Foundation Models

Chinese regulators approved Apple Intelligence built on Alibaba Qwen. The frontier model is becoming a licensed regional component, not a global product.

At Build 2026 Microsoft launched the MAI model family and made Project Polaris the default behind GitHub Copilot. What the in-house turn means for OpenAI, developers, and every hyperscaler weighing build versus buy.

On April 27 2026 Microsoft and OpenAI ended Azure exclusivity. OpenAI keeps a capped 20 percent payment to Microsoft through 2030 but is freed to ship production workloads on AWS, GCP, and Oracle. The same week, Anthropic locked in $40B from Google and $5B from Amazon against a $100B compute commitment. The cloud-and-foundation-model relationship that powered the post-2023 build-out is being unwound — and most enterprise AI architectures still assume the world that ended last Monday.

In thirty days every major frontier lab shipped an agentic-first foundation model. GPT-5.5 'Spud', Claude Mythos 5, Gemini 3.1, and DeepSeek V4 each arrived with two-to-three times the token economics of their chat-era predecessors and a benchmark portfolio that prizes operating-system control over multiple-choice trivia. The chat era did not end with an announcement; it ended with four near-simultaneous price hikes and a quietly retired benchmark suite.

The data wall the AI scaling discourse worried about for three years quietly fell in 2026. Synthetic data now generates the majority of new high-quality training tokens at every frontier lab. A deep look at what changed, why it works, what it costs, and what it means for the next 24 months of model capability.