Authors: Jean-Stanislas Denain, Campbell Hutcheson

Translation: Shenchao TechFlow

Senchao Introduction: Researchers from Epoch AI analyzed the public recruitment pages of OpenAI, Anthropic, xAI, and DeepMind to infer the strategic directions of these companies from the distribution of job postings.

The conclusions are quite interesting: The proportion of sales positions at OpenAI and Anthropic has surged over the past year, with the technical sales roles that help clients "learn to use AI" growing the fastest;

OpenAI is building a portable device with a camera that runs on self-developed chips, along with 7 robotics positions;

Anthropic does not produce chips but is negotiating data center contracts like crazy. The recruitment pages are one of the few publicly available strategic signal sources, and this analysis is very information-dense.

The full text is as follows:

AI companies are tight-lipped about their strategies, but recruitment pages are public.

These positions hide clues: what products a company is developing, who they want to sell to, and what they think will be bottlenecks. A position for a "Camera ISP Software Engineer" suggests a device with a camera. Posting for a "Frontline Deployment Engineer" indicates that implementing AI in enterprises is challenging. A bunch of robotics-related positions means ambitions go far beyond chatbots.

We analyzed the public positions of leading foundational model laboratories, including OpenAI, Anthropic, xAI, and Google DeepMind. Here are the findings.

Core Findings:

• The sales and related positions at OpenAI and Anthropic have significantly increased over the past year. The proportion of Anthropic's GTM (go-to-market) positions rose from 17% to 31%, while OpenAI's increased from 18% to 28%. The growth is concentrated in technical roles that help customers implement AI.
• Job postings can provide glimpses into the product roadmap. For instance, both OpenAI and DeepMind are investing in hardware products (robots and consumer devices), while Anthropic focuses more on improving core products.
• The recruitment pages also reveal different strategies for acquiring key resources (computing power and data) among these companies. For example, OpenAI has 21 positions related to self-developed chips, while Anthropic has none.

A few points to note: Job postings reflect the types of candidates companies want to hire, not the existing team. For example, if a team has 20 open positions, it could be a large team expanding or a completely new team that does not yet exist. A listing for a "Research Engineer" might be for hiring 1 person, 10 people, or not hiring anyone at all.

GTM Has Become the Largest Recruitment Category for OpenAI and Anthropic

Over the past year, the sales and related positions at OpenAI and Anthropic have grown significantly: Anthropic's GTM positions grew from 17% to 31%, while OpenAI's rose from 18% to 28%. This is not surprising for companies experiencing rapid revenue growth in a market that is far from saturated. Sales-related positions are currently the largest recruitment category for both companies. In comparison, research positions account for only 12% of Anthropic's open positions and 7% of OpenAI's.

Figure: Proportion changes of various positions at OpenAI and Anthropic

The fastest-growing segment is a subcategory: technical roles that help clients actually use AI. Both companies are hiring "AI Success Engineers," "Partner AI Deployment Engineers," "Solutions Architects," and "Frontline Deployment Engineers," roles responsible for helping clients identify AI use cases and complete integrations. Over the past year, the proportion of "Adoption and Implementation" positions at Anthropic increased from 5% to 11%, while OpenAI's rose from 11% to 17%. This indicates that clients are facing difficulties in fully utilizing AI products, and bridging this gap is crucial to teaching clients "what AI can do."

The geographical distribution of sales roles also reveals market focus. More than half of the sales positions at both companies are based in the United States (52% for Anthropic, 55% for OpenAI). Both have not disclosed regional revenue distribution, but the concentration of recruitment suggests that the United States remains the absolute main market.

Internationally, both companies are aggressively hiring in Europe and the Asia-Pacific region. Anthropic leans towards Europe (29% vs. OpenAI's 21%), while OpenAI leans towards Asia-Pacific (24% vs. Anthropic's 19%). Growth in the Asia-Pacific is concentrated in Japan, South Korea, India, Singapore, and Australia. Notably absent are China, the Middle East, Latin America, and Africa. Both laboratories focus on global sales, indicating they do not believe they will be pushed out of the market by local competitors in Europe and the Asia-Pacific.

