HWNEWS
The UK firm that licenses GPU designs to chipmakers has killed its RISC V CPU program (an open source chip architecture) and shelved its NPU (a chip purpose built for AI inference) designed before the ChatGPT era, narrowing to graphics as a new CEO
Imagination killed its RISC-V CPU program at the start of 2025, four years after launch, and shelved its pre-transformer NPU (a chip purpose-built for AI inference) a year earlier. The UK-based firm, which licenses GPU designs to chipmakers rather than selling chips itself, has spent two years narrowing its roadmap to a single product line: graphics.
The NPU IP is still on the books but has limited scope in the ChatGPT era, the company now concedes. The CPU line, which Imagination launched in 2021 as a RISC-V design (an open-source CPU architecture), never produced a high-volume design win. The decision to cut both is the strategic narrowing that defines the new CEO's inheritance.
New CEO Markus Mosen joined in February 2026, replacing Simon Beresford-Wylie, who retired late last year. Mosen is described as the company's seventh chief executive in a decade. The strategic cut was set in motion under his predecessor. The continuity claim Imagination wants to make, that this is not another reboot, lands against seven CEOs in ten years and two abandoned product lines.
The bet is on the next-generation F-series GPU architecture, officially targeted for launch around the end of 2026. Imagination has two large customers already signed, chief revenue officer Jake Kochnowicz told EE Times in an on-the-record interview at the company's Kings Langley headquarters. The F-series is positioned as a "significant refresh" that will let Imagination scale beyond the E-series' current ceiling of "around 300 mm² in TSMC N5," a 5-nanometer chipmaking process used by foundry partner TSMC. The architecture will not reach 1 PFLOPS, a measure of one quadrillion floating-point operations per second used to compare AI hardware. The generation after it could, which Kochnowicz called "huge for an IP-driven design."
Imagination underwrote that roadmap roughly two years ago with a $100 million loan, dedicated specifically to developing the next-generation architecture. The loan sized the bet. The R&D cost of a fresh GPU microarchitecture, plus the software and verification work needed to license it to external chipmakers, explains the figure. If the F-series slips or the two anchor customers falter, the loan becomes the operating story of 2027.
The customer mix is unusual for an IP licensor. Most F-series customers, Imagination says, will not sell chips. They will sell PCIe cards, a standard slot format for plugging accelerators into servers; pre-assembled servers; or run their own data centers on custom GPUs. Cloud gaming is the canonical example. The verticalization means Imagination's revenue tracks the deployment plans of a handful of large buyers, not the broader chip market.
The E-series already serves two big customers: one Chinese AI PC and graphics play, and one global consumer electronics company. Among the named accounts is a "large public EV company" in China that uses Imagination's GPU in an ADAS system, the onboard computer vision stack that helps a car drive itself. China exposure is described as growing again after a difficult 2019 to 2022 stretch, when customer failures and regulatory restrictions clipped deals. China accounted for roughly a third of Imagination's business in 2019 by prior EE Times reporting.
Imagination was bought in 2017 by Canyon Bridge Capital Partners, a private equity firm with Chinese backing that drew regulatory scrutiny at the time. Mosen's prior role was as CEO of WeEn Semiconductors, a European bipolar power startup with Chinese ownership. The executive who now runs a UK-headquartered GPU IP licensor, owned by Chinese-backed private equity, with growing Chinese customers and a Chinese AI PC design win, comes from a Chinese-owned European power-semiconductor firm. None of this is a smoking gun. It is the kind of cross-border ownership pattern the reader should see plainly.
"Sticking to strategy" is the line Kochnowicz used to frame the cuts. "We are a graphics company and will remain that way," he told EE Times, with AI treated as a "first-class citizen alongside graphics" on the same architecture. The decision to abandon two product lines and concentrate on one is, on its face, discipline. Two abandoned product lines and seven CEOs in a decade is also a story. The two readings do not cancel out, and a reader deciding whether the F-series launch lands should hold both.
The watch items are concrete. The F-series launch is targeted for late 2026, and the architecture that follows it is where the 1 PFLOPS ceiling could be reached. The two signed customers are the test of whether the loan was well-sized. Mosen's first year is a study in whether a seventh CEO in a decade can be the one who finally lands the bet.