The Deep Feed

01 — Not Boring

The Meta-Alpha Product: Why Data Centres are the New Industrial Engine

How the hunger for compute is accidentally funding the next generation of physical technology.

By Packy McCormick · 8 min read

Editor's note: The massive capital expenditure in AI infrastructure might be the accidental subsidy the physical world needs.

The public reaction to AI data centres is almost universally negative. People see them as energy gluttons, land-grabbers, and environmental liabilities. This view is understandable, but it misses a secondary, perhaps more significant, economic effect. We are witnessing the rise of a new kind of buyer: the extraeconomic buyer of capabilities. Historically, breakthroughs in difficult, expensive technologies like advanced nuclear reactors or high-voltage grids required a massive initial customer willing to pay a premium for a specific capability, regardless of the immediate cost-efficiency. This customer was usually a government body like NASA or the DoD. Today, the data centre has stepped into that role.

The Alpha Product Mechanism

To understand this, one must look at how technologies move down the cost curve. In the past, 'Alpha Products' provided the initial demand that allowed components to scale. The Sony Handycam drove the cost of lithium-ion batteries down; the 3.5-inch hard drive drove the cost of neodymium magnets down. These were specific products that needed a specific component so badly they were willing to pay the 'early adopter' tax. This scale then allowed those components to become cheap enough for the mass market. We are seeing a similar pattern now, but the scale is vastly larger.

The Data Center is the meta-Alpha Product. If you can sell them something they need, fast, they have an almost bottomless bid.

Because the demand for compute is so aggressive, data centres are becoming a massive, reliable source of revenue for companies working on things that have nothing to do with software. We are seeing interest from providers of supersonic turbines, enhanced geothermal energy, modular construction, and solid-state transformers. These companies are using the massive, immediate cash flows from the AI boom to fund the long-term research and development required to make their technologies commercially viable for the rest of the world. The AI boom is effectively subsidising the reindustrialisation of the planet.

Technologies being subsidised by AI demand

Advanced modular nuclear reactors
Enhanced geothermal energy systems
High-voltage direct current (HVDC) grids
Silicon photonics and advanced optical fibre
Solid-state transformers

This creates a strange tension. While the environmental cost of these data centres is a legitimate concern, the technological spillover could be the very thing that solves the energy crisis. If the capital required to build a next-generation nuclear plant can be de-risked by a decade of high-margin sales to hyperscalers, the transition to clean, firm power happens much faster. We are moving from a period of software-only growth into a period where software demand dictates the pace of physical engineering.

Key Takeaway

The massive capital flowing into AI infrastructure is acting as a de facto subsidy for the next generation of physical industrial technology.

02 — Lenny's Newsletter

From Prompting to Managing: The Rise of the Autonomous Agent

How the /goal command changes the fundamental relationship between humans and software.

By Claire Vo · 7 min read

Editor's note: The era of 'chatting' with AI is ending; the era of assigning it long-term objectives is beginning.

Most people use AI as a turn-based assistant. You ask a question, it gives an answer, and then you decide what to do next. This is 'babysitting' the model. It is a high-friction way to work because it requires constant human intervention to keep the momentum going. The shift toward autonomous agents—specifically through features like the /goal command in Codex—represents a move from prompting to managing. Instead of telling the AI what to do step-by-step, you give it a measurable outcome and let it run for hours.

The Death of the Prompt

When you move to a goal-based loop, the nature of the work changes. You are no longer writing instructions; you are defining constraints and verification methods. For example, instead of asking an AI to 'fix this error', you set a goal to 'eliminate all Sentry errors in this specific module within five hours, ensuring no new regressions are introduced'. The AI then enters a loop of observation, execution, and self-correction. It works while you sleep, tackling multi-step tasks that would otherwise require a human to sit and watch a terminal for half a day.

We are moving from babysitting the model to managing it.

This isn't just for engineers. The utility extends to any knowledge worker managing high-volume, low-complexity tasks. Cleaning up thousands of emails, organising project management tasks in Linear, or auditing vast amounts of documentation are all tasks that benefit from an autonomous loop. The key is the transition from 'outputs' (what the AI produces) to 'outcomes' (what the AI achieves). A successful goal is not measured by how many lines of code it wrote, but by whether the specific problem it was assigned to solve has actually disappeared.

Components of a strong autonomous goal

Measurable outcomes (not just tasks)
Clear verification methods
Strict operational constraints
A defined lifecycle (view, pause, resume)

However, this autonomy requires a new kind of literacy. If you give a weak goal—one that is vague or lacks a way to verify success—the agent will simply fail more efficiently. The bottleneck in productivity is no longer the speed of execution, but the clarity of intent. As these tools become more capable, the most valuable skill in the workforce will be the ability to decompose complex business problems into precise, verifiable objectives that an autonomous system can execute without supervision.

