The Scarcity You Were Paying For

Be glad if you live beneath your means, because what is coming will test even a conservative margin. The next three years will not be a gentle repricing of digital life. They will be a demolition of the layer of the economy that sits between people and the things they make, watch, hear, and know, and demolitions are hard on everyone standing near the building, including the people who were only renting space in it.

By 2028, households will begin cutting nearly every service they subscribe to, and two offerings that barely register in today's market will be the reason.

• The first is local AI capable of covering roughly 80% of everyday entertainment and digital consumption on hardware no more powerful than an ordinary laptop.
• The second is peer-to-peer infrastructure that removes the advertising model and returns ownership, control, and economic value to the people who create and consume content.

Neither one alone breaks the subscription economy. Together they remove the three scarcities it was built on.

A Bill That Keeps Rising for the Same Thing

The average American adult now spends $111 a month on subscriptions, up from $90 a year earlier, and roughly $21 of that goes to services nobody uses.¹ At the household level, the number is larger and the cause is plainer: average monthly digital subscription spending reached $169.70 in the second quarter of 2026, up 10.6% year over year, driven primarily by higher prices rather than by households adding services.² People are paying more for the same access. Netflix's standard ad-free plan now costs $19.99 a month and its premium tier $26.99, against $8.99 for the tier that carries advertising.³

Consumers have noticed.
Forty-three percent say they are likely to cancel at least one streaming service within three months, and 52% have already canceled or downgraded one because of a price increase.⁴ Only 24% of video streaming subscribers report no recent subscription fatigue, down from 37% in 2023.⁵

The platforms have noticed too.
Average price increases across Netflix, Disney+, and Prime Video fell from 24% of the prior price in 2023/24 to 14% in 2025/26, and the increases that did land fell unevenly: ad-free tiers rose $1.62 on average against $1.21 for ad-supported ones, widening the gap between them, because advertising has become the more important revenue line.⁶ Comscore's 2025 report put the ad tier at 45% of Netflix's US household viewing, against 34% a year earlier.⁷

That is what a mature toll booth looks like. The customer pays either in dollars or in data, and the intermediary keeps the spread either way.

Price increases have hit the ceiling of willingness to pay, so the operator shifts to charging advertisers for the attention of the people who refused to pay more. The model survives only as long as the customer has no other road.

Three Scarcities, All Expiring

The subscription economy was built on three scarcities, and every one of them is becoming a commodity.

Computing was scarce because serious processing lived in data centers only large companies could afford.

Distribution was scarce because reaching millions of people required networks, licensing deals, and storefronts owned by a handful of platforms.

Access was scarce because the catalog itself sat behind those platforms, licensed under terms the consumer never saw and could never own.

When the marginal cost of producing, processing, and distributing content approaches zero, the justification for charging a consumer every month for access collapses. A recurring fee makes sense for a recurring cost. It makes far less sense for something that costs nothing to deliver one more time and that the customer can produce, or retrieve, on a machine they already own.

Intelligence Moves Onto the Desk

The capability gap between what runs in a data center and what runs at home has been closing on a schedule. Epoch AI found that models small enough to run on a single consumer GPU typically match frontier capability after a lag of six to twelve months across four benchmarks, driven by open-weight models scaling compute at roughly the same pace as closed ones, by distillation compressing larger models with little loss, and by each new GPU generation fitting a larger model locally.⁸

The thresholds are concrete: about 28 billion parameters on an RTX 4090, about 40 billion on a 5090, at four-bit weights.⁹ Measured on Epoch's capability index rather than on hardware, the best open-weight models have run an average of four months behind the closed frontier since January 2026, a gap of eight index points.¹⁰

The hardware targets are now explicit. Google shipped Gemma 4 in April 2026 in four sizes, built to run and fine-tune on everything "from billions of Android devices worldwide, to laptop GPUs," with the two smallest intended to run offline on a phone.¹¹ Google's own memory table puts the smallest at 2.9 GB of weights at four-bit precision and the next at 4.5 GB.¹² Those are laptop numbers, published by the vendor.

The gap at the top is still real, and it is not a licensing gap.

The strongest coding model you can run on one consumer graphics card scores 68.0% on SWE-bench Verified; the top of the same leaderboard, measured on the same harness, scores 97.0%.¹³¹⁴ But the two models immediately behind that leader, at 96.4% and 95.4%, ship their weights openly and still need a data center to run. What separates the desk from the frontier is not who owns the model. It is how much memory the model needs, which makes the price of memory the whole question.¹⁵

That gap matters for building software and for frontier research. It does not matter for summarizing the news, generating a bedtime story, tutoring a child through algebra, narrating a long article aloud, composing background music, editing photos, answering a medical-terminology question, or holding a conversation.

Most everyday digital consumption is not frontier work. It is ordinary work done at volume, and ordinary work is exactly what a laptop-class model now handles.

The relevant threshold was never parity with the best model in the world. It was good enough for the thing a person was paying $10 a month to have done.

The irony is that AI subscriptions are themselves the fastest-growing line on the bill. Americans who pay for AI tools average four of them at almost $66 a month, and a quarter of those subscribers spend more than $100.¹⁶ That spend is the most exposed of all, because it is renting a capability that is being given away in open weights with a lag measured in months.

