The market doesn’t care about your narrative. It cares about where the next unit of attention—and compute—will flow. On Monday, Anthropic pushed Claude Cowork to all paid plans across mobile and web. The headline reads like a standard product update. The market reads it like a signal: the AI-agent war just moved from the desktop to the pocket, and the tokenization of that attention is the real story.
Context
Anthropic has been the theological king of AI safety—principled, cautious, and until now, functionally limited. Claude Cowork was a desktop-only feature that let users delegate long-running tasks: file editing, data analysis, tool orchestration. It was powerful but trapped in a single device. Meanwhile, OpenAI’s ChatGPT already lived on every screen, and Google’s Gemini was baked into Workspace. The gap was existential.
Now, Cowork runs on phones and browsers. Every paid user—Pro ($20/mo) or Max ($100/mo)—gets the same asynchronous agent capability. The engineering lift is non-trivial: cloud-based task queues, cross-device state snapshots, and low-latency inference orchestration. Anthropic solved the “session continuity” problem that kills most mobile agent products. s blind spot.
Core
Let’s cut through the press release. The real insight is not “mobile support.” It’s the compute architecture that enables it. Cowork agents run in the cloud, not on the device. Every task—reading a PDF, editing a spreadsheet, calling an API—consumes GPU cycles on Anthropic’s backend. The mobile client is just a thin shell. That means the unit economics of this product are defined by compute cost per session, not by subscription tiers.
Based on my experience designing tokenomics for AI-agent economies at an Abu Dhabi fund, I can tell you: this is a textbook case of “compute-for-equity” scaling. Anthropic is absorbing the variable cost of inference to capture user stickiness. The bet is that the data flywheel—task patterns, tool usage, session lengths—will compound into a defensible moat. But the risk is identical to the one we saw in DeFi summer 2021: if the cost of the input (compute) rises faster than the value of the output (subscription revenue), the model breaks.
We didn’t see this in the original Crypto Briefing piece. The article framed the expansion as a “flexibility” play. That’s marketing. The real driver is liquidity of attention. Every minute a user spends in Claude Cowork on mobile is a minute they are not in ChatGPT or Gemini. Anthropic is not just shipping a feature; they are buying mindshare at the cost of GPU time.
Contrarian
Here is the counter-intuitive angle: the mobile expansion weakens Anthropic’s narrative as the “safety-first” AI company. Mobile devices are inherently less secure than desktop environments. Public Wi-Fi, device theft, and permission creep are real attack surfaces. Claude Cowork holds access to emails, documents, and file systems. If a single agent session is compromised, the reputational damage to Anthropic’s brand could be severe. The market doesn’t price this risk yet because no major incident has occurred. But the odds are increasing.
Meanwhile, the enterprise procurement cycle will slow down. IT departments are already wary of agentic AI tools. Adding mobile access multiplies the compliance headache. GDPR, CCPA, and sector-specific regulations (HIPAA, SOC 2) require granular data flow mapping. Anthropic has not published a mobile security whitepaper. That silence is a red flag for any institutional investor.
Takeaway
The next narrative to watch is compute commoditization. If Anthropic can secure long-term GPU contracts (likely with AWS, their investor), the marginal cost of each Cowork session drops. That would unlock a freemium model or a usage-based pricing tier—both of which would pressure OpenAI’s margin structure. But if compute costs spike, Claude Cowork becomes a loss leader that burns cash faster than a retail trader in a bull market.
Follow the compute. Ignore the press release. The real alpha is in the infrastructure.