Key Takeaway

Productivity in the agentic era is defined by your ability to define outcomes rather than your ability to write instructions.

03 — Simon Willison

The Profitability Inflection: Why Enterprise AI is Finally Paying Up

The shift from consumer hype to high-margin professional utility.

By Simon Willison · 6 min read

Editor's note: The era of 'free' AI is ending as coding agents turn LLMs into essential professional infrastructure.

For the last two years, the narrative around AI has been dominated by massive user numbers and equally massive losses. ChatGPT became a global phenomenon, but turning 900 million weekly users into a sustainable business was a daunting mathematical challenge. Charging $20 a month to a fraction of those users would never cover the trillion-dollar infrastructure costs required to keep the lights on. But something changed in the spring of 2026. The industry has found its real product-market fit, and it isn't in the consumer chatbot; it is in the professional coding agent.

The API Pivot

Both OpenAI and Anthropic have recently moved to align their enterprise pricing with actual API token usage. This is a critical distinction. Previously, many enterprise plans offered 'unlimited' or heavily discounted usage. Now, companies are being billed for exactly what they consume. For a heavy user—specifically a software engineer using an agent like Claude Code or Codex—the costs are staggering. A single power user can easily consume over $1,000 worth of tokens in a month. When companies realize their staff are running these agents all day, the bills become a significant line item.

Coding agents changed everything. They are tools that burn vastly more tokens, but are quickly becoming daily drivers for the work of highly compensated professionals.

This shift marks the transition of AI from a 'toy' or a 'helper' to a core piece of professional infrastructure. When a company pays $200 or $1,000 per seat per month for an AI agent, they are no longer buying a subscription; they are buying a digital worker. This high-margin revenue model is what will allow these labs to fund the next generation of frontier models. The massive capital expenditure required for GPT-5 and beyond is being met not by consumer curiosity, but by the professional necessity of the global software engineering workforce.

Why agents drive revenue

Higher token consumption per user
Shift from per-message to per-token pricing
Integration into professional workflows
High willingness to pay from enterprise customers

While the 'AI failure' stories continue to circulate in the media, the economic reality is moving in the opposite direction. The companies that can provide reliable, agentic capabilities are seeing a massive ramp-up in revenue. The path to profitability for the AI giants is being paved by the very people they were originally designed to assist: the knowledge workers who are now using these tools to automate the most complex parts of their jobs.

Key Takeaway

The economic sustainability of AI rests on professional agents that consume massive amounts of compute, not on consumer chatbots.

04 — Simon Willison

The Theology of Intelligence: A Vatican Response to the Machine

Analyzing Pope Leo XIV's encyclical on the ethics of the AI revolution.

By Simon Willison · 7 min read

Editor's note: The Church's latest stance on AI provides a rare, structured ethical framework for a technology that currently lacks one.

The release of 'Magnifica Humanitas' by Pope Leo XIV is a significant moment in the ongoing debate over AI ethics. By choosing the name Leo, the Pope explicitly links this moment to the first Industrial Revolution. The encyclical is not merely a religious document; it is a philosophical critique of how rapid technological shifts can destabilise human dignity and social structures. It addresses the 'interpretability problem' not as a technical hurdle, but as a fundamental challenge to human agency and understanding.

The Cultivation of Intelligence

One of the most striking sections of the document is the description of how modern AI is 'cultivated' rather than 'built'. The Pope notes that because developers do not design every detail of these systems, but instead create a framework in which intelligence grows, we are effectively managing a black box. This lack of transparency is presented as a risk to human judgment. If we cannot understand the internal representations of the systems we rely on, we risk becoming subservient to processes that we can neither predict nor control.

Current AI systems are more 'cultivated' than 'built'... fundamental scientific aspects remain, at present, unknown.

The encyclical also warns against the 'simulated' nature of AI communication. As models become better at mimicking empathy, advice, and friendship, there is a danger that users—particularly the less discerning—will mistake these mathematical approximations for genuine human connection. The Pope argues that when words are simulated, they do not build relationships; they only build the appearance of them. This is a warning against the erosion of authentic human social bonds in an era of increasingly convincing artificial companionship.

Key ethical warnings in Magnifica Humanitas

The risk of excessive reliance on ready-made answers
The illusion of objectivity in culturally biased models
The environmental cost of massive computing power
The danger of simulated empathy in human relationships

Ultimately, the document argues that development is only 'truly human' if it places people at the centre rather than the accumulation of wealth. It challenges the tech industry to consider whether the efficiency gained through AI is worth the potential loss of human creativity, judgment, and social cohesion. It is a call for a development model that prioritises human flourishing over mere computational throughput.

