Luminal Raises $25M Series A on Claims of 90 Trillion Monthly Tokens
Luminal closed a $25M Series A led by Standard Capital and claims 90 trillion monthly tokens, but one outlet counts 45 trillion and its benchmarks lack independent verification.
By Daniel Okafor
3 min read

What's News
- Luminal raised a $25M Series A on Nov 12, 2026, led by Standard Capital.
- The company claims ~90 trillion tokens per month; SiliconANGLE cites ~45 trillion.
- Seed round was $5.3M in November 2025, led by Felicis, with angels Paul Graham and Guillermo Rauch.
- Published benchmarks show 36,000 tok/s vs vLLM's 26,000 — a 1.38× gap, not the claimed 3.2×.
- A large share of token volume runs on Amazon Trainium chips.
AI inference startup Luminal announced a $25 million Series A on November 12, 2026, led by Standard Capital, and says its inference cloud now processes roughly 90 trillion tokens per month. Seed investors Y Combinator, Felicis, SagaVC and Liquid 2 doubled down in the round. The San Francisco company launched its public inference cloud after a $5.3 million seed round in November 2025, led by Felicis, with angels including Paul Graham and Guillermo Rauch.
The token-volume claim is the strongest number in the story. It is also contested. SiliconANGLE reported roughly 45 trillion tokens per month in its November 12 profile, as noted by explainx.ai, while Technode and Technwz repeat the 90 trillion figure. The discrepancy has not been reconciled, and no third party has verified either number.
How does the compiler actually work?
Luminal's product pairs a search compiler with a hosted inference cloud. The compiler uses equality saturation — a technique that represents many mathematically equivalent implementations of a model and searches for the fastest one on a given chip — to generate optimized kernels ahead of time. Artifacts are cached for repeated inference. The company argues that most serving engines still interpret models at runtime, leaving throughput on the table.
Co-founder and CEO Joe Fioti, formerly at Intel AI hardware, said the team started from the observation that silicon capabilities sit idle because software does not exploit them. The pitch is software-led total cost of ownership: extract more tokens per dollar from chips customers already own, including non-Nvidia accelerators. A large share of Luminal's current token volume runs on Amazon's Trainium chips under what the company describes as a deepening partnership. The compiler also targets GPUs and other ASICs, with the stated ambition of letting customers choose hardware based on workload rather than software ecosystem.
What do the benchmarks show?
Luminal's website publishes a chart comparing tokens per second on GPT-OSS 120B across eight H100 SXM GPUs:
- Luminal: 36,000 tok/s
- TensorRT-LLM: 28,000 tok/s
- vLLM: 26,000 tok/s
The company also claims a 3.2× advantage over vLLM, sub-10ms p99 latency and zero runtime overhead. It offers a licensed deployment option for on-premises or VPC use.
Explainx.ai notes it has not independently reproduced these figures, and the arithmetic is awkward. The published tok/s numbers imply a 1.38× advantage (36,000 divided by 26,000), not 3.2×. It is unclear whether the 3.2× claim refers to a different configuration or model. Critical details are missing from the marketing page:
- Exact checkpoint hash
- GPU driver and power cap
- Batch size, sequence length and concurrency settings
- p50 versus p99 latency breakdowns
No public independent benchmarks exist. Pricing is not listed transparently for either the cloud or licensed tiers.
Why does it matter for buyers?
If Luminal's compiler delivers consistent gains across diverse hardware, it could give procurement teams leverage against Nvidia's CUDA ecosystem — not by replacing Nvidia silicon, but by making alternatives viable without rewriting agent stacks. The category is crowded: Nvidia's own TensorRT-LLM pushes compile-time search, Baseten's Rust engine targets one model family on Blackwell, and the SGLang consortium advances open runtimes.
For now, the 90 trillion token claim and the benchmark charts remain vendor statistics. Buyers should run A/B tests on their own prompts and hardware, with full reproducibility scripts, before changing autoscaling rules or latency SLAs.
Source: epyhia.com
More from Daniel Okafor
Show full bio
Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.
2,208 articles