MatX
Last reviewed
Jun 8, 2026
Sources
6 citations
Review status
Source-backed
Revision
v1 · 1,440 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 8, 2026
Sources
6 citations
Review status
Source-backed
Revision
v1 · 1,440 words
Add missing citations, update stale details, or suggest a clearer explanation.
MatX is an American semiconductor startup that designs AI chips purpose-built for large language models (LLMs). Founded in 2022 by two former Google engineers who had worked on the company's Tensor Processing Units (TPUs), MatX is based in Mountain View, California, and positions itself as a challenger to NVIDIA, whose GPUs dominate AI training and inference. Rather than building general-purpose accelerators, MatX concentrates its silicon on the matrix-heavy compute of transformer models, trading generality for efficiency and aiming for substantially higher performance per dollar on the largest models.[1][2][3] In February 2026 the company disclosed a funding round of more than $500 million, one of the larger raises among the wave of LLM-specific chip startups, and said it expected to ship its first product in 2027.[1][4]
MatX develops what it calls "high-throughput chips for LLMs," targeting the workloads of so-called frontier labs: organizations training and serving the largest AI models. The company states that its mission is to build "the best chips physically possible for the large model needs of frontier labs," and its first product is named MatX One.[3] The design philosophy is one of extreme co-design: where NVIDIA and other vendors sell silicon that can run a wide range of workloads, MatX deliberately optimizes only for large transformer models and explicitly does not target small models, convolutional networks, or recommender systems.[3] The company says this focus lets it dedicate more of each chip to dense matrix multiplication, which dominates LLM compute.
MatX pursues both training and inference on a single architecture, including training, reinforcement learning, inference prefill, and inference decode.[3] That breadth distinguishes it from several other AI-chip startups whose designs have been oriented primarily toward inference. The company reported a headcount of more than 100 people as of early 2026.[2]
MatX was founded in 2022 by Reiner Pope and Mike Gunter, who left Google roughly one week before the public launch of ChatGPT on November 30, 2022.[2] Pope, who serves as chief executive, came from Google's Brain research organization and had led AI software development for the company's TPUs; Gunter had been a lead designer of the TPU hardware.[1][2][5] Their shared background on Google's custom AI silicon, one of the few large-scale alternatives to NVIDIA GPUs deployed in production, underpins the company's pitch that purpose-built accelerators can beat general-purpose chips on frontier-scale models.
In interviews, Pope has framed the decision to start MatX around the economics of AI compute: because spending on chips has come to dwarf engineering salaries at the largest labs, he argues it is rational for those customers to invest in software for new hardware, and that the much-discussed lock-in of NVIDIA's CUDA software ecosystem is "pretty weak" for buyers who already operate across multiple hardware platforms.[5]
MatX One is built around a "splittable systolic array," a configurable matrix-multiplication engine that the company says delivers the highest FLOPS per square millimeter of any product it is aware of.[3] The chip uses a hybrid memory system: model weights are held in SRAM for low-latency access, an approach associated with inference-focused designs from Groq and Cerebras, while key-value (KV) caches are stored in high-bandwidth memory (HBM), the memory NVIDIA and Google use to support long contexts and high throughput.[1][3][5] MatX says this combination lets a single chip deliver throughput higher than any announced product while matching the low latencies of SRAM-first designs.[3]
The company emphasizes interconnect as a core differentiator, claiming the most scale-up interconnect of any announced product and the ability to scale out to clusters with hundreds of thousands of chips, which is intended to make large mixture-of-experts (MoE) models run efficiently.[3] MatX reports that MatX One can produce more than 2,000 output tokens per second on large 100-layer MoE models, and the architecture supports techniques such as speculative decoding and blockwise sparse attention.[1][3] The chip is programmed through a direct hardware-control model rather than a high-level abstraction layer.[3]
A central claim, repeated by the founders since the company's early funding announcements, is an ambition to make their processors roughly 10 times better than NVIDIA GPUs at training LLMs and serving results.[1][6] Around its Series A, MatX described its target customers as those running models with at least about 7 billion and ideally more than 20 billion activated parameters, with advanced interconnect to scale across large clusters.[6] As with most startup performance claims, these figures are company-stated and had not been independently benchmarked against shipping NVIDIA hardware as of mid-2026.
As of 2026, MatX had not yet shipped a chip. The company plans to manufacture MatX One at TSMC and said its first product is expected to ship in 2027, with a tape-out anticipated within roughly a year of the February 2026 funding announcement.[1][4][5] Pope has publicly acknowledged that scaling manufacturing to data-center volumes and building out the necessary software ecosystem remain significant challenges ahead.[5]
MatX has raised approximately $600 million across several rounds. The figures, leads, and dates below are drawn from press reporting; the company has at times declined to comment on specifics, and some details vary slightly between sources.
| Round | Amount | Date | Lead investors | Reported valuation |
|---|---|---|---|---|
| Seed | About $25 million | December 2023 | Nat Friedman, Daniel Gross | Not disclosed |
| Series A | About $80 million | November 2024 | Spark Capital | Over $300 million |
| Series B | More than $500 million | February 24, 2026 | Jane Street, Situational Awareness | Not disclosed |
The seed round was reported at about $25 million, led by investors Nat Friedman, the former chief executive of GitHub, and Daniel Gross, a former Apple AI and search lead.[6] In November 2024, TechCrunch reported a Series A of roughly $80 million led by venture firm Spark Capital, valuing the company at more than $300 million; MatX and Spark Capital declined to comment at the time.[6] Some later coverage described MatX as having "previously raised over $100 million" before its 2026 round, a figure that appears to aggregate earlier financings.[4]
On February 24, 2026, the company disclosed a round of more than $500 million, reported by Bloomberg and others.[1] The financing was led by quantitative trading firm Jane Street and by Situational Awareness, an investment fund founded by former OpenAI researcher Leopold Aschenbrenner. Additional backers reported included Marvell Technology, venture firms NFDG and Spark Capital, and Stripe co-founders Patrick and John Collison.[1][2] The round was first announced through a LinkedIn post by Pope; the valuation was not disclosed.[1]
MatX is one of a growing field of startups attempting to challenge NVIDIA's dominance of AI hardware by building silicon tuned specifically for transformer-based LLMs, a strategy sometimes described as a wave of "LLM-specific" or "transformer-native" chips. Its most direct comparison is Etched, which is designing chips hardwired for the transformer architecture and which raised $500 million at a reported $5 billion valuation in early 2026.[1] Other AI-chip companies in the broader landscape include Groq and Cerebras, known for SRAM-heavy, inference-oriented designs, as well as SambaNova and Tenstorrent.[3][5] MatX argues that its hybrid SRAM-plus-HBM approach and single architecture for both training and inference let it avoid the specialization that has confined some rivals largely to inference.
Beyond startups, MatX competes against NVIDIA's data-center GPUs and against the in-house accelerators of hyperscalers, including Google's TPUs, Amazon's Trainium and Inferentia, and Microsoft's Maia, all of which represent custom silicon efforts aimed at reducing dependence on NVIDIA.[1][5]
The significance of MatX lies less in any shipped product, since it had none as of 2026, than in what its founders' pedigree and its large funding round signal about investor appetite for purpose-built AI silicon. The company's bet is that as AI compute spending grows and a handful of frontier labs concentrate the largest workloads, hardware specialized for those workloads can win share from general-purpose GPUs on the basis of cost and performance. Whether MatX can convert that thesis into competitive, manufacturable chips remains to be demonstrated, with the company's product roadmap pointing to 2027.[1][3][5]