Key Takeaway

The ethical challenge of AI is not just about safety, but about whether we allow simulated intelligence to replace the authentic structures of human dignity and connection.

05 — Lenny's Newsletter

The Builder's Library: Distilling Decades of Wisdom

A curated guide to the books that define great product leadership.

By Lenny Rachitsky · 10 min read

Editor's note: In an age of ephemeral content, these books offer the only lasting signal for those building the future.

We live in an era of infinite, low-signal content. Newsletters, podcasts, and social media threads provide instant gratification, but they rarely offer the depth required to master a craft. For the product builder, the most efficient way to acquire high-level mental models is not to follow the latest trend, but to read the books that have already survived the test of time. A great book is the distillation of a single person's most valuable insights, refined over years and sold for the price of a lunch.

The Strategy of Execution

Great product management is often mistaken for a series of tactical decisions—prioritising a roadmap or talking to a customer. In reality, it is a discipline of strategy. To move from a feature-factory mindset to a strategic one, builders must study the frameworks of decision-making. Books like Richard Rumelt's 'Good Strategy/Bad Strategy' are essential because they strip away the fluff of 'vision statements' and focus on what strategy actually is: a coherent response to a specific challenge.

The smartest person in the world on a topic you care about spent years distilling their best ideas into a single read.

Beyond strategy, there is the human element: leadership and management. The transition from an individual contributor to a leader is often the most difficult jump in a career. The literature here focuses on the tension between driving high output and maintaining human empathy. The goal is to build systems and teams that can 'run through walls' without burning out the people inside them. This requires a deep understanding of both organizational design and the psychology of motivation.

Essential reading categories

Writing and Communication (Pressfield, Zinsser)
Management and Scaling (Mochary, Hughes Johnson)
Strategy (Rumelt, Martin)
Leadership and High Performance (Grove, Walsh)

The common thread among these works is their longevity. Many of the most impactful books on management were written decades ago. This is because while the tools change—from C++ to AI agents—the fundamental nature of human cooperation, competition, and decision-making remains constant. For the modern builder, the most effective way to stay ahead of the curve is to look backward at the principles that do not change.

Key Takeaway

Mastery comes from studying the enduring principles of strategy and human psychology, not the latest tactical trends.

06 — Stratechery

The Nvidia Delineation: A New Era of Reporting

How the world's most important chipmaker is separating hyperscalers from the rest of the world.

By Stratechery · 5 min read

Editor's note: Nvidia's new reporting structure reveals the two distinct battles being fought at the heart of the AI stack.

Nvidia is no longer just a chip company; it is the central clearinghouse for the AI economy. As its dominance has grown, so has the complexity of its business model. The company is currently engaged in two very different types of competition. One is a fight against commoditisation within the hyperscaler market—the massive cloud providers like AWS, Google, and Microsoft. The other is a race to own the entire stack for everyone else. Nvidia's recent decision to change how it reports its earnings is a direct response to this duality.

The Two Fronts

On the first front, Nvidia sells massive quantities of GPUs to the hyperscalers. In this arena, the threat is real: these customers are the ones most likely to design their own custom silicon to reduce their dependency on Nvidia. This is a battle of scale and efficiency, where Nvidia must constantly innovate to stay ahead of the internal chip projects of its largest customers. The margins here, while still high, are under constant pressure from the sheer volume and the competitive nature of the cloud market.

Nvidia is delineating between hyperscaler sales—where it fights commoditisation—and the rest of the market, where it runs the whole stack.

The second front is much more lucrative and much harder for competitors to enter. This is the 'everyone else' category—the enterprises, the research labs, and the specialized AI companies. For these customers, Nvidia doesn't just sell a chip; it sells an entire ecosystem of software, networking, and integrated hardware. When a company buys into the Nvidia stack, they are buying into a proven, end-to-end platform that works. In this segment, Nvidia is not just a component supplier; it is the architect of the entire computing environment.

The two faces of Nvidia

Hyperscaler Sales: High volume, fighting custom silicon, battling commoditisation.
The Rest of the Stack: High margin, ecosystem-driven, providing end-to-end solutions.

By separating these two segments in its reporting, Nvidia is giving investors a clearer view of where its growth is coming from and where its risks lie. It allows the market to see how much of its revenue is vulnerable to the vertical integration of the cloud giants, and how much is driven by the broader, more stable expansion of the AI ecosystem. It is a sophisticated move that acknowledges the company has outgrown the simple definition of a semiconductor manufacturer.

Key Takeaway

Nvidia's future depends on its ability to maintain its role as an ecosystem provider for the broader market while defending its hardware dominance against the cloud giants.