Problem

Mako, also presented in newer materials as Makora, addresses the difficulty and expense of making AI workloads run efficiently on GPUs. GPU kernels—the low-level code that performs the computation—are notoriously difficult to write and optimize, requiring scarce specialist engineering talent. When kernels are inefficient, AI operators waste expensive accelerator capacity, increase inference costs, and sacrifice speed and reliability. The central use case is optimizing kernels for large-scale model inference, particularly for novel model architectures, quantization schemes, or hardware where existing libraries do not deliver good performance.

The economics are substantial because better kernel performance can increase throughput while reducing the amount of GPU time required per request. M13 reports benchmark results claiming a 49% performance improvement and 70% cost reduction for Mistral-7B, and an 85% performance improvement with a 44% cost reduction for Qwen2-72B. These figures are presented as company benchmarks rather than independently verified results, but they illustrate the value proposition: turning low-level optimization work into direct infrastructure savings.

Product / Service

MakoGenerate, spelled MakoraGenerate in the company’s product materials, is an LLM-powered agent that autonomously generates, compiles, validates, and benchmarks GPU kernels. It currently supports CUDA and Triton for NVIDIA GPUs ranging from Hopper to Blackwell. The system combines LLM code generation with an automated feedback loop and evolutionary search: it creates candidate kernels, runs functional tests, collects compiler diagnostics and performance measurements, and feeds that information back into subsequent iterations while exploring choices such as thread-block sizes, memory access patterns, and asynchronous data movement.

The delivery model is a software platform rather than a consulting-only optimization service. The product has been offered as a free research preview and playground, with early access for its more advanced evolutionary-search capabilities; the company’s site also points to inference endpoints that use agents to optimize kernels, inference engines, algorithms, and orchestration. The intended benefit is an abstraction layer in which developers can write or deploy a workload once and obtain hardware-specific implementations without repeatedly hand-writing and tuning kernels across NVIDIA, AMD, or custom accelerators.

Market

Mako competes in AI infrastructure, specifically AI-native compilers, GPU kernel generation, inference optimization, and hardware-agnostic performance middleware. Its stated ambition is to sit between models and hardware, selecting and optimizing the combination of kernel, library, and accelerator for a workload. The closest named comparison in the research is OctoML/OctoAI, an earlier self-optimizing model-deployment platform later acquired by NVIDIA. Luminal is another adjacent competitor: it describes itself as an open-source ML compiler that generates CUDA kernels and simplifies production deployment. Vendor-native stacks such as NVIDIA’s TensorRT-LLM are also adjacent substitutes for teams that prefer a hardware-specific optimization stack, although they are not identical products.

Mako has meaningful financing and technical validation signals but limited publicly disclosed commercial traction in the evidence reviewed. Cornell reports an $8.5 million seed round led by M13, with participation from Neo, Flybridge, and others, while M13 also identifies angel participation including Jeff Dean. The company has published performance benchmarks and made its generator available in research preview, but the available evidence does not provide a customer count, recurring revenue figure, or confirmed production-scale revenue. It is therefore best characterized as an early-stage, funded commercialization effort with a live product and benchmark-led validation, rather than a company with publicly demonstrated scaled revenue.