MatX Raises $500M to Build LLM-Specific Chips That Challenge Nvidia's Dominance

Ex-Google Chip Engineers Bet $500 Million on LLM-Only Silicon

On February 24, 2026, MatX announced a $500 million Series B funding round to develop a semiconductor chip designed exclusively for running large language models. The round was led by Jane Street, the quantitative trading firm, and Situational Awareness, the investment fund formed by former OpenAI researcher Leopold Aschenbrenner. Additional investors include Marvell Technology, venture firms NFDG and Spark Capital, and Stripe co-founders Patrick and John Collison.

The company was founded in 2023 by Reiner Pope and Mike Gunter, both former Google engineers who worked on the search giant's Tensor Processing Unit (TPU) chips. Pope worked on software for Google's chips and AI models, while Gunter was a hardware engineer for the TPU program. They left Google with a specific thesis: a chip built exclusively for LLM workloads can fundamentally outperform general-purpose AI accelerators like Nvidia's GPUs.

The MatX One: Architecture Built Around LLM Constraints

The company's chip, called the MatX One, takes a radically different approach from Nvidia's general-purpose GPU architecture. Rather than designing silicon that handles a broad range of AI workloads, MatX has optimized every architectural decision for the specific computational patterns of large language models.

The core innovation is what MatX calls a "splittable systolic array." Traditional systolic arrays process data in fixed configurations. MatX's design can partition its computing modules into smaller configurations, allowing the chip to dynamically optimize performance based on the specific characteristics of the data being processed. This adaptability means the chip can adjust its compute topology to match different stages of LLM inference, from attention computation to feed-forward layers.

Memory Architecture

The memory strategy is equally specialized. MatX uses a dual-tier approach:

This combination addresses one of the fundamental bottlenecks in LLM inference: the tradeoff between latency and context length. SRAM-first designs deliver low latency but struggle with long contexts. HBM-based designs handle long contexts but introduce latency penalties. MatX's architecture claims to deliver both.

CEO Reiner Pope stated that the chip "combines the low latency of SRAM-first designs with the long-context support of HBM," delivering "higher throughput on LLMs than any announced system."

Performance Optimization Techniques

The MatX One incorporates two additional optimization techniques that target specific LLM inference bottlenecks:

Speculative decoding accelerates token generation by predicting multiple tokens simultaneously and verifying them in parallel, rather than generating one token at a time in a sequential autoregressive process.

Blockwise sparse attention improves the efficiency of the attention mechanism by selectively computing attention scores for relevant blocks rather than performing full quadratic attention across the entire sequence. This is particularly impactful for long-context inference where full attention computation becomes prohibitively expensive.

Manufacturing and Timeline

MatX plans to work with Taiwan Semiconductor Manufacturing Co. (TSMC) to manufacture the chip. The new funding will be used to complete the final chip design, reserve manufacturing capacity, and secure key components for rapid production scaling. MatX targets tape-out, the semiconductor manufacturing handoff, within one year, with initial chip shipments expected to begin in 2027.

This timeline places MatX in a competitive window where demand for LLM inference compute continues to outpace supply. If the company can deliver on its performance claims, it enters a market where hyperscale cloud providers and AI companies are actively seeking alternatives to Nvidia's supply-constrained GPU inventory.

Investor Composition Signals Market Confidence

The investor lineup is notable for its diversity and strategic significance. Jane Street, known for its quantitative trading operations and deep technical culture, brings both capital and potential demand as a heavy user of compute infrastructure. Leopold Aschenbrenner's Situational Awareness fund, formed after his departure from OpenAI, brings credibility from someone who has publicly argued that AI compute demand will grow dramatically in the coming years.

Marvell Technology's participation adds semiconductor industry validation, while the Stripe co-founders' involvement suggests interest from the broader technology infrastructure community. The previous round raised over $100 million from an overlapping investor consortium, bringing total funding to over $600 million.

The Competitive Landscape

MatX enters a market where Nvidia holds dominant market share in AI accelerators through its H100 and B200 GPU product lines. However, several dynamics create openings for challengers. Nvidia's GPUs are general-purpose designs that handle diverse AI workloads including training, inference, image generation, and scientific computing. This generality, while commercially powerful, means that workload-specific optimizations are necessarily constrained.

Other challengers include Groq, Cerebras, and SambaNova, each taking different architectural approaches. MatX's distinction is its narrow focus on LLM inference specifically, betting that the growing dominance of language models in enterprise AI creates sufficient market demand for specialized silicon.

Conclusion

MatX's $500 million raise represents a significant bet that the future of AI compute will bifurcate between general-purpose training accelerators and workload-specific inference chips. With ex-Google TPU expertise, a novel memory architecture, and a TSMC manufacturing partnership, the company has the technical foundations and capital to test this thesis. Whether the MatX One can deliver on its performance claims against Nvidia's entrenched ecosystem will become clear when chips begin shipping in 2027.