Together AI's $800M Series C Is a Hiring Round. 500 People Are the Target.
Together AI raised $800M led by Aramco Ventures on July 1. The scarce input is not GPUs. It is kernel engineers. Here is how to source ahead of them.
On July 1, 2026, Together AI closed an $800M Series C at an $8.3B post-money valuation, led by Aramco Ventures. If you read that as a compute story, you missed it. The 500 MW of capacity Together needs was capitalized separately by new investors. The $800M is payroll and R&D. That means Aramco money is pointed directly at the roughly two thousand people on earth who can write production CUDA/PTX and HIP monokernels, and Together's recruiters are already dialing them.
If you employ any of those people, your retention window is about one quarter. If you're trying to hire them, the pool just got 50x more expensive to reach through normal channels.
What the $800M actually buys
Together's own announcement is unusually honest about this. The equity round is separate from the 500 MW infrastructure commitment. Their public plan is a roughly 50-fold expansion of infrastructure footprint over the next five years. You cannot 50x an inference business by buying GPUs alone. You do it by shipping kernels that turn the same silicon into 10x more tokens per second.
The gap is quantifiable. Typical GPU decoding speed for 2B to 8B models on high-end hardware sits around 100 to 300 tokens per second. Kog, an 11-person Paris startup with $5M raised, is publicly benchmarking 3,000 tokens per second on 8x MI300X at batch size 1 FP16. That order-of-magnitude delta is the entire moat of an inference neocloud, and it is written by a specific kind of engineer.
Together's own job listings confirm the target profile. The Core ML (Turbo) role names "kernel backends, speculative decoding (e.g., ATLAS), quantization" and profiling across "GPU, networking, and memory layers." Their open reqs also include a Research Engineer for Frontier Speculative Decoding and a Systems Research Engineer for GPU Programming. This is not a general ML hiring wave. It is a laser at maybe five hundred people globally.
The pool is smaller than your Boolean thinks
Search "CUDA" on LinkedIn Recruiter and you'll get tens of thousands of hits. Most of them have touched CUDA the way most web developers have touched Kubernetes. The people Together will actually pay for have shipped memory-access patterns tuned to specific chiplet layouts. On the MI300X, latency changes depending on which die makes the request. Kog engineers map around that explicitly. The population that has genuinely done this work in production is measured in low hundreds.
When we filter our own index for CUDA-skilled engineers holding titles like GPU Engineer, Inference Engineer, Kernel Engineer, or Performance Engineer, the employer distribution collapses fast. Apple, Meta, Intel, Anduril, VMware, IMC, and StreamComputing dominate. That is a directional read, not a census, but the shape is right: this is not a LinkedIn-open-to-work population. It's a GitHub-commit-history population.
Which is why keyword sourcing fails here. You cannot type your way to a shortlist of people who have written a warp-specialized GEMM. You have to describe what they do. That's the friction we built Refolk to remove: you ask in plain English for the person you actually want, and it ranks across GitHub, LinkedIn, and the open web in one pass.
The non-obvious feeder pools
Two segments matter more than the obvious FAANG list and almost nobody targets them:
Prop trading. IMC shows up in our index for a reason. HRT, Jane Street, Jump, and Citadel Securities all quietly employ the microsecond-obsessed C++/CUDA profile Together needs. These engineers don't apply anywhere. They get paid enough not to. But they are visible through OSS: CUTLASS PRs, Triton contributions, occasional tinygrad or ggml commits under real names.
Defense. Anduril's presence in the same slice is not an accident. Autonomous systems teams have spent five years hiring exactly the profile that now maps cleanly onto inference kernels: sensor fusion at wire speed, custom CUDA graphs, ONNX and TensorRT deep in the runtime. These people are cleared, well-paid, and largely invisible to standard sourcing.
If you're building a list of names to defend or to poach, GPU MODE Discord, the CUDA subreddit, and the contributor graphs on PyTorch, CUTLASS, vLLM, SGLang, ROCm, and Triton will get you further in an afternoon than a week of InMail.
Together is not the only buyer
The bidding structure is worse than one flush competitor. In June 2026, Groq raised $650M to rebuild as an inference cloud. NVIDIA invested in Together's Series A and this Series C, and separately licensed roughly $20 billion of technology to Groq. NVIDIA is bankrolling multiple neoclouds fighting for the same engineers. That is not accidental. It is ecosystem strategy.
Add Abu Dhabi's MGX, which closed a $49B AI fund on the same day as Together's round (exceeding its $45B target) and has already backed Anthropic, OpenAI, and xAI. Aramco's move is not a one-off. It's the second wave of Gulf sovereign capital pricing this labor market.
The $800M is not for GPUs. It is for the five hundred people who can turn those GPUs into ten times the tokens. </pull>
pull The $800M is not for GPUs. It is for the five hundred people who can turn those GPUs into ten times the tokens.
If you are recruiting for llm inference engineer talent right now, assume every counteroffer conversation you have in the next ninety days is happening against a Together, Groq, or MGX-backed alternative. Price your retention accordingly.
