AI Market: Weekly Roundup — NVIDIA+Microsoft, GPT-5.5 on Agents, and Chinese Advances (Jun 03, 2026)

AI Market — Weekly Roundup

Hot Week: The AI Market Doesn't Stop

This was a dense week for announcements. While the market digests the implications of Microsoft Build, NVIDIA announces a strategic partnership with Microsoft, OpenAI releases general access to GPT-5.5, and China shows it's not falling behind on open models.

Let's get straight to what matters.

1. NVIDIA + Microsoft: AI Agents on Windows

NVIDIA and Microsoft announced a unified platform for AI agent deployment covering everything from Windows desktop to the cloud.

What was announced

NVIDIA NIM for Windows: Optimized models running locally on NVIDIA GPUs
Azure Local AI: Agent deployment on edge computing with Windows
DGX Cloud + Azure AI: Integrated training and fine-tuning

Why this matters

For the first time, a company can train an agent in the cloud (Azure/DGX) and run it locally on an employee's Windows machine — with the same weights, same accuracy, same latency.

For Brazilian companies that still hesitate to adopt AI due to data privacy concerns (LGPD), this is a turning point: agents running locally resolve most regulatory objections.

2. OpenAI Releases GPT-5.5 to All Plans

GPT-5.5, previously available only to Pro and Team subscribers, has been released to all ChatGPT plans, including the free plan (with limits).

What changes

Plan	Before	Now
Free	GPT-5 mini	GPT-5.5 (with usage limits)
Plus	GPT-5	GPT-5.5 (higher limit)
Pro	GPT-5.5	GPT-5.5 with no restrictions
Team/Enterprise	GPT-5	GPT-5.5 + Codex included

Practical impact

Free users now have access to far superior reasoning at zero cost
The difference between paid and free plans is now volume, not quality
Codex included in the Enterprise plan accelerates agent adoption in companies

OpenAI's message is clear: model quality is no longer a price differentiator — the differentiator is infrastructure, agents, and integration capability.

3. China: Open Model Qwen 3.5 Surpasses GPT-5 in Benchmarks

Alibaba released Qwen 3.5-72B, an open model that surpasses GPT-5 in mathematical reasoning (MATH) and code (HumanEval) benchmarks, and ties in general knowledge (MMLU).

Model specs

Parameters: 72B (active) + 256B (total MoE)
License: Apache 2.0 (commercial use allowed)
Training: 22 trillion tokens, 70% multilingual (including Portuguese)
Context: 256K tokens

Why this matters for Brazil

Qwen 3.5 has performance comparable to GPT-5 but:

It's free and open
Runs on local hardware (2x RTX 6000 Ada is sufficient)
Can be fine-tuned with Portuguese data

For Brazilian companies seeking independence from US APIs, Qwen 3.5 is the most viable alternative today.

4. Research: Physical Agents with Reinforcement Learning

Two relevant papers came out this week:

NVIDIA Research at CVPR

NVIDIA presented at CVPR 2026 advanced grasping techniques for robots using reinforcement learning with parallel simulation in Isaac Sim.

The breakthrough: robots that learn to pick up objects never seen before with 94% success after simulation training — without any real-world data.

DeepMind: Agents That Learn from Natural Language Instructions

DeepMind published a paper where agents in a simulated environment learn complex tasks using only natural language instructions, without reward engineering.

The model uses an LLM as internal judge: the agent proposes actions, the LLM evaluates whether the action is consistent with the instruction, and the agent adjusts its behavior based on the feedback.

5. What to Watch Next Week

WWDC 2026 (Apple): Rumors that Apple will announce its own agent platform integrated with iOS and macOS — on-device assistants with privacy as a differentiator
Databricks + MosaicML: New open model focused on data analysis and SQL
EU AI Act Regulation: First fines begin to apply for non-compliant companies

Weekly Analysis

Three trends are consolidating:

Open models catch up to closed models — Qwen 3.5 proves that open-source is no longer synonymous with inferior quality. This changes pricing dynamics and access.
Agents are priority number 1 — All big techs are allocating disproportionate resources to agents vs. pure models. The market has already understood that value lies in orchestration.
Brazil needs to move — With competitive open models, locally running agents (LGPD-friendly), and accessible platforms (Azure AI Agent Service), there's no more excuse not to start.

The question I leave you with: what is your company building with AI this week?

Marcos Luciano is Senior Media Buyer, AI & SEO Specialist at V4 Company. He writes daily about artificial intelligence, AI system architecture, and the technology market. Follow on LinkedIn for daily updates.