MemSilicon

Cover Block

Open sources

Field	Value
Name	MemSilicon
Tagline	Developing resistive RAM (RRAM) and compute-storage collaboration chips for large-model inference. [Xinhua, October 2026]
Headquarters	Tianjin, China [Crunchbase]
Founded	2015 [Crunchbase]
Stage	Series A [Xinhua, October 2026]
Business Model	Hardware + Software
Industry	Deeptech
Technology	AI / Machine Learning
Geography	East Asia
Growth Profile	Venture Scale
Founding Team	Academic Spinout [The Paper / Sina Finance, October 2026]
Funding Label	Series A
Total Disclosed	More than RMB 400 million [Xinhua, October 2026]

Links

Open sources

• Website: https://memsilicon.com/

What an Investor Needs First

PUBLIC MemSilicon is a Tianjin-based semiconductor startup focused on resistive RAM, or RRAM, and compute-storage collaboration chips for large-model inference, and it merits attention now because it disclosed a Series A of more than RMB 400 million in October 2026 while simultaneously elevating former Graphcore Greater China head Lu Tao to CEO and co-founder [Xinhua, October 2026] [36Kr, October 2026] [Donews, October 2026]. The public record is still thin, but the shape of the bet is clear enough: a 2015-founded academic spinout is trying to pair non-volatile memory technology with inference acceleration rather than compete head-on as a conventional GPU vendor [Crunchbase] [The Paper / Sina Finance, October 2026] [Reuters, September 2010].

The company says the new capital will support RRAM industrialization and commercialization, alongside development of a first-generation LPU, described in public coverage as a large-model inference processor aimed at the high-end inference market on cost-performance grounds [Xinhua, October 2026] [36Kr, October 2026] [Futunn, October 2026]. That positioning is interesting because it suggests differentiation at the memory-and-architecture layer, although no public benchmarks, customer names, deployment volumes, or revenue figures were surfaced in the cited materials [Tencent News, October 2026] [Tokenheart, October 2026].

On team quality, the strongest verifiable datapoint is Lu Tao's prior role at Graphcore as president and general manager for Greater China, which gives MemSilicon a more legible commercial and category profile at the moment of scale-up [NIUX, October 2026] [Donews, October 2026]. Separate coverage also ties the broader technical roots of the company to a Chinese Academy of Sciences team led by Academician Liu Ming, though that claim appears in company-linked or syndicated reporting and should be treated more cautiously than the funding announcement itself [The Paper / Sina Finance, October 2026].

The funding syndicate is notable for breadth. Xinhua reported Boyuan Capital as lead investor, with participation from Bosch Venture Capital, Chuangling Capital, miHoYo, Zhongguancun Science City Co., Ltd., and GigaDevice, alongside follow-on backing from Furoong Investment, CDB Capital, Hengyu Capital, Jinqiu Fund, Shunxi Fund, Xiaomi Strategic Investment, and CAS Star [Xinhua, October 2026]. Public taxonomy points to a hardware-plus-software model, but the investable question for the next 12 to 18 months is less about category labeling than whether MemSilicon can convert a credible architecture story into visible proof points on silicon readiness, commercial engagements, and performance evidence [Crunchbase] [Futunn, October 2026].

Partially corroborated -- Core funding details are corroborated by Xinhua, 36Kr, and Donews, but product specifics and team-origin claims rely partly on company-linked or syndicated coverage.

Taxonomy Snapshot

Axis	Value
Stage	Series A
Business Model	Hardware + Software
Industry / Vertical	Deeptech
Technology Type	AI / Machine Learning
Geography	East Asia
Growth Profile	Venture Scale
Founding Team	Academic Spinout
Funding	More than RMB 400 million disclosed in Series A [Xinhua, October 2026]

Inside the Company

PUBLIC

MemSilicon enters the public record with more ambition than operating history. Crunchbase describes the company as focused on RRAM storage and AI computing technologies, while the company website identifies the legal entity as Beijing Mingxin Qirui Technology Co., Ltd.

