Luminal’s $25M Series A: A Search Compiler for Multi-Chip Inference

Author

AI News Editorial

Published

2026-11-15 08:45

Axios Pro broke the story on 12 November: Luminal, a San Francisco inference startup, raised a $25 million Series A led by Standard Capital, with seed backers Y Combinator, Felicis, SagaVC and Liquid doubling down. Valuation is undisclosed; combined with the $5.3M seed from November 2025, the company says it has raised roughly $30.3M. Luminal was co-founded by CEO Joe Fioti, a former Intel chip designer, with Matthew Gunton (Amazon) and Jake Stevens (Apple), and came out of Y Combinator’s Summer 2025 batch.

The idea: compile by search, not by hand. Luminal calls itself a search compiler and inference cloud. The argument is that serving engines such as vLLM and TensorRT-LLM still interpret or hand-tune graphs per model-and-chip pair, leaving silicon on the table. Luminal instead explores a large space of mathematically equivalent implementations — the technique known as equality saturation — and picks the fastest one for the target accelerator. The pipeline runs from a PyTorch or Hugging Face import, down to a framework-free dataflow-graph IR, through hardware-aware passes (fusion, tiling, memory planning, scheduling), and emits native kernels for GPUs and ASICs alike.

The numbers are vendor numbers. Luminal’s own chart puts GPT-OSS 120B on eight H100 SXM at 36,000 tokens/s, against 28,000 for TensorRT-LLM, 26,000 for vLLM and 3,000 for PyTorch eager, plus claimed 3.2x over vLLM, sub-10ms p99 latency and zero runtime overhead. None of it is independently reproduced. Token throughput is also inconsistently reported: one outlet says the public cloud reached about 45 trillion tokens per month, another says 90 trillion, within four months of launch. Treat the number as a marketing range, not a measurement.

The strategy is chip diversification. Luminal says a significant share of its workloads runs on Amazon Trainium through a growing partnership, and the round funds broader compatibility across Nvidia, AMD and specialised chips. Fioti’s framing is software-led TCO: extract more tokens from chips you already own. The competitive field is heating up — Standard Kernel raised $20M in March 2026, SGLang spinoff RadixArk was valued near $400M in January, and Together AI raised $800M.

One caveat to price in: compile-first serving taxes your workflow. The first deploy of a model-plus-chip pair is slow, and any weight or architecture change triggers recompilation. For high-QPS agent traffic the amortisation works; for teams that fine-tune nightly, measure how often recompiles block rollouts.

Sources: Axios Pro · Tech Funding News · Unite.AI