The semiconductor bottleneck is reshaping AI's future while new interfaces bring intelligence directly into everyday workflows.

The Great Chip Shortage: When Conservative Foundries Meet Aggressive AI Buildouts

Here's a reality check that cuts through the AI hype. While everyone's talking about artificial general intelligence and trillion-dollar valuations, there's a much more immediate constraint throttling the entire industry—silicon supply.

Ben Thompson from Stratechery connects the dots on what he calls "the TSMC brake." Taiwan Semiconductor Manufacturing Company—or TSMC—makes the world's most advanced chips. But here's the kicker: they didn't buy into the AI hype early enough.

When ChatGPT launched in November twenty twenty-two, hyperscalers like Amazon, Microsoft, and Google immediately started pouring money into data centers. Their capital expenditures skyrocketed. But TSMC? Their spending stayed flat through twenty twenty-three and twenty twenty-four.

Now we're seeing the consequences. Amazon's Andy Jassy says demand exceeds supply across all their AI services. Microsoft's CFO reports the same story—they're monetizing capacity as fast as they can build it. Google expects to remain in a "tight demand-supply environment" through twenty twenty-six.

TSMC's CEO finally admitted the obvious: silicon from TSMC is the bottleneck, not power or cooling or data center space. They're scrambling to catch up with fifty-two to fifty-six billion dollars in capital expenditure this year. But here's the brutal math—new fabs take two to three years to build. The shortage isn't getting fixed until twenty twenty-eight or twenty twenty-nine.

Thompson argues this creates a massive hidden risk. TSMC is being rational—they're managing their downside by avoiding overinvestment. But that risk doesn't disappear. It gets transferred to the hyperscalers who are now forgoing billions in potential revenue.

The only real solution? Competition. Samsung and Intel need to become viable alternatives to TSMC. But that requires the exact companies getting squeezed by TSMC to take the risky step of diversifying their chip suppliers.

AI Moves Into Mainstream Workflows: From Code to Spreadsheets

While the chip shortage plays out in the background, AI is quietly infiltrating the tools everyone actually uses every day.

Rowan Cheung from The Rundown highlights Anthropic's expansion of Claude for Excel. After three months in beta for enterprise customers, it's now available to Pro-tier users. This isn't just another AI integration—it's Claude embedded directly in the spreadsheet sidebar where it can work across multiple files without session limits.

Think about the implications. Excel has over one billion users worldwide. Most knowledge workers spend hours every week wrestling with formulas and data manipulation. Now they can just describe what they want in natural language.

Cheung frames this as part of getting "Claudepilled"—first developers with Claude Code, then teams with Claude Cowork, now spreadsheet users. The pattern is clear: AI is moving from specialized developer tools into the workflows that define modern office work.

This connects to broader adoption patterns. Gallup's latest workplace AI report shows the divide is getting starker. Tech workers lead with sixty percent using AI frequently. Remote-capable roles hit sixty-six percent adoption. But manufacturing and retail workers? They're barely touching these tools.

The Browser as AI Sandbox: Rethinking Where Intelligence Runs

Simon Willison picks up on a fascinating technical angle from Google's Paul Kinlan. Instead of running AI agents in heavyweight containers like Docker, what if we used the web browser itself as the sandbox?

Kinlan's insight is elegant: browsers already solve the hardest sandbox problem. They run untrusted code from anywhere on the internet safely. Over thirty years of security evolution went into making that work.

His demo called Co-do proves the concept. It uses the File System Access API to work with local files, content security policies to control network access, and WebAssembly in web workers for safe code execution. The result feels similar to Claude Cowork but runs entirely in the browser without multi-gigabyte containers.

Willison highlights a particularly useful discovery—the webkitdirectory input type that works across Chrome, Firefox, and Safari. It gives browser applications read-only access to entire directory trees, opening up new possibilities for AI tools that need to understand project structure.

The broader theme here is infrastructure convergence. As AI capabilities standardize, the real innovation happens in delivery mechanisms. Whether that's chips, interfaces, or sandboxing approaches—the platforms matter as much as the models.

The semiconductor shortage reminds us that even the most advanced AI is constrained by physical reality. But the browser sandbox work suggests we might be overthinking some of our infrastructure assumptions. Sometimes the solution isn't building something new—it's recognizing what we already have.