Microsoft Build 2026: Seven Proprietary AI Models and the Dream Machine

On June 3, 2026, Microsoft held its annual developer conference, Build 2026, where it announced seven proprietary large language models in the “MAI (Microsoft AI)” series. Covering diverse modalities such as reasoning, code generation, image creation, voice recognition, and transcription, these models signal Microsoft’s full-scale entry into foundational model development. Alongside this, Microsoft unveiled the Surface RTX Spark Dev Box, a local AI workstation for developers. NVIDIA’s CEO Jensen Huang emphasized in a video message, “The PC is evolving from a personal computer to personal AI,” supporting this transformative move.

Meanwhile, OpenAI introduced ecosystem expansions for its AI agent platform “Codex,” adding role-specific plugins, website sharing features, and annotation functionality to attract non-developer users. Tencent Cloud announced price cuts of up to 97.5% for its DeepSeek-V4 series, while Alphabet revealed plans to raise $80 billion for AI infrastructure investment. These developments showcase the dynamic shifts occurring across the AI industry.

In this article, we summarize the key announcements from Build 2026 alongside other significant news from the AI sector revealed on the same day.

Microsoft Unveils Seven Proprietary Models

The highlight of Build 2026 was Microsoft’s simultaneous launch of seven AI models in its new “MAI” series—marking the company’s transition away from relying heavily on OpenAI’s models to its own in-house development.

Here is an overview of the models:

• MAI Thinking 1: A reasoning-focused model employing sparse MoE architecture with active parameters of 35 billion and a total parameter scale of approximately 1 trillion. It supports a context size of up to 256,000 tokens.
• MAI Code 1 Flash: A code-generation model trained end-to-end by Microsoft, achieving a 51.2% score on SWE Bench Pro, surpassing Claude Haiku 4.5’s 35.2%. It is already available for GitHub Copilot’s individual users.
• MAI Image 2.5: An image-generation model integrated with PowerPoint, with plans for expansion into OneDrive.
• MAI Transcribe 1.5: A transcription model supporting 43 languages.
• MAI Voice 2: A voice model supporting 15 languages, capable of adapting to voices with short samples.

These models cover five key areas—reasoning, code generation, image creation, voice processing, and transcription—highlighting Microsoft’s aim to strengthen its platforms such as Copilot and Azure AI with proprietary technology.

The Developer’s “Dream Machine”

On the hardware front, Microsoft unveiled the local AI workstation Surface RTX Spark Dev Box, dubbed the “dream machine” by CEO Satya Nadella. Built on NVIDIA’s RTX Spark platform, the device promises unparalleled AI capabilities.

Key specifications:

• AI computation power: 1 petaflop
• CPU cores: 20
• Unified memory: 128GB
• Power consumption: 100W
• Release date: Fall 2026

This device enables developers to perform large-scale model fine-tuning and inference processing locally, reducing reliance on cloud-based solutions. With Jensen Huang emphasizing the “personal AI” concept, Microsoft’s strategy to transform PCs into AI workstations becomes increasingly evident.

Project Solara Reference Devices

Microsoft also unveiled two reference devices under the codename “Project Solara.”

• Desktop work terminal: Equipped with MediaTek chips, this device supports identity verification, voice interactions, and standard access to Microsoft 365 Copilot.
• Digital worker badge: Featuring Qualcomm’s wearable chips, this badge is designed for mobile work scenarios such as healthcare and on-site data collection.

While it remains uncertain whether these devices will be commercially available, their introduction aims to facilitate the development of AI-native devices by OEM manufacturers.

OpenAI Expands Codex Ecosystem

On the same day, OpenAI hosted the “Intelligence at Work” event, announcing major enhancements to its AI agent platform “Codex.” The company plans to integrate Codex’s core features directly into the ChatGPT app within weeks, allowing users to execute Codex capabilities within ChatGPT.

According to official data, Codex has over 5 million weekly active users, with non-developer users accounting for 20% and growing at a rate three times faster than developers. The new features aim to accelerate this trend.

