AN-031 Company lists · list
AI Chip and Semiconductor Startups
AI chip startups in the HardwareMap index: data center inference, edge NPUs, analog, photonic and neuromorphic chips. HQ, stage, disclosed raise, status.
by the HardwareMap editors3 min read
From the HardwareMap index, REV 26.09, refreshed 3 Oct 2026
Contents
AI chip startups in the HardwareMap index number 32 as of REV 26.09, from wafer-scale data center processors to milliwatt edge chips. This page lists each with HQ hub, stage, disclosed capital raised and status. Sorting rule: where the chip runs (data center or edge), then architecture, then disclosed funding. It sits under the hardware startups index.
The table is live, drawn from the same records as each company datasheet. The full silicon category adds RISC-V core makers, photonic interconnects, quantum computing and server builders.
How we sort AI chip startups
- Data center inference and training. Accelerators sold as cards, servers or cloud capacity.
- Edge AI. Chips for cameras, robots, cars and devices, measured in watts or less.
- New compute physics. Analog in-memory, photonic and neuromorphic chips.
Disclosed capital raised orders the names inside each group. "Raised" means disclosed equity only.
Data center AI chip startups
Cerebras Systems (Sunnyvale) builds wafer-scale processors and the CS-3 systems around them, and sells inference through its own cloud. It priced its IPO at $185 per share on May 13, 2026, and trades on Nasdaq as CBRS (Cerebras).
Tenstorrent (Santa Clara), led by chip architect Jim Keller, sells accelerators such as Blackhole and licenses RISC-V CPU and AI designs to other chipmakers. SambaNova Systems (San Jose) builds reconfigurable dataflow chips and systems for enterprise inference. Positron AI (Reno) builds inference chips and systems and announced an $875M Series C in September 2026, per its index record.
Three bet on narrow targets. Etched (San Jose) designs Sohu, an ASIC that runs only transformer models (Etched). MatX (Mountain View) designs chips for training and running large language models. d-Matrix (Santa Clara) uses digital in-memory compute for inference.
Korea has its own data center group: Rebellions, FuriosaAI and HyperAccel, all in Seoul. Europe adds Fractile (London), and Israel NeuReality (Caesarea), which builds an AI-CPU to keep GPUs fed. Tensordyne (Sunnyvale), formerly Recogni, builds inference systems around its Napier processor.
Edge AI chip startups
Edge chips sell into cameras, robots, cars and earbuds, where power and cost matter more than peak speed. Hailo (Tel Aviv) makes accelerators and vision processors, including the Hailo-10. Axelera AI (Eindhoven) ships its Metis AIPU on M.2 and PCIe cards. SiMa.ai (San Jose) builds system-on-chips for robots, drones and cars. Blaize (El Dorado Hills) builds graph-streaming processors and is listed.
Smaller edge makers include Kneron (San Diego), DEEPX (Seongnam), Mobilint (Seoul), EdgeCortix (Tokyo), MemryX (Ann Arbor), Netrasemi (Bengaluru) and two ultra-low-power specialists, Syntiant (Irvine) and Aspinity (Pittsburgh). Robot makers are a growing customer base for this group; see top robotics companies.
New compute physics: analog, photonic, neuromorphic
EnCharge AI (Santa Clara) and Mythic (Austin) compute in analog inside memory arrays to cut power. Q.ANT (Stuttgart) and Lumai (Oxford) compute with light. Innatera (Rijswijk), SynSense (Chengdu) and Prophesee (Paris) build brain-inspired processors and event-based vision sensors that react only to change.
These are the earliest-stage names on the list. Lumai is at seed and Innatera at Series A, per their index records.
Semiconductor startups beyond AI accelerators
The rest of the silicon category serves the same data centers from other angles. Lightmatter (Mountain View) and Ayar Labs (San Jose) move data between chips with light instead of copper, and Salience Labs (Oxford) builds optical switches. Enfabrica (Mountain View) designs networking chips for AI clusters. SiFive (Santa Clara) licenses RISC-V processor cores, and SiPearl (Maisons-Laffitte) designs CPUs for European supercomputers. Axiado (San Jose) builds security silicon for data center platforms.
One exit sits in the category: NeuroBlade (Tel Aviv) was acquired by Amazon Web Services in September 2025, per its index record. And one startup makes chips rather than only designing them: Pragmatic Semiconductor (Cambridge, UK) builds flexible integrated circuits in its own UK fabs.
What the index shows about AI chip startups
- Inference is where startups compete. Most data center names in the table describe their product as inference-first.
- The Bay Area leads, but not by much. 16 of the 32 are American, and nine are in the Bay Area. See hardware startups in the SF Bay Area for the wider cluster.
- Strategic investors are common. Samsung, Arm, Marvell, Dell, DENSO and Mercedes-Benz appear in the investor lists recorded in the index, alongside venture funds.
Chip startups need more capital per product than most hardware. Investors who back them are in deep tech VC firms. The next REV adds new entries and marks status changes here first.
Frequently asked questions
How many AI chip startups are in the HardwareMap index?
The index lists 32 AI chip startups as of REV 26.09, inside a wider silicon category that also covers RISC-V cores, photonic interconnects, quantum and servers. Half of the AI chip startups are American. South Korea is the second-largest group with five, followed by the Netherlands, Israel and the United Kingdom with two each. Two are publicly listed.
Which AI chip startup has raised the most?
Cerebras Systems has the largest disclosed raise in the index at about $10.2 billion, and it listed on Nasdaq in May 2026. Among private companies, Tenstorrent leads at about $3 billion, followed by SambaNova Systems at about $2.5 billion and Positron AI at about $1.2 billion. Figures are disclosed equity only and never converted into valuations by the index.
What is the difference between a training chip and an inference chip?
Training builds a model by running huge batches of data through it many times, which needs high memory bandwidth and fast links between many chips. Inference runs the finished model to answer requests, which rewards low cost per answer and low latency. Most AI chip startups in the index describe their data center products as inference-first, and the edge chips on this page run inference only.
Where are AI chip startups located?
The Bay Area is the largest cluster, with nine AI chip startups in San Jose, Santa Clara, Sunnyvale and Mountain View. Seoul has four, with a fifth in nearby Seongnam. Europe has a group in the Netherlands, Oxford and London, Paris and Stuttgart. Israel has two. Location tracks where chip design talent and foundry relationships already exist, not where chips are made.
Do AI chip startups make their own chips?
No. Almost all are fabless: they design the chip and pay a foundry to manufacture it, then buy packaging and test services. That keeps capital needs far below those of building a fab, but each new chip still needs design tools, a large engineering team and mask sets paid for before any revenue. Pragmatic Semiconductor, a silicon company in the index outside this list, is an exception: it runs its own fabs.
HardwareMap catalogues every AI chip startup that meets the bar. Open any company above for its datasheet.
The index
All 219 startups, filterable by category, hub, stage and funding.