Staff Reporter |
Editor | Wen Shuqi
Axera (0600.HK), the AI inference system-on-chip (SoC) company, has just dropped its first half-year report since going public — and the numbers are eye-catching. Revenue for the first half of 2026 hit ¥402 million, a staggering 181.8% jump year-over-year. Gross margin climbed from 20% to 29%, and every single business line posted triple-digit growth. Not too shabby for a company still finding its footing in the public markets.
But here’s the thing: the single biggest wildcard in the semiconductor world right now is the explosive surge in memory prices. According to TrendForce, contract prices for mainstream DRAM skyrocketed 58% to 63% quarter-over-quarter in Q2 2026, while NAND flash contract prices leapt 70% to 75%. With AI compute demand going through the roof and manufacturers funneling capacity toward HBM, the supply of general-purpose memory keeps shrinking.
For most downstream device makers, pricier memory usually translates to squeezed margins and headaches. But here’s the plot twist we uncovered: the memory price surge is actually reshaping demand in a positive way — it’s speeding up the adoption of high-end controller chips.
“Since memory now eats up a much bigger slice of the BOM (bill of materials), end products face pricing pressure, and customers are far more eager to upgrade their controller chips. So ironically, the memory price hike is accelerating the shift toward higher-end products,” the management team revealed. Axera’s terminal business, spearheaded by its Black Light technology, is enjoying both higher volumes and better prices as a result.
Meanwhile, the company’s inventory ballooned from around ¥300 million at the end of 2025 to over ¥700 million. Management explains this as a deliberate move: with advanced process node capacity tight and lead times stretching out, they locked in supply after securing order visibility from “alpha customers” — the early-adopter headliners who commit to orders upfront.
Back in 2019, Qiu Xiaoxin walked away from her role as CTO of Unisoc to found Axera. With a Tsinghua bachelor’s and master’s, a PhD from USC, and a resume that includes stints as VP at AT&T and Broadcom, she had a clear conviction: the next big AI inference battleground wouldn’t be in the cloud — it would be on the edge and in end devices.
In its very first year, Axera moved at a pace rarely seen in the industry. It took just nine months to tape out its first chip, the AX630A, which packs two proprietary core IP blocks: the Axera SmartEye AI-ISP (delivering full-color imaging in near-darkness — that’s the “Black Light” magic) and the Axera TongYuan NPU (a power-efficient inference processor). Notably, the NPU natively supported the Transformer architecture back in 2021 when it was designed — and all the major IP inside the SoC is developed in-house.
These two technologies form the company’s reusable foundation. Since then, it’s been roughly one new generation per year, and the terminal computing business remains Axera’s bread and butter. In H1 2026, terminal computing revenue reached ¥341 million, up 169.5%, representing 85% of total revenue, with the “Black Light” series seeing sales surge over 200%.
What’s driving this growth? A few things: the continued rollout of smart solutions in industrial automation and smart city applications, escalating demand for on-device AI compute, and yes, that memory price surge pushing the whole industry toward high-end products. The latest generation, the AX615 series, is already branching into emerging fields like embodied robots and industrial vision.
In the smart vehicle arena, Axera timed its entry perfectly — right as China’s AEB (Automatic Emergency Braking) mandate prepares to take effect next year, meaning every new car will need to come with assisted driving chips as standard. Its core product, the M57, targets L2-level assisted driving with a focus on high integration, low power consumption, and strong cost-effectiveness. It entered mass production at a leading automaker in March. According to the earnings report, the company shipped 420,000 automotive chips in the first half, secured 24 new design wins, and now works with 25 automaker brands, including seven international names.
Axera told us that large models are rapidly making their way into vehicles, and AI agents deploying in smart cockpits have become a major industry trend — the golden window for large-scale adoption has arrived. To capitalize, the company’s first smart cockpit large-model compute acceleration chip has successfully taped out, completing its automotive product matrix. The M97, targeting high-level assisted driving, came back from fabrication and lit up successfully in February 2026, with engineering samples delivered in the first half. Several top automakers have already kicked off evaluations and selection.
Speaking of which, top carmakers are increasingly pulling algorithm development in-house. Industry data shows that among the top 20 NOA-equipped brands, over half have self-developed solutions. When asked whether chips and algorithms are moving toward decoupling or integration, management’s take: “Both business models will coexist for the long haul.”
Axera’s read on the situation: in the industry’s early days, the integrated software-hardware approach offers clear advantages, helping customers get to market fast. But as the sector matures, decoupling gains momentum — customers want to build their own algorithms and system capabilities on top of a chip platform. Axera positions itself as a Tier 2 chip supplier, offering an open, flexible, and composable platform that lets customers port their algorithms and do differentiated development — without trying to own the algorithm definition.
And then there’s edge AI — the sleeping giant Axera is betting on for the future.
As large models migrate from the cloud down to devices, local inference demand is exploding across AI PCs, intelligent agent boxes, and beyond. IDC projects that by 2027, inference will account for over 70% of global intelligent computing power, with edge infrastructure growing faster than core data centers.
Axera has already launched its edge chip, the AX8850, and in August 2026 established a subsidiary called “Axera Computing” focused on AI inference solutions. Its next-generation high-compute chip has completed tape-out, supports multi-chip cascading, and can run full-capability large models. The compute acceleration card built on this chip delivers over 1000 TOPS of inference power.
This aligns with the broader trend we’ve observed across the AI chip industry — at the WAIC computing exhibition, it was clear that everyone’s pivoting from just selling chips to selling full-stack solutions: systems, compute power, and services. The competitive battleground has expanded from single chips to accelerator cards, entire racks, and even clusters.
In the first half of this year, Axera’s edge AI business generated ¥27.69 million in revenue, up 251.9% year-over-year — the fastest-growing segment, though still under 7% of total revenue. Axera’s take: “The growth in AI inference demand is still in its early innings. The explosive potential ahead is massive.”
But potential comes at a cost — and that cost is R&D spending. In H1, R&D expenses hit ¥516 million, far exceeding the revenue figure, with over 80% of its 700-person team dedicated to R&D. Multiple advanced-node chips are being taped out this year, with mass production ramping up next year. As a result, the company posted a net loss of ¥688 million for the period, widening from ¥562 million a year earlier.
Management frames it as a natural timing mismatch between investment and returns in the chip industry. Their confidence is palpable: “We’re very optimistic. The multiple advanced-node chips we’re taping out this year will progressively scale up next year and the year after, laying a rock-solid foundation for revenue growth.”
