MANUFACTURING SERIES
The GPU Capacity Problem
What you will take away: A map of the GPU workloads a manufacturer actually runs, the analyst picture of why GPU compute has stopped behaving like a commodity, an honest account of when owning GPU hardware still wins and when it does not, and a capacity decision framework you can apply this quarter.
IDC describes AI as no longer a demand catalyst but the demand foundation, consuming silicon at a pace the industry has never seen, with high-bandwidth memory now the primary constraint in the accelerator supply chain and most capacity pre-committed through 2026. For a manufacturer buying GPU hardware for CAE, rendering or AI inference, that means higher prices, longer lead times, and a queue you do not sit at the front of.
The instinct is to buy ahead of the shortage. The better question is whether you should be buying peak-sized GPU hardware per seat and per project at all, when most of it idles most of the time, in a market where that capital is now scarce and expensive.
Why GPU compute is repricing now
The supply is constrained and pre-committed. IDC forecasts worldwide semiconductor revenue rising 52.8% to $1.29 trillion in 2026, with high-bandwidth memory the primary bottleneck and most capacity pre-committed through 2026 and allocated into 2027. The largest buyers have locked supply. Everyone else queues and pays.
The shortage reaches the desktop too. IDC forecasts PC average selling prices rising 18.3% in 2026, with no meaningful relief before the end of 2027. A like-for-like hardware refresh in this window buys the same architecture at a materially worse price.
The workload mix is shifting to continuous. Deloitte’s TMT Predictions 2026 highlight inference, the steady serving workload, overtaking training as the dominant driver of AI compute demand. Continuous workloads reward fixed, owned or reserved capacity running at high utilisation, not on-demand bursts priced for occasional use.
The manufacturing GPU workload map
Sort your GPU workloads before sizing any hardware. Two axes, burst versus continuous and data sensitivity, decide the cost curve and where the workload should run.
| Workload | Pattern | What it demands |
|---|---|---|
| CAE solving (Ansys, Abaqus, CFD) | Burst, schedulable | Right-sized compute and overnight scheduling. Reserved or shared capacity beats per-seat hardware; nobody perceives latency in a four-hour solve. |
| Batch rendering and visualisation | Burst, schedulable | Centralises trivially. Pay for the run, not for standing hardware that renders nothing most of the week. |
| Interactive GPU workstations (CAD viewports) | Continuous (working hours) | A steady GPU profile sized to real concurrency. A shared pool serves the team’s real day, not everyone’s worst hour. |
| Simulation surrogate or training on your data | Burst, compute-heavy | Reserved private capacity suits the burst. Data sensitivity constrains where it can run before cost is even discussed. |
| AI inference on proprietary design data | Continuous, steady | The strongest case for fixed private infrastructure: predictable cost on a steady workload, and the data never leaves your control. |
| AI inference on public or low-sensitivity data | Continuous or burst | Public platforms are fine on cost. Not every GPU workload needs to be brought in-house. |
The stranded-asset problem
A physical GPU workstation, or a deskside compute box, is sized for its owner’s worst hour and idles the rest of the year. The machine specified for the quarterly full-vehicle CFD run spends most of its life rendering a parts browser. Multiply that across the fleet and the waste is structural: peak-sized GPU hardware, average utilisation, every seat.
Shared GPU infrastructure is planned around the opposite assumption. Interactive workstations dominate working hours; the same GPU capacity runs CAE and rendering queues overnight. One pool, two shifts of value. Capacity is sized to realistic concurrency and grows in increments rather than full-fleet buys timed badly against a shortage.
Where the GPU sits becomes a sovereignty question
For CAE and rendering, placement is mostly an economics question: put the compute where it is cheapest to run at the utilisation you actually achieve. The moment a GPU runs inference on your design IP, the question changes. That workload is high-sensitivity and continuous, the exact profile that belongs on infrastructure you control, under a jurisdiction you choose.
A GPU pool that will eventually serve inference on your crown-jewel data should be planned for that from the start. Retrofitting the data posture after the workload is live is far harder than designing for it before.
The ownership fork
| Dimension | Owning per seat | Shared or reserved capacity |
|---|---|---|
| Capital timing | Full cost up front, repriced into the 2026-27 shortage. | Provisioned for measured concurrency; capacity added in increments. |
| Utilisation | Peak-sized, average-used; idle capacity is paid-for and stranded. | Sized to real concurrency; day-night cycle puts interactive and batch work on the same hardware. |
| Peak handling | The job that outgrows the box waits for the next refresh. | Heavy jobs draw heavy resources for their duration, then release them. |
| Supply exposure | Exposed directly to lead times and price rises on scarce hardware. | Provider absorbs procurement and refresh; you buy capacity, not silicon. |
| Data posture | Controlled if on-premise, but tied to one location and refresh cycle. | Depends on the provider; choose one whose jurisdiction matches your sensitive workloads. |
| Obsolescence | You carry the refresh and disposal cost per device, on a short cycle. | Refreshed centrally on a longer asset life, without touching the desk. |
Six questions before you buy
Run every GPU workload, and every hardware purchase the team proposes, through these questions. The answers size the capacity; the capacity sets the cost.
- What is the real utilisation across a month, not its peak hour? The average, measured, is the number that matters.
- Is the workload burst or continuous? And does the way you are buying it match the shape of the work?
- What are lead times and price trajectory on the hardware? And can the business afford to wait for it in a constrained market?
- If this runs inference on sensitive data, under whose jurisdiction does the compute sit? Sensitivity decides placement before cost does.
- What is the cost per job or per inference at production volume? And who owns that number?
- If demand grows, can you add capacity in increments? Or does it mean another full hardware buy timed against the next shortage?
Forty-five minutes with our technical team, working through your GPU workloads against utilisation, compute pattern, cost and data sensitivity. You leave with each workload placed on the right side of the buy-versus-share line, a clear view of what belongs on owned hardware and what belongs on shared capacity, and a plan you can act on this quarter.
Sources
- IDC, Semiconductor Applications Forecast, April 2026.
- IDC, Worldwide Personal Computing Device Market: 2026 Outlook, 2026.
- Deloitte, TMT Predictions 2026, December 2025.