Hello hello!
A useful way to understand a new technology is to ask what became cheap, then ask what became scarce as a result. The internet made publishing cheap, so attention became scarce. Cloud computing made servers easier to rent, so distribution and software became more important than owning hardware. AI is now making several kinds of production much cheaper at once: code, content, analysis and, increasingly, decision-making.
That does not remove bottlenecks. It moves them.
This week, the interesting stories are mostly about where those bottlenecks are going. Social feeds are deciding which kinds of stories humans learn to tell. Nvidia is trying to turn compute into something Wall Street can finance like infrastructure. Anthropic is heading toward the public markets, where the economics of frontier AI will become much harder to hide. Vibe coding is making software creation accessible faster than it is making software creation safe. And commerce is beginning to encounter customers that do not have eyes, fingers or much patience for a badly structured website.
1. The algorithm eventually teaches you what kind of life is worth reporting
Sociologist Kathryn Jezer-Morton’s new book, The Story of Your Life, examines a subtle consequence of spending two decades telling our lives through social platforms. In a recent conversation with The Atlantic, she argues that feeds tend to reward stories with recognizable movement: redemption, improvement, transformation. What struggles to travel is the “flat story,” where life continues without a satisfying before-and-after.
This sounds like a media observation, but it is really an incentives observation. Any system that rewards some behavior will eventually produce more of that behavior. The clever part about social media is that the reward does not stop at what we publish. Once people know which stories perform, they start noticing their lives through the same template. A holiday needs a highlight. A career setback needs a lesson. A bad year needs to become the year that changed everything. The feed starts as an editor and ends up as a co-author.
Companies do this too. Startup storytelling has converged on an oddly small number of plots: we were rejected, then we won; nobody believed us, then everyone did; we nearly died, then revenue exploded. These stories work because compressed narratives are easy to remember. But the danger is that founders start confusing what is narratively legible with what is strategically important. A business that quietly compounds for six years is less interesting on LinkedIn than one that almost collapses and recovers in six months, even though you would probably rather own the first one.
There is a useful marketing lesson here, but it is not simply “use anecdotes.” Anecdotes are powerful because human beings can simulate another person’s experience more easily than they can feel a percentage. That makes them good communication devices and terrible substitutes for evidence. The better move is to use the anecdote to make someone care, then use the evidence to determine whether the anecdote is actually representative.
The feed rewards the clean story. Reality rarely has one. Good operators should be able to tell the difference.
2. Nvidia has realized that financing can be a product feature
Nvidia announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create financing platforms intended to mobilize more than $500 billion for AI infrastructure. Jensen Huang described compute as an “investable asset,” and Nvidia said the platforms are meant to give customers access to long-duration financing for what it calls AI factories. Huang also said Nvidia could backstop up to $125 billion of the potential deals.
At first glance this is a financing story. It may actually be a product strategy story.
Suppose you manufacture a machine that costs $10 million and your customer can clearly make $15 million with it. You still have a problem if the customer does not have $10 million. Lowering the price is one solution. Lending them the money is another. If capital becomes abundant enough, you can increase demand for the machine without making the machine cheaper.
That is roughly what Nvidia is helping build around compute.
For the first phase of the AI boom, access to GPUs itself was an advantage. Companies signed large contracts because capacity was scarce and possessing it created optionality. But if Nvidia and Wall Street succeed in making compute financeable, fungible and rentable, owning GPUs becomes less strategically interesting for most companies. The important question becomes what return you can produce from each dollar of compute.
This is what financialization often does to productive assets. It separates ownership from use. Airlines do not need to manufacture or necessarily own the plane to build an airline. A retailer does not need to own the building to operate a store. If compute follows the same path, the company with the largest GPU pile may not have the advantage. The company that can turn rented intelligence into the highest-value output might.
There is an obvious risk in a supplier helping finance demand for its own products. If financing becomes necessary to keep sales accelerating, it becomes harder to tell how much demand is natural and how much has been manufactured by cheap capital. But bubbles and useful infrastructure are not opposites. It is possible to overfinance something the world genuinely needs.
The more interesting change for everyone downstream is simpler: compute is slowly becoming a financing problem instead of a procurement problem. If that continues, buying lots of hardware because it feels strategically scarce may age badly. The financial system is being assembled specifically to make scarcity rentable.
3. Anthropic’s IPO may finally tell us what kind of business AI really is
Anthropic confidentially submitted a draft S-1 to the SEC in June. The filing itself is not public yet, so anyone claiming to know exactly what its prospectus contains is guessing. But the eventual public filing will matter for a reason much bigger than Anthropic’s valuation: it will give us one of the first detailed looks at the economics of a frontier AI lab.
People casually talk about AI companies as software companies. That comparison may turn out to be misleading.
Traditional software has wonderful economics because the millionth copy costs almost nothing to produce. AI is different. The millionth user creates inference cost. A customer who uses Claude twice as much actually consumes more resources. Better models can also require enormous continuing expenditure on training, chips, power and data centers. This does not make AI a bad business. It means the shape of the business matters.
The Anthropic prospectus should help answer whether frontier AI behaves more like SaaS, a cloud provider, a semiconductor company or some new mixture of all three. Gross margins will matter, but the really interesting numbers will be underneath them: how inference costs change as models improve, how much revenue comes from a small number of large customers, how long contracts last, how much compute has already been committed, and whether rapidly falling token prices are being offset by rapidly rising usage.
