Microsoft's "Delivering one Copilot" super app leaked before Build 2026 with a summer launch. MiniMax released M3 — 1M context, open weights, ~59% on SWE-Bench Pro, $0.30/M input tokens. Nvidia unveiled RTX Spark / N1X at Computex with 128GB of unified memory. The agent stack is consolidating into a small number of platform players.

Copilot super app leaks , Minimax M3 , Nvidia N1X

Microsoft's "Delivering one Copilot" super app leaked four days before Build, the open-weight frontier got a new name with the M3 release, and Nvidia announced the RTX Spark / N1X PC chip on the same Taipei Computex stage. Three releases in one week, all in service of the same argument: the agent stack is consolidating, the open model tier got sharper, and the laptop is back as an inference target.

What You Need to Know: A Fortune exclusive on May 29, 2026 revealed Microsoft is building a unified "Copilot super app" under the slogan "Delivering one Copilot" — combining GitHub Copilot, Copilot chat, Copilot Cowork, and a new agentic workflow capability internally called Autopilot. MiniMax released M3, an open-weight multimodal model with a 1M token context, scoring ~59% on SWE-Bench Pro. Nvidia unveiled the RTX Spark (also called N1X) at Computex — a Blackwell + Arm Grace PC superchip with 128GB of unified memory, shipping this fall on laptops from Microsoft, Dell, HP, ASUS, Lenovo, and MSI.

Why It Matters

For developers: Microsoft's super app is a strategic bet that one Copilot identity can hold the agent stack together. The fact that GitHub Copilot and Claude Code are in the same sentence should tell you how serious the Cursor/Claude Code threat is.
For builders choosing a model: M3 with 1M context and 56–57% on SWE-Bench Pro changes the open-weight calculus. "Frontier vs open" is now "open vs open, at different price points."
For ML engineers: Native multimodality (trained on 100T+ tokens of interleaved data) is the architecture that's winning in 2026, not bolted-on vision adapters.
For PC buyers: 128GB of unified memory is the threshold for running real local agent stacks. The Apple-Silicon-style "local-first" pitch is now an Nvidia pitch too.
For IT: The Microsoft + Nvidia partnership on RTX Spark means the Windows-on-Arm story now has both a credible chip and a credible OS vendor. That's a third mobile-class ecosystem, not just a chip.

What Actually Happened

Microsoft "Delivering one Copilot" leaks before Build

On May 29, 2026, Fortune's exclusive detailed a project being run by Jacob Andreou, Microsoft's recently appointed head of Copilot. The internal slogan is "Delivering one Copilot." The super app would unify the GitHub Copilot coding assistant, Copilot chat, Copilot Cowork (task management), and a new agentic workflow capability internally called Autopilot, into a single interface. A toggle would let users move between personal and enterprise Microsoft 365 Copilot accounts. Microsoft plans to launch the app by the end of summer 2026; at Build 2026 in San Francisco, Satya Nadella briefly teased the project but did not demo it. Mary Jo Foley wrote up the no-show for GeekWire.

The strategic context is rough. Less than 4.5% of Microsoft's 450M Microsoft 365 customers pay for Copilot features. GitHub Copilot has more than 4.7M paid subscribers — and is now in direct competition with Cursor and Anthropic's Claude Code. The consumer Copilot chatbot lags OpenAI and Google in active users. The super app is, in part, an attempt to consolidate the brand and reduce the "which Copilot do I open?" friction that has been Microsoft's biggest AI-product problem. Microsoft declined to comment to Fortune.

MiniMax M3 lands with 1M context and open weights

On June 1–2, 2026, MiniMax released M3, an open-weight native multimodal model pre-trained on more than 100 trillion tokens of interleaved multimedia data. Headline numbers: 1M token context, ~56–59% pass@1 on SWE-Bench Pro (above GPT-4.5's reported score), and $0.30 per million input tokens at inference. Fireworks AI's launch post covers the cost-per-quality angle; datanorth.ai summarizes the benchmark; MindStudio's analysis walks through the SWE-Bench Pro score in detail. The NVIDIA developer forums already have a thread on running M3 with NVFP4 quantization on quad DGX Spark nodes — the developer ecosystem is moving fast.

