TwelveLabs Tripled to 178. The Marengo Research Org Lives in Seoul.
TwelveLabs closed a $100M Series B on July 1. The video-AI talent pool is in Seoul, not SF, and US-centric LinkedIn searches miss it entirely.
On July 1, 2026, TwelveLabs closed a $100M Series B co-led by NEA and NAVER Ventures, with Amazon coming in to lock AWS as preferred cloud. If you read that press release as an SF story, you missed the actual hiring map. The company tripled from ~58 to ~178 in twelve months, and the research org that ships Marengo and Pegasus lives in Seoul.
Every US recruiter running "video understanding engineer" through LinkedIn right now is fishing the wrong pond. The people who have actually shipped a production video foundation model at scale mostly hold Korean passports, publish out of KAIST and SNU, and answer to advisors most Bay Area sourcers cannot name.
The "San Francisco HQ" is a marketing artifact
Go to twelvelabs.io/careers. The location filter offers exactly two options: Seoul, South Korea and Remote US. That is not a rounding error. Seoul roles span Research Science, ML Engineering, ML Data, and Recruiting, including a Senior HRBP partnering with technical leaders in Korea and a Senior Talent Sourcing Partner contract role that only makes sense if the org's center of gravity is in Gangnam, not SoMa.
The tell is in the SF job descriptions themselves. TwelveLabs' San Francisco product roles now explicitly require "availability until approximately 8pm PT on most weekdays" because product staff collaborate daily with the Seoul research team. When your California hires are working Korean hours, the org chart has already voted.
FinSMEs reporting on the round confirms the Series B capital funds R&D in both San Francisco and Seoul, plus new offices in New York and London. Seoul is not the satellite. Seoul is the model factory.
What Marengo and Pegasus actually need
Marengo 3.0 is TwelveLabs' video embedding model, distributed via Amazon Bedrock. Pegasus 1.5 structures video into scene boundaries, entities, and temporal segments. Building either at production quality requires a specific skill stack that thin US searches will not surface:
- Cross-modal contrastive learning at video scale. Not image-text. Video-text, with temporal windowing.
- Video encoder architectures past the standard TimeSformer/ViViT literature.
- Long-context temporal modeling for scene boundary detection.
- AWS Neuron SDK and Trainium kernel work, because the models launch first on Trainium under the new Amazon commitment.
That last bullet is a hiring signal a lot of recruiters are still reading as an infra story. It is not. When a foundation model company optimizes new releases for Trainium first, they need engineers who have actually written accelerator-specific kernels, not people who list "distributed training" on LinkedIn. Search AWS Neuron SDK, Trainium, Inferentia, and SageMaker HyperPod as literal keyword strings. Expect Samsung Research and Coupang alumni to over-index on the underlying systems chops.
The problem is that stitching those five signals together (video-language pretraining plus Trainium plus Korean-language interview capability plus Pangyo commute plus available in the next 90 days) is exactly the kind of query that breaks Boolean sourcing. Which is why we built Refolk: describe the person in plain English, get a ranked shortlist across GitHub, LinkedIn, and the open web, no keyword acrobatics required.
The Seoul talent pool concentrates in four labs
If you are not sourcing by lab and advisor lineage, you are guessing. The Seoul video-AI talent pool is not a broad market. It is roughly four institutions and three universities:
NAVER AI Lab (Seongnam). The most directly adjacent research to Marengo. Names to know: Sangdoo Yun, Sanghyuk Chun, Jin-Hwa Kim (joint with SNU AIIS), Jiyoung Lee. Publications on cross-modal retrieval, video contrastive learning, and multimodal embeddings. NAVER Ventures led the Series B alongside NEA. GP YJ Park has publicly said TwelveLabs was NAVER Ventures' first-ever investment. That relationship is not just capital. It is a talent conduit that a US-based recruiter cold-messaging on LinkedIn cannot replicate.
