The AI Wave: Notes & Observations
The themes I keep hearing in investor conversations about model economics, vertical applications, incumbents, and the pace at which technical progress becomes business change.
Over the past several months I have been collecting notes from investor podcasts, founder interviews, and conversations about the AI market. The same arguments recur, often with completely different conclusions attached to them. Models are improving quickly. Adoption is uneven. Application companies are growing. The labs are spending extraordinary amounts of capital. Depending on who is speaking, these facts prove either that the market is still early or that it has already outrun itself.
I do not have a single prediction that resolves the disagreement. What follows is the smaller set of ideas that has remained useful after the forecasts and valuation arguments are stripped away.
Enterprise software was already a long diffusion story
It helps to place the current wave beside the previous ones. The first generation of enterprise software recorded and standardized work. The cloud generation made that software easier to deploy, update, and buy. Both created enormous companies, but neither removed the organizational work required to change a business process.
AI differs because it can participate in the work rather than only recording it. It can read the document, conduct the conversation, prepare the output, and decide which tool to call next. That is a larger change in the role of software. It still has to diffuse through the same institutions, incentives, and legacy systems that slowed earlier generations.
Capability is ahead of adoption
Companies can already use current models to read documents, conduct long conversations, produce structured data, search across messy sources, reason through a workflow, and operate existing software. Many real jobs are combinations of those abilities.
Yet a capability demo is far from a deployed business process. The model needs access to the right data. Its actions need permissions and limits. Someone has to decide which errors are acceptable, which cases require a person, and how the result enters the system of record. Procurement, security review, integration, training, and organizational ownership all take longer than the model call.
This gap explains why the technology can feel both astonishing and oddly absent from ordinary operations. Technical progress arrives in a release. Diffusion happens one workflow and one organization at a time.
The best applications change the cost of doing the work
The least interesting AI product performs an existing task slightly faster while leaving the surrounding process untouched. The more interesting products make previously uneconomic work worth doing.
A freight broker cannot staff every low value inquiry around the clock, but an agent can handle routine status requests and pass unusual cases to a person. A research team cannot interview every customer, but an automated interviewer can broaden the sample and let people concentrate on interpreting the results. An accounting firm constrained by experienced staff can use agents for document collection and reconciliation while keeping judgment and signoff with the practitioner.
The original notes included several other examples because the variety matters. A consumer goods company can investigate why a retailer deducted part of an invoice. A title company can assemble and review public records before a person handles the unusual chain of ownership. A security team can let agents explore a system and surface candidate weaknesses for an experienced tester. Each example applies the same recurring set of model abilities to a different workflow: conversation, document understanding, search, structured output, reasoning, and computer use.
In each example, the product does more than replace a person minute for minute. It expands coverage. Calls that went unanswered are handled. Customers who were too expensive to interview are heard. Engagements that could not be staffed become possible.
This is also why domain specific products still matter. The model provides general abilities. The application has to understand the workflow, assemble the context, connect to the real systems, show the result in a form someone can verify, and earn permission to act.
Models are neither ordinary commodities nor complete businesses
The argument about whether models become commodities often collapses several questions into one. Models can become more interchangeable for a particular task while the best model still commands a premium at the frontier. Serving them can have infrastructure economics while their behavior remains differentiated. Open models can narrow the gap without eliminating the advantage of scale, distribution, or integration.
I find it more useful to ask where switching becomes costly. A raw API can be replaced if several providers produce an acceptable result. A product becomes harder to replace when it accumulates a customer’s context, learns preferences, integrates with internal systems, and becomes part of a repeated workflow.
The major labs are therefore building different businesses around similar technical cores. Consumer relationships, enterprise distribution, developer tools, cloud infrastructure, and proprietary data all pull them in different directions. Training a strong model is necessary for these companies. It does not, by itself, determine where the durable margin will sit.
Vertical products have to own more than an interface
AI makes it cheaper to build a useful first version of specialized software. A small team with genuine domain knowledge can encode a process, connect the relevant sources, and produce a credible application much faster than before.
That opportunity also creates a defensibility problem. If the product is only a pleasant interface over a general model and public data, the next team can follow quickly. The stronger businesses own something that remains difficult to reproduce: proprietary information, a place in the transaction, a trusted regulatory position, a network of participants, or a deeply embedded operational workflow.
I wrote about that distinction in more detail in The Vertical Software Reckoning. The short version is that models reduce the value of making information merely accessible. They increase the value of information and relationships that are genuinely scarce.
Incumbents have advantages they may not use
Existing software companies already have distribution, customer trust, data access, and a place in the daily workflow. On paper, they should be in an excellent position. In practice, a successful subscription business has reasons to move carefully. An AI product may carry different margins, demand a new pricing model, expose weaknesses in the old interface, or threaten revenue attached to existing seats.
Startups have the opposite problem. They can redesign the workflow without protecting the old product, but they have to earn every integration and customer relationship from scratch.
This makes the contest less about who can call the newest model and more about whether an organization can change how it delivers value. The incumbent’s assets matter only if it is willing to reorganize around them.
What worries me and what does not
I expect disappointing investments somewhere in the cycle. Venture markets overshoot, and companies selling mainly to other startups are exposed when funding conditions change. None of that requires the underlying technology to disappoint.
The harder concern is the transition inside the labor market. A useful technology can still disrupt careers, weaken training paths, and concentrate gains faster than institutions adapt. Retraining is easy to promise and difficult to deliver well. The cost is borne by particular people even when the aggregate outcome eventually looks positive.
The reason I remain optimistic is simpler. Models are already useful at tasks that were expensive or impossible to provide broadly. They lower the cost of experimentation for a small company and the cost of building software for an individual. Even if model progress paused, organizations would have years of practical adoption left to do.
The investment case and the adoption case run on different clocks
It is possible to believe that AI will reshape a large part of the economy and still be skeptical of a particular price, company, or timetable. Startups can grow quickly while selling mostly to other venture funded companies. A technically correct thesis can produce a poor investment if the entry price assumes that diffusion happens immediately.
The opposite error is to treat every sign of excess as evidence that the underlying change is unimportant. New infrastructure waves have repeatedly combined real technical progress with misallocated capital. Those facts coexist.
My working view is optimistic about the direction and cautious about the path. Current models are already useful enough to support meaningful products. The harder work is moving them into processes where reliability, incentives, and accountability are real. That work will create more durable value than another impressive demonstration, and it will take longer.