The clearest milestone visible from those sources is the company website’s archived news post about an angel financing round, which indicates MemSilicon had begun raising external capital before its October 2026 Series A announcement, even if the exact timing and size of that earlier round are not established in the sources permitted for this section [MemSilicon] Crunchbase also maintains a company profile for MemSilicon, corroborating the existence of the business and its positioning around RRAM and AI compute, but the profile does not by itself resolve the company’s full chronology, legal structure, or location history [Crunchbase]. Taken together, the evidence supports a narrow overview: MemSilicon is a Chinese deeptech startup building around resistive memory and AI inference infrastructure, with a public milestone trail that remains relatively thin outside financing disclosures [Crunchbase] [MemSilicon].

Partially corroborated -- Based on Crunchbase and the company website, with partial chronology and limited corroboration on legal and historical details.

Under the Hood

MIXED

MemSilicon is being presented publicly as a memory-centric AI chip company rather than a conventional accelerator vendor. The consistent thread across the financing coverage is resistive RAM, or RRAM, paired with what Chinese-language reports describe as compute-storage collaboration for large-model inference [Xinhua, October 2026] [36Kr, October 2026] [Tencent News, October 2026]. That framing matters because RRAM is a non-volatile memory technology, meaning it can retain data without power, and has long been discussed as a lower-power alternative in parts of the memory stack, though broad commercial replacement of flash remains a category-level possibility rather than a MemSilicon-specific achievement [Reuters, September 2010].

The more specific product claim is that the new Series A proceeds will support RRAM industrialization and commercialization, alongside development of a first-generation LPU, described in coverage as a large-model processing unit for inference [Xinhua, October 2026] [36Kr, October 2026] [The Paper / Sina Finance, October 2026]. Public reporting also attributes to the company an ambition to target the higher-end inference market with strong token throughput and cost performance, but no benchmark data, model compatibility list, tape-out milestone, customer deployment, or verified demo has surfaced in the cited materials [Futunn, October 2026]. The company website, as captured in the provided research, offers little operating detail beyond the corporate identity, so the technical picture here rests mostly on financing-announcement coverage rather than on product documentation or independent testing [MemSilicon] [Xinhua, October 2026].

Claim stands unchecked -- Material product details rely largely on company-linked financing coverage, with limited independent technical corroboration and no verified public demo in the cited evidence.

Market Research

PUBLIC

The market matters now because inference demand is rising faster than conventional memory and power budgets, which puts non-volatile memory and compute-near-memory architectures back into the investable conversation, even when company-specific adoption data remains thin [Reuters, September 2010] [Xinhua, October 2026].

Public evidence does not support a clean TAM, SAM, or SOM for MemSilicon itself, and that gap is material. What the record does support is the company’s chosen wedge: RRAM commercialization and an inference-oriented processor program, both positioned around large-model inference rather than general-purpose training workloads [Xinhua, October 2026] [36Kr, October 2026]. In practice, that places MemSilicon at the intersection of two better-documented adjacent markets: non-volatile memory and AI inference silicon. The analytical discipline here is to treat those as analogous markets, not direct revenue pools for the company.

The demand driver is straightforward. Large-model inference scales with token volume, latency requirements, and energy constraints, and that combination favors architectures that can reduce data movement and improve memory efficiency. Reuters’ earlier technical framing of RRAM is still relevant at the category level: it described RRAM as a non-volatile memory that retains data with power on or off and cited low power consumption as a core characteristic, while noting its potential as a flash-memory substitute [Reuters, September 2010]. MemSilicon’s own financing announcement then ties that memory category to present inference demand, with proceeds allocated to RRAM industrialization and commercialization and to development of a first-generation inference processor [Xinhua, October 2026] [Tencent News, October 2026].

The adjacent markets are at least as important as the core one. If RRAM adoption in AI inference proves slower than expected, substitute demand could still come from broader embedded memory, storage-class memory, or specialized accelerators that compete on bandwidth-per-watt rather than on raw model-training throughput. Conversely, if frontier-model deployment continues shifting from centralized training clusters toward distributed inference at the edge or inside enterprise systems, memory-centric architectures could gain relevance beyond a narrow semiconductor niche. The current public record does not show customer wins, benchmarks, or deployment volumes, so that remains a conditional upside case rather than an observed market position [Tokenheart, October 2026] [Crunchbase].

