex textures, and more detailed real-time rendering. Those demands required more GPU memory year over year, and at manageable costs for individual consumers, to be competitive.
Enterprise buyers, meanwhile, were purchasing on a different axis. Professional workstation GPUs were often valued for certified drivers, application support, error-correcting code (ECC) memory, validated thermal designs, and workstation warranties rather than raw gaming performance per dollar.
Those priorities made sense for computer-aided design (CAD), visualization, modeling, and design workflows. They also meant that some lower- and mid-tier workstation cards carried less VRAM than flagship gaming cards at similar or lower street prices. In recent generations, for example, several professional cards have shipped with 8 GB to 16 GB of GDDR6, while high-end gaming cards such as the GeForce RTX 4090 and Radeon RX 7900 XTX reached 24 GB.
AI CHANGED WHAT ENGINEERING SOFTWARE DEMANDS
More than 90% of software engineers now report using AI-assisted tools, but the infrastructure burden is not limited to cloud-hosted coding assistants. As companies bring more AI inference, model testing, simulation, and generative design workflows closer to engineering teams, GPU memory has become a more common constraint inside ordinary technical work.
Enterprise software engineering is undergoing a dramatic shift that tracks the widespread adoption of AI. Hardware designed prior to the onset of the AI revolution is often at a disadvantage. A workstation purchased in 2021 for CAD, visualization, modeling, or software development may still be functional, but it was not bought for the AI-heavy workflows now arriving inside those same teams.
AI coding tool adoption went from negligible to 90% among professional developers in roughly three years, within a single hardware refresh cycle. As a result, the hardware procurement system is struggling to keep up.
THE UTILIZATION PROBLEM
This is where a teenager’s gaming rig becomes instructive.
The more revealing comparison is about utilization rather than specs. AI and simulation workloads are spiky: An engineer might need substantial GPU compute for 45 minutes of model inference, then nothing GPU-intensive for the rest of the day. Enterprise GPU utilization averages around 5%, which means that companies are paying for hardware that sits idle 95% of the time. A gaming PC gets pressed into service at near-capacity whenever it runs.
The gaming rig, in other words, wins twice. It has more of the memory that matters, and it uses the capacity it has.
WHAT TO ASK BEFORE THE NEXT REFRESH CYCLE
Before approving the next hardware procurement, three questions are worth raising with your information technology team:
- What GPU workloads are your engineers running today, and how often?
- How will those workloads change in the next 18 months as AI tools become standard across engineering workflows?
- Would on-demand cloud compute, matched to teams’ actual needs rather than physical machines headed for depreciation after three years, better serve this new reality of AI-era workloads?
The $487 billion expected spend on AI infrastructure this year runs the risk of following the same procurement logic that created the current mismatch.
The lesson is not that companies should buy gaming PCs for engineers. Rather, it’s that GPU strategy must stop treating demand as a fixed workstation spec. AI workloads are intermittent, memory-intensive, and spreading faster than refresh cycles can absorb. Companies that keep buying hardware for yesterday’s workflows may keep finding that the most AI-ready machine in the building belongs to someone’s teenager.
Haroldo Jacobovicz is founder of Arlequim Technologies, S/A.