Kakao Brain. Karlo, Kanana, multimodal work. Advisory ties to KAIST through Prof. Kim Seung-ryong and Prof. Shin Jin-woo signal where the next cohort trains.
LG AI Research. EXAONE team. Deeper on foundation model systems than video specifically, but a serious source of engineers who have shipped at scale.
Samsung Research. Refolk's index shows Samsung Electronics as the top current employer of ML/CV researchers in South Korea (10 in the visible cohort), followed by LG AI Research, Hyperconnect, and 당근마켓 (Karrot). Samsung Research alumni are also where you find the Trainium-adjacent systems people, because they have already worked on custom silicon paths.
The universities are narrower than most people expect: KAIST (visual recognition group), Seoul National University (the Department of AI, established 2021, has roughly 30 faculty), and POSTECH, with Korea University as the fourth. Advisor names that meaningfully change candidate quality: Jaegul Choo at KAIST, Kwanghoon Sohn at Yonsei, Seungryong Kim at KAIST.
The compensation arbitrage is real and closing
This is the part most US founders miss until it is too late. Senior AI researcher packages at Samsung, NAVER, and Kakao run KRW 80 to 150 million, roughly $60K to $115K, versus $300K to $600K+ for equivalent US roles. That is a 60 to 70 percent discount for buyers willing to hire in-country.
But this arbitrage is narrowing, not widening. The Brain Pool Programme supplements and Samsung/NAVER's differentiated AI tracks (up to KRW 200 to 300 million for elite researchers) have already pulled top-tier compensation up. TwelveLabs' Series B, backed by NAVER Ventures for local credibility plus AWS-aligned equity, is a package that domestic Korean AI startups without foreign capital cannot match. Nucamp, citing Korea Herald reporting, notes ~70 percent of Korean firms are actively recruiting AI talent and report a shortage of "job-ready" candidates. Everyone is bidding for the same 1,400-per-year output.
The recruit-from-Seoul window for US buyers is 12 to 18 months, not indefinite.
If you are running a video, robotics, or multimodal startup and you have not opened a Korean entity or a Korea-remote hiring path, the calculus in 2027 is going to be materially worse than it is right now.
The Amazon-locked risk to the rest of the ecosystem
Techfundingnews framed a pattern worth taking seriously: Amazon named itself preferred cloud for Odyssey weeks before the TwelveLabs deal, under a similar structure. Cloud-provider capture of AI startups tends to funnel technical talent into AWS-adjacent orbits: Bedrock partnership engineers, Trainium optimization specialists, SageMaker HyperPod internal tooling.
For competitors, the practical consequence is that engineers who work at TwelveLabs for 18 months become AWS-fluent in a way that makes them more attractive to AWS and to Amazon's AI-adjacent acquisitions than to your Nvidia-native infra. That gravitational pull increases the longer the partnership runs.
Two operating conclusions:
- Source the current TwelveLabs cohort now, before Trainium specialization narrows their outbound reply rate to non-AWS employers.
- Source the Kakao Brain and Samsung Research adjacent pool now, because they are the last large group of Korean video-AI talent not yet inside an AWS-locked contract.
What US-centric sourcing gets wrong, specifically
Three failure modes we see repeatedly when recruiters try to hire against TwelveLabs-adjacent profiles:
Wrong platform. Korean researchers are dramatically under-indexed on LinkedIn relative to their publication output. Many maintain more current presences on Google Scholar, GitHub, personal academic pages, and Korean-language platforms. A LinkedIn-only search returns maybe 15 percent of the real pool.
Wrong keywords. "Video understanding" is a US marketing term. Korean researchers publish under "video-language pretraining," "temporal grounding," "video moment retrieval," and "video question answering." Different literature, different keyword sets.
