A 3-Person AI Startup Is Paying $350K for Kernel Work

AI infrastructure jobs are where the money in AI has quietly moved, and here's the receipt: a 3-person startup just listed a role paying up to $350K plus 0.40% equity. If you're reading this from Sri Lanka, you can't apply. The listing says "US citizen/visa only."

That's exactly why it's worth reading. A job ad is a company publishing, in public, the shortest possible statement of what it thinks is scarce. Morph's tells you something specific about where the money in AI has moved — and most of it is not where the bootcamps are pointing.

---

📊 What the listing actually says

The role is Member of Technical Staff at Morph, tagged S23 on Y Combinator's jobs board. Here's the original posting. I'm commenting on it, not republishing it — go read the source for the full text.

Field	What the listing says
Role	Member of Technical Staff
Compensation	$175K – $350K
Equity	0.40%
Experience	6+ years
Location	San Francisco, CA
Eligibility	US citizen/visa only
Team size	3
Founder	Tejas Bhakta
Interview	2-day work trial

The company description is one sentence: Morph "builds specialized code-generation models and serves them on a custom inference stack." The work named includes autoresearch for kernels and custom speculative-decoding models. The stated success metrics are tokens per second, tokens per dollar, and correctness.

Three people. One of them is being hired at up to $350K. That ratio is the whole story.

---

⚡ The scarce skill isn't "AI" — it's making GPUs cheaper

Notice what is not in that listing. No prompt engineering. No RAG pipeline. No agent framework. No fine-tuning a model on a customer's PDFs.

What's there instead:

1. Kernels — writing the low-level GPU code that actually executes matrix operations, rather than calling someone else's.
2. Serving, routing, autoscaling, capacity optimization — the unglamorous plumbing that decides whether a GPU sits idle at 30% utilisation or runs hot.
3. Speculative decoding — using a small cheap model to guess ahead, and a big expensive model to verify the guesses in one pass.

All three point at the same target: cost per token. When you serve a model, the model itself is a fixed asset and the GPU hours are the variable one. A team that cuts inference cost 40% has done something no amount of clever prompting achieves.

If you want to feel that maths rather than read about it, our speculative decoding speedup calculator implements the standard improvement-factor formula. Put in an acceptance rate and a draft-to-target cost ratio and watch how fast the speedup collapses when the draft model guesses badly. Ten minutes with it will teach you more about why this role exists than a week of reading job ads.

Key takeaway: The premium is on the layer below the model, not above it. Everyone is building on top of LLMs. Very few people can make them run cheaper, and those people are being paid like it.

---

🚫 "US citizen/visa only" is a fact, not an insult

I want to be direct about this, because a lot of career advice for Sri Lankan engineers is dishonest by omission.

You cannot apply to this role. Not "it'll be hard." The listing states the eligibility rule, requires San Francisco, and runs a 2-day work trial. A 3-person company has no immigration lawyer and no relocation budget. That's not gatekeeping, it's arithmetic.

What follows from that:

• Don't build a career plan around US-onsite seed-stage startups. The failure rate on that path from here is near total.
• Do build a plan around companies that already hire distributed. Contractor-first infra companies, open-source foundations, and Series-B-and-later firms with an existing EOR setup are where the door is actually open.
• Do treat the skill list as portable. The eligibility rule is local to this job. The skills are not — kernel work and inference optimisation are priced high everywhere, including at companies that will hire you as a contractor.

---

🛠️ How to build this skill set on a Sri Lankan budget

The obvious objection: this work needs GPUs, and GPUs cost money I don't have. Partly true, but less true than it looks. Here's what's genuinely reachable without a hardware budget.

What you want to learn	Free/cheap way in	What it costs
CUDA / Triton kernel basics	Colab or Kaggle free GPU sessions	Free, time-limited sessions
Reading real serving code	vLLM and llama.cpp source on GitHub	Free
Quantisation and its accuracy cost	Run a small model locally on CPU, measure output quality yourself	Free
Batching and throughput intuition	Benchmark a small model at batch size 1 vs 16 and plot it	Free
Speculative decoding theory	Read the Leviathan et al. paper, then model it numerically	Free
Sustained GPU time for real experiments	Hourly rental from a spot-instance provider	Cheapest line item you'll actually pay for

A concrete first project, in order:

1. Pick one small open model you can run end to end.
2. Measure its tokens per second at batch size 1. Write the number down.
3. Change exactly one thing — batch size, quantisation level, attention implementation — and measure again.
4. Publish the numbers, the method, and the thing that surprised you.

That fourth step is the one people skip, and it's the one that gets you hired. A public benchmark with an honest methodology section is worth more than a certificate, because it demonstrates the exact instinct this job is screening for: you measured, you didn't guess.

A GitHub repo titled "I made this model 2.1× faster and here's the profiler output" is a stronger application than any CV I've seen.

---

What this means for you

Read job ads from companies you can't join. They're the cleanest market signal available, they're free, and nobody is trying to sell you anything in them.

The specific reads from this one:

• Infrastructure skills are outpricing application skills right now. If you're choosing what to go deep on, the low-level side has less competition and higher ceilings.
• Small teams pay top-of-band for narrow depth. A 3-person company doesn't hire generalists at $350K. It hires the one skill it can't fake.
• Geography still binds hard at seed stage. Aim your applications at companies structured to hire remotely, and treat SF-onsite listings as free curriculum rather than opportunities.
• Measurement is the differentiator. Tokens per second, tokens per dollar, correctness. Every one of those is a number. Get comfortable producing numbers.

You can't take this job. You can take the reading list it implies, and that part is free.

---

Commentary on a public Y Combinator job listing. All figures quoted are from the listing itself as of 2 August 2026; roles and terms change, so check the source before relying on any of it.