Daniel Kim on why GPU orchestration is the boring story of the decade

Daniel Kim doesn't enjoy interviews. He's polite about it, but the discomfort is obvious — a man who would clearly rather be in a code review than a conversation about himself.

We're in Cortex Systems' Seoul headquarters, which looks more like a research lab than a Series B fintech-equivalent unicorn. Whiteboards, half-empty coffee cups, exactly one ping-pong table that nobody appears to use.

Cortex builds GPU orchestration software. That description, Kim acknowledges, is so boring that it has been a recurring obstacle to his fundraising.

"Every pitch meeting I do, I have to spend the first fifteen minutes explaining why what we do is interesting. The answer is: GPU orchestration is the bottleneck for every enterprise that's actually serious about AI in production. The companies that don't believe me are the ones still trying to train models on whatever GPUs they can find."

What Cortex does

The product is conceptually simple. Cortex sits between an enterprise's AI workloads (training jobs, inference services, batch processing) and its compute resources (a mix of in-house GPUs, hyperscaler instances, and increasingly, specialized AI infrastructure providers). The software dynamically allocates workloads across the available infrastructure based on cost, latency requirements, and capacity constraints.

The engineering is harder than it sounds. The state-of-the-art schedulers built inside the hyperscalers — which is where Kim worked before founding the company — represent decades of accumulated complexity. Cortex's bet is that it can deliver 70-80% of that capability as a third-party product, available to companies that don't operate at hyperscale themselves.

The customers betting on that, based on the names Kim was willing to share, are increasingly the F500. Two of the world's three largest banks. A US auto manufacturer running predictive maintenance at scale. A pharmaceutical company doing real-time drug-discovery workloads.

On the $165M Series B

The round, closed quietly in March, was led by an APAC-focused growth fund with participation from two US firms. Kim is unusually open about why the company raised when it did.

"Enterprise AI spend is going to compound at rates the market doesn't yet appreciate. The window for us to build out the GTM capacity to serve that demand is short. We raised because we'd rather be over-capitalized for the demand wave than under-capitalized."

He's also explicit about what the capital won't be used for. Cortex will not, he said, be expanding into adjacent product categories — no model serving, no model registry, no AI gateway. The bet is on becoming the definitive orchestration layer.

The competitive picture

Cortex isn't alone in this space. Kim names two competitors he respects (he won't name the ones he doesn't) and is matter-of-fact about the dynamic. "There will be one to three winners in this category globally. We intend to be one of them. The market is large enough to support multiple winners, and frankly the category needs them — a single winner would be a regulatory problem."

For more on the AI infrastructure layer and our broader technology coverage, see those sections.