Kog Wants a Monokernel Writer. LinkedIn Returns 6 Profiles Worldwide.
Kog's 3,000 tok/s post lit up HN's July 2026 thread. The CUDA/PTX pool is tiny, invisible on LinkedIn, and about to get expensive. Here is where to actually find it.
Kog AI's July 4 post on the Hacker News "Who is Hiring" thread reads like a benchmark flex: 3,000 tokens/s per request on 8x AMD MI300X, 2,100 on 8x NVIDIA H200, batch size 1, FP16, no speculative decoding. The whole decode runs as a single persistent GPU kernel written in CUDA with PTX inline assembly on NVIDIA and HIP with CDNA ISA on AMD. It is also a warning to every recruiter and founder trying to hire the same person this quarter: the pool of engineers who can actually ship that is smaller than your candidate spreadsheet.
What Kog actually asked for
Read the post carefully. Kog is hiring a GPU Engineer, remote in a Europe-compatible timezone with one week per month in Paris, to "own low-level kernel work in CUDA/PTX or HIP/CDNA ISA, the monokernel pipeline, profiling infrastructure inside it, scaling to the frontier MoE models that run in production."
The word that matters is "or." Kog is willing to take either an NVIDIA-side PTX assembly writer or an AMD-side CDNA ISA assembly writer. That "or" is not generous. It is a tell. They know the AND is essentially unfillable at any comp band. The people who have shipped both, at production quality, on a persistent monokernel, are countable on two hands.
For context on why the low-level work is unavoidable, Kog published their own math: a kernel launch and cleanup costs about 4.5 microseconds on MI300X. Ten kernels per Transformer layer across twenty-five layers is 1,125 microseconds of overhead per token before any useful work, capping achievable speed at roughly 890 tokens/s. To break past that, you fuse the entire decode into one persistent kernel and hand-write the memory dance in assembly. There is no framework shortcut.
Kog's technical writeup is explicit about what they threw out to get there: "our critical decoding path does not rely on third-party frameworks, libraries, and abstractions (like PyTorch, Triton, CUTLASS, NCCL, ROCm CK, AITER, or RCCL)." No PyTorch. No Triton. No CUTLASS. The stack every "senior GPU engineer" on LinkedIn actually knows was deleted from the hot path.
That is the profile every inference startup now wants. Together AI just raised $800M. AMD co-signed Kog's result on its own engineering blog. NVIDIA has its own posts up. At least a dozen July 2026 HN listings are hunting the same shape of person.
The pool is not "thousands of CUDA engineers"
Here is the empirical anchor. Query a major professional-network index for the actual skill combination Kog is asking for: CUDA AND PTX AND GPU kernel inference context. You get six profiles worldwide. Six.
Not six hundred. Not six thousand. Six.
number: 6
label: LinkedIn profiles worldwide matching CUDA + PTX + kernel inference
note: Refolk index query. Titles include "Software Engineer, LLM Systems, CUDA Kernels and GPU Performance" at AMD.