## Read the ATLAS paper for the target list
Together's proprietary inference stack is called ATLAS (AdapTive-LeArning Speculator System). The paper's author list is a public map of the technical center of gravity: Tri Dao, Percy Liang, Ce Zhang, Ben Athiwaratkun, Junxiong Wang, Yineng Zhang, Avner May. If you are defending against Together, these are the people whose networks their recruiters will work first. If you are trying to hire from Together, these are the anchors whose second-degree connections you should be mapping.
This is where cuda kernel engineer recruiting stops looking like traditional sourcing and starts looking like intelligence work. You are not searching for keywords. You are tracing coauthor graphs, commit histories, and Discord handles back to real identities.
What Kog tells you about the ceiling
Kog is the case study nobody is writing up. Eleven people. Five PhDs. $5M raised from Varsity VC and BPI France. Founded in 2023 by Gaël Delalleau, an École Polytechnique engineer whose career spans cybersecurity research and high-performance GPU work. And they are publicly hiring against Together for the same profile.
How? By putting the benchmark in the job post. "3,000 tok/s on 8x MI300X and 2,100 on 8x H200 at batch size 1 FP16" is not marketing copy. It is a resume filter. Ninety-nine percent of applicants self-select out because they cannot parse batch-size-1 memory-bandwidth-bound decode. The one percent who can are exactly the people Kog wants to interview.
The lesson for anyone doing gpu engineer sourcing: your job description is currently doing zero filtering. Rewrite it to name a benchmark, a kernel, or a specific piece of hardware. If your JD could describe both a Kubernetes SRE and a CUDA kernel engineer, it will attract neither.
Delalleau's background is also instructive. He came in through security and HPC, not through an ML PhD. The best kernel engineers rarely do. If your intake pipeline only recognizes ML research pedigree, you are excluding half the pool by construction.
A ninety-day playbook
Here is what to do this quarter, before Together's recruiters finish their first sweep.
Week 1-2: Map the ATLAS graph. Pull the coauthor lists from the ATLAS paper, Tri Dao's Mamba work, and the recent speculative decoding literature. Cross-reference against GitHub commits to vLLM, SGLang, CUTLASS, and Triton in the last twelve months. You'll surface maybe 300 names. Half of them will not be findable on LinkedIn.
Week 3-4: Work the OSS graph, not the job boards. Anyone whose last three commits are to ROCm or CUTLASS is a live candidate. Anyone with a Kaggle profile and a "learning CUDA" repo is not. This is the single largest quality lift in gpu engineer sourcing and it is invisible to LinkedIn Recruiter.
Week 5-8: Approach through the work. These engineers do not respond to "great opportunity" InMail. They respond to specific technical questions about their commits. If you cannot have that conversation, hire an engineer to send the first message. Or use a tool that surfaces the specific commit in the candidate record so your recruiter can reference it. This is exactly the workflow Refolk was built to shorten: you get the person and the reason they matter in the same view.
Week 9-12: Reprice retention on your existing team. If you employ any of the ~500 target profiles, assume Together, Groq, or an MGX-backed competitor will offer them 40% to 80% more in the next two quarters. Sovereign capital does not care about your comp bands.
The structural shift
Together AI's series c round is a signal about how AI infrastructure hiring works from now on. Compute is capitalized separately. Talent is capitalized through equity. And the talent that matters is small enough to name, public enough to trace, and expensive enough that sovereign wealth funds are now the marginal bidder.
The recruiters who win the next twelve months of together ai hiring competition are the ones who stopped thinking in job titles and started thinking in commit histories, benchmark numbers, and coauthor graphs. Everyone else will spend Q3 wondering why their InMail response rates collapsed.
FAQ
How many engineers globally can actually do MI300X kernel work?
Low hundreds. HIP and CDNA ISA fluency is a much smaller pool than CUDA/PTX because AMD's inference market share only recently made the work commercially interesting. Kog's public benchmarking is a good proxy for who is credible: if someone cannot discuss chiplet-aware memory access, they are not in the pool regardless of what their LinkedIn says.
Is Together AI actually going to outbid us, or is this a headline?
Assume yes for the specific niche. Their public reqs name ATLAS, speculative decoding, and kernel backends explicitly, and $800M with 500 MW of compute capitalized elsewhere leaves an enormous payroll runway. For general ML engineers, they are one of many bidders. For CUDA/PTX/HIP kernel writers, they are the price setter for at least the next quarter.
Where do I actually find these people if not LinkedIn?
GitHub contributor graphs on CUTLASS, Triton, vLLM, SGLang, ROCm, tinygrad, and ggml. GPU MODE Discord. The CUDA subreddit. Coauthor lists on recent inference and speculative decoding papers. The HN "Who is hiring?" thread, which is now a functional sourcing tool for this exact profile. Almost none of these people run open-to-work signals on LinkedIn.
What benchmark should our JD name?
Whatever your team actually cares about. Tokens per second per request at a specific batch size and precision on specific hardware is the current standard, because it filters out anyone who has not done real inference optimization. If you cannot write that sentence honestly, get an engineer to write it before you post the role. A JD that could describe five different jobs will attract none of the right people.