I’ve been thinking about money. Specifically, the kind of money where your total comp has a comma after the first digit and six more digits after that. $2M/year. The number that separates “doing well in tech” from “building generational wealth in a single job.”
This isn’t abstract. I’m a mid-20s engineer at Microsoft AI, working on Bazel migration, CI/CD, Docker, Kubernetes GPU clusters. Infrastructure and developer experience. I’m good at it. But “good at infra at a big tech company” tops out somewhere around $500-700K if you grind to Staff+. The $2M ceiling exists at exactly one type of employer right now: frontier AI labs.
So I mapped out the three archetypes that actually reach $2M+ in AI. And I’m trying to figure out which bet to make.
Who actually makes $2M?
Let’s be specific. $2M total comp means:
- Staff+ Research Scientist at OpenAI, Anthropic, DeepMind, xAI — base ~$300-400K, equity doing the rest
- Senior Staff/Principal ML Engineer at the same places — the people who bridge research and production
- Infra leads running training clusters at scale — the ones keeping $100M+ training runs alive
The comp is equity-heavy. Like, 60-70% equity. Which means it’s tied to company trajectory. An OpenAI Staff researcher’s $2M package is $2M because OpenAI’s valuation keeps climbing. Same role at a lab that stalls? Maybe $800K.
This is important: you’re not just picking a skill path. You’re picking a company bet. The skill gets you in the door. The equity makes you rich.
Three archetypes can get through that door.
Type 1: The Researcher
The person who conceives the next breakthrough. Novel architectures, new training paradigms, the RLHF innovations, the scaling laws paper. These people get $2M+ because they can think of things others literally cannot.
What you need:
- PhD, almost certainly. Or equivalent depth — but “equivalent” means you independently produced novel research that got published at top venues. The self-taught research path exists the way climbing Everest without oxygen exists: documented but not recommended.
- Publication record. Not just “I have papers” — papers that other researchers cite, that changed how people think about a subfield.
- A specific research identity. “I do ML” won’t cut it. “I’m the person who figured out [specific technique] for [specific problem]” will.
What to study:
Deep learning theory, optimization (not just “use Adam”), whatever subfield you’d bet your career on — reasoning, multimodal, alignment, efficient training, interpretability. Pick one. Go deep enough that you can explain what’s wrong with every paper in that area from the last two years.
Timeline: 4-6 years if starting a PhD now. Maybe 3 if you’re already in a research-adjacent role and have extraordinary output.
Honest take for someone like me: This is the hardest path to enter from where I am. I don’t have a PhD. I don’t have a publication record. Starting a PhD at 25-26 means I’m 30-31 when I finish, competing against people who’ve been doing research since undergrad. Not impossible. But the opportunity cost is brutal — that’s 5 years of lost industry comp, lost equity, lost career momentum in the fastest-moving field in history.
The researcher path pays $2M because of a genuine ideas moat. You can’t grind your way to novel insights. Some people just see patterns others don’t, and most of that ability is cultivated through years of deep, focused research training. If you have that spark and you know it — this is the path. If you’re not sure? That uncertainty is probably your answer.
Type 2: The Applied ML / AI Engineer
The person who takes a paper and ships it. Understands transformers well enough to modify them, knows training loops well enough to debug them, builds evals well enough to trust them. The bridge between “this is theoretically possible” and “this is running in production.”
What you need:
- Solid ML fundamentals — not research-depth, but deep enough to read a paper, understand the key insight, and implement it without hand-holding
- Exceptional engineering — the research is table stakes, but the reason this role pays $2M is that most researchers can’t ship and most engineers can’t research. You do both.
- Taste — knowing which research matters, which shortcuts are safe, which corners you absolutely cannot cut
What to study:
Transformer internals (attention mechanisms, positional encoding, KV cache). Training pipelines end to end (data loading, distributed training, loss curves, checkpointing). RLHF/DPO/GRPO — the alignment techniques that turn a base model into a product. Eval frameworks — how do you know if your model is actually better? Inference optimization — quantization, speculative decoding, batching strategies.
Timeline: 2-3 years of focused upskilling from a strong SWE base. You’re not starting from zero — you already know how to build systems. You’re adding the ML layer on top.
How to get there:
Ship things. Open source contributions to training frameworks. Fine-tune models and document what you learned. Build an eval system and open-source it. Contribute to vLLM, or Axolotl, or whatever inference/training framework is hot when you read this. The portfolio matters more than credentials here — frontier labs hiring for this role want to see that you’ve done it, not that you studied it.
Honest take: This is the most “learnable” path to $2M. The knowledge is all public — papers, code, tutorials. The barrier isn’t access, it’s discipline and depth. The risk is that you end up as a “good ML engineer” but not exceptional enough for frontier lab comp. The difference between a $400K ML engineer and a $2M one isn’t 5x the skill — it’s being in the top 100 people in the world at bridging research and production. That’s a small club.
Type 3: The AI Infrastructure Engineer
No need for algorithmic brilliance. You build the platform that makes Types 1 and 2 productive. GPU cluster management. Training infrastructure at scale. Serving systems that handle thousands of requests per second on $50K GPUs. The distributed computing that makes a 10,000-GPU training run actually converge.