That last question matters enormously to anyone building on top of these models. Developers have spent the past few years benefiting from an extraordinary competition among labs. When companies are fighting to establish a platform, they have good reasons to make the platform attractive. Cheap API pricing can be part of customer acquisition.
Public companies eventually acquire another constituency: shareholders.
That does not automatically mean prices rise. Better chips and better models can keep lowering the cost of intelligence. But it does mean the economics become less philosophical. If serving a token costs more than the market thinks, the gap has to be paid by someone. Today that someone may be an investor financing growth. Tomorrow it may be the customer.
The most interesting question in Anthropic’s S-1 will therefore not be “How fast is Claude growing?” We already know AI demand is growing quickly. It will be how much economic value Anthropic keeps each time Claude does another unit of work.
That number will tell us a lot about the businesses being built on top of it too.
4. Vibe coding removed the gatekeeper, not the responsibility
A recent preprint studying the security of vibe-coded applications assembled more than 10,000 real-world repositories and performed deeper audits on 200 deployed applications. The researchers found at least one exploitable vulnerability in 90% of those audited apps, with 76.7% of the identified vulnerabilities rated high or critical. Because this is a preprint and the detailed audit covered 200 applications, the numbers should not be treated as a census of every AI-built app. The failure patterns are still worth paying attention to: broken access control, cryptographic problems, injection, exposed secrets and temporary logic that quietly survived into production.
The obvious conclusion is that AI writes insecure code.
The more interesting conclusion is that we have changed who is allowed to create software without changing who is responsible when the software fails.
Traditional software development had an inefficient but useful feature: expertise acted as a gate. To put a database, authentication system and payment flow on the internet, somebody usually had to know enough about software to understand that these things contained danger. Vibe coding removes much of that gate. That is precisely why it is powerful. It also means people can now deploy systems containing security boundaries they do not know exist.
An AI agent can fix a requirement you give it. The harder problem is the requirement you never knew to give.
A founder can ask, “Add login.” An experienced engineer hears several unstated questions inside that sentence. How are sessions stored? What happens after password reset? Can one user access another user’s object by changing an ID? Are secrets exposed client-side? How long do tokens live? What happens after repeated failed attempts? The novice sees a login box. The engineer sees a collection of trust boundaries.
This is why the long-term security solution probably will not be telling millions of new builders to become security engineers. The platforms themselves will have to convert experienced engineers’ tacit knowledge into defaults. Lovable is already moving in this direction with security scanning and dedicated trust centers for published apps. That is the right abstraction. The user should not need to remember every dangerous thing the system could do.
Cars became mass-market products because drivers did not need to understand combustion, metallurgy and crash physics. The safety knowledge moved into the machine, the road and the rules around them.
Vibe coding will have to do the same thing.
The opportunity created by AI coding is not merely that more people can build software. It is that software platforms can eventually encode enough good judgment that people can safely build things they do not fully understand.
We are not there yet.
5. Your next customer may never see your website
On its first-quarter earnings call, Shopify said AI-driven traffic to its stores had grown eightfold year over year while orders from AI-powered searches had increased nearly thirteenfold. Shopify has also structured more than one billion products with attributes, current pricing and inventory data for AI systems, and said traffic from catalog-powered AI searches converts about twice as well as traffic from general AI searches built from scraped web information.
The temptation is to call this another acquisition channel.
It could become something more important: a change in who websites are designed for.
The web we know assumes a human is on the other side. Product photography builds desire. Navigation helps someone browse. Reviews create confidence. Checkout pages reassure a nervous buyer. Entire careers in conversion optimization exist because humans hesitate, become distracted and misunderstand things.
An agent has different problems. It wants an exact price. It wants to know whether the product is in stock. It wants machine-readable specifications. It wants to know whether shipping arrives before Tuesday. It wants a transaction path it can reliably complete.
In other words, much of the modern ecommerce site is a persuasive interface built over a database. Agents increasingly care about the database.
This creates an interesting inversion. For twenty years, online sellers optimized information so search engines would send them humans. In agentic commerce, software may increasingly consume the information and make part of the purchasing decision itself. SEO was about convincing an algorithm that a human should see your page. Agent commerce may partly be about convincing an algorithm that your product is safe to buy.
That makes structured data, accurate inventory, clean APIs and predictable checkout behavior more than technical housekeeping. They become merchandising.
There is also a bigger strategic risk. If people begin delegating product discovery and comparison to AI, the agent becomes a new intermediary between the merchant and the customer. We have seen this movie before. Google aggregated websites. Amazon aggregated sellers. App stores aggregated developers. Whoever controls discovery eventually acquires bargaining power over everyone supplying the thing being discovered.
So merchants should absolutely make themselves easy for agents to understand. But they should also notice what they are trading away.
The first stage of a new distribution channel feels like free growth. The second stage is usually somebody discovering they own the customer relationship.
That may be the more important agent-commerce story.
If you haven’t yet, subscribe, like, leave a comment, and share Token Economy. It helps us keep digging past the headlines and bringing you the most interesting ideas, shifts, and second-order effects across AI, technology, startups, and business each week.