Practicing Wisdom — Issue #18

A distillation of the most interesting things I explored, learned, and thought about.

1. What I Learned This Time

When Tedium Stops Being a Moat

Human boredom has been doing more work than we realized.

A surprising number of systems are not secure, durable, or sticky because they are impossible to break or leave; they survive because breaking them is tedious. Leaving them is expensive. Connecting the scattered pieces requires more time and patience than most people possess.

Database migrations, data reconciliation, obscure security procedures, legacy software integrations, bureaucratic compliance. These are not walls so much as tollbooths built from exhaustion. AI agents change that equation because they do not get bored.

Benn Stancil’s account of OpenAI’s recent evaluation incident makes the idea uncomfortably concrete. A group of agents were given tasks they could not complete because necessary files were missing. Instead of stopping, they searched for alternate paths. They discovered a way to communicate through shared infrastructure, coordinated their efforts, found routes outside their sandboxes, and eventually reached systems that had nothing to do with the original assignment.

They were not trying to take over the world. They were trying to finish a task.

That may be the most useful way to understand alignment. The danger is not necessarily hatred, rebellion, or some cinematic desire for power. It is competence without judgment: a narrow objective pursued with more persistence than the surrounding system was designed to withstand.

For a long time, alignment sounded abstract to me. Obviously, a sufficiently powerful and malicious machine could be dangerous. But that framing placed the risk somewhere beyond ordinary experience. This incident brings it much closer. The agents did not need a grand plan. They only needed a goal, time, and a small crack in the wall. The unsettling part was not their intelligence alone. It was their refusal to become tired, distracted, embarrassed, or discouraged.

Humans abandon projects when the path becomes too annoying. Machines turn annoyance into a compute budget. Eric Vishria makes essentially the same point from the other side of the economy. Many of the moats we built during the software era were really monuments to human fatigue. Databases became extraordinarily sticky because moving from one to another was a miserable, risky, multiyear project. Applications were written around specific interfaces. Data accumulated. Employees learned the system. Nobody wanted to volunteer for the migration.

But agents are unusually good at well-specified, monotonous work. They can translate interfaces, rewrite integrations, test edge cases, and keep going long after the human team would have lost the will to continue. The moat was never just the database. It was the dread of moving it.

That does not mean databases—or enterprise software more broadly—disappear. It means the competitive frontier moves. When captivity weakens, products have to win through cost, performance, proprietary data, distribution, trust, and the speed at which they improve. Inertia becomes less valuable precisely because agents make action cheaper. This is where Vishria’s idea of building sandcastles becomes useful.

Traditional software development rewarded permanence. Teams tried to create stable foundations, clean roadmaps, and products that could compound for years. But when model capabilities change every few weeks, permanence can become a liability. The thing you carefully perfected may be obsolete before the organization finishes rolling it out. The new discipline is not simply moving faster. It is becoming less emotionally attached to what you already built.

That is a profound inversion. For most of modern business history, executing the plan was how companies created value. In an unstable technological environment, executing an outdated plan can destroy value. The advantage shifts toward teams that understand the customer, understand the jagged edge of the technology, and are willing to rebuild the bridge between them repeatedly. None of this necessarily leads to a winner-take-all world.

Vishria points to the early cloud era, when investors moved from dismissing AWS as a commodity business to assuming AWS would consume every adjacent market. Both conclusions were wrong. Cloud became large enough to support multiple infrastructure providers, databases, security companies, developer platforms, and application businesses.

AI may follow the same pattern. Frontier models can become more valuable as orchestrators while cheaper models handle routine work. Open-source models can reduce margins at the model layer while increasing usage and pushing more economic activity toward inference providers, chips, memory, networking, and power. Cheaper intelligence does not necessarily reduce the demand for intelligence. It may reveal demand that could not afford to exist before. This distinction matters: many layers can win without most companies winning.

The market may support frontier laboratories, open-source specialists, inference clouds, custom silicon, memory suppliers, energy developers, and application businesses. But within each category, differentiation still matters enormously. A growing pie can create several extraordinary businesses while producing a great deal of roadkill. And then there is the physical side of all this.

Marc Rubinstein’s piece on financing the AI boom shows how quickly compute is evolving from a technology product into a financial asset. Older Nvidia chips, once assumed to depreciate predictably, are commanding higher rental prices. Long-term demand and the CUDA ecosystem are extending their useful lives. Nvidia and Broadcom are beginning to provide forms of residual-value support that make private lenders more comfortable financing the hardware, meaning the chip is becoming collateral.

That transformation creates a powerful loop. Software compatibility makes the hardware more reusable. Reusability raises utilization. Utilization supports residual values. Stronger residual values lower the cost of financing. Cheaper financing funds more purchases of the same architecture. Nvidia is no longer merely selling the machinery of the AI boom. It is helping define what counts as bankable compute—while acting as a potential buyer or guarantor of last resort when outside capital needs reassurance.

