Builder's Briefing — April 28, 2026

Good morning and welcome to Builder's Briefing for April 28th, 2026. I'm Alex, here with Sam, and we have a packed show today. The Microsoft-OpenAI exclusive partnership is officially over, Copilot is moving to usage-based billing, and there's a silent Postgres crisis you need to know about.

Yeah, it's one of those days where you read the headlines and think — okay, the landscape is actually shifting under our feet. Let's get into it.

So the big story: Microsoft and OpenAI have ended their exclusive partnership and their revenue-sharing arrangement. This is huge. OpenAI is no longer locked to Azure for distribution, and Microsoft no longer has to share revenue on AI compute. This is arguably the biggest structural shift in the AI platform landscape since GPT-4 launched.

Right, and what's wild is how immediately actionable this is. If you've been building on Azure-only OpenAI endpoints, you should be abstracting your inference layer like, today. Because OpenAI models are going to start showing up on GCP, AWS, independent inference providers — the whole deal.

Exactly. And the pricing angle is fascinating. Microsoft just lost its moat as the sole enterprise channel for OpenAI's best models. So Azure AI pricing has to get more aggressive now, which honestly is great for everyone.

And if you're a startup that's been going back and forth between OpenAI-via-Azure and the direct OpenAI API — the direct path just became way more strategically sound. OpenAI now has every incentive to make their own platform stickier, since they can pursue enterprise deals on their own terms.

The takeaway here is that model access is getting commoditized at the distribution layer. The winners are going to be builders who stay model-agnostic and treat inference as a swappable commodity. And that segues perfectly into some of the AI tooling that's trending today.

Oh, you're talking about Shimmy. I love this one.

Yes! Shimmy is a single-binary Rust inference server — no Python runtime needed — that runs GGUF and SafeTensors models with hot-swap and auto-discovery. And crucially, it speaks OpenAI's API dialect. So in a world where we're all trying to be model-agnostic, this is a drop-in replacement for local inference.

That's interesting because if you pair that with the Microsoft-OpenAI decoupling story, you see the whole picture. You can point your app at OpenAI's API today, swap to a local model via Shimmy tomorrow, and your application code doesn't change. That's the dream.

Another one I want to flag — Chrome is shipping a built-in Prompt API that lets web apps call on-device models directly. Client-side summarization, classification, form assistance, all without any API costs or round-trips to your server.

So between Shimmy for local server inference and Chrome's Prompt API for client-side, the trend is unmistakable — AI inference is decentralizing away from the big cloud endpoints. Builders who get ahead of this are going to have a real cost advantage.

Also worth a quick mention — Dirac, an open-source terminal coding agent, just topped TerminalBench using Gemini 3 Flash Preview. Gemini's flash-tier models are showing up as genuinely viable for agentic workloads at a fraction of the cost of frontier models.

That's a big deal for anyone budgeting agentic workflows. You don't always need the biggest model if the agent harness is good enough.

Alright, dev tools. GitHub Copilot is moving to usage-based billing, ditching the flat rate. If you're on a team plan, audit your actual usage now. Heavy users were getting a bargain, light users were subsidizing them.

This also tells you something about where GitHub thinks things are going. They expect completion volume to spike as agents generate more and more code, and they want to capture that upside. Flat rate doesn't work when your power users are running agents that generate ten X the completions.

Smart read. And speaking of constraining agents — there's a neat little open-source tool called EvanFlow that wraps Claude Code in a test-driven development loop. You write your tests first, and the agent iterates until they pass.

That's exactly the right pattern. TDD as a guardrail for AI code generation. It's a small project, but honestly, if you're using Claude Code for anything beyond trivial tasks, constraining it with a TDD harness should be standard practice.

And here's a signal I want to highlight: two separate multi-agent coding orchestration SDKs — agtx and GasCity — both trending on the same day. The market is clearly screaming for better ways to manage agent-to-agent handoffs in code generation pipelines.

When you see two tools in the same niche trending simultaneously, that's not a coincidence. That's a market signal.

Okay, infrastructure — and this one is urgent. pgbackrest, which is the most popular PostgreSQL backup tool, is no longer being maintained.

Ohhh, this is the kind of thing that causes incidents six months from now when nobody's thinking about it. If your production Postgres depends on pgbackrest — and a lot of shops do — you need to start evaluating alternatives immediately. Forks, Barman, managed backup solutions, whatever. But don't just sit on this.

Silent infrastructure risk is the most dangerous kind. Also, fun story — the Dutch Central Bank is ditching AWS for Schwarz Group's European cloud. That's the Lidl parent company.

Wait, Lidl as in the grocery store? They have a cloud now?

They absolutely do, and it's being chosen by a major central bank for data sovereignty reasons. If you're building B2B SaaS for European customers, having a non-US-hyperscaler deployment option is becoming a competitive advantage, not a nice-to-have.

The European sovereign cloud market is very real. I think a lot of US-based builders underestimate how much momentum this has.

Quick security note — there was a massive breach at Mercor, an AI contractor platform. Four terabytes of voice samples stolen from forty thousand contractors. If you're collecting AI training data through contractor platforms, your training pipeline is an attack surface. Treat it with the same security posture as production user data.

Four terabytes of voice data. That's not a small leak, that's a catastrophic one. And voice data is biometric — you can't just rotate it like a password. This should be a wake-up call for anyone in the training data pipeline.

Couple of quick hits before we wrap. Someone bought Friendster for thirty thousand dollars — there's a great post-mortem on acquiring dead social networks, link in the briefing. The FDA approved the first gene therapy for genetic hearing loss. And a runner named Sawe ran a sub-two-hour marathon in a competitive race — first time in history.

Okay, the Friendster thing I need to read. And the sub-two-hour marathon — that's a genuinely historic human achievement. Love seeing that alongside all the AI news.

So pulling it all together — today's signal is clear. The AI inference layer is decoupling from platform lock-in. Microsoft-OpenAI exclusivity ending, Shimmy for local inference, Chrome shipping on-device AI, multi-agent orchestration tools popping up everywhere. If you're building AI-powered products, invest in abstraction layers that let you swap models and providers without touching application code.

And check your pgbackrest dependency. Seriously. Today, not next quarter.

That's the briefing for April 28th. All the links and details are in the show notes. We'll be back tomorrow — until then, keep building.

See you tomorrow, folks.