Government sales are also a key focus area. Both OpenAI and Anthropic have 10 government sales positions, covering federal civilian, defense, and state and local governments. Among them, OpenAI has 1 position dedicated to national security, while Anthropic has 2. xAI has 2 positions focused on international government sales, located in London and Dubai, plus 1 targeting the U.S. government. These positions indicate that government will be an important revenue source for foundational model laboratories.

Unlike Anthropic, OpenAI, and xAI, DeepMind's recruitment page hardly shows any sales dynamics because the distribution of Gemini is handled by Google's existing sales organization.

The Recruitment Pages Reveal New Product Directions for OpenAI and DeepMind

Job postings also indicate what each company is building. Anthropic has 5 product and engineering positions aimed at improving Claude Code, while OpenAI has 10 similar positions for Codex enhancement. Both companies have one engineering position focused on financial services. OpenAI also has 3 positions related to new features for ChatGPT Health and OpenAI for Healthcare.

However, from the perspective of existing products, the view from recruitment pages is not perfect. It's difficult to determine whether a position is for expanding existing functionalities or building entirely new products, as platform or infrastructure roles often span multiple product lines. Hence, recruitment information is most valuable when revealing "new bets."

First, OpenAI is building a consumer hardware device. This project has 15 open positions. These positions reveal numerous details: a "Camera ISP Software Engineer" position describes building an imaging system for a battery-powered portable device; a "Research Engineer" position focuses on running transformer models directly on the device; an "Operating System Engineer" position mentions self-developed chips. Taken together, it appears to be a portable device with a camera that runs self-developed AI chips and can execute AI models at the edge. Additionally, two hardware and operations positions based in Singapore suggest that manufacturing is being prepared. DeepMind is also betting on hardware, with two XR glasses development positions, one of which suggests that voice commands will be the core interaction method.

Second, OpenAI and DeepMind are both betting on robots. OpenAI has 7 positions related to robotics, focusing on large-scale simulation training and enhancing simulation realism. The positions also hint that some robots may have flexible components or shells, and production is scaling up. DeepMind has 9 robotics positions, focused on building a humanoid robot with dexterous hands.

Beyond hardware, OpenAI has 2 positions for social products in the incubation phase and 1 position for an "employment platform," helping users train skills, obtain certifications, and match with employers. Anthropic has 1 research product manager position specifically exploring new product categories, along with another focused on consumer new products.

The Recruitment Pages Also Reveal Strategies for Acquiring Computing Power and Data

The recruitment efforts of various laboratories also showcase their different routes regarding core inputs (computing power and data). The most apparent division is: self-built computing infrastructure vs. outsourced procurement.

OpenAI has 21 positions related to self-developed chips (mostly engineering roles), accounting for 3% of all current positions. Anthropic has not taken the self-developed chip route but has opted for another path: multiple positions focus on collaborating with external partners to design and build data centers, including a "Data Center Mechanical Engineer" responsible for guiding the cooling and mechanical system design of outside companies, and a "Data Center Design Execution Lead" responsible for connecting Anthropic's technical requirements with third-party delivery partners. Anthropic also has 3 legal positions dedicated to negotiating data center or co-location hosting contracts.

Another highlighted direction in recruitment is the reinforcement learning training environment. Anthropic has several positions focused on building environments for training new model capabilities, including a "Environment Expansion" team responsible for constructing RL environments and managing vendor relationships, and a "Universes" team creating ultra-realistic long-duration training scenarios for agents. OpenAI is also hiring researchers for the "Synthetic RL" team, developing reinforcement learning training methods based on self-play, simulators, and synthetic feedback.

Unlike OpenAI, Anthropic, and DeepMind (which do not have dedicated human annotation positions), xAI's recruitment shows a different data strategy. It has 27 positions for human data annotation, suggesting that xAI prefers to keep the data annotation business in-house. It is also interesting that xAI openly recruits for these positions. Other laboratories similarly rely on large-scale human annotation but typically choose to outsource or not publicly advertise.

Conclusion

Job postings are an imperfect signal source, but they are one of the few public windows to observe the evolution of leading AI laboratories. The current picture is that these companies are heavily investing in sales and product implementations, expanding into new product categories, and competing for key resources like computing power and data. As the laboratories continue to expand and their strategies gradually diverge, their recruitment pages will remain one of the best observation windows.