OpenAI publishes research on how AI is expanding work: ChatGPT users take on tasks that cross role boundaries. The research accompanies the formal expansion of ChatGPT Work enterprise, with the agentic platform Presence (voice and chat) and ChatGPT for Small Businesses that includes a skill builder to automate recurring tasks. On the infrastructure front, Project Camellia deploys 3.2 gigawatts of data center capacity in Georgia.

Why this matters to you. For teams using AI agents at work, the choice is shifting from model selection to the complete package: platform, training program, and compute. The compute race is the thread connecting the labs’ moves, as we reported on July 23: Anthropic secured 2 GW from AMD, Google invested 40 million in AI tokens for research, and now OpenAI responds with 3.2 GW of its own. Whoever has more computing power can run more agents for more users, and that determines who stays competitive on speed and pricing. The research also shows that role boundaries are shifting: teams using AI take on work that wasn’t theirs before.

In detail

The context.

Until June, OpenAI talked about models. Since July, it talks about platforms. Presence formalizes a single product for deploying voice and chat agents in enterprise: the shift from model to platform is explicit. Now comes the research to back it up.

The post “How AI is expanding what people do at work” shows that ChatGPT users are taking on tasks that cross role boundaries, redrawing the lines of work. Someone who previously only did data analysis now handles reporting, or a salesperson manages part of technical support. AI lowers the cost of crossing those lines.

What’s changing.

Three pieces fit together. Presence is the agentic platform: voice, chat, and unified presence for enterprise deployment. ChatGPT for Small Businesses brings the skill builder to small businesses, automating recurring tasks without coding. Project Camellia is the 3.2 GW data center in Georgia, with commitments on energy, water, and local community.

The third piece is what shifts the competitive equation. Compute is needed to run agents in production, serving millions of users in parallel, not just to train models.

The compute thread.

As we reported on July 22, AMD guaranteed Anthropic compute credits and hardware for 2 GW of GPU MI450. Two days later, Google invested 40 million in AI tokens for research. Now Camellia is the third signal: 3.2 GW, more than Anthropic’s 2 GW.

The pattern is clear: labs compete on available computing power, not just model quality. Whoever has more gigawatts can serve more agents simultaneously, keep prices low, and response times fast. For teams choosing AI tools, this means the question isn’t just “which model is better” but “which platform has the compute to handle the load as the team grows”.

Limitations.

OpenAI’s research is a blog post, not a peer-reviewed paper: methodology isn’t detailed, and sample size isn’t disclosed. Camellia’s 3.2 GW is an announcement: construction timeline and actual compute availability aren’t published. Presence is formalized as a product, but specific APIs and connectors aren’t yet publicly documented. The skill builder is described at program level, not as a verifiable tool.