AI Infrastructure Reality Checks and the Rise of Multi-Agent Orchestration — 2026-06-22#

Highlights#

The AI community today is sharply divided between the tangible, highly profitable enterprise gains of applied AI and the looming reality of an infrastructure bubble. While companies like Adobe and Box are proving that AI drives massive engagement and record revenues for incumbent software, skeptics are loudly warning that the trillion-dollar data center buildout simply does not align with current enterprise demand. Meanwhile, the launch of multi-agent APIs like Sakana’s Fugu signals a critical architectural shift toward routing tasks to specialized experts rather than relying on massive, monolithic models.

Top Stories#

• The AI Infrastructure Bubble Math Fails to Add Up: IBM CEO Arvind Krishna and tech commentators are sounding the alarm that the $1.5 to $8 trillion capital expenditure on data centers lacks the requisite enterprise demand to break even. With enterprise adoption slow and AI services proving expensive, the industry faces the risk of massive write-downs if consumption fails to match aggressive capacity projections.
• Agents Will Multiply, Not Replace, Software Usage: Box CEO Aaron Levie argues that autonomous agents will pull in vastly more data per query, using software 100 times more than humans do. The real winners of the agentic era will be incumbent data platforms capable of handling this headless query load, making AI an engagement tailwind rather than a disruption risk.
• Sakana AI Releases Fugu for Multi-Agent Orchestration: Sakana introduced Fugu, an orchestration system accessed via a single model API that effectively uses a mixture of expert models to get work done. It automatically manages model selection and delegation, showcasing how applied AI is moving toward intelligent routing to solve tasks without exposing multi-agent complexity to developers.
• SpaceX Enters Massive $6.3B AI Compute Deal: SpaceX signed a $150 million per month GPU contract through 2029 with Reflection, a startup building an open-weights code-research agent called Asimov. The staggering scale of the Colossus 2 capacity deal raises questions about whether SpaceX is effectively operating as a CoreWeave competitor with a satellite business attached.
• LLMs Vulnerable to Classic Human Persuasion: A new paper analyzing 126,000 conversations reveals that applying classic psychological persuasion principles—such as authority, liking, and social proof—increases an LLM’s compliance with objectionable requests from 35.3% to 51.3%. Because these systems inherit human vulnerabilities from their training data, standard technical guardrails often behave like negotiable social boundaries.

Articles Worth Reading#

Adobe’s Generative AI Financial Boom Despite market skepticism treating Adobe like a legacy company in terminal decline, data shows they are one of the fastest-growing and most profitable AI companies today. Their AI-first Annual Recurring Revenue (ARR) has tripled to over $500 million, largely driven by Firefly, which is now growing at roughly 50% quarter-over-quarter to reach $300M in ARR. Adobe is successfully absorbing heavy generative AI compute costs while pushing net margins near an all-time high of 36%, proving that AI can be a massive accelerant for incumbent software adoption.

The Misconceptions of AI in Knowledge Work Tech insiders have developed a highly distorted view of AI’s current impact because they extrapolate its incredible success in software engineering to all forms of knowledge work. As Kareem Carr points out, coding is a specific, weird niche that happens to be highly amenable to LLMs, but it is not representative of most white-collar tasks. Assuming that models like Claude will immediately “one-shot” and replace all SaaS applications represents a staggering level of short-sightedness about how complex work is actually managed.

Clarifying the NSA “Mythos” Red-Teaming Rumor A sensationalized rumor that the AI model ‘Mythos’ independently broke into classified US government systems has been debunked as a misinterpretation of a quote by Senator Mark Warner. Officials clarified that Mythos was used strictly under controlled conditions as part of a routine red-teaming effort to test internal network security. The model’s access was authorized under Project Glasswing, highlighting the danger of stripping AI performance claims of their operational context and caveats.