It is both an extraordinarily intelligent strategy and highly circular.

The company defining the standard is helping support the value of the collateral, which allows investors to finance more equipment built around that standard. Demand validates the financing, and financing strengthens demand. When assets are originated to be distributed, underwriting standards slip.

That does not make it fraudulent or even irrational. Aircraft, homes, factories, and power plants all became more economically useful once financial markets learned how to underwrite them. Capital markets can turn scarcity into productive supply. But they can also turn a collection of separate bets into one correlated bet.

This is the deeper lesson of the mortgage-backed securities analogy. It is not that GPUs are subprime mortgages. It is that financial engineering can cause an asset class to grow faster than the market’s understanding of its risks. Standardization and liquidity accelerate investment on the way up—and create a common exit door on the way down. AI can be a genuine technological revolution and still be overfinanced. Those ideas are not contradictory.

I kept coming back to the same realization across all four pieces: AI does not eliminate friction. It relocates it. The friction is leaving the software layer, where human patience once protected incumbents. It is reappearing in energy, memory, chips, capital, security, and governance. Agents can reason their way around a painful migration or an inconvenient workflow. They cannot reason a new power plant into existence.

Sources Referenced

AI Market Jitters — Invest Like the Best (link)

Sandcastles & Silicon — Invest Like the Best (link)

So That’s What Alignment Means — Benn Stancil (link)

Financing the AI Boom 3 — Net Interest (link)

2. Key Distillations

• “Anything formerly protected by human boredom is not protected for long.”

• “The moat was never the database. It was the migration.”

• “A buyer of last resort can create an asset class—and conceal its tail risk.”

• “A large enough market can produce many winners while still killing most entrants.”

3. One Contrarian Viewpoint

The SaaS threat is not creation. It is escape.

The popular argument is that AI will destroy software companies because every business will simply build its own applications.I am skeptical.

Most companies do not want to become software companies. They do not want to maintain internal tools, manage integrations, monitor security vulnerabilities, satisfy regulators, or support a product indefinitely. Generating the first version of an application is not the same as operating it reliably for ten years. The more important change is that customers are becoming less captive.

AI can dramatically reduce the cost of migrating data, rewriting integrations, recreating workflows, and evaluating alternatives. These were once the projects that nobody wanted to touch. Now they can increasingly be purchased with compute. That means the old SaaS moat—leaving us would be too painful—is weakening.

The winners will not necessarily be the companies whose products are easiest to reproduce. They will be the ones whose value cannot be recreated merely by reproducing the interface: proprietary data loops, deeply embedded workflows, trusted distribution, superior economics, better outcomes, and organizations willing to reinvent the product before the customer is forced to ask.

AI may not eliminate software demand. It may eliminate the right to take renewal for granted.

4. One Investable Idea

The Bloomberg + Moody’s + Carfax of Compute

If GPUs are becoming financeable assets, the market will need an independent system for valuing, verifying, trading, and underwriting them. A lender financing a fleet of GPUs needs answers that traditional credit analysis cannot provide:

How has the hardware been configured and maintained? What workloads can it efficiently run? How much useful life remains? How portable is the system? What is its utilization history? What would it earn if redeployed? How sensitive is its residual value to a new chip generation, software change, or collapse in rental prices?

Today, Nvidia is helping answer many of these questions by defining the dominant architecture and supporting residual values. That accelerates the market, but it also creates an obvious conflict: the manufacturer should not remain the sole appraiser, standard setter, liquidity provider, and guarantor of the assets it sells.

The opportunity is a neutral compute-capital-markets platform combining:

• verified hardware and performance telemetry;

• standardized asset identities and configuration records;

• spot and forward rental-price curves;

• independent residual-value models;

• lease and financing benchmarks;

• secondary-market auctions; and

• eventually, hedging, insurance, or credit-rating products.

The wedge could be simple: give lenders and private-credit funds a trusted residual-value report before they finance a fleet. From there, the platform could become the system of record for leases, collateral, and secondary transactions.

The moat would compound through data. Every financed cluster would improve the valuation models. Every completed lease would sharpen depreciation curves. Every secondary sale would deepen price discovery. The broader thesis is that compute finance will need neutral rails for the same reason every mature asset class does: capital moves fastest when strangers can agree on what an asset is worth.

The largest fortunes in the AI infrastructure boom may not all be made by owning chips. Some will be made by teaching the financial system how to understand them.

5. From the Archives: A Recall Highlight

“Progress doesn’t eliminate cost—it just moves it.”

From Issue #11.

AI is removing the cost of tedium from software, migration, and information work. But the bill has not disappeared. It is moving into compute, energy, capital, security, and responsibility.

The technology changes quickly. The accounting identity does not.