Lola Vision Systems Wants to Cut the Cost of Running AI on Any Chip
Lola Vision Systems, a Washington, D.C.-based AI infrastructure company founded in 2024 by Tayo Adesanya, is building software and chips for running AI models on devices. TechCrunch
Its core product is software described as a "compiler toolchain." It translates an AI model and client code into instructions a specific chip can run. Think of it as an automated translator between models and hardware. The pitch is portability without manual setup work. Adesanya said setting up an AI model on new hardware by hand can take roughly 200 hours just to begin testing.
Edge deployment teams will recognize the pain. The target ISA, the instruction set a chip understands, changes. Memory budgets change. Developers fall back to reference kernels, generic and slow backup code, and schedules slip. Lola is betting that automating the model-to-machine-code path lowers that cost enough to matter when customers choose accelerators.
Adesanya came to the problem from the selection side. He was a student at Purdue University and previously worked with microchips and AI processors helping large manufacturers decide which chips to use in hardware. That work involves evaluating performance, software maturity, and integration effort across vendors, the same friction Lola now claims to reduce.
The company is also developing its own semiconductor chips. It says a dozen corporate customers have signed letters expressing interest in buying its chips once available. It has one signed customer. It plans to license its software for use on existing hardware to generate revenue sooner, a sequencing decision that keeps software cash flow ahead of silicon lead times.
On labs and ecosystem, Lola has partnered with SCALE, a microelectronics workforce development program, to work with more semiconductor labs. It has raised just over $1 million in total funding to date. It was selected for the TechCrunch Battlefield 200. Its own site, lolavisionsystems.com, describes the company as a full-stack edge AI platform with a focus on Edge AI for Mission-Critical Applications. The site lists AI model training, optimization and edge deployment among its offerings, along with multi-modal sensor fusion among its capabilities.
The broader context here is the split in edge AI between fast-moving models and fragmented hardware. Foundation models and vision models iterate quickly. Edge silicon does not. Each new accelerator typically needs graph lowering, conversion of the model graph for hardware, plus operator coverage, quantization handling to shrink numerical precision, and careful memory scheduling before throughput and latency stabilize. A toolchain that handles that reliably across targets would shorten evaluation cycles and reduce lock-in to a single vendor stack.
In my view, the licensing plan is the part to watch. Building compilers for existing parts forces compatibility with real customer models, toolchains, and runtimes. It also creates a tighter feedback loop than waiting for first silicon. The risk is scope. Supporting N chips times M model families is an ongoing maintenance burden, especially with just over $1 million raised. Focus on a narrow set of workloads and a small set of ISAs would matter more than breadth at this stage.
Worth flagging for practitioners is what success would enable if the approach holds. Faster retargeting means hardware teams can test more parts against production models, and model teams can ship to deployed devices without rewriting for each backend. That does not remove verification, safety qualification, or system integration for mission-critical use. It could make those later steps start earlier, which is often where edge programs gain or lose quarters.
