Your newsletter digest — June 4, 2026

Ben Thompson filed a sharp Daily Update from Microsoft's Build 2026 conference in San Francisco, covering three things that happened to arrive in the same week: Nvidia's new AI PC chip, Microsoft's Project Solara concept, and Microsoft's first homegrown AI model family. The thread running through all three is a question Thompson has been circling for months — where does the actual intelligence live in an AI-native world?1

At Computex in Taipei, Nvidia CEO Jensen Huang unveiled the RTX Spark (also called the N1X) — a new PC chip made in partnership with Microsoft, debuting in Windows laptops from Dell, HP, Asus, Lenovo, and MSI this fall. The chip packs up to 20 Arm CPU cores, a Blackwell GPU with 6,144 CUDA cores, 128GB of unified memory, and up to 300 GB/s of bandwidth.1

Thompson's take: underwhelmed. He traces the conceptual problem back to when the chip was conceived — three years ago, when local inference looked like the future. But the reasoning era changed what matters. KV cache sizes exploded, decode became the bottleneck, and now in the agentic era, CPU performance is critical. The RTX Spark devotes heavy die area to GPU cores that will lose to the cloud on memory size and bandwidth, while shortchanging the CPU work that agents actually depend on. His verdict: "suitable if you just want a chatbot circa 2023."

What struck Thompson at the Build keynote is that Satya Nadella seemed equally unimpressed — and Nadella, who famously ended Windows' reign as Microsoft's organizing principle back in 2018, has no real attachment to the PC form factor. Local inference for light chatbot use is fine. But it's not where the AI that matters is going.

The more compelling announcement at Build was Project Solara — a new Microsoft platform, currently vaporware but with real working devices shown, targeting a device ecosystem where agents rather than apps are the primary tenant. Built on Android instead of Windows, with Qualcomm and MediaTek as chip partners, it envisions a constellation of purpose-built devices — glasses, wearables, purpose-built terminals — all spoked off a cloud hub.1

The key insight Thompson pulls out: the problem with wearables has always been that they require the human to stay in the loop, which is awkward and inefficient. Project Solara's model flips that — you give a brief instruction, then the agent handles the work in the cloud without you involved. The device is just the point of contact, not the compute.

Thompson is careful: this is still vaporware, and Microsoft has plenty of motives to push cloud-centric narratives given that it doesn't own a phone platform. But he finds the architectural logic genuinely compelling, particularly for enterprise, where context and compute already live in the cloud. "Agents work best in the cloud, and across apps and devices; yes, the phone might be one of those devices, but when it comes to agents it shouldn't be the hub."

This connects directly to Thompson's earlier Thin Is In piece, which argued that AI is bringing back the thin-client paradigm — intelligence in the data center, interaction at the edge.

Multi-device workspace — the kind of setup Project Solara's cloud-hub model is designed around

Also at Build: Microsoft's AI Superintelligence Team unveiled seven models built from scratch. The flagship, MAI-Thinking-1, reportedly matches Anthropic's Claude Sonnet 4.6 in blind human testing and ties Claude Opus 4.6 on a widely used coding benchmark.1

CEO of Microsoft AI Mustafa Suleyman's pitch was enterprise ownership: these models can be fine-tuned with reinforcement learning environments (RLEs) on your own company data, the resulting model belongs entirely to you, and Microsoft claims a McKinsey-tested MAI model delivered higher win rates than GPT-5.5 at 10x lower cost.

Thompson draws a parallel to AWS's Nova Forge offering. The concept — letting cautious enterprises embrace AI on their own terms without betting on cutting-edge performance — is classic Microsoft: it may not win on raw capability, but helping enterprises that are more worried about data sovereignty and vendor lock-in than model benchmarks is exactly how Microsoft has held its enterprise position for decades.

The real story from Build isn't any single announcement — it's the picture they form together. The AI PC that Nvidia and Microsoft co-designed three years ago reflects what people thought would matter: local inference, GPU power on the device. Project Solara reflects what actually matters now: agents that run in the cloud, with devices as thin, purpose-built interfaces. And the MAI models reflect Microsoft's hedge: build credible in-house models so you're not permanently dependent on OpenAI or Anthropic, while giving enterprises a reason to stay inside the Microsoft stack.

Whether Project Solara becomes a real product or a concept deck is an open question. But the direction it points — cloud as the hub, thin devices as spokes, agents doing the work without a human in the loop — is where Thompson has been pointing for most of 2026.