Technological equality is getting further away from ordinary people.

Previously, it was difficult for individuals to access advanced computers.

Back then, computers were called mainframes because they were truly massive. They filled an entire room, requiring dedicated power, air conditioning, and administrators. Only governments, universities, and large corporations had such computer rooms. If an individual wanted to use one, they had to submit a request and wait for others to schedule time.

At that time, the machines were far away from people.

Later, the most touching part of tech history was how these distant machines gradually moved closer to ordinary people.

Desktop computers landed on office desks, the internet entered homes, laptops were stuffed into backpacks, and smartphones were slipped into pockets. You no longer needed to be on good terms with the computer room admin, apply for approval, or put on shoe covers before using a computer. You could use advanced technology for your own purposes.

This was certainly not achieved through goodwill.

Computers became cheaper not because capitalists suddenly had compassion. Smartphones became widespread not because the supply chain suddenly believed in technological equality. What really worked was scale. Hundreds of millions of computers and over a billion smartphones created a massive consumer market.

Ordinary people were once the tech industry's biggest customers.

The biggest customers are usually taken care of.

So production lines expanded, processes improved, and costs were driven down round after round. The best screens, chips, storage, and batteries were always expensive at first, but a few years later, they often trickled down to cheaper devices. If you couldn't afford the best machine today, you might be able to buy something more powerful at a lower price in a few years.

This is the most straightforward form of technological equality. An ordinary person, starting from a sufficiently cheap and open entry point, could gradually approach the future.

Now, that path is narrowing.

On one side, memory prices are rising, pushing up the costs of computers, phones, and game consoles little by little. On the other, the most capable AI large models are being surrounded by sovereignty, computing power, and security reviews.

Ordinary people are paying the cost for this round of technological leap, yet technological equality is moving further away from them.

Who Gets the Wafer First?

Apple's price hike should not be seen merely as an Apple price hike.

Tim Cook told the media that Apple was experiencing a "once-in-a-century flood." Apple sells over 200 million iPhones a year, holds the strongest bargaining power in the global supply chain, and has sufficient profit margins to absorb many component price increases. For a long time, Apple was the one that made suppliers bow.

But this time, Apple also raised prices.

By 2026, DDR5 and other mainstream memory prices have rebounded significantly from their lows a few years ago, with some retail channels seeing prices double or more. The memory cost for the next-generation iPhone Pro could be nearly $100 more than before.

For Apple, this cost can still be shuffled between selling price, profit, and product lines. For low-end PCs, budget smartphones, and cheap game consoles, there is little room to maneuver.

Today, the prices of almost all consumer electronics are driven by memory.

Globally, the companies that can mass-produce memory are mainly Samsung, SK Hynix, and Micron. The DRAM in iPhones and the HBM next to NVIDIA's AI chips are all produced by them. Together, these three companies control over 90% of the global DRAM market.

For many years, this market revolved around phones and computers. The demand for phones and computers was large enough that suppliers had to cater to them. Ordinary consumers individually have no bargaining power, but over a billion phones and hundreds of millions of computers together represent one of the strongest global demands.

Scale was once on the side of ordinary people.

It made computers cheaper, phones cheaper, and memory sticks into ordinary commodities. If someone couldn't afford a new computer, they could at least add a stick of memory to their old one. The machine could last a few more years, and people wouldn't be eliminated by the next round of technology so quickly.

Many people who have built their own computers know Crucial.

It's not particularly a shiny brand. It's just a memory stick. Many people bought it when they first upgraded their old computer's memory. An old computer got slow, and they couldn't afford a new one, so they added a stick of memory.

That's the significance of low-end technology entry points.

Now, Micron has exited the Crucial brand. The reason is not hard to understand. The same wafer is more profitable when sold to data centers.

AI companies need HBM, which is placed next to GPUs for high bandwidth, supporting model training and inference capacity. They can sign multi-year contracts, lock in supply in advance, and accept higher prices.

Consumer electronics start to give way.

HBM is not ordinary memory. It stacks multiple layers of chips and connects them through complex packaging. An HBM chip occupies about twice the wafer area of a regular DDR memory chip. It is more expensive, harder to make, and more worth the supplier's priority.

In 2020, HBM accounted for only about 2% of DRAM wafer capacity. By 2026, it has risen to about a quarter. Demand is still growing at 70% annually. TrendForce data shows that in the first quarter of 2026, DRAM contract prices rose nearly 90% in a single quarter. SK Hynix says shortages could last until 2030. Jefferies predicts memory prices will rise another 40% to 50% in the second half of this year, with the uptrend possibly until 2028.

This price hike may not be purely a natural market outcome.