Regulatory and macro forces cut both ways. China’s policy support for domestic semiconductors and strategic compute infrastructure has created a financing backdrop in which memory and AI chip startups can raise meaningful capital, as reflected in MemSilicon’s investor syndicate and the size of the disclosed Series A [Xinhua, October 2026] [36Kr, October 2026]. At the same time, advanced semiconductor commercialization is shaped by long design cycles, manufacturing dependencies, and a procurement environment that can turn quickly if incumbents improve inference economics on established platforms. That is why the market case here is more about architecture fit than headline category excitement.

Analogous market claim	What it indicates	Relevance to MemSilicon
RRAM is a non-volatile memory technology with low power consumption and potential to replace flash memory [Reuters, September 2010]	Memory category can matter where persistence and power efficiency are valuable	Supports the logic for a memory-led semiconductor wedge
MemSilicon is directing new capital to RRAM industrialization and commercialization [Xinhua, October 2026]	The company is targeting commercialization of a specific memory category, not only research	Suggests near-term market focus is narrower than full-stack AI compute
MemSilicon is also developing a first-generation inference processor for large-model workloads [36Kr, October 2026]	The addressable opportunity extends into inference silicon if product execution holds	Expands exposure from memory IP toward accelerator economics

The table points to a market thesis built on category convergence rather than a single large disclosed market estimate. Investors should read this as an architecture bet on memory efficiency inside inference systems, with market size still easier to argue qualitatively than to quantify from public data.

Partially corroborated -- Section relies on named public sources, but no third-party TAM/SAM/SOM study or direct market-size figure for MemSilicon’s target segment was confirmed.

Competition and Substitutes

Competitive map

MIXED MemSilicon is positioning against a broad stack of inference alternatives, not against a single direct peer, because its public story combines new memory technology with an inference chip roadmap rather than a conventional accelerator pitch [Xinhua, October 2026] [Tencent News, October 2026] [Crunchbase].

That distinction matters because the immediate competitive field breaks into three layers. The first is incumbent accelerator vendors, where NVIDIA remains the default reference point for large-model inference infrastructure by virtue of software ecosystem, deployment maturity, and customer familiarity, even though those advantages are not specific to MemSilicon's disclosed materials. The second is AI chip challengers, where Graphcore is relevant by name because MemSilicon's new CEO, Lu Tao, was reported as Graphcore's former president and Greater China general manager, which gives MemSilicon some visible operating overlap with the inference-accelerator category even if no direct product comparison has been published [Donews, October 2026] [NIUX, October 2026]. The third is adjacent memory and compute-in-memory approaches, where the real substitute is not one disclosed company but the broader practice of pairing mature memory components with standard GPUs or inference ASICs rather than adopting an RRAM-centered architecture.

The public evidence suggests MemSilicon's edge today is mostly on talent formation and architectural positioning, not on proven go-to-market control. Multiple reports say the company is focused on RRAM commercialization and compute-storage collaboration for large-model inference, and one profile describes the team as originating from Academician Liu Ming's group with long experience across materials, devices, process, and chip design [Xinhua, October 2026] [The Paper / Sina Finance, October 2026]. If that technical bench is real, it is a plausible early moat because non-volatile memory commercialization tends to depend on accumulated process knowledge that is harder to assemble quickly than a model-serving software layer. Still, the durability of that edge is not yet established publicly. The same sources do not disclose shipping products, benchmark data, named customers, or production deployments, so the current advantage reads as promising but perishable until it is translated into manufacturable parts and repeatable design wins [Xinhua, October 2026] [36Kr, October 2026].

The main exposure is that MemSilicon appears to be contesting categories where incumbents already own key channels. NVIDIA controls the de facto software and developer standard for AI inference in many markets, and challengers such as Graphcore have spent years trying to prove differentiated hardware can overcome ecosystem inertia. MemSilicon may also face a specific substitution risk from large cloud buyers and model operators who prefer known GPU or ASIC supply chains plus mature memory stacks over adopting an unproven RRAM-based path. Public reporting does not yet show that MemSilicon owns a distribution channel, a captive customer base, or a standards position that would force market attention on its preferred architecture [Xinhua, October 2026] [Tokenheart, October 2026].

