1. Nvidia and Wall Street Build a $500B AI Financing Platform — Substack

  • Why read: How Wall Street is turning AI data centers into securitized assets like power plants.
  • Summary: Nvidia signed MOUs with Apollo, BlackRock, and Blackstone to raise $500 billion for AI data centers. The plan is to finance computing equipment like aircraft or power plants, backed by steady GPU lease cash flows. Even older A100 chips maintain high utilization, making them securitizable. This creates a new asset class for pensions and sovereign wealth funds. For context on the scale, building a 100 MW data center takes up to 3 million labor hours.
  • Read more

2. Anthropic’s $6B Push to Lower AI Costs — Substack

  • Why read: Anthropic's strategy to cut compute costs and expand into robotics through a massive acquisition.
  • Summary: Anthropic is in talks to buy Decart for $6 billion. Decart builds real-time world models and software that improves chip efficiency, which would cut Anthropic's training and inference costs. Integrating Decart's hardware-agnostic stack allows Claude to run better across Nvidia, Google, Amazon, and AMD chips. Decart's visual models also offer a path into video and robotics. This would be Anthropic's biggest acquisition, signaling aggressive growth before an anticipated $2T IPO.
  • Read more

3. The App Layer for AI Continues to Expand — Substack

  • Why read: Tracking the massive capital flowing into AI developer tools and software generation.
  • Summary: Money is pouring into the AI app layer, especially developer tools. River AI raised $1.1 billion to help companies fine-tune open-weight models. Lovable, which generates web apps from text prompts, raised $400 million at a $13.3 billion valuation. Devin creator Cognition is targeting a $40 billion valuation. Blacksmith raised at a $550 million valuation to validate AI-generated code.
  • Read more

4. Revising — proofofconcept.pub

  • Why read: What Michael Mann's film "Heat" teaches founders about iterating on a core product thesis.
  • Summary: Michael Mann's film "Heat" was a revision of his shorter TV movie "L.A. Takedown." For founders, this shows the value of shipping a compressed version to test a concept. "L.A. Takedown" proved the chase dynamic worked but lacked character depth. Iteration means diagnosing exactly what succeeded and failed in the first version, then adjusting the specs while holding onto the core thesis.
  • Read more

5. Terafab is bringing its own power plant(s) — Substack

  • Why read: Why SpaceX is building a $16.8B chip factory entirely off the grid.
  • Summary: SpaceX and Tesla's new $16.8 billion Texas chip campus, Terafab, will run entirely off-grid. It relies on on-site natural gas plants and battery storage. Leading-edge fabs need 400-700 megawatts of steady power, and waiting for a grid connection takes years. By generating its own power, SpaceX avoids grid delays and aims to produce chips by late 2027.
  • Read more

6. Anthropic gets a landlord (two, actually) — Substack

  • Why read: Anthropic's creative infrastructure deals to lock down power and data centers.
  • Summary: Anthropic partnered with Macquarie Asset Management and GIC to launch Theseus Infrastructure, a platform to build and lease data centers. The funds provide project equity, and Anthropic is the anchor tenant, agreeing to cover any consumer electricity price hikes caused by the sites. Anthropic also signed a 20-year lease with crypto miner Riot Platforms for 191 MW in Texas. AI labs are turning to crypto miners because they already hold grid interconnections.
  • Read more

7. 2/3 of the gigawatts are imaginary — Substack

  • Why read: The truth about AI power demand: most grid requests are phantom load.
  • Summary: Wood Mackenzie audited 1,066 gigawatts of data center power requests on US grids and found only 28% will likely be built. The rest is phantom load from speculative land grabs and duplicate applications. Many proposed projects lack financing. This overestimation confuses grid planners and hides the actual energy requirements of the AI sector.
  • Read more

8. AI security saga, week 4: arming the defenders — Substack

  • Why read: The rapid deployment of specialized AI models to defend against cyber attacks.
  • Summary: After finding vulnerabilities in unreleased models, OpenAI expanded its cyber defense program, Daybreak. It gave vetted blue and red teams access to GPT-5.6-Cyber, a model trained for defensive security. OpenAI warned that defenders have a closing window to prepare for AI-driven attacks. This follows a similar program by Anthropic, showing the industry's rush to prevent its tools from being weaponized.
  • Read more

