Microsoft Slows Down on AI Chips: Braga to Arrive in 2026 | Types of cpu | Computer hardware components and their functions | Cpu hardware list and functions | Turtles AI

Microsoft’s “Braga” AI chip, scheduled for production in 2026, is delayed by six months due to design complexity and team turnover. It is expected to underperform NVIDIA’s Blackwell.

Key points:

• Delay moved to 2026
• Project Braga will not surpass Blackwell
• Technical issues and team exits
• Autonomy risk from NVIDIA

Microsoft had launched the Maia 100 chip in November 2023, the first incarnation of an internal silicon line, with the aim of making a full debut as early as 2024; The roadmap had called for Braga in 2025, followed by Braga-R and Clea by 2027, but now the launch has been pushed to 2026. Reasons for the delay include significant design changes made mid-stream, shortages of skilled labor, and a higher-than-expected turnover rate, with losses of up to 20% of key resources.

Preliminary estimates suggest that Braga will underperform NVIDIA’s Blackwell GPU, launched in late 2024, whose architecture offers significant advantages in terms of HBM3e memory, compute efficiency with fifth-generation Tensor Cores, and superior capabilities in generative AI workloads. By 2026, NVIDIA is expected to have already released the Rubin generation, with even greater performance, further deepening the gap.

Rivals in the cloud, such as Amazon with Trainium3 and Google with seventh-generation TPUs, are well ahead: the Trainium3 chip is expected by the end of 2025, while the TPU‑7 has been operational since April 2025, confirming that Microsoft remains behind in large-scale deployment. This leads to a continued dependence on the supply of NVIDIA solutions, with impacts on negotiation capabilities for Azure and on operating costs, given that NVIDIA chips guarantee not only superior performance but also efficiency and consolidated uptime.

The full picture shows how the race for proprietary AI chips involves strategic, technical and organizational obstacles: the internal project could undergo further adjustments in the roadmap and design with last-minute requests from OpenAI, already emerged in the reports on the development of Maia, causing instability and continuous revisions. In the absence of autonomous alternatives, Microsoft could evaluate intermediate solutions, further purchasing NVIDIA technology or resorting to external outsourcing, raising questions about the economic sustainability of its hardware strategy.

Braga’s delay thus becomes a mirror of the effort needed to compete in the AI ​​domain: developing a chip that can compete with established giants like NVIDIA proves to be a complex challenge, requiring a balance between innovation, organizational continuity and tight deadlines.

A final reflection: the future of Microsoft’s AI infrastructure will remain shaped by the evolution of internal projects like Braga, but also by the ability to integrate external solutions in a strategic and scalable way.