Wednesday, June 3, 2026
Microsoft announces a seven-model MAI family backed by a rare, transparent training report; Alphabet raises about $80 billion, including Berkshire's first big Google stake, to fund the compute race.
Microsoft becomes a frontier lab
Microsoft used its Build conference to announce seven in-house MAI models, led by the reasoning model MAI-Thinking-1, casting itself as a first-party frontier lab rather than a reseller of OpenAI’s and Anthropic’s. Microsoft says MAI-Thinking-1 is a mixture-of-experts model with about 1 trillion total parameters and 35 billion active, a 256K-token context window, pre-trained on 30 trillion tokens across 8,192 Nvidia GB200 GPUs. Its headline scores are self-reported: 97% on AIME 2025, 53% on SWE-Bench Pro, and blind human raters at Surge preferring it to Claude Sonnet 4.6, with no methodology released.
The durable artifact is a 109-page technical report that researchers called the most transparent at this scale. It discloses the scaling-ladder metric Microsoft used to promote model sizes, model-FLOPs-utilization figures, data curation, and an explicit no-distillation, no-synthetic-data design. One disclosed recipe begins reinforcement learning from a checkpoint with no prior reasoning exposure, with AIME 2025 accuracy rising from under 20% to over 95%.
The transparency has limits. Reading the paper, Simon Willison found Microsoft’s claim of “clean and appropriately licensed data” overstated: training drew on a proprietary crawl of about 1.2 trillion pages filtered to 794 billion, plus 24.2 billion Common Crawl pages, the same web-crawl posture as its peers.
Beyond models, Microsoft paired the family with its own MAIA 200 silicon, which it claims delivers 30% more performance per dollar and 1.4x the performance per watt of the GB200, plus a GitHub Copilot desktop app and Windows agent APIs. MAI-Code-1-Flash, a 137-billion-parameter code model with 5 billion active, is rolling out to Copilot and VS Code users.
Google’s $80 billion question
Alphabet is raising about $80 billion in equity, Ben Thompson writes at Stratechery, citing Bloomberg’s account of one of the largest equity deals ever: a $40 billion at-the-market program slated for Q3 2026, $30 billion in underwritten and mandatory-convertible offerings, and a $10 billion investment from Berkshire Hathaway, its first major Google stake, struck near the stock’s record high.
Thompson reads the raise as Alphabet turning from an asset-light advertising business into a capital-intensive “capital company” funding its AI and cloud buildout. His puzzle: why issue equity when Google holds about $126 billion in cash against $81 billion in debt and has room for tax-advantaged borrowing? His answer is that large debt issuance is also coming, because compute demand is underestimated, or that Google wants to share the risk on its capital spending. Google Services ran a 45% operating margin last quarter; the faster-growing Cloud, at $20.0 billion in revenue and a 33% margin, is the capital-hungry engine the raise feeds. The thesis: if compute supply runs short, the firm with the most cash capacity buys the most compute, and Google’s in-house TPUs give it a cost edge if compute commoditizes. This is interpretation, not disclosed rationale.
Local computer-use agents
H Company released Holo3.1, an open-weight family of computer-use agents based on Qwen in four sizes, from 0.8 billion to 35 billion parameters, its first with quantized checkpoints (FP8, Q4 GGUF, NVFP4) for local execution. The company says its NVFP4 weights run 1.74x the throughput of BF16 on an Nvidia DGX Spark while scoring about 2 points lower on the OSWorld benchmark, and that the 35-billion-parameter model’s AndroidWorld score rose from 67% to 79.3%. The benchmarks are vendor-defined, with no methodology released.
What to watch today
- MAI-Code-1-Flash rolling out to GitHub Copilot and VS Code, the first MAI model in general developers’ hands.
- Independent replication of MAI-Thinking-1’s AIME 2025 and SWE-Bench Pro scores, now that access is limited to select early partners.
- Alphabet’s $40 billion at-the-market program, slated for Q3 2026, and whether the debt raise Thompson expects follows.
- Third-party OSWorld and AndroidWorld numbers for Holo3.1’s quantized weights.