Big Bank 💩 On AI Infra; Meta Muses; China May Hide Models

Today's AI Outlook: 🌥️

The AI Boom Gets A Visit From The Ghost Of Bubbles Past

The Bank for International Settlements, basically the central bank for central banks, warned that the AI boom could curdle into a long bust if the enormous infrastructure buildout fails to produce matching returns. The concern is not that AI is useless; the concern is that history is full of useful technologies, from canals to railways to the dot-com internet, that attracted more money than they could repay on schedule.

Why it matters

This warning hits harder because it comes from the sober-suit crowd, not a Reddit thread with a chart and 11 flame emojis. Hyperscalers are expected to spend more than $1T through 2026, and if returns wobble, the freeze could reach beyond Silicon Valley because households hold large amounts of stock exposure.

The Deets

• BIS compared today’s AI buildout to earlier investment manias around canals, railways and dot-com companies.
• The concern centers on massive capex, cheap credit and uncertain payback timelines.
• AI infrastructure has become a market-wide bet, not just a tech-sector side quest.
• A weak quarter from a major AI spender could quickly turn optimism into tighter capital and harsher cost controls.

Key takeaway

The AI boom’s scariest ghost story is not that the tech fails. It is that the tech works, but the math still comes up wanting.

🧩 Jargon Buster - Capex: Money companies spend on major long-term assets, like data centers, chips and infrastructure.

⚡ Power Plays

China May Put Best Models Behind A Bigger Wall

Beijing is reportedly weighing new limits on foreign access to China’s strongest AI models, including systems from ByteDance, Alibaba and Zhipu AI. The talks reportedly cover both closed and open models, along with startup funding limits and penalties for leaking proprietary AI.

Why it matters

Chinese models have gained global traction because they are cheap, efficient and increasingly competitive. Restrictions would turn model access into another geopolitical pressure point, especially after the U.S. tightened oversight around frontier AI releases.

The Deets

• Officials reportedly met with ByteDance, Alibaba and Z AI about foreign-use limits.
• Models such as Qwen, Doubao and GLM-5.2 are central to the discussion.
• The potential rules could affect both open and closed models.
• The move would make AI access more vulnerable to political shifts on both sides of the U.S.-China divide.

Key takeaway

The model wars are getting a customs desk. Cheap, powerful AI may soon come with more borders.

🧩 Jargon Buster - Frontier model: A cutting-edge AI model near the top of current performance, usually expensive to build and strategically sensitive.

Microsoft Starts Counting The AI Bar Tab

Microsoft is cutting about 4,800 jobs, roughly 2.1% of its global workforce, while reportedly moving some Excel and Outlook AI tasks away from outside models and onto its own MAI systems. The company’s AI business is large, but the cost of running it may now be forcing a harder look at staffing, vendors and margins.

Why it matters

Microsoft has been one of the biggest winners of the AI boom, but even winners have to pay for GPUs, CPUs and inference. The company’s AI business reportedly has a $37B annual run rate, while last quarter’s capex hit $31.9B, with much of that tied to compute.

The Deets

• Job cuts are affecting sales, Xbox and other teams.
• Microsoft is reportedly using more in-house MAI models for Excel and Outlook prompts.
• The move could reduce reliance on OpenAI and Anthropic for some tasks.
• Cost control is becoming part of the AI strategy, not a back-office footnote.

Key takeaway

The AI boom has reached its spreadsheet era. Microsoft is still building, but now every prompt has to justify the bill.

🧩 Jargon Buster - Inference: The process of running an AI model after it has been trained, such as answering a prompt or generating an image.

🛠️ Tools & Products

Meta’s Muse Image Brings Leaderboard Heat And Privacy Heartburn

Meta released Muse Image, a free in-house image generator rolling out across Meta AI, Instagram and WhatsApp. The model opened at No. 2 on Arena’s text-to-image and editing leaderboards, trailing only OpenAI’s GPT Image 2, and Meta also teased a Muse Video model that ranked No. 3 in preview testing.

Why it matters

Meta has billions of users, massive creative surfaces and an ad business that loves anything that turns ideas into assets faster. But one feature also raised privacy concerns: users can reportedly tag a public Instagram user and pull that person’s photos into new AI images, with opt-out as the default and no notification.

The Deets

• Muse Image combines with Muse Spark for agentic features such as web search, tool use and self-editing.
• It is free in Meta AI and is rolling out across Instagram and WhatsApp, with Facebook, Messenger and Meta’s ad platform coming later.
• Meta previously leaned on outside creative AI tools from companies such as Midjourney and Black Forest Labs.
• The privacy debate centers on public Instagram photos becoming usable AI inputs unless users opt out.