The most plausible 18-month scenario is a sorting process between companies that can demonstrate inference economics in production and those that remain at the architecture narrative stage. NVIDIA is the clearest winner if software compatibility and deployment certainty continue to outweigh hardware novelty, because buyers under delivery pressure usually default to the stack they already know. MemSilicon could still emerge as a category winner if RRAM industrialization progresses and the first-generation LPU shows a measurable cost or power advantage in inference workloads, but that condition remains unproven in public. Graphcore is a useful named loser if differentiated inference hardware again struggles to convert technical promise into broad commercial adoption in China-facing enterprise channels, though that is a category inference rather than a reported outcome for its current business [Xinhua, October 2026] [Donews, October 2026].

Partially corroborated -- Relies on multiple public reports for MemSilicon's positioning and leadership, but the section's comparative analysis is constrained by the absence of named direct competitors and by limited public evidence on product benchmarks, customers, and deployments.

Opportunity

Upside case

PUBLIC The size of the prize here is straightforward: if MemSilicon can turn RRAM from a lab-centered memory technology into a deployable inference stack, it could matter not as another AI chip vendor, but as a new hardware layer for running large models at lower cost and power in markets where conventional accelerators remain scarce or uneconomic [Xinhua, October 2026] [Tencent News, October 2026] [Reuters, September 2010].

The headline opportunity is to become a domestic inference platform built around memory-centric architecture rather than a direct GPU substitute. That outcome is reachable, not merely theoretical, because the public record already shows three ingredients in place. First, the company has raised more than RMB 400 million in Series A financing, a meaningful amount for a pre-commercial semiconductor effort in China [Xinhua, October 2026] [36Kr, October 2026]. Second, the round brings in a syndicate that spans financial capital and semiconductor-adjacent strategic names, including Bosch Venture Capital, GigaDevice, Xiaomi Strategic Investment, and miHoYo, which suggests interest beyond a narrow academic project [Xinhua, October 2026] [Futunn, October 2026]. Third, management has tied the proceeds to two specific milestones, RRAM commercialization and first-generation LPU development, which gives the company a clearer path than a generic claim to build AI chips [Xinhua, October 2026] [The Paper / Sina Finance, October 2026].

The logic of the bet rests on architecture. RRAM is a non-volatile memory technology that has long been discussed for its low-power characteristics and potential to displace some legacy storage use cases, even if that displacement has not happened at scale yet [Reuters, September 2010]. MemSilicon is trying to connect that memory approach to large-model inference through compute-storage collaboration, and public coverage consistently describes the company in those terms [Tencent News, October 2026] [Crunchbase]. If that integration works in production, the company could sell into a part of the AI infrastructure market where performance per watt and system cost matter as much as absolute benchmark leadership.

Scenario	What happens	Catalyst	Why it's plausible
RRAM wedge into inference systems	MemSilicon first commercializes RRAM into niche AI infrastructure or adjacent memory deployments, then uses that footprint to pull through its LPU inference processor.	Successful RRAM industrialization funded by the current Series A [Xinhua, October 2026].	The company has explicitly said the new financing will be used mainly for RRAM industrialization and commercialization, which suggests a staged go-to-market rather than a single all-or-nothing chip launch [Xinhua, October 2026] [36Kr, October 2026].
Domestic inference alternative	MemSilicon becomes a credible China-based option for large-model inference buyers seeking architectures outside the mainstream GPU stack.	Launch of the first-generation LPU positioned for large-model inference [The Paper / Sina Finance, October 2026].	Multiple outlets describe the company as focused on compute-storage collaboration for large-model inference, and the appointment of Lu Tao, previously Graphcore's president and Greater China general manager, adds operating relevance to that move [Tencent News, October 2026] [NIUX, October 2026].
Strategic ecosystem supplier	One or more investors become commercial advocates, helping MemSilicon secure design-ins, manufacturing support, or early deployment pathways across devices, industrial systems, or AI infrastructure.	Investor syndicate alignment after the Series A close [Futunn, October 2026].	The round includes strategic and ecosystem-linked names such as Bosch Venture Capital, GigaDevice, Xiaomi Strategic Investment, and Zhongguancun Science City, which can matter in semiconductors where customer access and supply-chain credibility often arrive through networks before scale revenue does [Xinhua, October 2026] [Futunn, October 2026].

What compounding looks like, if it appears, is less about network effects in the software sense and more about design credibility feeding commercial access. A first win in RRAM commercialization could validate process know-how, attract more system-level partners, and create a practical distribution path for the LPU effort. That matters because semiconductor startups rarely compound from brand alone. They compound from proof that the architecture can be manufactured, integrated, and supported at customer standards. The early signs are limited but directionally relevant: the company has framed a two-step roadmap, and the investor base includes parties with reasons to care about chips beyond a pure financial return [Xinhua, October 2026] [36Kr, October 2026].