The architectural choice that matters: native multimodality, not bolted-on. M3 was trained from the start on interleaved text/image/audio/video, which is why the long-context window and the coding score aren't fighting each other. For teams building agent stacks, that means fewer compromises when you want a single model handling retrieval-augmented generation, vision, and tool use at the same time.

Nvidia RTX Spark / N1X at Computex 2026

On June 1, 2026, Jensen Huang unveiled RTX Spark (also called N1X) in his Computex keynote in Taipei. The chip is two flagship Nvidia parts fused onto one package: a Blackwell GPU paired with a new Arm-based custom Grace CPU, plus 128GB of unified memory, manufactured on TSMC 3nm with help from MediaTek. The first wave is more than 30 laptops and 10 desktops from Microsoft, Dell, HP, ASUS, Lenovo, and MSI, with units as thin as 14mm and a fall 2026 ship date. Huang pegged the broader CPU TAM at $200B. CNBC covered the keynote in detail, including the announcement that Nvidia's data-center Vera CPU is now in volume production with Anthropic, OpenAI, xAI, Dell, Oracle, and CoreWeave as early customers. Tom's Guide's tracker is the best one-stop summary of the long road to launch.

The Take

These three stories are not three different stories. They're three loads on the same structural beam. The Copilot super app is Microsoft's bid to own the agent runtime identity. RTX Spark is Nvidia's bid to make the laptop a credible local inference target. M3 is the open-weight tier's bid to make the model layer swappable. The throughline is that the AI stack is consolidating into a small number of platform-shaped players — Microsoft, Nvidia, the open-weight labs — and each of them is trying to be the default for a different layer of the stack. The companies that lose are the ones that got stuck in the middle of any of these layers.

The M3 release is the most important for rank-and-file developers. A 1M context open-weight model that hits above GPT-4.5 on SWE-Bench Pro at $0.30/M input tokens is the kind of release that resets the cost curve for everyone. If you're a startup picking a model today, the answer is no longer "frontier or open." It's "which open model at which price point, with which latency profile." The frontier API is now a feature of the open model — not the other way around.

The Copilot super app story is the one with the most execution risk. Microsoft is good at platform consolidation, but the GitHub Copilot vs. Claude Code vs. Cursor fight is being decided on developer mindshare, not on bundling. If the super app ships and GitHub Copilot is the default coding tab, Microsoft gets to keep that surface. If developers keep alt-tabbing to Claude Code or Cursor, the super app is just another Microsoft 365 ribbon. The summer 2026 launch will tell us which way it's going.

The RTX Spark number to watch isn't the GPU equivalence (RTX 5070-ish) — it's the 128GB of unified memory. That's the threshold for running a serious local agent stack: a 7B–13B model, the context, the tool router, and the embedding index, all in one process. If RTX Spark scales, the "everything must go to the cloud" assumption starts to crack for agent stacks, the same way Apple's M-series cracked it for creative apps. The strategic implication is that Microsoft + Nvidia, not just Microsoft + OpenAI, is the binding partnership to watch for the next 18 months.

Quick Summary

Microsoft's "Delivering one Copilot" super app leaked a week before Build, with a summer 2026 launch. MiniMax released M3 — 1M context, open weights, above GPT-4.5 on SWE-Bench Pro, $0.30/M tokens. Nvidia unveiled RTX Spark / N1X at Computex — 128GB unified memory, fall 2026 ship date. The AI stack is consolidating; the open-weight tier just got sharper; the laptop is back as an inference target.

Sources:

Source: TLDR | mr.technology — The Master Skill Index