On June 25, 14 consumers and 3 small PC retailers filed a class action lawsuit in California federal court, accusing Samsung, SK Hynix, and Micron of colluding to manipulate DRAM supply and pricing since 2022, inflating memory prices by about 700% over the past four years. The plaintiffs claim the three companies coordinated to reduce DDR3 and DDR4 capacity under the guise of shifting to HBM, artificially creating a shortage.

The lawsuit has not been decided; the allegations need court review.

But this industry has prior precedent.

In 2005, Samsung pleaded guilty to manipulating DRAM prices with the U.S. Department of Justice and was fined $300 million. SK Hynix also pleaded guilty the same year and was fined $185 million. Total related fines in that round reached $731 million, and several executives were imprisoned. The plaintiffs' lawyers now say this is the same group of companies, the same market, facing the same issue for the third time.

AI demand is real. Data centers pay more, orders are more stable, so suppliers are naturally willing to shift resources there.

Market concentration is also real. Three companies control the vast majority of DRAM capacity. They can claim they are just responding to demand, and while doing so, they can also raise the water level a bit.

For ordinary consumers, the two things don't make much difference in the end. Both become bills.

They don't need to know the packaging structure of HBM or how DRAM contract prices are negotiated. They will only find that upgrading the memory in their old computer is more expensive, new computers have less configuration, phones are pricier, and low-end products are harder to buy.

Who gets the wafer first sounds like a factory issue.

But it's actually about who gets close to advanced technology first and who gets pushed out.

The First to Be Pushed Out

For those who can afford the new MacBook Pro, memory price increases just mean spending a bit more money. They might frown, curse Apple for raising prices again, and then proceed to order.

Many more people in the world buy laptops under $500, smartphones under $100, and used previous-generation machines from second-hand markets. IDC predicts that memory price hikes will significantly raise the average selling price of smartphones, putting pressure on entry-level models and small manufacturers. Behind this number, it's not that low-end users suddenly have more money, nor that everyone no longer needs cheap phones.

With memory prices rising, these cheap entry points are hit first.

HP said in its Q1 earnings call that memory's share of PC material costs jumped from 15% to 18% last quarter to 35%. For low-cost PCs, a slight increase in material costs could directly erase profits.

Gartner analysts even said that low-end laptops under $500 could disappear from the market within two years.

The disappearance of these products isn't because nobody needs them.

On the contrary, the people who need them the most. Students, small merchants, temporary remote workers, and people just entering digital life all need a "good enough" computer. These machines have little profit margin to spare. When memory prices rise, the whole machine becomes awkward. Sell it too high, nobody buys; sell it too cheap, the manufacturer loses money.

The significance of cheap machines is not just cheapness. It is an entire low-cost entry point into modern life. Over the past decade, smartphones costing a few hundred yuan pulled over a billion people from offline to online. Someone used one for the first time to collect payments, someone for online classes, someone to find a job, someone to send out their voice.

If someone has to pay a high cost to access new technology, they won't try it easily. What cheap devices truly change is a person's posture toward technology. You can first buy a cheaper one to try.

High-end products can raise prices, add a Pro, Max, or Ultra suffix, and package the increase as an upgrade. Low-end products cannot. If they raise prices even a little, users may walk away.

The bankruptcy of a low-end consumer electronics brand won't become headline news in tech media. A student missing an online class, a small shop owner delaying replacing a laggy cash register, an elderly person waiting longer for an app to open—these are hard to become public events on social media.

The cruelest part of technological inequality is often not slamming the door shut.

It just makes some people slower. Slower to master, slower to enter a new era, slower to update their understanding of the world.

Then the gap widens.

AI in the Electricity Bill

It's not just memory that's getting more expensive; it's also electricity.

Data centers are going up one after another in Northern Virginia—gray boxes without windows, surrounded by substations, fences, and parking lots. They don't look like traditional factories; they have no chimneys and emit no smoke. But they consume a lot of electricity.

The U.S. Energy Information Administration, in its annual outlook, expects commercial building electricity consumption to grow faster than residential and industrial sectors, with data centers as a key driver. Goldman Sachs estimates that AI data center expansion will continue to push up electricity prices, with lower-income households hit harder because electricity makes up a larger share of their spending.

The electricity consumed by data centers comes from the public grid. The grid doesn't grow out of nowhere. Substations need to be built, transmission lines expanded, peaking capacity added, local governments grant land, and utility companies spread the investment back into bills.

Data centers can sign long-term contracts and receive more favorable industrial electricity rates. The bills eventually fall on more diffuse places. Into household electricity bills, small business electricity bills, and onto people who have never trained a model or know what HBM is.

This is a very subtle form of redistribution in the AI era.

