Together AI's $800M Series C Needs 194 People. LinkedIn Shows 13.
Together AI raised $800M to grow GPU capacity 50x. The CUDA kernel engineers it needs live in vLLM and SGLang commits, not LinkedIn titles.
On July 1, 2026, Together AI closed an $800M Series C at an $8.3B post-money, led by Aramco Ventures, and told the market it will grow its infrastructure footprint roughly 50-fold over the next five years. The careers page is already loaded with roles like "Systems Research Engineer, GPU Programming" and "LLM Inference Frameworks and Optimization Engineer." If you are the recruiter or founder tasked with filling those seats from a LinkedIn search, you are about to lose the race before you start.
The title-first search returns an empty room
Run the search yourself. "CUDA Engineer," "GPU Kernel Engineer," "Performance Engineer," filtered on CUDA, Triton, PTX, in the United States. Add "vLLM," "SGLang," "TensorRT-LLM," or "inference kernel" as a keyword.
Here is what that search returns inside Refolk's index:
Thirteen people. One if you actually want inference framework experience. Meanwhile the top current employers on that list are Apple, Qualcomm, Modular, Meta, and Rivos. Not a single neocloud in the top ten. Together AI's Series C, Baseten's $1.5B raise two weeks earlier at a $13B valuation, Crusoe reportedly negotiating $3B at $18B, Switch's $2B round, Groq's continued build-out: all of them are hiring from a pool that, on paper, does not exist.
It does exist. It just does not describe itself the way ATS taxonomies want it to.
Where the kernel engineers actually live
The population you want is on GitHub, and mostly under two projects.
vLLM now has more than 2,000 contributors from dozens of companies and universities. Its kernel stack alone spans FP8, MXFP8/MXFP4, NVFP4, INT8, INT4, GPTQ/AWQ, plus optimized attention kernels including FlashAttention, FlashInfer, TRTLLM-GEN, FlashMLA, and Triton. Anyone shipping non-trivial PRs against that surface area can do the work in Together's JD.
SGLang, hosted under LMSYS, is deployed on over 400,000 GPUs worldwide. It was originated at Stanford by Ying Sheng and is now maintained by an aggressive contributor community that overlaps heavily with vLLM. Deployed-at users read like the entire hiring competition list: xAI, AMD, NVIDIA, Intel, LinkedIn, Cursor, Oracle Cloud, Google Cloud, Microsoft Azure, AWS, Atlas Cloud, Voltage Park, Nebius, DataCrunch, Novita, Modal.
Two names to know as anchors: Woosuk Kwon and Lianmin Zheng, the core initiators of vLLM, still active in maintenance. Every recruiter working this brief should be able to draw the graph out from them.
The 194-person shortlist nobody has built
Here is the non-obvious cut. As many as 194 developers have shipped code to both vLLM and SGLang. That is roughly 30% of SGLang's total code contributors. Cross-contribution is a strong proxy for A-tier: these engineers are opinionated enough about inference internals to work on two competing frameworks and get their PRs merged in both.
A founder could work through 194 profiles by hand in a week. Most recruiting teams will spend that week writing Boolean strings against LinkedIn titles and end with zero replies. Public examples of the pattern: CatherineSue at Oracle shipped 76 PRs to SGLang as a core contributor after starting on vLLM. ShangmingCai at Alibaba Cloud's Feitian Lab went from 18 vLLM PRs to 52 SGLang PRs. Neither shows up in a title search. Both are exactly the person the JD is written for.
The geography problem nobody wants to talk about
Roughly 33% of vLLM's contributors are based in China. For SGLang, the number is closer to 52%. If you are a US or EU neocloud recruiter, that math forces a decision:
- Sponsor visas aggressively and pay legal to move fast.
- Hire remote in APAC and build a real APAC engineering org.
- Compete for the 47 to 67% of contributors outside China, against OpenAI, xAI, Anthropic, Meta, and every hyperscaler doing the same.
Most in-house teams are set up for none of these. The default posture (US-based, on-site preferred, no visa transfer) leaves you fishing in the smallest possible pond in a market where Together's own JD already lists San Francisco, Singapore, and Amsterdam. Note the Singapore posting. That is not an accident. Together posts a base band of $160,000 to $230,000 plus equity, which almost certainly sits below what OpenAI and Anthropic pay for the same skill set. If you cannot outbid, you have to out-locate.
Recruiters filtering on "CUDA Kernel Engineer" are searching an empty room.
What a real sourcing brief looks like
Stop writing the brief around titles. Write it around behavior.
A usable brief for Refolk or any commit-graph tool looks like this:
- Ships merged PRs to vLLM, SGLang, TensorRT-LLM, or FlashInfer in the last 12 months.
- Contributions touch kernels, attention, quantization (FP8, MXFP4, NVFP4, INT4), or scheduler internals, not just docs or CI.
- Currently at a company that is not already a top-tier inference employer, or is at one but has been there long enough to be a plausible mover.
- Timezone or work-authorization compatible with your open reqs.
