Subscribe
Analysis

Long-form pieces

Theses, vendor breakdowns, category surveys, vertical deep-dives. New analysis lands roughly weekly during the build phase.

routingopenrouter

Tokenmaxxing is over. Routing is the shape.

Chinese-origin AI models now handle nearly half of the enterprise token traffic on OpenRouter in a peak week, up from 4.5% eighteen months ago. That number does not mean the frontier moved to Hangzhou. It means the frontier stopped being the point. The operating pattern of enterprise AI in mid-2026 is a router in front of a portfolio of models, and the router picks the cheapest one that clears the bar. Anthropic and OpenAI still write the ceiling. They just don't write most of the tokens.
nvidiacustom-silicon

Every frontier lab is quietly drawing its own chip, and Nvidia has about eighteen months to notice

Five weeks. Four custom-silicon announcements. Meta and Qualcomm on Dragonfly. OpenAI and Broadcom on Jalapeño. Google's next-gen TPU already committed to Anthropic. And now Anthropic itself, according to TechCrunch, in early talks with Samsung. The frontier labs have collectively decided that renting Nvidia at Nvidia's margins forever is not the plan. The interesting question is what that means for the shape of the industry in 2027.
openaianthropic

The federal government quietly became the AI sector's product manager

In a single week, OpenAI shipped its biggest model under a government-vetted preview list, Anthropic asked the Senate to sanction a Chinese rival for distillation, and the Pentagon's six-month Anthropic phase-out kept ticking. None of these events were on anyone's roadmap eighteen months ago. Read together, they describe a sector that has stopped being a free market and started being a regulated industry, without anyone formally announcing the change.
humanoid-robotsfigure

Every humanoid robot pilot is in a car factory. That is not a coincidence.

Figure at BMW Spartanburg. Apollo at Mercedes Tuscaloosa. Atlas at Hyundai Georgia. The humanoid deployments cluster at automotive OEMs for reasons that have nothing to do with the robots and everything to do with the only assembly line that still has the capex budget, labor demographics, and ROI math to make a two-hundred-thousand-dollar biped pencil out.
anthropicclaude

Anthropic is the only frontier lab the US is trying to ban, and also the one everyone else is racing to integrate

The same week the Pentagon is six weeks into a project to replace Claude in classified workflows, Microsoft has Claude in 11,000-model Foundry, Apple is reportedly making Claude an iOS Extension, and Anthropic's web-traffic share grew 306 percent in a quarter. The split is not random. It is what happens when one lab holds a policy line and everyone gets to vote on whether they like the line.
anthropicclaude-partner-network

The model is now table stakes. The consultant is the product.

Anthropic just formalized a three-tier consulting ladder with a top rung that requires 1,000 certified practitioners. Accenture is training 30,000, Cognizant is routing 350,000, Deloitte has 470,000 in scope. The AI lab business model just stopped being software and started being Salesforce, on a 24-month timeline instead of a 25-year one.
anthropicmemory

Your HBM supplier is now your shareholder, and that is how you know the compute crunch is permanent

Samsung, SK Hynix, and Micron all wrote checks into Anthropic's $65 billion Series H. The memory layer of the AI stack just promoted itself from commodity input to strategic equity holder, and the implications run further than most analyst notes are pricing in.
anthropiccompute

Your biggest rival is now your landlord: Anthropic's strange new compute portfolio

Anthropic just stitched together a roughly $85 billion compute portfolio across Google, Amazon, and Musk's Colossus data centers. The map of AI alliances is being rewritten by megawatt availability, not loyalty.
weekly-digestroundup

Weekly digest: May 12-17, 2026

This week's signal across models, tools, applied AI, and policy. Claude Opus 4.7 ships, Cursor-Windsurf closes, EU Act compliance window advances, and the enterprise RAG conversation gets more honest.
enterprisecoding-agents

The coding-agent procurement cycle

Why enterprise rollouts of AI coding tools are running 9-12 months from pilot to seat-license, and what that timeline tells you about the next leg of the category.
open-sourceopen-weights

The open vs. closed model debate is the wrong frame

Framing the AI model landscape as 'open source vs. proprietary' obscures the question that actually matters: who controls the training data, and what are the implications of that control.
long-contextretrieval

The long-context economics question

1M-token context windows are standard at the flagship tier now. The real question isn't capability, it's whether stuffing everything into context is actually cheaper than retrieval for your workload. Usually it isn't.
ragproduct-strategy

RAG is not a product strategy

Retrieval-augmented generation is a useful technique. It is not a moat, not a differentiator, and not a strategy. The number of pitch decks that don't seem to know this is unsettling.
inferenceeconomics

The inference cost curve is the most important chart in AI

Compute cost per token has dropped by roughly 100x over 18 months. The trajectory of that decline, and where the floor might be, matters more for AI application economics than any individual model capability.