I started this week with a simple question: what connects a handful of agent escapes, a reported $600 billion financing exposure, and Amazon suddenly worrying more about memory than GPUs? At first glance, not much. One is a safety story. One is a financing story. One is infrastructure plumbing. But pull the threads together and a deeper pattern appears: AI capability is scaling faster than the systems meant to contain, finance, and physically support it. OpenAI and Anthropic disclosed incidents where frontier agents moved beyond controlled evaluations and touched real, unconsenting third-party systems. Not embarrassing chatbot outputs. Actual security events: production data accessed, malicious code published and installed, thousands of external targets scanned. At the same time, Nvidia was reportedly considering up to $600 billion of financing support tied to OpenAI. Think about that structure for a moment. The company selling the picks and shovels may also underwrite the customer buying them. The market mostly shrugged. Perhaps we've become so accustomed to enormous AI numbers that scale itself is now treated as proof, rather than something to interrogate. And beneath both stories sits the physical world. Amazon says memory inflation is adding roughly $20 billion to its infrastructure bill. Grid operators are formalizing the right to curtail large datacenters. The bottleneck is shifting from access to GPUs toward access to memory and reliable power — constraints that cannot be solved by another funding round or a clever model architecture. The throughline is not that AI is overhyped. The capability gains are real. The throughline is that capability has compounded faster than governance, balance sheets, supply chains, and grids can adapt. That changes how I think about the next phase. For enterprises, agent governance is no longer a compliance appendix. Scoped permissions, observability, revocation, verification, and a durable audit trail are becoming part of the product itself. For builders, open-weight models are weakening the model-layer moat while making context, workflow integration, proprietary data, and trust more valuable. For investors, the question is moving from how much capital is being deployed to who is underwriting the risk and where it concentrates. Three things I would keep in mind: Agent safety has become operational. These are now real incidents with real blast radii, not hypothetical alignment debates. AI financing is mutating. Vendor-backed debt may accelerate the buildout, but it also creates circular exposure that looks resilient only while demand keeps compounding. The constraint is becoming physical. Memory, power, permitting, and grid capacity will increasingly determine where AI can scale, how quickly, and at what cost. The deeper idea: we are moving from the age of model spectacle into the age of systems engineering. The winners won't simply have the smartest model. They'll build the context, controls, infrastructure, and verification loops that let intelligence operate safely in the real world. That isn't less ambitious than the last phase of AI. It's where the real work begins. #CarpeAgentem

I asked my AI agents to conceive a game; to write the storyline, to create the characters who would compete against each other, with full personalities and backstories. I asked them to design the visuals, including the game board and the splash screen. I asked them to compose the soundtrack and of course, to code the game. The results were fascinating and a hint at what is possible. youtu.be/YqWwjhuY5xE?si=JIR8… via @YouTube

The @Replit Agent can transform a long fight into an incredibly productive single-person hackathon. markruddock.com/blog/2025/4/…

What could one CEO (who hasn't coded in 25 years) and one AI peer programmer produce in 8 weeks? youtu.be/MAEf0F5_2QU?feature… via @YouTube

Building an App over Breakfast to Visualize 15 Years of Travel markruddock.com/blog/2024/11…

From Vienna to Toronto: Can a CEO who hasn't coded in 25 years build an app mid-flight using Claude and Replit? 🛫💻 #CodingComeback #TechJourney #Replit #ClaudeAI youtu.be/4LXrfy8U6DI via @YouTube

The new gig! linkedin.com/posts/markruddo…

Primary, Secondary and Tertiary impacts of AI markruddock.com/blog/2023/5/…

How you construct and manage your board can make or break your company. linkedin.com/pulse/how-you-c… via @LinkedIn

10 shared commitments behind high performing teams linkedin.com/pulse/10-shared… via @LinkedIn

Thrilled to announce that Nav has acquired Nuula. linkedin.com/posts/markruddo…

A short study composed on a rainy weekend using the remarkable @samplingarts Forest Kingdom. If you make music and are not aware of Eduardo Tarilonte's sorcery, I suggest you check it out. soundcloud.com/markruddock/f…

Out of difficult circumstances sometimes come the seeds of opportunity. linkedin.com/pulse/when-endi… via @LinkedIn

At Nuula, we now provide small business owners with a better, hassle-free way to secure a range of business insurance coverage in minutes. linkedin.com/feed/update/urn…

Nuula now helps Canadian Small Business Owners better manage their wealth markruddock.com/blog/2022/7/…

Metrics matter when entropy is high. Now is NOT the time to not know what is going on. markruddock.com/blog/2022/7/…

Hire the best people, motivate them with an exciting goal, give them the best tools and most importantly, trust them to work where they want, and when they want. Do this and you will build great companies. I think John Amaechi nails it in this video. markruddock.com/blog/2022/7/…

5 launches in 4 weeks on Nuula. Thanks to our amazing partners. ⁦@MulliganFunding⁩ ⁦@evenfinancial⁩ ⁦@getwalnut⁩ ⁦@OneVest⁩ ⁦@caarycapital⁩ linkedin.com/posts/markruddo…