What you need:
- Distributed systems expertise — not “I read the DDIA book” but “I’ve debugged a NCCL hang at 3am across 512 GPUs”
- GPU/CUDA programming — understanding memory hierarchies, kernel launches, the actual hardware you’re optimizing for
- Kubernetes at serious scale — not “I deployed a pod” but “I manage a fleet of GPU nodes with custom schedulers and preemption”
- Operational experience with large training runs — this is the real moat. You can’t simulate this. You either have war stories or you don’t.
What to study:
NCCL internals and collective communications. Distributed training frameworks — FSDP, DeepSpeed ZeRO, Megatron-LM parallelism strategies. GPU memory optimization (activation checkpointing, mixed precision, memory-efficient attention). Serving systems — vLLM, TensorRT-LLM, Triton Inference Server. Container orchestration for ML workloads. Observability for training — how do you know when a training run is going sideways before you’ve wasted $500K in compute?
Timeline: 1-2 years from where I am right now. I’m already doing Kubernetes, Docker, CI/CD for an AI org. The gap is specificity — I need to go from “general infra” to “AI training infra.”
Honest take: This is my closest path. The moat here is operational — frontier labs will pay $2M for someone who’s kept large-scale training runs alive because that experience is genuinely scarce. You can’t get it from a textbook. You get it from being on-call when a $10M training run hits a NCCL timeout and you have 30 minutes to figure out if it’s a network issue, a GPU failure, or a software bug.
The downside: infra is less “sexy” than research or applied ML. Your work enables $2M breakthroughs, but the narrative credit goes to the researcher. You have to be okay with that. Also — the $2M infra roles are very specifically at frontier labs running massive training. Not every infra job qualifies. You need to be at a company where the training cluster is the product.
My optimal path: Type 2.5
Here’s my actual bet.
Pure Type 3 gets me close, but the hybrid is what’s most in-demand right now. The person who can manage training infrastructure AND understands what the training is actually doing. Who can debug a distributed training hang AND tell you whether the loss curve looks healthy. Who builds the platform AND can have an intelligent conversation with researchers about what the platform needs to support next.
I’m calling this Type 2.5 — infra-first, ML-literate. And I think it’s arguably the highest-leverage profile at frontier labs right now because:
- Pure infra people don’t understand the workload well enough to make good platform decisions
- Pure ML people build brittle infrastructure because they don’t know distributed systems
- The person who does both makes better decisions at every layer
My 2026 study plan:
Q1 (Jan-Mar) — Strengthen the foundation I already have:
- CUDA programming fundamentals — write kernels, understand memory coalescing, profile with nsight
- NCCL internals — read the source, understand ring allreduce, tree allreduce, when each is optimal
- Hands-on with DeepSpeed ZeRO stages — not just “use it” but understand the memory math
Q2 (Apr-Jun) — Add the ML layer:
- Transformer architecture deep dive — implement one from scratch, understand every component
- Training loop internals — learning rate schedules, gradient accumulation, loss scaling, checkpointing
- Fine-tune a model end to end — LoRA, full fine-tuning, understand when each makes sense
Q3 (Jul-Sep) — Bridge the gap:
- RLHF/DPO pipeline implementation — the alignment stack that turns base models into products
- Eval frameworks — build or contribute to one. Understand what “better” means for an LLM.
- Inference optimization — quantization, KV cache optimization, batching strategies
Q4 (Oct-Dec) — Ship and position:
- Open source contribution to a training or serving framework (vLLM, Megatron, etc.)
- Build something visible — a tool, a benchmark, a write-up that gets attention
- Start interviewing. The market for this profile is hottest at year-end when labs are planning next year’s training runs.
The meta-game
Skills get you in the door. But $2M comp is about leverage and scarcity, not just skill.
Company selection matters more than anything. $2M at a frontier lab is realistic. $2M at Microsoft or Google for the same work? Much harder — the equity upside is smaller, the levels are more bureaucratic, the comp bands are tighter. You need to be at a company where AI is the core bet, not a division.
Timing matters. Pre-IPO equity at a frontier lab that 10x’s is how $2M packages become $5M packages. OpenAI, Anthropic, xAI — these companies are in the growth phase where early employees’ equity is still appreciating rapidly. Joining post-IPO, when equity is liquid but stable, caps your upside.
Network into the pipeline. Frontier labs don’t hire through job boards. They hire through referrals, through people they’ve seen ship impressive work, through open source contributions they’ve noticed. The meta-game is being visible in the right circles. Contribute to the frameworks these labs use. Write about problems they care about. Build relationships with people already inside.
Open source is your resume. A meaningful contribution to vLLM or DeepSpeed or Megatron-LM is worth more than any credential. It’s proof you can do the work, visible to exactly the people who hire for these roles.
The honest ending
I don’t know if I’ll get to $2M. Putting a number on a career goal feels crass until you do the math on what that money actually means — paying off your parents’ house, never worrying about healthcare, having the freedom to take risks later. It’s not about the number. It’s about what the number buys.
What I do know: the path from “AI infra at big tech” to “AI infra lead at frontier lab” is the shortest line I can draw. And adding ML depth on top makes that line thicker — harder to miss, more valuable when it lands.
The bet for 2026 is clear. Strengthen the infra core, add the ML layer, ship something visible, and position for the move. Not “follow your passion” — follow the scarcity. Right now, the scarcest person in AI is the one who understands both the infrastructure and the intelligence running on it.
That’s what I’m building toward.