Inside data centers, there are GPUs, HBM, liquid cooling pipes, electricity contracts, land permits, network fiber, and a whole generation of companies hungry for intelligence.

All of this is expensive and in demand.

Samsung's semiconductor profits rebounded, Micron's storage business gross margin rose, and SK Hynix employees were reported to potentially receive extremely high profit-sharing bonuses.

Consumer electronics haven't returned to an exciting era, but upstream is making money. The best capacity, highest profits, and tightest schedules are all given to data centers.

Production lines adjust for them, the grid expands for them, capital opens up for them.

And then, the models were regulated.

We all know what happened with Fable 5 and GPT-5.6. From the government's perspective, these moves are not hard to understand.

The more powerful the frontier model, the more it resembles infrastructure and also weapons. It might write code, do biological design, manipulate information, affect financial transactions. When a model becomes powerful enough, it is no longer just a company's product but part of national capability.

Chips are subject to export controls, computing power is reviewed, and model releases are naturally included in approvals.

Ordinary people are now placed in a strange position.

Memory prices rise, so they have to spend more on electronics. Data centers expand, so they pay more electricity each month. Public grids, land, supply chains, capital markets—all make way for the strongest models. Yet the capabilities ultimately trained by these resources may, due to region, identity, payment tier, enterprise qualification, and government approval, keep ordinary people outside.

Ordinary people share the cost but may not share the capability.

It's very close to you, so close that it's in your information feed every day. It's also very far, so far that when you truly need it, you find you can only observe from a distance.

In the past, a technology first served institutions, then individuals. First served the rich, then ordinary people. It was also unfair, with long price and identity barriers, but the general direction was downward. The larger the output, the lower the price, the more alternatives, the richer the gray paths.

Today's AI doesn't fully follow this order.

The more advanced the model, the more it needs concentrated computing power. The more concentrated, the easier it is to be controlled by a few platforms. The more risky, the easier it is to be reviewed by governments. The more commercially valuable, the more it flows first to customers with stronger paying ability.

It may not naturally trickle down along the consumer market like computers and phones did.

Technology is still advancing.

But progress no longer automatically brings equality.

The Mainframe Is Back

The most important word in "personal computing" is not "computing," but "personal."

The most crucial change in the personal computing era was never just that computers got smaller, phones got lighter, or screens got brighter. More importantly, an ordinary person could finally buy a portion of computing power, bring it home, put it on their desk, and later in their pocket. That machine might not be very powerful, might become obsolete quickly, might heat up when opening certain software, but it belonged to you.

Belonging to you—that matters greatly.

AI changed this relationship.

You still hold a computer and a phone, even more advanced than before. But the capabilities of AI are in remote data centers, in GPU clusters, in HBM, in liquid cooling pipes, in electricity contracts, in platform permissions, and policy reviews.

You haven't bought that capability into your hands. You are still like in the mainframe era, sending requests for use to a distant place. Whether the request can be fulfilled depends on many things. Which country you are in, how much you pay, whether you are a business customer, whether the platform has computing power that day, whether the model has passed approval, whether a certain function is available in your region.

Of course, AI hasn't completely left ordinary people behind. Free chatbots, open-source weights, and on-device models still bring some capabilities to more people.

But the frontier capabilities that truly define the boundaries of scientific research, coding, content production, and enterprise efficiency are increasingly concentrated in a few data centers, platform accounts, and national licenses. Ordinary people can use AI, but they may not be able to access the most important capability layer of this AI era.

This is not the same problem as an old computer being slow.

In the past, the main contradiction between a person and a machine was insufficient performance, high price, or difficult software. There were many troubles, but most could be solved over time, through second-hand markets, low-end products, open-source software, and a bit of folk wisdom.

But today, the most powerful AI model is not a device; it is a continuously burning resource system. It consumes electricity, chips, and memory every day, depends on data center schedules, and also on regulatory approvals and platform strategies.

The stronger it becomes, the less it resembles a consumer product that naturally trickles down. It looks more like a centrally dispatched capability.

The direction of personal computing was to break down institutional capabilities and sell them to individual people. The direction of AI, at least at the frontier, is to reconcentrate capabilities and then distribute them through accounts, subscriptions, APIs, regional policies, and enterprise contracts. What you see is a chat window, but behind it is actually a rationing system.

Ordinary people are no longer the biggest customers that the consumer market fights over. They can no longer gain capabilities by buying a device, and cannot just pay for their own machine.

They start paying for machines far away.

It's in their electricity bills, in memory prices, in phone price hikes, and in the disappearance of low-end computers. But when they truly want to use that most powerful machine, they still have to wait for an account, a region, a permission, and a policy that might change at any time.

The mainframe is back, back as an order.