That is a query, not a title. And it is exactly the kind of query that title-based tools cannot answer. This is the specific reason we built Refolk: you describe the person in plain English, including the repos they ship to, and get a ranked shortlist across GitHub, LinkedIn, and the open web in a single pass.
The LMSYS funnel most recruiters miss
LMSYS is the non-profit that hosts SGLang and runs Chatbot Arena/LMArena. If you are hiring GPU inference talent and you have never sourced from LMSYS's contributor pool, Discord, or affiliated academic labs, you are missing a channel that competitors are already working. Same goes for the academic origin of these projects. Stanford (via Ying Sheng), Berkeley Sky Computing Lab (via the vLLM lineage), CMU, and a handful of Chinese universities produce this cohort. PhD advisors and lab pages are legitimate sourcing surfaces. Nobody on the recruiting team has to click into a title filter to find them.
Why Aramco leading changes the EVP
Aramco Ventures leading, plus commitments for over 500 MW of compute capacity to be capitalized independently, is not a pure financial signal. It is a deployment mandate. Expect a named Gulf-facing partnership announcement before the end of July, which means Together AI is about to hire not just for SF inference work but for on-prem Saudi and UAE deployments. That is a wildly different pitch than "come work on inference in San Francisco."
For a recruiter, this creates two distinct EVPs to sequence separately:
- Open-source scope pitch: ship kernels to vLLM/SGLang forks that will run across a 500 MW footprint growing 50x. Bookings already at $1.15B, customers include Cursor, Cognition, Decagon. Mission-plus-scale for people who could otherwise take an OpenAI offer.
- Frontier deployment pitch: stand up bare-metal inference clusters in the Gulf. Different profile, often more infra-heavy, often more open to relocation, often ex-hyperscaler SREs who have crossed into kernel work.
You cannot run both pitches from the same sequence. And you cannot find the second cohort at all from a title search. This is again where commit-graph sourcing pays off. Refolk lets you split the 194-person cross-contributor list against secondary signals (bare-metal experience, cluster-scale infra work, Gulf ties or willingness) without rewriting your search from scratch.
The competitive clock
OpenAI has a live public role, "Software Engineer, Inference - CUDA / Kernels," that asks for exactly the same person: writing, porting, and optimizing GPU kernels for inference workloads, deep familiarity with CUDA or equivalent. Baseten just raised $1.5B. Crusoe is reportedly in talks for $3B. Every one of these companies has recruiters running the same broken title search this week.
The team that wins the next 90 days is the team that:
- Builds the 194-person cross-contributor list before anyone else does.
- Layers on the next ring: single-project contributors with 10+ merged kernel PRs, roughly a few thousand people.
- Segments by work authorization and timezone before the first message goes out.
- Runs different EVPs for open-source scope versus frontier deployment.
- Treats LinkedIn as a verification layer, not a discovery layer.
If you are staffing a neocloud right now, the discovery layer is GitHub, the arXiv author graph, LMSYS, and the commit histories of four repos. LinkedIn confirms who someone is. It does not tell you they are the person you want. Refolk stitches the commit graph to the LinkedIn profile to the current employer in one query, which is the workflow this hiring cycle actually requires.
Together AI just told the market it needs to grow 50x. The hiring math to support that growth runs through a shortlist of a few hundred people. Whoever builds that list first, and messages it well, ships the roadmap. Whoever keeps searching for "CUDA Kernel Engineer" will be reading about the winners in next year's funding announcements.
FAQ
Why doesn't LinkedIn surface CUDA and inference-kernel engineers well?
LinkedIn indexes on job titles and self-reported skills, and the actual population you want rarely uses the canonical title. Most vLLM and SGLang contributors carry titles like "Software Engineer," "Research Engineer," "Member of Technical Staff," or an academic affiliation. Their kernel work lives in commit histories, not in headline text. That is why a US-wide title search returns only 13 people even though the underlying population is orders of magnitude larger.
How do I actually build the 194-person cross-contributor list?
Pull the contributor graphs of vLLM and SGLang from GitHub, intersect by GitHub username, filter to authors of merged code PRs (not docs or CI), then enrich each profile with current employer, location, and contact path. You can do this manually in a week, or you can describe the shortlist in plain English to Refolk and get it back ranked. Either way, the important part is that this is a finite list. Treat it that way and work through it person by person.
What compensation range do I need to compete with Together AI?
Together's public band for the LLM Inference Frameworks and Optimization Engineer role is $160,000 to $230,000 base plus equity and benefits, in San Francisco, Singapore, or Amsterdam. That is likely below OpenAI and Anthropic total comp for the same skill. If you cannot outbid, lead with scope (500 MW footprint growing 50x), mission (open-source kernel work that ships to real customers like Cursor and Cognition), and equity upside at an $8.3B valuation. Comp alone will not win this cohort.
Should I be hiring in China or building a remote APAC team?
If your company can support it operationally and legally, yes. Roughly 33% of vLLM contributors and 52% of SGLang contributors are based in China. Ignoring that pool means competing with every US neocloud, OpenAI, Anthropic, and the hyperscalers for the smaller half of the population. A Singapore hub (as Together itself has posted) plus real remote hiring in APAC gives you access to the larger half and often better retention than an SF-only posture.