The added functionalities include:

• Role-specific plugins: Six plugins designed for data analysis, creative production, sales, product design, stock market investment, and investment banking are integrated with 62 applications and 110 skills. Future expansions will include corporate finance, private equity, and legal domains.
• Sites: Previewed for enterprise and business users, this feature generates interactive web pages that can continuously update content, enabling sharing of analysis results or planning via URLs. Early integrations with platforms like Vercel, Figma, Replit, and Webflow have begun.
• Annotation functionality: Previously limited to code and websites, annotations now extend to documents, spreadsheets, and slides, allowing users to select areas and provide specific adjustment instructions for Codex to act upon.

These enhancements aim to evolve AI from chatbots into platforms capable of managing a wide range of knowledge work.

Tencent Cloud Slashes DeepSeek-V4 Prices by 97.5%

Tencent Cloud announced price reductions of up to 97.5% for its DeepSeek-V4 series models.

Detailed pricing adjustments:

• DeepSeek-V4-Pro: Input and output inference costs reduced by 75% to 0.003 yuan and 0.006 yuan per 1,000 tokens, respectively.
• Cache hit price: Reduced by 97.5% to 0.000025 yuan per 1,000 tokens, down from 0.001 yuan.
• DeepSeek-V4-Flash: Cache hit price adjusted to 0.00002 yuan per 1,000 tokens, a 90% reduction.

This dramatic price cut reflects the intense competition among Chinese cloud providers like Alibaba, Tencent, Huawei, and ByteDance. Reducing inference costs is crucial for accelerating the adoption of AI agents.

Alphabet Plans $80 Billion AI

Infrastructure Funding

According to Bloomberg, Google’s parent company Alphabet plans to raise a massive $80 billion through equity financing for AI infrastructure, computational capacity, and related capital expenditures.

The funding will consist of:

• $30 billion through underwritten public offerings.
• $40 billion through at-the-market offerings of Class A or C shares.
• $10 billion in direct investment via a private placement by Berkshire Hathaway.

Alphabet also intends to allocate part of the revenue from its ATM program to cover tax obligations related to employee stock compensation. With its capital expenditure forecast for this year already raised to $180–190 billion, the company expects sustained investment growth in AI infrastructure next year.

This unprecedented scale of funding underscores the capital-intensive nature of AI infrastructure investment in the tech industry.

Other Noteworthy News

• Possible Leica Camera Acquisition: Bloomberg reported that Asian investment entity HSG (affiliated with Sequoia Capital China) is leading bids to purchase Blackstone’s 45% stake in Leica Camera AG. If Austrian billionaire Andreas Kaufmann, who owns the remaining 55%, decides to sell, HSG may acquire full ownership. Leica’s valuation is estimated at €1 billion ($1.2 billion).
• ByteDance Robotics Team Restructuring: According to “LatePost,” ByteDance has reorganized its robotics team under the leadership of Zhou Chang. This move appears to align robotics development more closely with the company’s core management, reflecting ByteDance’s focus on integrating AI with robotics.

Editorial Insights

Short-Term Impact: Microsoft’s entry into proprietary model development is a significant step towards reducing reliance on OpenAI. The competitive performance of MAI Code 1 Flash, which outperformed Claude Haiku 4.5, hints at intensifying competition in code generation. Microsoft’s introduction of the Surface RTX Spark Dev Box may prompt a shift from cloud reliance to local inference.

Long-Term Perspective: Alphabet’s $80 billion funding plan signals that the AI infrastructure race has become a constant battle rather than a one-time event. Over the next 1–3 years, competition among cloud providers is likely to drive prices down further, reducing barriers to AI adoption. Microsoft’s proprietary model strategy could diversify the market, breaking the current oligopoly. OpenAI’s focus on non-developer users indicates a broader application of AI, extending well beyond software engineering to other areas of knowledge work.

Editorial Questions: Tencent Cloud’s 97.5% price reduction raises questions about the sustainability of such aggressive pricing in a fiercely competitive market. Will Chinese cloud providers maintain profitability under these conditions? Additionally, Microsoft’s shift towards proprietary models prompts speculation about its relationship with OpenAI—will their collaboration persist or evolve into a rivalry? These developments will be closely observed in the latter half of 2026.

References

• ifanr: Microsoft Releases Windows “Dream Machine”/Tencent Cloud Cuts DeepSeek-V4 Prices by 97.5%/Leica Camera May Be Acquired — Published June 3, 2026