What frontier AI labs
are actually building

Building its own silicon. 60 Hardware/Silicon roles alongside 70 Model Infra — unusual for a pure software lab. They're de-risking dependence on Nvidia while simultaneously running the largest enterprise GTM in AI (195 roles). Trust & Safety (58) is a compliance requirement for government contracts, not mission.

The Trust & Safety (48) + Legal & Compliance (25) + Safety & Alignment (18) cluster = 91 roles. That's 21% of all hiring. This isn't idealism — it's the price of selling to Fortune 500 and government. Finance & Corp Dev at 43 roles signals major partnership/M&A activity. They're building the compliance moat nobody else wants to pay for.

29% of roles are Data — the highest share of any lab. This is a deliberate data acquisition and annotation play: building proprietary training data at scale using X's real-time social corpus. Hardware (40) and Model Infra (32) are secondary. xAI is betting that data quality wins the next model generation, not architecture.

Applied Research (15) + Fundamental Research (5) + Safety & Alignment (6) = 39% of all hiring. Product (16) exists but GTM is barely 1 role. DeepMind isn't trying to win enterprise revenue — they're publishing the science that others commercialise. Robotics (3) is early signal of Gemini Robotics investments.

GTM is 31% of roles, with EMEA-specific sales tags dominating. Applied Research (21) is unusually high for a commercial lab — they're maintaining research credibility to justify open-weight releases while selling proprietary APIs to European enterprises who can't use US-hosted models for regulatory reasons. Geographic Expansion (7) is all EU.

Applied Research (22) and Model Infra (15) alongside GTM (28) — they're positioning as the "safe enterprise choice" with on-prem deployment capability. Data (13) roles signal RLHF/fine-tuning pipelines for customer-specific models. Geography (8) is English-speaking enterprise markets only.

Model Infrastructure is their #1 category (19 of 65 roles = 29%) — more than GTM. For a consumer search product, this is a clear signal they've moved beyond API calls and are training and serving their own models end-to-end. Applied Research (11) confirms active model development. Search tags suggest RAG infrastructure at scale.

GTM (25) + Geographic Expansion (12) = 61% of all hiring. APAC expansion tags are prominent. Cognition is past the demo phase and in full enterprise sales mode — they're trying to land dev-tool contracts before coding assistants commoditise. Almost no research hiring = they're shipping, not discovering.

Public sector and government tags dominate their GTM (31) and Product (46) roles. Robotics/Embodied AI (10) alongside Applied Research (30) shows they're building data pipelines for autonomous vehicles and defense robotics — not just LLM RLHF. Scale is positioning as the data infrastructure layer for the US military's AI programs.

Geographic Expansion is 29% of all roles — an extraordinary share. Combined with Voice/Audio (5), this reads as aggressive multilingual voice cloning expansion into non-English markets before competitors. GTM (70) is the largest absolute count in the dataset relative to size. They're signing enterprise deals for voice AI in every major language simultaneously.

91 Model Infrastructure roles — more than any other company in this dataset. CoreWeave isn't an AI lab; it's the compute substrate that AI labs rent. GPU, Kubernetes, data center, and cloud infra tags dominate. Every lab that can't afford or justify building its own compute (everyone except OpenAI and xAI) is a CoreWeave customer.

66% of all roles are Hardware/Silicon — ASIC, signal integrity, HPC, supercomputing tags. Etched is building a chip that only runs transformers, betting that the architecture is locked in and specialised silicon will be 10× cheaper and faster than general GPUs. Model Infra (19) roles are writing the software stack for their custom hardware.

Robotics/Embodied AI (9) + Hardware (5) = 70% of all roles. The hardware here isn't silicon — it's actuators, sensors, and mechanical systems. π is building foundation models for robot manipulation, betting that the same scaling laws that worked for language will work for physical control. Tiny company, very focused signal.

Hardware (8) + Product (9) = 52% of roles with supply-chain and manufacturing tags. Sesame is building a physical consumer device — not just software. Voice/Audio (2) alongside hardware suggests a smart speaker / ambient AI form factor. The supply-chain hiring is a tell: they're going to manufacturing, not just prototype.

Product (13) is their top category, with Android/Kotlin/mobile tags prominently in their tech stack. They're building a mobile-first music creation product — not a B2B API. Voice/Audio (5) + Data (4) signal active model training. Trust & Safety (4) suggests content moderation for user-generated music at scale.

GTM is 69% of all roles — the highest concentration in the dataset. Customer success, AI transformation, and AI adoption tags suggest they're not just selling seats but running implementation projects for large enterprises deploying AI. Applied Research (6) maintains model quality while the sales machine drives revenue.

Applied Research (4) and Safety/Alignment (1) alongside GTM (6) = they're commercialising interpretability research. Goodfire is the only lab in this dataset where positioning and product marketing appear explicitly in tech tags — they know the hard part isn't the science, it's convincing enterprises that interpretability is worth paying for.

Applied Research (4) dominates, with fine-tuning, post-training, multimodal, VLM, and vision tags. Liquid is betting on alternative architectures (liquid neural networks) that are more efficient than transformers for edge deployment. Voice/Audio (2) suggests multimodal expansion. Very research-heavy for a company of this size.