There is also a talent flywheel available if execution starts to show. Lu Tao's move from Graphcore to MemSilicon gives the company a public signal of ambition in inference hardware, and the broader team is described in local coverage as originating from Academician Liu Ming's Chinese Academy of Sciences group with deep experience across materials, devices, process, and chip design, although that team claim remains company-linked and should be treated cautiously [NIUX, October 2026] [The Paper / Sina Finance, October 2026]. In practice, one taped-out product milestone or one visible deployment could make future hiring, supplier negotiations, and follow-on fundraising meaningfully easier.

The size of the win is hard to quantify from public data because no credible market-sizing source is in the record and no direct public comparable is named in the sourced materials. Even so, a useful framing is this: if the "domestic inference alternative" scenario plays out and MemSilicon ships a commercially relevant architecture spanning memory and inference processing, the company could grow into a strategic semiconductor asset rather than a single-product startup (scenario, not a forecast). That would imply value creation driven by platform control over a difficult part of the AI stack, not just by one chip cycle. The evidence does not support a precise valuation band yet, but it does support an upside case where successful RRAM commercialization becomes the entry point to a broader inference hardware franchise [Xinhua, October 2026] [Tencent News, October 2026] [Crunchbase].

Partially corroborated -- This section relies mainly on Xinhua, 36Kr, Tencent News, Futunn, Reuters, Crunchbase, and NIUX. Core financing and positioning claims are publicly corroborated, but product roadmapping and team-depth details remain only partially verified in public sources.

Sources

Open sources

1. [Xinhua, October 2026] MemSilicon完成超4亿元A轮融资 宣布战略升级 卢涛出任CEO 面向大模型推理 加速存算协同落地 | http://www.news.cn/finance/20261020/4c658288a276478eb73ed062510d47fe/c.html

2. [Crunchbase] Memsilicon | Crunchbase Company Profile & Funding | https://www.crunchbase.com/organization/memsilicon

3. [The Paper / Sina Finance, October 2026] MemSilicon完成超4亿元A轮融资，宣布战略升级 | https://portal.sina.com.hk/technology/sina/2026/10/19/1959466/memsilicon%E5%AE%8C%E6%88%90%E8%B6%85%E4%BA%BF%E5%85%83a%E8%BC%AA%E8%9E%8D%E8%B3%87%EF%BC%8C%E5%85%AC%E5%8F%B8%E6%88%B0%E7%95%A5%E5%85%A8%E9%9D%A2%E5%8D%87%E7%B4%9A/

4. [36Kr, October 2026] Mingxin Qirui completes Series A financing of over RMB 400 million | https://eu.36kr.com/en/newsflashes/4032399614481033

5. [Donews, October 2026] MemSilicon完成超4亿元A轮融资 芯片老将卢涛出任CEO | https://www.donews.com/news/detail/1/6745188.html

6. [Reuters, September 2010] Micron says phase change memory close to market | https://www.reuters.com/article/us-micron-technology-idUSTRE68O5EA20100925

7. [Futunn, October 2026] MemSilicon Announces Completion of Over RMB 400 Million in Series A Financing; Industry Veteran Lu Tao Appointed as CEO | https://news.futunn.com/en/post/1001511496/memsilicon-announces-completion-of-over-rmb-400-million-in-series

8. [Tencent News, October 2026] 存算协同芯片公司铭芯启睿完成超4亿元A轮融资 | https://news.qq.com/rain/a/20261019A09ZDH00

9. [Tokenheart, October 2026] Memory Startup MemSilicon Raises Over 400 Million Yuan | https://www.tokenheart.ai/article/lcig-portfolio

10. [MemSilicon] 研发下一代智能存算芯片，「铭芯启睿」完成近亿元天使轮融资，多家战投出资-北京铭芯启睿科技有限公司 | https://memsilicon.com/newsinfo/8730091.html

11. [NIUX, October 2026] 存算协同芯片公司铭芯启睿完成超4亿元A轮融资 卢涛出任CEO | https://niux123.com/article/24375.html