The 10 Hottest Semiconductor Startups Of 2025
CRN rounds up 10 semiconductor startups developing silicon products, including AI chips and networking solutions, that either compete with Nvidia’s or complement them.
As Nvidia demonstrated during its most recent earnings, demand for computer chips to fuel AI workloads remains voracious, even as critics continue to question whether the industry is nearing the peak of a dotcom-like bubble.
This is potentially good news for semiconductor startups who hope there will be enough room for them to grab a decent chunk over time of the AI infrastructure market, which Nvidia expects to reach up to $4 trillion by the end of the decade.
[Related: AI Chip Startups Seek An Edge By Enlisting The Channel]
While some startups like Untether AI have come and gone, there are plenty of others who continue to raise hundreds of millions of dollars from investors and develop silicon products that either compete with Nvidia’s or complement them.
Direct competitors in the startup world include Axelera AI, which is focused on high-performance edge computing, as well as those who are focused on data center deployments, such as d-Matrix, FuriosaAI, NextSilicon and Rebellions.
A few of these startups—like NextSilicon, Tenstorrent and Tsavorite Scalable Intelligence—are not only developing rival AI chips but also general-purpose CPUs and, in some cases, other complementary technologies to deliver new paradigms in computing.
Then there is Cornelis Networks, which is building scale-out networking solutions to challenge Nvidia in this area, as well as Celestial AI, which is developing optical interconnect technology to enable seamless scale-up networking in AI data centers.
Xsight Labs, on the other hand, is looking to shake up data centers and edge environments with its high-bandwidth, low-latency DPU, server and Ethernet switch products for use cases that range from storage to security.
What follows are descriptions of CRN’s 10 hottest semiconductor startups of 2025, which include some of their biggest milestones from the past 12 months.
Axelera AI
Top Executive: Fabrizio Del Maffeo, CEO and Co-Founder
Axelera AI is challenging Nvidia with AI chips that it says can run faster while using less energy and at lower cost for edge computing workloads.
The Netherlands-based startup in October announced its new Europa chip, saying it will 629 trillion operations per second (TOPS) in a 45-watt power envelope when it debuts early next year in a big step up from the chip’s predecessor, which can reach a maximum of 214 TOPS while using four to eight watts depending on the form factor.
This came after Axelera AI in June launched its global Partner Accelerator Network with more than 15 partners of several types, including OEMs such as Lenovo, Dell Technologies and Advantech; electronics distributors like Astute Group, Rutronic and Silicon Applications Group Corp; and other companies such as Macnica ATD Europe and Arduino.
The company also promoted CMO Alexis Crowell in July to grow the company’s business in North America and South America, with a major U.S. emphasis.
Celestial AI
Top Executive: David Lazovsky, CEO
Celestial AI says its optical interconnect technology is uniquely equipped to meet the demands of large-scale AI workloads in data centers “while setting new standards for bandwidth, latency, energy efficiency and total cost of ownership.”
The Santa Clara, Calif.-based startup in August that it closed its $255 million Series C1 funding round from a wide range of investors, including Fidelity Management and Research Company, BlackRock, Samsung Catalyst Fund, AMD Ventures and Penguin Solutions. This brought the startup’s total funding to $520 million.
Celestial AI has said it will use the new funding to “expand and qualify its volume manufacturing supply chain to serve customer demand” for its products, which consist of connectivity, switching and packaging solutions for optical scale-up networks thar are all based around the company’s Photonic Fabric technology platform.
Cornelis Networks
Top Executive: Lisa Spelman, CEO
Cornelis Networks, an Intel spin-off, says its family of scale-out networking solutions offer “the best performance with devastatingly good price-performance” for AI data centers compared to offerings based on InfiniBand or Ethernet.
The Wayne, Pa.-based startup in November revealed its 800-Gbps CN5000 family of scale-out networking solutions, saying it will deliver the “industry’s highest performance with up to 1.6 million messages per second” with support for Ethernet RoCEv2 and Ultra Ethernet.
Cornelis Networks also announced this month that its 400-Gbps CN5000 products, which launched this year, have been qualified and integrated across Lenovo’s ThinkSystem V3 and V4 servers. In addition, the vendor said it has partnered with the U.S. Department of Energy as part of a select group of companies to support the agency’s new Genesis Mission initiative to boost scientific research and discovery with AI.
D-Matrix
Top Executive: Sid Sheth, CEO and Co-Founder
D-Matrix is betting on digital in-memory computing to provide what it calls the “world’s most efficient AI inference computing platform for data centers.”
The Santa Clara, Calif.-based startup in November announced that it raised a $275 million funding round from several investors, including Microsoft’s M12 venture fund, at a $2 billion valuation. This was done to “advance the company’s roadmap, accelerate global expansion and support multiple large-scale deployments,” according to d-Matrix.
The funding round came after d-Matrix in October revealed SquadRack, which it called the “industry’s first blueprint for disaggregated standards-based rack-scale solutions for ultra-low latency batched inference.” Powered by its Corsair chips, SquadRack is built with the Supermicro X14 server platform, Broadcom PCIe switches and Arista leaf ethernet switches that connect to its JetStream NICs.
