Anthropic held acquisition discussions to buy artificial intelligence semiconductor startup MatX for approximately $7 billion before talks stalled, according to reporting from Reuters. The negotiations, which have since transitioned into discussions surrounding a potential commercial partnership, highlight the increasing urgency among frontier AI laboratories to secure in-house silicon engineering capabilities.
Following the breakdown of active acquisition talks, MatX is currently seeking to raise fresh venture capital at a valuation of approximately $4 billion.
MatX Architecture and TPU Pedigree
MatX was founded in late 2022 by former Google engineers Reiner Pope and Mike Gunter. Both founders played central roles in Google's Tensor Processing Unit (TPU) development programs: Pope served as the efficiency lead for Google PaLM and contributed to TPU v5e optimizations, while Gunter served as chief hardware architect for Google ML accelerators.
Unlike general-purpose GPUs optimized across diverse scientific and graphics workloads, MatX designs application-specific integrated circuits (ASICs) tailored exclusively to transformer-based large language models:
- Splittable Systolic Arrays: The compute core optimizes matrix multiplication density specifically for dense and mixture-of-experts (MoE) transformer architectures.
- Hybrid Memory Hierarchy: The architecture balances high-bandwidth memory (HBM) with on-chip static random-access memory (SRAM). Model weights can reside in low-latency SRAM for fast token generation, while key-value (KV) caches utilize external HBM to support extended context windows.
- Inference and Training Economics: MatX targets sub-10ms per-token latencies on 70-billion-parameter models and claims up to a 10x compute-density improvement over traditional architectures for large-scale training clusters.
Anthropic's Multi-Track Silicon Strategy
The $7 billion acquisition attempt reflects Anthropic's broader push to mitigate hardware bottlenecks and control inference infrastructure economics as model usage scales.
The company has pursued a multi-pronged compute strategy across several vendors and architectural approaches:
- Internal Silicon Team: Anthropic recently recruited veteran chip executive Amir Salek, formerly a senior Google TPU architect, to lead its internal semiconductor design efforts.
- Hyperscaler Custom Silicon: The lab maintains deep infrastructure ties with Amazon Web Services, training and serving workloads across AWS Trainium and Inferentia clusters, alongside custom TPU deployments via Google Cloud and Broadcom.
- Alternative Accelerator Deployments: Anthropic secured an agreement with Advanced Micro Devices (AMD) to acquire up to 2 gigawatts of Instinct MI450 server infrastructure starting in 2027, backed by up to $5 billion in milestone-based investment from AMD.
- Third-Party Cloud Deals: The lab finalized a $45 billion cloud compute commitment with UK infrastructure provider Nscale for 460 megawatts of Nvidia Vera Rubin compute capacity.
Shifting Economics of Frontier Model Serving
The pursuit of custom silicon by frontier labs mirrors moves by rivals OpenAI and Meta, both of which have assembled dedicated internal hardware groups alongside foundry partnerships with TSMC and Broadcom. As frontier model deployment transitions from bursty experimental chat interfaces to sustained, continuous agent execution workloads, serving economics become heavily dependent on memory bandwidth and FLOP utilization efficiency.
While Anthropic and MatX did not complete the $7 billion transaction, the reported transition toward a commercial partnership indicates Anthropic remains focused on diversifying its hardware foundation beyond standard merchant silicon.