Key takeaway

Meta cooked up a serious image model, then seasoned it with the company’s favorite spice: a privacy setting most people will never find.

🧩 Jargon Buster - Agentic: AI that can use tools, search the web or take steps toward a goal instead of only generating a single response.

Claude Cowork Wants To Manage The Manager

Claude Cowork is being pitched as a way to improve workplace one-on-ones by turning past transcripts into templates, rubrics and performance dashboards. Anthropic is also adding the platform to mobile and web in beta, which gives users more ways to continue tasks across devices.

Why it matters

AI is moving from drafting memos to shaping management rituals. That can help managers create more consistent conversations, but it also turns routine workplace check-ins into another data-heavy AI workflow.

The Deets

• Users can create a Claude Cowork project and upload past one-on-one transcripts.
• Claude can generate a call template and evaluation rubric tailored to a manager’s team.
• Teams can update templates as new transcripts are added.
• The tool can also be used to build a performance dashboard.

Key takeaway

Your next one-on-one may come with a rubric, a transcript and a chatbot quietly judging everyone’s follow-through.

🧩 Jargon Buster - Rubric: A structured scoring guide used to evaluate performance or progress against defined criteria.

💰 Funding & Startups

Figma Picks Up Bud’s Agent-Building Talent

Figma acqui-hired the team behind Bud, formerly known as Orchids, an AI app-builder and agent platform. The move gives Figma more AI product-building firepower as design tools increasingly blur into prototyping, workflow automation and app creation.

Why it matters

Design software is becoming more competitive as AI turns prompts into interfaces, apps and working prototypes. Figma does not need every feature to be homegrown when it can bring in teams already building at the edge of agentic software.

The Deets

• Bud focused on AI app-building and agent workflows.
• Figma is bringing in the team through an acqui-hire.
• The move fits the broader trend of design platforms adding more AI-native creation tools.

Key takeaway

Figma is staffing up for a world where designers do not just mock up apps. They prompt them into existence.

🧩 Jargon Buster - Acqui-hire: A deal where a company buys another company mainly to hire its team, not necessarily to keep the product alive.

DoorDash Builds A Benchmark For Its AI Code Reviewer

DoorDash created DashBench, an internal benchmark that tests AI code reviewers against 105 past company code changes. The goal is simple and very practical: count not just what the AI flags, but what it misses.

Why it matters

AI code review sounds great until the missed bugs become production problems. DoorDash’s work shows that companies may need their own benchmarks to safely use AI inside critical engineering workflows.

The Deets

• AI models reviewing code alone caught about 20% to 30% of problems.
• A pairing of two Claude models found just over half.
• A combination of Kimi K2.6 and Claude Fable 5 caught about two-thirds of issues and 8 of 10 critical bugs.
• That setup reportedly cost $3.81 per review.
• DoorDash co-founder Andy Fang said the benchmark gives the company confidence to use open models in the pipeline.

Key takeaway

DoorDash is treating AI code review less like magic and more like QA. That is how the grown-ups enter the room.

🧩 Jargon Buster - Benchmark: A test used to compare how well models or systems perform on specific tasks.

⚡ Quick Hits

• Vercel says the next developer fight is separating models from agents, so teams can swap intelligence layers without rebuilding workflows.
• SK Hynix will make it easier for U.S. investors to access its stock as AI memory demand keeps boosting chip suppliers.
• Google is using more consumer activity to train AI, with opt-outs spread across several surfaces.
• Nvidia’s Kyber AI rack system has reportedly been delayed to 2028 because of manufacturing issues.
• Discord admitted an AI moderation bug wrongly banned users over harmless images.
• Anthropic said Fable 5 will remain available in subscription plans until July 12 after initially planning to move it off plans and onto usage credits.
• Tufa Labs won ARC-AGI-3’s $37.5K milestone contest by wrapping a small Qwen model in an open-source coding agent.

🧰 Tools Of The Day

• Hy3: Tencent’s open-source model is aimed at cheap, reliable agent deployment.
• Willow Frontier Mini: A free, unlimited AI dictation model for turning speech into text.
• Gemini Managed Agents: Google’s agent API supports background tasks and remote tools.
• Claude Cowork: Anthropic’s workplace assistant can turn one-on-one transcripts into templates, rubrics and team dashboards.

Today’s Sources: The Internet, AI Secret, The Rundown AI