In partnership with
In Today’s Issue:
🤑 Microsoft turns Build into an agent-first stack
💵 OpenAI pushes Codex into role-specific workflows
📈 MAI-Thinking-1 shows Microsoft’s in-house model ambitions
📉 Tennessee makes big data centers pay for power infrastructure
✨ And more AI goodness…
⚡ The Signal
Microsoft Build 2026 gave developers a map of Microsoft’s agent architecture.
Microsoft used the event to show the stack it wants around agents: in-house MAI models, a context layer in Microsoft IQ, secure execution containers for agent work, Project Solara devices, Discovery for science, and even quantum hardware progress through Majorana 2. The strategy is plain: Microsoft does not want Office to be a wrapper around someone else’s model. It wants models, tools, operating boundaries, enterprise context, devices, and infrastructure to reinforce one another.
All the best,
Kim Isenberg
🛠️ OpenAI Turns Codex Into a Work OS
OpenAI says more than 5 million people now use Codex every week, and non-developers already make up about 20% of overall Codex users while growing more than 3x as fast as developers. The commercial pull is just as steep: The Information reports that revenue from enterprise Codex customers has been growing about 50% week over week, with overall usage up around 5% day over day. The update adds six role-specific plugins, 62 popular apps, 110 skills, Sites for shareable internal apps, and annotations for refining Codex output in place.
👉 tl;dr: Codex is moving beyond coding into role-specific knowledge work, with plugins and shareable sites as the new surface area.
💾 AI RAM Crunch Hits PC Builders
Tom’s Hardware reports that 32GB DDR5 kits now start at $375, with lower-priced kits disappearing as AI demand keeps squeezing memory supply. The consumer side effect is ugly: ordinary PC builds and upgrades are now paying for the same memory crunch powering the data-center boom.
👉 tl;dr: AI infrastructure costs are spilling into consumer hardware, turning RAM from a cheap upgrade into a serious budget line.
⚡ Tennessee Makes Data Centers Pay for the Grid
Tennessee’s new law requires data centers with at least 50 megawatts of peak demand in their first three years to pay for their own electricity infrastructure. The bill’s backers say ratepayers should not carry the cost of AI-scale power upgrades, while critics worry utilities can still spread some costs when upgrades benefit the wider system.
👉 tl;dr: AI infrastructure is becoming a ratepayer fight, not just a cloud-capacity story.
Ask your AI assistant to turn any big platform keynote into four buckets: model capability, product surface, developer infrastructure, and physical infrastructure.
Why it helps: Microsoft Build is a clean example. If you only track the model names, you miss the real strategy: context layers, secure execution, agent devices, enterprise workflows, and the power/compute base underneath.
Try this: Paste a launch recap and ask: “Separate the announcements into model upgrades, user-facing products, developer/infrastructure changes, and physical infrastructure dependencies. For each bucket, explain what changes for a startup building on this ecosystem.”
🎬 Watch This
Project Solara: A new vision for agent-first computing shows Microsoft’s attempt to move agents off the laptop and into dedicated devices. The video centers on a chip-to-cloud platform, including an access badge built with Qualcomm silicon and a desk-style agent device tied to workplace context. Microsoft’s bet is that agents can get more autonomy if the hardware, identity layer, and enterprise controls stay inside systems it can govern.
“The goal here is to build what we think of as a hill-climbing machine: an organization that can continuously improve, cycle after cycle, as we apply more compute, better data, and sharper evaluation.”
– Mustafa Suleyman, Microsoft AI; Building a hill-climbing machine: Launching seven new MAI models
The dustup sitting next to Microsoft’s Build week is a policy one: on June 2, President Trump signed an executive order asking AI companies to voluntarily give the government early access to "covered frontier models" up to 30 days before release, so federal agencies can assess cybersecurity risks. The move follows alarm over Anthropic’s Claude Mythos, which can reportedly find software vulnerabilities at unprecedented speed, prompting Anthropic to limit it to a small set of trusted partners and Treasury and Fed officials to brief Wall Street CEOs on the risk. Participation is voluntary, so the real question is how many labs opt in. If they do, frontier-model launches start to look more like security reviews than ordinary product drops.
Microsoft Build Became the Agent Stack Keynote
The Takeaway
👉 Microsoft announced a seven-model MAI family, including MAI-Thinking-1, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5, and MAI-Voice-2.
👉 Project Solara points to agent-first devices, including an access badge built with Qualcomm silicon and a desk-style device built with MediaTek silicon.
👉 Microsoft IQ adds Work IQ, Web IQ, Fabric IQ, and Foundry IQ as context plumbing for agents across enterprise and web knowledge.
👉 Discovery, secure execution containers, GitHub Copilot updates, and Majorana 2 show Microsoft trying to own the agent stack from model to infrastructure.
Microsoft Build 2026 was packed, but the through-line was simple: Microsoft wants agents to have their own stack. The company used the keynote to introduce a broader MAI model family, led by MAI-Thinking-1, and to show how those models could plug into Foundry, GitHub Copilot, VS Code, PowerPoint, OneDrive, and enterprise workflows. Mashable framed the day as Microsoft reducing its reliance on OpenAI; our live coverage showed the more practical version of that story, with models, devices, workplace context, developer tools, and security controls arriving together.
