Anthropic Merges Claude Chat and Cowork for Documents and Slides

01Anthropic merges Claude chat and Cowork, letting paid users create documents and slides in the same conversation

Anthropic is merging Claude’s chat interface with Cowork, a separate workspace originally built for larger, multi-step jobs. The company said users had found it frustrating to decide which interface should handle a task, while work begun in one could not continue in the other. Starting with a rollout announced today, users can instead begin in any conversation and let Claude determine which capabilities the task requires.

For someone preparing a weekly report, that could mean asking Claude to review what changed, flag delays, write the report in an established format and turn its highlights into five presentation slides. According to Anthropic, the document and deck can emerge from the same conversation, allowing the slides to reflect the report without the user transferring context between separate modes. Claude can ask for clarification, continue working after the user closes a laptop and, if scheduled, start the recurring report every Monday.

The consolidation introduces Claude Docs and Claude Slides, while bringing Claude Design, its visual-work tool, into conversations. In Docs, users can write and edit a document together with Claude. Slides can draft presentations that users edit directly, present from Claude or download as PowerPoint or PDF files. Design, Docs and Slides output can be opened on a phone through a single shareable link; users can move individual elements, leave comments or tell Claude what to change. The standalone version of Claude Design will remain available.

Anthropic says existing Cowork users will retain their chats, projects, artifacts, connectors and skills, so they can resume work where they stopped. People who mainly use ordinary chat do not have to change their workflow: Claude is meant to invoke the additional capabilities when a request grows into a larger assignment.

Users also choose how much autonomy Claude receives. By default, it asks for confirmation before taking an action. An optional setting allows it to continue independently and check in only when something requires closer review, while leaving the user with final approval.

The combined experience will reach existing and new Pro and Max subscribers first across Claude’s web, desktop and mobile apps over the next few weeks, with no manual activation required. Docs, Slides and Design are in beta on paid plans. Team and Free plans will follow, but Anthropic gave no specific date; Enterprise administrators will receive at least 30 days’ notice and can decide when to enable the features.

02AI Infrastructure Spending to Near $1.1 Trillion by 2027, With Hyperscalers Needing a Major Productivity Gain to Break Even

A small group of hyperscale technology companies is pouring unprecedented sums into the data centers that power artificial intelligence, despite a wide gap between current spending and revenue. Alphabet, Microsoft, Amazon, Meta and Oracle, which partners with OpenAI, are expected to spend about $750 billion on data centers this year, while total AI revenue is estimated at only $150 billion to $200 billion.

Jessica Wachter, a finance professor at the University of Pennsylvania’s Wharton School, and a collaborator used that mismatch to calculate the returns required to support the buildout. Rather than predict how capable or widely adopted AI will become, they estimated that relevant expenditure will approach $1.1 trillion by 2027 and asked how quickly the companies’ earnings must grow to justify it.

Their accounting suggests the companies would need to increase their own productivity to 2.7 times its current level to break even by 2030. The calculation includes the cost of capital, a 15% return and depreciation of the assets. Wachter said such growth is not impossible, but it would compress gains comparable to roughly a decade of the US information-technology boom into only a few years.

Financing is making that hurdle higher. The companies have begun borrowing heavily for new data centers, increasing their interest obligations as debt becomes more expensive. Free cash flow—operating cash flow after capital expenditure—is expected to turn negative for the group. Alphabet reported nearly $120 billion in revenue in its latest quarter but ended with a free-cash-flow deficit of about $5.9 billion after infrastructure spending.

Rapid chip improvement creates another recurring cost. Graphics processing units, or GPUs, supply the computing power behind AI systems and account for about 60% of a data center’s cost. Their performance roughly doubles every two years, meaning operators may have to spend billions more on newer chips by the decade’s end to remain competitive. Without those upgrades, facilities could become stranded assets that no longer generate adequate returns.

Future AI demand, model efficiency, computing requirements and data-center profitability remain unknown. Wachter warned that companies unable to meet their profit targets could fall behind on interest payments and face bankruptcy, but the research does not establish that any individual hyperscaler will do so—or that the required productivity boom cannot occur.

03Firefox Tests Mistral-Powered AI Browsing Assistant, With France and North America First

Mozilla is testing a Mistral-powered AI assistant inside Firefox, initially making the tool available to users in France and North America. Called Firefox Smart Window, the beta feature is designed to help people work through complex searches, retrieve information they previously opened and organize relevant material from their browser tabs.

Mistral, an AI company that has traditionally focused on enterprise customers, said its models now power Smart Window. The partnership gives Mistral a route to consumers while allowing Mozilla to embed AI assistance directly in the browser. The United Kingdom and Germany are expected to receive the feature later this year, although the companies did not provide exact launch dates or explain who qualifies for the test.

The assistant is intended to use the context of a person’s open tabs to source information and help recover something important that the user clicked away from. The announcement did not identify the specific Mistral model involved or say where requests will be processed.

Mistral said it is training and fine-tuning its systems for regional languages, dialects and cultural context so responses better reflect local usage. That localization is part of the companies’ broader pitch that AI products should be adapted to individual markets instead of offering the same experience everywhere.

Mozilla and Mistral are also emphasizing data handling as a point of differentiation. According to their announcement, Smart Window conversations are not saved on Mozilla’s servers by default, while participating providers such as Mistral agree to zero data retention. The source did not describe an independent audit of those commitments or provide further technical details about the privacy safeguards.

Mozilla framed the partnership as part of a design in which the browser can support competition among AI providers. Mozilla Corporation CEO Anthony Enzor-DeMeo said browsers should not funnel users into a single company’s product pipeline, but should leave room for different providers and open-source technology. Both organizations describe themselves as advocates of open ecosystems, with Mistral pointing to its releases of open-weight models.

The next announced expansion is to the United Kingdom and Germany later this year; the precise timing, model choice and broader rollout remain unspecified.