Key Takeaways
- • Luminal raised a $25 million Series A led by Standard Capital, announced November 12, 2026, to expand its AI inference compiler and hosted inference business across Nvidia, AMD and specialized chips.
- • Luminal reports 36,000 tokens per second for GPT-OSS 120B on eight Nvidia H100 SXM GPUs, versus 28,000 for TensorRT-LLM and 26,000 for vLLM, according to its website comparison.
- • Luminal offers value-based compiler licensing: in Fioti's illustrative example, a fixed fraction of measured throughput gains is charged, so a customer doubling throughput on a $1.50 GPU saves 75 cents and pays 25 cents.
- • The round follows a $5.3 million seed led by Felicis Ventures, with angel investors including Paul Graham and Guillermo Rauch.
- • Named production customers include real-estate media company AutoHDR (about 5 million images a month) and Freebuff (trillions of tokens weekly).
Summary
Luminal has raised a $25 million Series A led by Standard Capital, according to Unite.AI, which reported the round on November 12, 2026. The AI infrastructure startup is building two businesses around inference, the computation that produces an AI model's output: a managed hosted service on its own cloud, and a compiler it licenses to customers that already operate their own hardware. Co-founders Joe Fioti, Jake Stevens and Matthew Gunton lead the company.
The core technology is a compiler that translates models written in frameworks such as PyTorch into low-level code for GPUs and application-specific integrated circuits. According to Luminal's technical overview, the pipeline applies hardware-aware fusion, tiling, memory planning and scheduling. Rather than relying solely on handwritten heuristics, the company spends substantial computation at compile time searching possible implementations, using equality saturation to represent search spaces that can contain trillions of options on a given chip. A January 9, 2026 technical post describes automatically generated megakernels that fuse a model's forward pass into a single kernel, reducing repeated kernel-launch delays. Luminal reports 36,000 tokens per second for GPT-OSS 120B on eight Nvidia H100 SXM GPUs, compared with 28,000 for TensorRT-LLM and 26,000 for vLLM in its own website comparison.
Luminal Cloud offers managed serverless endpoints, automatic batching and scale-to-zero capacity, with customers paying for usage. Licensed deployments can run in a customer's cloud or on-premises. For some compiler licenses, Fioti described pricing as a fixed fraction of the measured improvement over a customer's existing baseline, tying license revenue to delivered value rather than hosted usage alone. Production customers include AutoHDR, which processes about 5 million images a month with Luminal handling almost all of that workload while shifting a growing share to AMD chips, according to the founders, and Freebuff, which sends Luminal trillions of tokens each week. The founders said Freebuff's coding requests can contain around 100,000 input tokens, making cache-hit rates a significant cost factor.
Luminal began as a side project Fioti built to run models on a laptop. It later added an Nvidia backend and an open-source community. The company keeps its core compiler open while charging for hosting, a model the founders compared with Vercel's pairing of Next.js and its cloud. The Series A follows a $5.3 million seed round announced November 17, 2025, led by Felicis Ventures, with angel investors including Paul Graham and Guillermo Rauch.
The company is hiring compiler engineers and its first non-engineering employee, a head of growth, and has so far been engineering-only, working in person in downtown San Francisco. The founders expect inference workloads to split increasingly between chips specialized for prompt processing and those specialized for token generation, a coordination challenge they see as growing for software. The key test ahead is whether Luminal can sustain its reported cost advantages as inference spreads across specialized hardware.
The core technology is a compiler that translates models written in frameworks such as PyTorch into low-level code for GPUs and application-specific integrated circuits. According to Luminal's technical overview, the pipeline applies hardware-aware fusion, tiling, memory planning and scheduling. Rather than relying solely on handwritten heuristics, the company spends substantial computation at compile time searching possible implementations, using equality saturation to represent search spaces that can contain trillions of options on a given chip. A January 9, 2026 technical post describes automatically generated megakernels that fuse a model's forward pass into a single kernel, reducing repeated kernel-launch delays. Luminal reports 36,000 tokens per second for GPT-OSS 120B on eight Nvidia H100 SXM GPUs, compared with 28,000 for TensorRT-LLM and 26,000 for vLLM in its own website comparison.
Luminal Cloud offers managed serverless endpoints, automatic batching and scale-to-zero capacity, with customers paying for usage. Licensed deployments can run in a customer's cloud or on-premises. For some compiler licenses, Fioti described pricing as a fixed fraction of the measured improvement over a customer's existing baseline, tying license revenue to delivered value rather than hosted usage alone. Production customers include AutoHDR, which processes about 5 million images a month with Luminal handling almost all of that workload while shifting a growing share to AMD chips, according to the founders, and Freebuff, which sends Luminal trillions of tokens each week. The founders said Freebuff's coding requests can contain around 100,000 input tokens, making cache-hit rates a significant cost factor.
Luminal began as a side project Fioti built to run models on a laptop. It later added an Nvidia backend and an open-source community. The company keeps its core compiler open while charging for hosting, a model the founders compared with Vercel's pairing of Next.js and its cloud. The Series A follows a $5.3 million seed round announced November 17, 2025, led by Felicis Ventures, with angel investors including Paul Graham and Guillermo Rauch.
The company is hiring compiler engineers and its first non-engineering employee, a head of growth, and has so far been engineering-only, working in person in downtown San Francisco. The founders expect inference workloads to split increasingly between chips specialized for prompt processing and those specialized for token generation, a coordination challenge they see as growing for software. The key test ahead is whether Luminal can sustain its reported cost advantages as inference spreads across specialized hardware.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.