FuriosaAI
Top Executive: June Paik, CEO and Co-Founder
FuriosaAI is taking on Nvidia with energy-efficient, high-density AI chips that are designed to deliver “sustainable, software-driven performance scaling” in data centers.
After reportedly rejecting an $800 million offer to get acquired by Meta earlier this year, the South Korean startup announced in July that it raised a $125 million Series C funding round to “meet growing demand from global enterprise customers” for its flagship RNGD AI chip.
Among those customers is the AI research wing of South Korean electronics giant LG, which found that RNGD provides 2.25 times better performance per watt for LLM inference compared to GPUs, namely Nvidia’s A100 that debuted in 2020, according to FuriosaAI.
The startup then revealed in September its NXT RNGD Server, which it said can provide 3.5 times more compute per rack than GPUs—with the main comparison point being Nvidia’s H100 SXM system—thanks to the product using just 3 kilowatts per system.
NextSilicon
Top Executive: Elad Raz, CEO and Founder
NextSilicon is mounting a challenge to Nvidia, Intel and AMD with AI chips and CPU cores it says can run faster, with the former not requiring any code changes for software.
The Tel Aviv, Israel-based startup in October claimed that its Maverick-2 AI chip, set for volume production in the fourth quarter, can provide up to 10 times greater performance than “leading” GPUs using 60 percent less power for “algorithmically complex workloads.”
NextSilicon also revealed Arbel, which it called an “enterprise-grade” CPU core based on the RISC-V instruction set architecture. The company claimed that Arbel can surpass competing RISC-V products as well as Intel’s Lion Cove P-core architecture—currently used in Core Ultra Series 2 chips but coming in a next-generation Xeon line-up—and AMD’s Zen 5 core architecture—used now in server, desktop and laptop product lines.
Rebellions
Top Executive: Sunghyun Park, CEO and Co-Founder
Rebellions is going after Nvidia with an AI chip it says can enable rack-scale performance for frontier AI models with high utilization and low power.
The South Korean startup in November announced the hiring of two former SambaNova Systems and Oracle executives: Marshall Choy as chief business officer to lead its newly established U.S. entity and global business, and Jennifer Glore as executive vice president of product management to align Rebellions’ development teams and end users.
The same month, Rebellions announced that it closed a $250 million Series C funding round from several investors, including Arm and Samsung, to expand its global footprint and expand through partnerships in the United States.
Tenstorrent
Top Executive: Jim Keller, CEO
Tenstorrent is shaking up the AI computing space with a business model that mixes selling specialized processors with providing open-source software, licensing out chip technologies to third parties and working with other firms to develop computing solutions.
The Information reported in November that the Toronto, Ontario-based startup is in discussions to raise a new $800 million round led by Fidelity Management and Research Company that would give it a roughly $3.2 billion valuation. Last December, it raised $693 million at a $2 billion pre-money valuation from several investors, including Samsung Securities, Bezos Expeditions, Hyundai Motor Group and LG Technology Ventures.
Tenstorrent has made several announcements this year, including the launch of a “scalable, cost-efficient AI data center solution” with South Korean AI software company Moreh as well as its Blackhole PCIe cards for handling “massive AI workloads efficiently.” The company also shared details about its licensable Ascalon RISC-V CPU technology for data center and AI applications along with the Open Chiplet Atlas specification.
Tsavorite Scalable Intelligence
Top Executive: Shalesh Thusoo, CEO and Founder
Tsavorite Scalable Intelligence is shaking up the computing world with its Omni Processing Unit (OPU) that unifies CPU, GPU, memory and connectivity in a single device that can be configured for different power, performance and scaling requirements.
The Milpitas, Calif.-based startup announced in November that it has more than $100 million in pre-orders for its silicon and enterprise-class Helix AI appliances, set for launch next year, from various customers, including Fortune 500 companies, sovereign cloud providers and leading systems integrators across the world.
Tsavorite said that its Agentic Operating Stack allows developers to deploy models on its OPU “without code rewrites, quantization adjustments or proprietary software dependencies,” noting its support of popular frameworks such as PyTorch, vLLM and Triton.
Xsight Labs
Top Executive: Yossi Meyouhas, CEO
Xsight Labs is challenging the paradigm for end-to-end connectivity in data centers and at the edge with its high-bandwidth, low-latency DPU, server and Ethernet switch products.
The Kiryat Gat, Israel-based startup last month announced multiple partnerships, with EdgeCore Networks using its E1-SoC to provide the “industry’s first 800G DPU” in a PCIe form factor, Hammerspace choosing the E1 DPU to fulfill AI storage needs with the Open Flash Platform, and Cyber Forza using the E1 DPU for an 800G security platform.
These partnerships were announced after Xsight Labs in May launched its Arm-based E1-SoC system-on-chip for cloud and edge AI data centers, saying that it’s the “industry’s highest performance, software-defined DPU” and the “only product of its kind to provide full control plane and data path programmability.”
Xsight Labs said the E1-SoC powers the E1-Server, which it called the “first-to-market 800G DPU.” The single rack-unit system can be used as a “high-performance, stand-alone edge server system” or as “comprehensive product development platform,” according to the company. It said the E1-SoC is “perfect for developing compute, networking, security, storage, smart switch and air-cooled inference in smaller form factors.”