DensityAI
Novel memory architecture for frontier LLM inference · DensityAI
DensityAI builds AI accelerators around a novel memory architecture combining near-memory compute topology of SRAM with HBM density, specifically for frontier-scale LLM inference. They explicitly target beating GPUs on energy and speed for large-model, long-context workloads.
Why: Extremely close match: novel memory architecture for frontier LLM inference only, near-memory compute topology. Nearly identical technical bet and wedge to Fractile.
Etched
Transformer-specific ASIC for frontier inference · Etched
Etched builds specialized chips exclusively for transformer inference at frontier scale, co-designing chips, racks, software, and manufacturing for best-in-class throughput. Like Fractile, they target the same wedge (frontier model inference bottleneck) with hardware purpose-built for transformers only, not general compute or training.
Why: Nearly identical wedge and workload specificity (transformer-only inference). Full-stack approach from chip to cluster. Main difference: less explicit about memory-compute interleaving architecture, more emphasis on co-designed clusters.
d-Matrix
In-memory compute for low-latency AI inference · d-Matrix Corporation
d-Matrix builds digital in-memory compute architecture with integrated performance memory for ultra-low latency batched inference. Their Corsair platform explicitly addresses memory-compute proximity similar to Fractile's approach, targeting datacenter inference deployment at scale.
Why: Very similar technical bet on memory-centric compute architecture. Targets generative AI inference specifically. Strong match on low-latency focus and datacenter deployment model.
Positron AI
Purpose-built transformer inference hardware · Positron AI, Inc.
Positron builds purpose-built hardware specifically for transformer model inference at any scale, emphasizing highest performance and lowest power for generative AI. They target the same frontier model inference bottleneck with production-ready hardware shipping now.
Why: Identical workload specificity (transformer inference only) and wedge. Less explicit about memory architecture innovation but clear focus on performance/power tradeoffs for generative AI.
Archality
Memory wall solution for AI inference · The Archality Corporation
Archality addresses the memory wall problem in AI inference systems, focusing on eliminating wasted power and capital from constantly moving model data. Very early stage (3 employees, founded 2025) but targeting the same memory-compute bottleneck.
Why: Similar technical bet on solving memory wall for AI inference. Very early stage with less clarity on full-stack approach or specific workload focus, but clear memory-compute architecture angle.
Phynomy
Memory wall elimination for AI inference · Phynomy
Phynomy targets the memory wall problem, claiming GPUs waste 90% of energy moving data between memory and compute. Building adaptive, energy-efficient AI inference hardware. Very early stage (2 employees) but focused on same memory-compute bottleneck.
Why: Clear focus on memory-compute bottleneck for inference. Very early stage with limited detail on architecture or workload specificity, but aligned on core technical bet.
MatX
High-throughput chips for frontier LLMs · MatX
MatX builds chips for large model needs of frontier labs, delivering higher throughput than announced products while matching best latencies. Targets both training and inference (prefill, decode, RL), excelling on FLOPS and latency for frontier-scale workloads.
Why: Targets frontier model scale but includes training, not inference-only. Less specific about memory architecture innovation. Strong on performance tradeoffs and market timing.
Lamb Labs
Custom inference chips with parallel decoding · Lamb Labs
Lamb Labs builds custom chips for AI inference, post-training models into diffusion architectures that decode in parallel, targeting 20,000+ tokens per second. Different architectural approach (parallel decoding vs. memory-compute interleaving) but similar inference-only focus.
Why: Inference-only focus with custom silicon, but fundamentally different technical bet (parallel decoding via diffusion architecture vs. memory-compute interleaving). Less emphasis on frontier model scale.
