I build multi-agent AI systems on AWS and write about what I actually find out. Most of the hot takes you're reading are wrong. Some of mine are too. That's what makes it interesting.
This Week in The Stack
The one that matters is below. Along with the context the headlines keep leaving out.
Three frontier model releases in less than a month. The labs are moving fast enough now that by the time you integrate the new hotness, there's a newer hotness behind it. Here's the thing nobody's saying: the capability gap between these models is narrowing while the infrastructure gap — tooling, evals, deployment, orchestration — is widening. That's where the actual leverage is.
Amazon committed $50B and became their exclusive cloud partner. That's not an investment — that's a power move dressed as a check. The infra layer is eating the model layer in real time. AWS didn't win by having the best AI model. They won by being the pipes. This deal says the next round goes the same way.
The Model Context Protocol is quietly becoming the HTTP of agentic AI. 97M installs isn't a trend — it's a foundation getting poured. I've been building on MCP since early beta and the developer experience has gone from "you have to want it" to "this is obviously how you connect agents to tools." That shift matters more than another benchmark.
Grok + orbital infrastructure + Starlink compute + X's data moat in one company targeting a $1.5T IPO. Say what you want about Elon — and there's plenty to say — but vertical integration at that scale is a different category of company than what we've built before. The question isn't whether it's overvalued. It's whether anyone else can match the surface area.
The distance between frontier closed models and high-quality open source is compressing faster than the hyperscalers want to admit. If you're building products on top of expensive API calls right now and you haven't stress-tested open-source alternatives in the last 90 days, you're leaving money and optionality on the table. Run the eval.
Remember when Sora was going to change everything? Now it's discontinued quietly mid-press-cycle. This is how AI hype actually works: big launch, big promises, eighteen months of real usage, quiet sunset while something else is getting launched. The video AI space didn't die — it moved to Runway, Kling, and a dozen others who weren't trying to win the PR game first.
Featured Essay
I've been thinking about this all week and I finally figured out what's actually going on. GPT-5.4 dropped. Then Gemini 3.1. Then Grok 4.20. Three frontier model releases in 23 days. The headlines called it an arms race. What it actually is: a commoditization event happening in slow motion.
Here's the tell. When you look at the benchmarks for these releases — and I did, because someone has to — the gap between first place and third place is a rounding error on most real-world tasks. Writing, coding, reasoning, summarization: they're all within noise margin of each other at the frontier. The models are getting better. They're also getting more interchangeable.
I've seen this before. Not in AI — in cloud infrastructure. Remember when AWS, Azure, and GCP were fighting over compute specs? Virtual CPUs, RAM ratios, storage IOPS? And then at some point you realized the real differentiation was never the commodity — it was the managed services, the developer experience, the integrations, the data gravity. The model layer is running the same playbook right now, just faster.
The money is in the pipes, not the water.
Look at what Amazon just did. They didn't try to out-train OpenAI. They wrote a $50B check and became the exclusive cloud partner. That move says: we already know who won the model war. It doesn't matter. What matters is where the workloads run, where the data lives, and who owns the deployment surface. AWS has been building that for 20 years. One wire transfer and they captured a generation of AI workloads.
Meanwhile MCP hit 97 million installs. Quietly. Without a press conference. That's a protocol becoming infrastructure. That's the kind of boring, foundational thing that everyone builds on and nobody notices until it's already everywhere. HTTP. TCP/IP. OAuth. REST. Now MCP. Add it to the list.
If you're building in AI right now, the question to ask yourself isn't "which model should I use." It's "what layer am I actually competing on, and is that layer defensible." The model layer isn't — at least not for long. The integration layer, the orchestration layer, the data layer, the trust layer: those are where companies get built.
The model wars are over. The infrastructure wars just started. I know which one I'd rather be fighting.
Newsletter Archive
The first issue. Read it to understand what this is and isn't.