TSMC Just Told You Who's Winning. Follow the Capex.

TSMC posted its fourth consecutive record quarter. Net income jumped 58% to $18.2 billion. Revenue hit $35 billion. AI-related high-performance computing now accounts for 61% of total sales, up from roughly half a year ago. CEO C.C. Wei raised full-year revenue growth guidance above 30% and signaled capital expenditures will hit the top end of the $52-56 billion range.

The numbers confirm what the labs have been signaling through behavior: demand for AI compute is not cooling. It is accelerating. When TSMC raises capex guidance, it is not speculating. It is responding to binding purchase orders from the largest technology companies on Earth. This is not sentiment. This is concrete being poured.

The question worth sitting with: if the picks-and-shovels supplier is investing at this pace, what does that tell you about what its customers expect to sell? And if those customers are wrong about demand, who absorbs the loss? TSMC gets paid either way. The companies ordering those chips do not.

Today’s Movers

Anthropic Quadruples London Headcount, Eyes $800B Valuation Anthropic announced an 800-person London office, quadrupling its UK presence from 200 (CNBC). The move comes days after OpenAI opened its first permanent London office. Investor offers now value Anthropic near $800 billion, narrowing the gap with OpenAI’s $840 billion. The London race between these two is no longer about talent. It is about regulatory proximity and enterprise trust in markets that will write AI governance law. If you are building on either platform, the stability of your provider just became a geopolitical question (we covered the competitive pressure building on April 14).

OpenAI Ships Sandboxed Execution for Its Agents SDK OpenAI released a significant update to its Agents SDK: sandboxed execution environments, a model-native harness for file and tool interaction, and configurable memory (Help Net Security). Agents can now run code, edit files, and execute commands in isolated containers with credential separation. This is the infrastructure layer that turns agent demos into deployable software. If you evaluated the Agents SDK before and passed, re-evaluate now. Sandbox isolation was the missing piece for production use.

Google Launches Gemini Desktop Apps for Mac and Windows Google shipped native Gemini apps for both macOS and Windows, with Option+Space quick access on Mac and file upload support (The Verge). The play is distribution, not capability. Google wants Gemini to be the thing you reach for before you open a browser. If you are building tools that compete for the same reflex, your window just got narrower.

Adobe Firefly AI Assistant Orchestrates Across Creative Cloud Adobe launched Firefly AI Assistant, a conversational agent that chains multi-step workflows across Photoshop, Premiere, Lightroom, Illustrator, Express, and Frame.io (TechCrunch). This is not image generation. This is agentic orchestration across a product suite. Adobe just showed what it looks like when an incumbent actually integrates AI instead of bolting it on. If you serve creative professionals, the bar for what “AI-assisted” means just moved.

Allbirds Abandons Shoes for AI, Rebrands as ‘NewBird AI’ Allbirds, once a $4 billion sustainable footwear brand, announced a pivot to AI compute infrastructure. The stock surged 582% in a day (CNN Business). The company will drop its B Corp certification and acquire GPUs with $50 million in convertible financing. A shoe company buying GPUs tells you everything about where speculative capital is flowing. This is not a signal to follow. This is a signal to watch carefully.

Resolve AI Hits $1.5B Valuation for Production Ops Resolve AI raised $40 million at a $1.5 billion valuation for AI-driven incident response and debugging, with customers including Coinbase, DoorDash, and Salesforce (PR Newswire). Production operations is where AI agents face real consequences for being wrong. If your team runs on-call rotations, this category is worth watching.

One Tool Worth Knowing

OpenAI Agents SDK now includes sandboxed execution with credential isolation and persistent state. Agents can run code, edit files, and execute commands in isolated containers with manifest-based workspace setup. If you have been hand-rolling agent sandboxing, this removes significant infrastructure burden. The harness pattern, where the model gets a structured interface to its environment rather than raw tool calls, is worth studying even if you do not use the SDK.

Wisdom Speaks

“Waste no more time arguing about what a good man should be. Be one.” — Marcus Aurelius, Meditations, Book X.16

Marcus Aurelius wrote that in his tent, likely after a day spent managing an empire that was cracking at its edges. The instruction is not about moral perfection. It is about the gap between deliberation and action.

The AI industry is deep in its own version of this gap. Anthropic and OpenAI are arguing about whose revenue numbers are real. A shoe company is rebranding as an AI infrastructure play. Capital is flowing toward the idea of building rather than the act of it. The debate over who should control AI infrastructure risks becoming the whole conversation.

Solomon understood this. He built the most impressive structure in the ancient world, and then wrote the psalm that undermined every builder’s pride: “Unless the LORD builds the house, the builders labor in vain. Unless the LORD watches over the city, the guards stand watch in vain.” He was not against building. He was the builder. But he had learned, through years of accumulated wisdom and accumulated failure, that the foundation matters more than the facade. You can pour $56 billion in capex into the most advanced fabrication on Earth. You can quadruple your London office. You can pivot your entire company to chase the wave. But if the foundation is self-glorification, if the building became the point, Solomon already wrote the ending.

Build. But know what you are building on.