Imagination Technologies appointed its seventh CEO in a decade — Markus Mosen, former VP at NXP Semiconductors and Infineon Technologies Asia-Pacific — and is consolidating around GPU IP. The leadership churn presents a challenge for licensing partners betting a decade on the roadmap.
Imagination killed its RISC-V CPU at the start of 2025, citing long investment timelines and a consolidating market. Its NPU went on the shelf in 2024; the architecture predates the transformer era and doesn't fit current workload profiles. The GPU business — the only profitable segment — remains, alongside a $100 million loan from two years ago funding the F-series architecture.
| Product Line | Status | Action Taken | Stated Reason |
|---|---|---|---|
| RISC-V CPU | Discontinued | Killed early 2025 | Long investment timelines; consolidating market |
| NPU | Shelved | Shelved 2024 | Pre-transformer architecture; doesn't fit current workload profiles |
| GPU (E-series / F-series) | Active / In development | Only segment retained | Only profitable segment; $100M F-series architecture rewrite funded |
F-series is a platform reset. The current E-series tops out at roughly 300 mm² on TSMC N5. F-series scales beyond that for automotive and data center, with launch slated for end of 2026. Two anchor customers are signed. The generation after F-series could hit 1 PFLOPS. That matters for a pure-play IP licensor, where scaling depends entirely on licensee hands. Imagination treats graphics and AI as co-equal workloads on the same architecture, avoiding programming model fragmentation.
| Attribute | E-series (current) | F-series (upcoming) |
|---|---|---|
| Max die size | ~300 mm² on TSMC N5 | Scales beyond 300 mm² |
| Target markets | Existing licensees | Automotive, data center |
| Launch | Current generation | End of 2026 |
| Peak compute (next gen roadmap) | — | ~1 PFLOPS |
| Architecture investment | — | $100 million loan |
| Workload model | Graphics-primary | Graphics + AI as co-equal workloads |
| Anchor customers signed | — | 2 |
Most F-series licensees will not build chips for external sale. The expected customer profile is verticalized: PCIe card vendors, server builders, and operators running private data centers on custom GPU silicon. Imagination is infrastructure IP for closed stacks, not a merchant GPU supplier. F-series silicon will appear in specific verticals — EVs, embedded data center appliances, cloud gaming hardware — not commodity inference cards.
China represents the largest operational risk. China accounted for roughly one-third of Imagination's business in 2019; the current figure varies with intermittent licensing deals. Mosen's experience leading WeEn Semiconductors (a European bipolar power startup with Chinese ownership) and his Asia-Pacific roles at NXP and Infineon show preparation for managing that dependency. Export controls create ongoing uncertainty for any U.S.-aligned licensee evaluating Imagination's roadmap.
The balance sheet is structurally cleaner than twelve months ago. A lawsuit from former CEO Ron Black, who sought approximately £200 million in damages, settled for a low single-digit million figure, removing a material acquisition liability.
For architects evaluating edge or automotive inference, Imagination is not a near-term alternative to vertically integrated NPU vendors for transformer workloads. F-series is the bet. It won't reach 1 PFLOPS in its first generation, and its customers are building closed stacks. The end-of-2026 F-series launch will signal whether the $100 million architecture rewrite delivered competitive performance-per-watt for edge inference.