Private AI chip companies, by latest reported valuation

Etched builds Sohu, an ASIC that hard-wires the transformer architecture into silicon, trading a GPU's flexibility for speed and efficiency on transformer-based models. It now says its systems can run any frontier model, including mixture-of-experts designs (DeepSeek, Qwen) and non-transformer ones (Mamba), pushing back on the view that the chip is locked to one architecture. Its valuation rose roughly fourfold across 2026, from about $5 billion in December to $10.3 billion in July and $21 billion in a $700 million round led by Jane Street, which led after running one of Etched's racks in its own datacenter. The main risk is concentration: silicon optimized for today's dominant architectures has less room to adapt if they are displaced.

SambaNova sells whole systems rather than merchant chips: its reconfigurable dataflow processor (RDU) runs inside turnkey racks for enterprises and governments that want frontier-scale inference behind their own firewall. After cutting staff and shifting from training to inference, it was reported in late 2025 as an Intel acquisition target at about $1.6 billion. Instead it deepened the Intel relationship into a Xeon-plus-RDU inference design and raised a $1 billion Series F at an $11 billion valuation, a large step up from that acquisition mark.

Groq's Language Processing Unit (LPU) is built for deterministic low-latency inference, which through 2025 drove its valuation to $6.9 billion alongside a global data-center build-out. In December 2025 Nvidia struck a roughly $20 billion non-exclusive licensing deal and hired founder Jonathan Ross and other staff; Groq then raised $650 million to re-staff and refocused on running an inference cloud. A $350 million round in August 2026 valued it at $3.5 billion, about half its 2025 peak, with planned participation from Nvidia. Groq now operates 13 data centers and says its cloud combines Nvidia's LPX platform with its own technology, so the question is less rival-versus-incumbent than how a neocloud running significant Nvidia silicon sets itself apart.

Tenstorrent, led by Jim Keller, pairs RISC-V CPUs with its own AI cores and licenses the designs as well as selling chips, positioned against Nvidia's CUDA lock-in. Its figure here is the oldest on the list, about $2.7 billion from a Series D that closed in December 2024, and likely understates its current value: in 2026 it drew reported takeover interest from Intel and Qualcomm, which Keller played down. Recent output spans a compact edge accelerator with Razer and a pay-per-use chip service in Japan.

Rebellions is South Korea's national challenger to Nvidia, backed by Samsung and major domestic institutions. It raised a $250 million Series C at $1.4 billion in late 2025, then a $400 million pre-IPO round at about $2.34 billion, with JPMorgan lined up for a Seoul listing. It leans into sovereign AI through a partnership with SK Telecom and Arm; by August 2026 Nvidia was reported to be in early talks with the company.

d-Matrix bets that inference is limited by memory bandwidth rather than raw math, building digital in-memory compute into its Corsair cards and extending it with JetStream I/O accelerators. A $275 million Series C in November 2025 valued it at $2 billion. It positions as a full stack beside the incumbents: its accelerators run alongside Nvidia GPUs in Parasail's inference cloud, and it acquired deployment-software firm Wallaroo.ai.

Positron competes on power efficiency, an increasingly binding constraint on data-center economics. It says its Atlas accelerator matches or beats an Nvidia H200 on inference at about a third of the power, though this is a company figure without a standardized public benchmark (the workload, precision, and power-measurement basis are not disclosed). A $230 million Series B in February 2026 pushed it past a $1 billion valuation, following a $51.6 million Series A in mid-2025.

Tensordyne (formerly Recogni) builds the Napier inference chip and rack-scale systems around a logarithmic number system, its "Pareto" math, aimed at higher tokens-per-watt. It has raised about $209 million to date (most recently a $100 million Series C in February 2024), with a Series D in preparation. The roughly $1 billion shown is the company's own estimate rather than a priced round, so it is the least comparable figure on this list.

FuriosaAI, based in South Korea, makes the RNGD chip for efficient inference on large models. It reportedly turned down an $800 million acquisition by Meta in 2025 to stay independent and won LG AI Research as a customer for its EXAONE models. It has been raising a roughly $500 million round ahead of a planned IPO; its last clearly closed valuation is near $735 million (July 2025).