The Toll Moves From Monthly to Upfront

The obstacle is hardware, and in 2026 it is a serious one. Gartner projects the combined price of DRAM and solid-state storage will rise about 130% by the end of 2026 against 2025 levels, pushing average PC prices up 17% and smartphone prices up 13%, with memory climbing to 23% of a PC's bill of materials from 16%.¹⁷ TrendForce revised its first-quarter DRAM contract forecast up to a 90–95% quarterly increase and expects prices to stay high through 2028.¹⁸ The same AI buildout that makes local intelligence possible is making the machine to run it more expensive.

That does not stop the shift. It changes the shape of the payment.

A household spending $111 a month per adult on subscriptions is spending more than $1,300 a year per person on access it will never own. A capable laptop, even at inflated 2026 prices, is a one-time purchase that keeps working after the payment stops. The cost of digital life moves from a perpetual rent to a capital expense, and capital expenses favor households with savings.

This is where living beneath your means stops being a virtue and becomes a position.

The household with cash can buy the machine and stop the bleed. The household living at the edge of its income keeps renting, keeps paying the annual increases, and keeps funding the intermediaries the longest.¹⁹ The machine that ends the rent is getting more expensive at the cheap end faster than anywhere else, and the cheap end is where that household shops.

The exit is being priced out from the bottom. Gartner expects worldwide PC shipments to fall 10.4% in 2026 as memory costs pass through to retail, and its senior director analyst Ranjit Atwal puts the end point plainly: "Ultimately, we expect the sub-$500 entry-level PC segment will disappear by 2028."²⁰

That is the same year the subscription cuts start. The household that most needs a one-time purchase to replace a monthly one is the household whose price tier is scheduled to stop existing.

Distribution Without a Landlord

Local intelligence handles what a person consumes alone. Peer-to-peer infrastructure handles everything that moves between people, and it attacks the second pillar of the current model: the intermediary's cut.

Music is the cleanest example of what that cut costs. Spotify pools subscription and advertising revenue and, by its own account, pays out roughly two-thirds of music revenue to rights holders.²¹ It also insists that no service pays a fixed amount per stream, because listeners do not pay per stream — the per-stream figure is total payouts divided by total streams after the fact, and in 2026 that arithmetic lands between $0.003 and $0.005.²²

Spotify paid more than $11 billion to the industry in 2025, bringing lifetime payouts to nearly $70 billion, with roughly half generated by independent artists and labels.²³ Its head of music, Charlie Hellman, says the platform "accounts for roughly 30% of recorded music revenue"; 13,800 artists cleared $100,000 on it last year and 1,500 cleared a million.²⁴

Important

Before any label takes its share, a million streams is worth a few thousand dollars to the people who made the work. The platform does not compose, perform, or record anything. Its product is the index, the delivery, and the payment rail, and every one of those is something a protocol can do without a company in the middle.

The protocols already exist. The AT Protocol behind Bluesky makes the account an identifier the host does not own: the stated design goal is that "changing a provider should not cause a user to lose their identity, social graph, or content," and the migration path moves the repository, the media, and the preferences to a new server while the identifier itself keeps pointing at the user.²⁵ Bluesky passed 47 million registered users in October 2026.²⁶ Its mobile monthly active users fell 27% year over year to 10.4 million by June, and the quarterly average has halved since the end of 2024.²⁷

Important

Portable identity is solved technology. What has not arrived is an economic reason to leave, because a decentralized feed is still, to the user, just another feed.

The flip comes when the protocol carries money as well as posts: when a listener's payment reaches the artist directly, when a reader's subscription goes to the writer and not a platform, when the creator's audience is an asset they hold rather than a list the platform rents back to them. At that point the advertising model has nothing to sell, because there is no centralized audience to package.²⁸

What Disappears

Put the two together and the list of things consumers no longer need to pay for gets long.

• No centralized gatekeeper deciding what is available.
• No mandatory subscription to reach a catalog.
• No intermediary extracting a percentage of every transaction.
• No business model that depends on harvesting attention and personal information to compensate for prices that have hit their ceiling.

The next major disruption will not be AI replacing jobs one role at a time.

It will be AI eliminating entire layers of economic intermediation that people have been conditioned to treat as permanent: the aggregator, the storefront, the licensing middleman, the ad network, the recommendation engine that exists mainly to keep a subscriber from canceling.

Warning

Those layers employ a great many people and anchor a great deal of market value. Their contraction will be felt in paychecks and portfolios well before it shows up as savings on a credit card statement.²⁹ That is the second reason to keep a margin.

The disruption arrives as an income shock for some households at the same time it arrives as a cost reduction for others, and nobody gets to choose which side they are on in advance.³⁰³¹

The Monopoly Is Not the Moat

Companies that mistake their distribution monopolies for durable competitive advantages are about to learn how quickly consumers leave when the road reopens. Streaming already shows the behavior in miniature: subscribers churn out, wait for a full season, binge it, and cancel again. That is not loyalty to a platform. It is tolerance of a toll for one piece of content, and tolerance ends the day the content is reachable another way.

The durable businesses on the other side will be the ones that sell something a protocol cannot replicate: the work itself, the hardware it runs on, or the expertise to make the local stack useful. The ones that sold access will find that access was the scarce thing, and that it is not scarce anymore. Take the Netflix plan that rose from $17.99 to $19.99 this year. When the household that pays it can generate most of an evening's entertainment on a machine it already owns, and pay the few creators it actually values directly, the question will not be whether $19.99 is too much.

It will be what, exactly, the $19.99 was ever buying.