Wrong geography filter. "Seoul" misses Seongnam (where NAVER and Kakao actually sit, in Pangyo Techno Valley) and Suwon (Samsung). Refolk's index shows the ML/CV researcher distribution as Seoul 12, Gyeonggi 4, Seongnam 2 in the visible cohort, and the Gyeonggi/Seongnam numbers are the ones most sourcers accidentally exclude.
The compounding effect of these three mistakes is that a "diligent" search returns 8 to 15 candidates when the real addressable pool is meaningfully larger. If you would rather describe the person in English (say, "video-language researchers in Pangyo who have shipped on Trainium or Neuron") and let the system do the translation and geographic expansion, that is exactly the query shape Refolk was built for.
A tactical shortlist for the next 90 days
If you are hiring against this pool right now, work in this order:
- Map current TwelveLabs Seoul employees by publication history. LinkedIn is optional. Google Scholar is not.
- Pull the last three years of NAVER AI Lab papers on cross-modal retrieval and video contrastive learning. Trace co-authors back to advisor labs.
- Filter for AWS Neuron/Trainium exposure through Samsung Research, Coupang, and Bedrock partnership engineers, not just through TwelveLabs alumni.
- Physically show up in Pangyo. The density of NAVER, Kakao, and adjacent AI startups within a two-kilometer radius rewards in-person coffees at a rate that video calls do not replicate.
- Get a Korean entity or EOR arrangement done now. By the time you have offers to make, the setup lag will cost you 6 to 8 weeks per hire.
The Series B closes a specific window. TwelveLabs will finish its planned 2026 hiring by end of Q1 2027 at the current pace. The Seoul pool that funded TwelveLabs' tripling is the same pool that funds everyone else's video, robotics, and embodied AI ambitions. It is finite, it is concentrated, and it now has a credible AWS-aligned employer sitting on top of it with $100M in fresh capital.
Move first, or hire second.
FAQ
Why is TwelveLabs' research team in Seoul instead of San Francisco?
Founder network and talent density. TwelveLabs was NAVER Ventures' first-ever investment, and the company has been structurally Korea-anchored since the seed round. Seoul offers ~1,400 AI PhDs per year, top-15 global publication rank via KAIST, and a compensation structure that lets a Series-B-stage company hire senior researchers at 30 to 40 percent of Bay Area cost. The July 2026 Series B explicitly funds R&D in both cities, but Marengo and Pegasus development lives in Seoul.
How do I actually source Korean ML researchers if I do not speak Korean?
Start with Google Scholar and paper co-authorship graphs, not LinkedIn. Filter for advisors (Jaegul Choo, Seungryong Kim, Kwanghoon Sohn) rather than keywords. Use GitHub for engineering-heavy candidates and CVPR/NeurIPS/ICCV proceedings for research candidates. For outreach, English is generally fine at the senior researcher level, but response rates roughly double when you can reference the specific lab and advisor. Tools that let you query in plain English and pull across GitHub, LinkedIn, and the open web (Refolk does this) collapse most of the platform-hopping friction.
What does "AWS Trainium optimization" actually mean as a hiring signal?
It means the company needs engineers with hands-on AWS Neuron SDK experience, custom kernel authoring for Trainium/Inferentia, and familiarity with SageMaker HyperPod for distributed training. That skill set is thin globally. In Korea, it over-indexes at Samsung Research (custom silicon exposure), Coupang (large AWS deployments), and Bedrock partnership engineers. Standard "distributed training on GPUs" experience does not transfer directly.
Is the compensation arbitrage still worth it if we open a Korean entity?
Yes, but the window is 12 to 18 months, not indefinite. Base senior AI researcher packages still sit at $60K to $115K USD equivalent, versus $300K to $600K+ in the US. Even with Brain Pool supplements and Samsung/NAVER premium tracks pushing top-of-market to KRW 200 to 300 million, the delta remains 40 to 50 percent for equivalent-caliber researchers. The economics stop making sense when the top of the Korean market crosses roughly KRW 400 million, which most credible forecasts put in the 2027 to 2028 window.