AMD may invest up to $5 billion in Anthropic while Anthropic buys up to 2 gigawatts of AMD's next-generation AI chips for deployments beginning in the first half of 2027, according to Bloomberg's report.

This is more than a chip supply agreement. AMD is selling compute while placing capital behind the model lab that will generate demand for it. Anthropic locks in capacity and another option beyond Nvidia; AMD gets exposure to both hardware revenue and its AI partner's growth.

Sundar Pichai says Alphabet revenue grew 24% year over year, Google Cloud grew 82%, Gemini reached 950 million monthly active users, and Google's model APIs are processing 22 billion tokens per minute. He also says Gemini Enterprise is used by 90% of Fortune 100 companies.

These are figures reported by Alphabet's CEO, but they point to another front in the AI race: users, API volume and enterprise penetration. Model quality still matters; distribution and revenue conversion increasingly determine who can sustain investment at scale.

OpenAI is reportedly lifting projected compute spending through 2030 to roughly $750 billion. In Project Camellia in Georgia, the company plans an initial $20 billion investment, has contracted 3.2 GW of power for 2028–2032, and will lead the campus's design and development.

OpenAI also says it will pay the infrastructure costs. At this scale, model-lab advantage increasingly depends on power, campuses, rack-scale systems and the ability to coordinate the full infrastructure stack.

OpenAI introduces Presence for enterprise AI agents: OpenAI Presence is a platform for deploying voice and chat agents that can answer questions, access internal systems, take permitted actions and escalate to people when needed. The product is built around policies, guardrails, simulations, evaluations and a post-deployment improvement loop

Claude turns a research dataset into an interactive product — Claude can query the Anthropic Economic Index, a public dataset on how AI is used across the economy. Users can ask which occupations use AI most or what tasks are being automated, while the full dataset remains free to download.

ToFu puts the agent harness alongside the model — The paper proposes an inspectable, token-efficient research-agent harness that can run locally. Its central argument is that agent performance depends on the LLM, orchestration code, tool use and trajectory coordination.

Hugging Face adds tools for checking agent attack surfaces — Hugging Face has introduced tools focused on OAuth and session-token exposure, attack-surface checks and structured results. This is the testing layer agents need as they begin touching real systems.

OpenAI adds Nubank and BNY Mellon leaders to its boards — David Vélez of Nubank and Robin Vince of BNY Mellon have joined OpenAI's boards, bringing more fintech and traditional-finance experience into its governance.

Teams often add a chatbot without changing the workflow around it. The result is another destination to check, another tool to maintain and another handoff for people to manage. The AI may be capable, but it is disconnected from the documents, systems and decisions that make the work valuable.

OpenAI's newsroom examples show a more useful pattern across document search, verification, translation, newsletter production, audience analysis and prospecting: connect AI to the team's existing information and tools, then keep human judgment at the point where it matters. This lets an agent remove one repetitive step without forcing the team to rebuild its entire operating system.

How to set it up:

The takeaway: An agent becomes useful when it shortens a loop that already exists. Start with one bounded workflow, connect it to the right source material and keep the final judgment with a human before adding more automation.

OpenAI examines how agents can be used across functions and run multiple tasks in parallel. The useful shift is from what a model can do to how an organization allocates work when several processing loops run at once.

Why read it: it helps founders and operators see agents as an operating capability, not only a personal tool.

McKinsey separates AI transformation into three horizons and highlights the gap between deploying tools and changing workflows, roles and operating methods.

Why read it: higher adoption does not automatically produce higher impact. The harder step is redesigning the system of work around new capabilities.

In an analysis connected to the OpenAI–Hugging Face incident, Chamath Palihapitiya focuses on attacker–defender asymmetry: an attacker needs one effective path, while defenders must protect every device, account, vendor and software system every day.

Why read it: the piece places agent security inside a broader cyberwarfare context and shows why faster attack automation could change the defense problem.