The most concrete agent hardware idea was Project Solara. I would describe it as a chip-to-cloud platform for AI agent devices, including an access badge built with Qualcomm silicon and a stationary MediaTek device tied to workplace context. That context comes from Work IQ, part of the bigger Microsoft IQ layer, alongside Web IQ, Fabric IQ, and Foundry IQ, which Microsoft pitches as the context layer grounding agents in workplace data, live web knowledge, semantic data spaces, and retrieval.
The model news is an important step because Microsoft is no longer only presenting Copilot as a wrapper around outside frontier models. MAI-Thinking-1 is Microsoft AI’s first reasoning model, MAI-Code-1 is tuned for GitHub and available in Copilot and VS Code, MAI-Image-2.5 is available in PowerPoint and Foundry, and MAI-Transcribe-1.5 supports 43 languages. Add Microsoft Discovery for scientists, Microsoft Execution Containers for safer agent execution, and Majorana 2’s claimed jump in qubit lifetime, and the keynote becomes a stack map rather than a product list.
Why it matters: Microsoft is trying to make the agent race about integration: models, context, tools, devices, security boundaries, and compute. That is a harder story to copy than a single model release, but it also raises the execution bar across almost every Microsoft product surface.
Sources:
🔗 https://mashable.com/tech/microsoft-build-2026-keynote-everything-we-learned
🔗 https://www.cnet.com/news-live/microsoft-build-2026-news-ai-copilot/
🔗 https://microsoft.ai/news/building-a-hillclimbing-machine-launching-seven-new-mai-models/
Someone just spent $236,000,000 on a painting. Here’s why it matters for your wallet.
Late last year, a Klimt sold for the highest price ever paid for modern art at auction.
An outlier sure, but it wasn't a fluke. U.S. auction sales grew 23.1% in 2025. The $1-5mm segment even grew 40.8% YoY.
Meanwhile, Apollo’s chief economist Torsten Slok said to expect ‘zero in return in the S&P 500 over the coming decade.’
Each environment is unique, but after dot-com, post war and contemporary art grew about 24% annually for a decade. After 2008, about 11% for 12 years.
It’s also had near-zero correlation with the S&P 500 since ‘95.*
Now, Masterworks lets you invest in shares of artworks featuring legends like Banksy, Basquiat, and Picasso.
$1.3 billion invested across over 500 artworks.
28 sales to date.
Net annualized returns on sold works held 12 months+ like 14.6%, 17.6%, and 17.8%.
Shares can sell quickly, but my subscribers can skip the waitlist:
*Investing involves risk. Past performance is not indicative of future returns. See important Reg A disclosures at masterworks.com/cd.
The chart: MAI-Thinking-1 benchmark results put Microsoft’s 35B-active / ~1T-total sparse MoE reasoning model in the same conversation as much larger frontier systems. Microsoft reports 97.0% on AIME 2025, 94.5% on AIME 2026, 87.7% on LiveCodeBench v6, and 52.8% on SWE-Bench Pro.
The lesson: Microsoft is arguing that the important metric is not raw size, but whether a model can deliver strong reasoning and coding performance at an inference footprint enterprises can actually use every day.
The caveat: The chart is Microsoft’s own presentation of public benchmarks and side-by-side evaluations. Useful signal, but still a launch-day benchmark view rather than independent production evidence.
MAI-Thinking-1 Is Microsoft’s Model-Factory Thesis
⚡ Bottom line: Microsoft’s MAI-Thinking-1 paper uses one model to explain a repeatable “hill-climbing machine” for improving models through data, RL, evaluation, and infrastructure.
💡 Why it matters: Microsoft is trying to prove it can build serious reasoning models in-house while still distributing AI through OpenAI-backed products.
🔎 What it means: The model is a 35B-active / 1T-total sparse MoE trained from scratch, with no third-party model distillation, on 30T pre-training tokens plus 3.55T mid-training tokens, then pushed through STEM, coding, helpfulness, and safety RL climbs.
The MAI-Thinking-1 technical report frames progress as a systems problem. Microsoft describes the model as the first output of a process that combines in-house data pipelines, training infrastructure, reinforcement-learning environments, rewards, evaluations, and safety tests into a loop for repeated improvement.
The technical payload is substantial. MAI-Base-1 is a 35B-active / 1T-total sparse Mixture-of-Experts model trained on 8K GB200 GPUs in Azure. The paper says pre-training used 30 trillion tokens from publicly available and licensed human-generated data, followed by 3.55 trillion mid-training tokens, while avoiding synthetic pre-training data generated by language models and avoiding third-party distillation.
The headline numbers are strong for the model’s size: 52.8% on SWE-Bench Pro, 97.0% on AIME 2025, 94.5% on AIME 2026, and 87.7% on LiveCodeBench v6. Microsoft also says the model matches Claude Opus 4.6 on SWE-Bench Pro coding and is competitive with Sonnet 4.6 across a wide range of benchmarks, with professional raters from Surge, its independent rating partner, preferring it over Sonnet 4.6 in blind side-by-side tests.
The sharper claim is methodological. Microsoft is betting that clean data, learned capabilities, enterprise controls, and an internal RL/infrastructure loop can become a durable advantage. The open question is whether that loop keeps climbing once the model moves from launch benchmarks into messy enterprise workloads.
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.