Imagination drops CPUs and NPUs to consolidate its GPU business
A company redefining itself
Imagination Technologies has undergone a radical transformation over the past year. With Markus Mosen as the new CEO (the seventh in a decade), the company has made a clear strategic decision: abandon its CPU and neural accelerator lines to consolidate as a GPU IP vendor.
This is not an improvised change of direction. According to Jake Kochnowicz, Imagination's chief revenue officer, the decision follows a cold analysis of the market and of the company's real capabilities.
Why it is exiting CPUs and NPUs
Imagination launched a RISC-V CPU in 2021 that it discontinued in early 2025. The reason is economic: "The timeline we would have had to keep investing before becoming profitable was very long," Kochnowicz told EE Times. The RISC-V market showed signs of consolidation and some companies were already struggling. Maintaining two simultaneous bets, GPU and CPU, when only the GPU generated revenue, was inefficient.
The neural accelerator Imagination had developed was frozen in 2024. The company retains the intellectual property, but that design was created before the transformer era. In a market dominated by ChatGPT and similar architectures, the value of that IP has eroded. Kochnowicz was blunt: "It's a transformer world, and that changes the game."
GPUs that do AI and graphics
The current strategy is different. Imagination positions itself as a graphics company that integrates AI as a first-class citizen, not the other way around. The idea is to avoid software fragmentation and programming models by keeping graphics and AI on the same architecture.
The reasoning makes practical sense: some customers want to do both. A cloud gaming server can process graphics during the day and run AI workloads at night. A unified architecture allows that flexibility without duplicating design or investment.
The bet on the future: F series
Imagination has invested a 100-million-dollar loan in developing a next-generation GPU architecture. The F series will be a "significant refresh" able to scale from automotive to data center, and from graphics workloads to AI.
The official launch is planned for late 2026, but two major customers are already committed. The F series will enable much larger GPUs than the current E series, which tops out around 300 mm² on TSMC's N5 process. The F series will not reach 1 PFLOPS, but the following generation could.
Who buys GPU as IP
Most F-series customers will not sell chips directly to the mass market. Instead, they are verticalized: they sell PCIe cards for existing server infrastructure, prebuilt servers, or even their own data centers based on custom GPUs.
Cloud gaming is a classic example. Customers want to add video codecs, upscaling, and compression to their GPU silicon, customizing the product. Imagination's IP enables exactly that: a GPU with all those capabilities integrated into a single product.
The China factor
The company has faced controversy over technology transfer to China after its 2017 acquisition by Canyon Bridge Capital Partners, backed by Chinese capital. Kochnowicz clarified the current picture: Imagination has integrated teams in China (customer support and sales) that report to headquarters in the United Kingdom. It does not develop IP in China or the US, precisely to avoid trade restrictions. A major Chinese electric-vehicle maker uses its GPUs in ADAS systems.
The current regulatory environment is more manageable, although Chinese customers remain blocked from advanced-node technology due to US export restrictions. Imagination also does not authorize military applications for small GPUs in drones.
Back to what works
With seven CEOs in a decade, instability has been real. But senior technical leaders remain in place, many with decades at the company. The new leadership has introduced operational changes ("flattening" the hierarchy for agility) without diverting the fundamental technical course.
What Imagination is doing is, in essence, recognizing its limits and strengths. GPUs are its core business. Trying to compete in CPUs and specialized neural accelerators was a costly detour. Returning to focus is not a defeat: it is maturity.
Source: EE Times.
