- The GPU squeeze that is starving AI teams now also hits engineering simulation.
- Siemens can’t fix the shortage, so it makes where compute runs the customer’s choice.
The GPU squeeze that has defined the AI build-out for two years is now pressing on a quieter corner of enterprise computing: the simulation workloads that engineers run to design cars, aircraft and industrial hardware. And the executive who runs that business at Siemens is candid that availability, not price, is the constraint his customers keep hitting.
“Availability is one of the critical factors for our customers,” said Sam Mahalingam, who leads simulation and test, high-performance computing and data analytics at Siemens Digital Industries Software, speaking at the sidelines of RealizeLIVE Asia-Pacific in Bengaluru. The problem is structural, and familiar to anyone watching the AI infrastructure market. “When a new GPU comes out from Nvidia, the hyperscalers are immediately capturing the complete supply,” he said. “The customer organisations are finding it difficult to get.”
Why the GPU squeeze lands on engineering teams last
The same shortage throttling AI teams puts industrial engineering departments near the back of the queue. Each new Nvidia generation is spoken for by the largest cloud providers before general buyers can reach it, with backlogs running into millions of units through 2026. The practical result is that enterprises take what is left: older, previous-generation chips that the market has moved past. “The older GPUs are available,” Mahalingam said, naming Nvidia’s earlier H100 as the kind of hardware customers fall back to when the latest parts are out of reach.
He is clear about where he thinks responsibility sits. “This is an Nvidia problem,” he said. “They need to increase their supply, and they need to make sure the quotas are provided appropriately.” It is a pointed thing for a partner to say on the record, and a reminder that even a company of Siemens’ scale is downstream of the same bottleneck as everyone else.
The workaround his customers use is the one the broader market has settled on. Enterprises with existing cloud commitments can borrow the hyperscalers’ priority position. “If our customers have enterprise agreements with any of the cloud providers, then they do get some quota where they can run our software on those GPUs,” Mahalingam said. Access, in other words, increasingly runs through the cloud contract rather than the hardware order.
On whether there is a genuine alternative to Nvidia, Mahalingam was deliberately guarded. AMD, he said, is “a tough next close” and “a little bit more cost-effective,” but he declined to be drawn on whether it matches Nvidia’s performance, pointing instead to customers as the people who should answer that. It was a careful non-answer, and a telling one about how settled Nvidia’s position remains even among those who would benefit from a rival.
The deployment answer: let the customer decide
If Siemens cannot fix the supply of GPUs, its strategy is to make where the compute runs a matter of customer choice rather than a constraint the software imposes. Asked where his customers’ simulation workloads actually run, Mahalingam did not point to a single answer. “We leave it to our customers in terms of where they want to have their compute resources,” he said. “We support all the implementation models.”
Those models span the full range. A customer can simulate its own data centre on its own HPC cluster. It can take what Siemens now calls Simcenter Unlimited, a hyper-converged appliance that arrives as sealed hardware and software, which Mahalingam likened to a domestic appliance. “We provide the hardware and the software, but the hardware is locked,” he said. “We push it to their data centre, we connect it to the network, we power it up. It is actually a software solution, but we are like an appliance, like a dishwasher.”
Or the workload can run in the cloud; Siemens supports Azure, AWS, Google Cloud and Oracle Cloud Infrastructure, either bursting from on-premise or running entirely off-site. The point of the spread is to make the location of compute invisible to the engineer using it. In Simcenter X, the company’s cloud-based, single-interface environment, “you don’t even need to know where your job runs,” Mahalingam said. Provided a company’s IT has granted access to on-premise and cloud resources, the system routes each job to whatever capacity is free and returns the results without the user tracking where the work happened.
That abstraction is also how Siemens hedges the shortage. A workflow that can move fluidly between an owned cluster, a locked appliance and four clouds can chase available capacity wherever it sits, a meaningful advantage when the newest silicon is spoken for, and the fallback is a scramble for whatever quota a cloud agreement can unlock.
The constraint that outlasts the shortage
The GPU squeeze will ease as Nvidia’s supply catches up with demand, and the fallback to previous-generation hardware is a manageable inconvenience rather than a wall. The more durable point in Mahalingam’s account is the shift in how enterprises secure compute at all: less through buying hardware, more through the cloud relationships and flexible deployment that decide who gets access when the newest chips are scarce.
Siemens is not solving the shortage. It is building for a world in which the shortage, in one form or another, is permanent.
Want to experience the full spectrum of enterprise technology innovation? Join TechEx in Amsterdam, California, and London. Covering AI, Big Data, Cyber Security, IoT, Digital Transformation, Intelligent Automation, Edge Computing, and Data Centres, TechEx brings together global leaders to share real-world use cases and in-depth insights. Click here for more information.
TechHQ is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