9. Muse Glimmer: Meta goes open again (manifesto included) — Substack

  • Why read: Meta's new open-weight model optimized for agents on consumer GPUs.
  • Summary: Meta released Muse Glimmer, a 30-billion parameter agentic model with open weights under an Apache 2.0 license. It is sized to run on a single consumer GPU. The model includes native vision for reading screens and uses speculative decoding for fast text generation. It is trained end-to-end to handle agent loops and recover from failed tool calls.
  • Read more

10. “The Future is for Everyone”: the Zuck manifesto — Substack

  • Why read: Mark Zuckerberg's argument against closed AI systems and his plan for open-source superintelligence.
  • Summary: Mark Zuckerberg published a 6,500-word manifesto arguing that concentrating superintelligence in closed institutions is society's biggest risk. He advocates for "personal superintelligence," where AI serves users rather than corporations. Meta plans to release weights for its Muse Spark 1.2 model, making it the largest American open-weight model. Meta also announced a $1 billion fund for communities hosting its data centers.
  • Read more

11. Also shipping: agent workhorses everywhere — Substack

  • Why read: The flood of cheap, reliable models built to power continuous AI agents.
  • Summary: AI labs are releasing inexpensive models designed for continuous agent workflows. Nvidia launched Nemotron 3.5 Lightning, a 30B mixture-of-experts model, and a router to assign tasks. xAI released Grok 4.6, updated for task persistence. Google introduced Gemini 3.7 Flash at half the cost of its predecessor, targeting coding and agent applications. These models lower the cost of deploying autonomous agents.
  • Read more

12. B2B For Physical Products Is Crushing It — SaaStr

  • Why read: Why software tied to physical goods is outperforming digital SaaS.
  • Summary: Pure digital SaaS companies face AI disruption, but software servicing the physical world is growing. Shopify grew 34%, Toast 23%, and Samsara 30%. Their business models handle real-world transactions and hardware that AI cannot easily replace. They also do not rely on seat licenses, which autonomous agents might make obsolete.
  • Read more

13. The Zombie Executives of B2B + AI — SaaStr

  • Why read: The risk of hiring legacy tech executives in fast-moving AI startups.
  • Summary: B2B AI startups are hurting themselves by hiring "recycled" executives from legacy tech giants. These leaders often lack the adaptability for the current AI landscape. They rely on old playbooks, over-delegate, and add bureaucracy that slows product iteration. This approach slows growth and drives away top engineering talent. Founders need flexible leaders with current technical skills, not just prestigious resumes.
  • Read more

14. Stop chasing success — Substack

  • Why read: Why personal development compounds faster than forcing goals into reality.
  • Summary: Drawing on Jim Rohn's advice, the author argues that success is attracted by who you become, not chased. Deep relationships and insights come from accumulated trust and continuous learning. Over time, opportunities naturally find those with the capacity to handle them. Instead of asking how to force a goal, ask who you need to become to achieve it.
  • Read more

15. OpenAI’s Head of Design: This is the best time in history to be a designer | Ian Silber — Substack

  • Why read: How AI makes design the new bottleneck and biggest opportunity in product development.
  • Summary: Ian Silber, OpenAI's head of product design, argues AI makes this the best time to be a designer. While AI multiplied engineering output, design hasn't seen the same acceleration. He advises designers to do less and focus on user understanding and building a point of view—areas where human intuition beats AI. As technical execution becomes a commodity, human taste and invention matter more.
  • Read more

Themes from yesterday

  • Massive infrastructure investments: AI companies are skipping the grid to build their own power plants and raising billions to finance data centers.
  • Open vs. closed models: Meta's open-weight models and Zuckerberg's manifesto directly challenge the closed strategies of companies like OpenAI.
  • B2B and the AI divide: Software tied to physical goods is thriving, while traditional SaaS companies face disruption and must update their leadership.