According to LightSource Capital, ICY Tech has completed a Pre-A+ funding round worth hundreds of millions of yuan. In addition to financial investors, this round introduced leading domestic GPU companies and leading robotics companies. The entry of these industrial shareholders implies that ICY Tech may enter the stages of product validation, joint testing, and commercialization earlier.
Founded in 2023, ICY Tech's core team comes from the Applied Magnetism Center at Peking University. The company primarily focuses on developing inference processors based on MRAM (Magnetoresistive Random Access Memory), targeting edge and end-side AI inference scenarios.
| What Problem Does It Aim to Solve?
A real bottleneck in large model inference is that after model deployment, weights remain largely static, yet every time a Token is generated, the system must re-read massive amounts of weights. Whether users perceive responses as faster or slower often depends not just on computing power strength, but more critically on data transfer efficiency between storage and computation. Data movement itself generates no computational value but consumes power and increases latency.
Traditional solutions involve trade-offs. DRAM offers large capacity but requires periodic refreshing, leading to additional power consumption; SRAM provides high speed without refresh needs but has large cell areas, creating significant pressure on chip area and costs when scaling up capacity. ICY Tech chose the MRAM route: it uses magnetic states to store data, which is non-volatile (data persists after power loss), requires no refresh, and potentially allows for smaller cell sizes than SRAM. By placing model weights closer to computing units to reduce repeated data movement, they propose "Weight-Resident Computing."
In inference scenarios, weights are typically read-only rather than written, which aligns well with MRAM's non-volatile nature and high-bandwidth read characteristics.
| Challenges Lie in Materials and Engineering
ICY Tech's distinguishing feature is that it does not start solely from chip architecture to find storage solutions but builds up from the material and device layers. The team has long experience in condensed matter physics, magnetism, and spintronics, focusing on how magnetic materials can achieve stable reading and writing, how devices can realize low power consumption, and how devices integrate with inference architectures.
This cross-layer capability from underlying materials to chip architecture is uncommon in chip startups. Most teams excel in one specific link while relying on partners for others. ICY Tech keeps the most challenging material and device links under its own control, resulting in a longer path but greater controllability over underlying parameters.
Currently, ICY Tech has completed tape-outs for multiple validation chips, with some back-end testing underway. Their validation chips have passed third-party inspections and end-to-end model inference verification, achieving 24-hour continuous stable operation.
| FPGA First, ASIC Follow-up
ICY Tech's product roadmap progresses through four layers: At the bottom is SpinPU-M01, an FPGA-based chip and accelerator card, currently used mainly for validation and available for early customer trials. Above this is SpinPU-M02, the true ASIC chip, which is still undergoing tape-out and engineering development—a key factor for future mass production scale. Based on this, the company plans heterogeneous servers combining SpinPU and CPUs, delivered as complete systems. At the top level is cloud computing services, allowing customers to access compute resources on a pay-per-use basis without purchasing hardware. This tiered approach is pragmatic: FPGA solutions allow customers to validate scenarios first, while ASIC tape-out and mass production take time; servers and cloud services enable the business model to run before large-scale chip production begins.
This tiered arrangement is quite practical: using FPGA cards first lets customers interact with the product and validate scenarios. While ASIC tape-out proceeds, servers and cloud services activate the business model, ensuring revenue streams during the gap period before chip mass production.
| What Industrial Shareholders Bring
Robotics enterprises care more about whether models can respond with low latency on integrated machines; scenarios like robotic arms struggle to accept latency caused by cloud round-trips. GPU enterprises focus more on how dedicated inference capabilities can be embedded into existing computing ecosystems to address energy efficiency shortcomings in inference scenarios.
The shared value of these two types of shareholders is bringing real-world constraints into chip design early: what target models are needed, what the power budget is, and what the deployment environment looks like. Chip startups fear working in isolation, only to discover after tape-out that products do not match customer scenarios—a single trial-and-error cycle can waste one to two years. Customer involvement at the design stage significantly reduces this risk.
For hardware startups, obtaining access to real scenarios is often harder than securing funding. In this round, ICY Tech made progress in both aspects.
| A Splash of Cold Water
ICY Tech is not the only one exploring MRAM inference chips. Whether this path ultimately succeeds depends on mass production yield, cost control, software ecosystem, and system adaptation. The barrier to entry for new storage media into mainstream mass production systems will not be low.
Another variable is timing. The premise for this route to work is that demand for end-side and edge inference truly scales up, and power constraints become tight enough to require new solutions. If cloud computing continues to become cheaper and easier to use, the window of opportunity may not be as wide as imagined.
ICY Tech itself stated that the next one to two years are critical for transitioning from technical validation to scaled implementation. For startups pursuing material-based routes, sample performance and customer feedback speak louder than funding amounts. This round of capital primarily buys time; ultimate success depends on engineering progress over the next one to two years.
Editor: Shuo Xin Yu-W²