EN / 中文

Thoughts

Conception is power

Because of AI, the World Is Getting Slower

The World Did Not Get Faster. The Speed Was Reserved for AI.

For the past two years, almost everyone has been saying that AI will make the world faster.

Faster coding, faster writing, faster image generation, faster customer support, faster search, faster office work. It sounds as if connecting AI to every piece of software will automatically move the world into a lower-friction, higher-efficiency era.

I am starting to think the opposite may be happening.

Because of AI, the world is getting slower.

This is not just a complaint about bad software. There is a very concrete supply-chain logic behind it: AI is absorbing the best memory, storage, wafer capacity, electricity, servers, and capital spending. A small number of data centers are becoming faster and faster, while ordinary users and ordinary companies face a different reality: more expensive computers, more expensive phones, more expensive RAM, more expensive SSDs, longer replacement cycles, older devices staying in service, and software that keeps getting heavier.

The result is simple: the model runs faster in the cloud, while the person waits longer on the device.

Memory Prices Can Change How Often People Replace Devices

In February 2026, Gartner published a blunt forecast: surging memory costs will cause worldwide PC shipments to fall 10.4% in 2026 compared with 2025, while smartphone shipments will fall 8.4%. Gartner also estimated that combined DRAM and SSD prices would rise 130% by the end of 2026, pushing PC prices up 17% and smartphone prices up 13%. More importantly, Gartner expected PC lifetimes to increase by 15% for business buyers and 20% for consumers.

Put those numbers together, and the meaning becomes clear. People did not suddenly stop needing computers. Companies did not suddenly stop needing devices. They are starting to delay upgrades because upgrades are getting too expensive.

In the past, when a computer felt slow, you could add memory, replace the SSD, or buy a new machine. Now 32GB or 64GB of RAM increasingly feels like a premium choice. SSD upgrades no longer feel like a casual purchase. For companies, the pressure is even more serious. Servers, office PCs, laptops, and phones all compete for procurement budgets, and rising memory and storage prices directly eat into those budgets.

So the practical result is boring but powerful: when a user’s computer gets slow, they tolerate it for a while; when a company’s servers are underprovisioned, the team squeezes more out of them; when an employee’s laptop is old, it stays in service for another year; when a phone runs out of storage, the user deletes a few apps.

Slowness does not arrive all at once. It accumulates through thousands of small decisions to make do.

AI Is Reordering the Memory Supply Chain

Why are memory prices rising like this?

IDC’s explanation of the 2026 memory shortage is direct: the rapid expansion of AI infrastructure and AI workloads is putting significant pressure on the memory ecosystem. Instead of expanding conventional DRAM and NAND used in smartphones, PCs, and other consumer electronics, major memory makers have shifted production toward higher-margin products for AI data centers, such as HBM and high-capacity DDR5. IDC says this has restricted the supply of general-purpose memory modules and pushed prices higher across the board.

That is the key point.

AI does not merely consume GPUs. AI consumes an entire physical supply chain: HBM, DDR5, LPDDR, enterprise SSDs, packaging capacity, advanced nodes, mature nodes, power management ICs, copper, gold, electricity, and data center space. It is not just a software boom. It is a hardware sinkhole that reorders real-world resources.

Gartner gives the macro background. Worldwide AI spending is forecast to reach $2.52 trillion in 2026, up 44% year over year, and AI infrastructure will add $401 billion in spending in 2026. In another Gartner IT spending forecast, data center systems spending is expected to exceed $650 billion in 2026, up 31.7%, while server spending is projected to grow 36.9%.

That means when an ordinary person tries to buy memory, the competitor is not just another ordinary person. The competitor is an AI data center, a hyperscaler, a cloud provider, a company that can sign long-term agreements, reserve capacity early, and use enormous capital spending to buy priority in the supply chain.

In that competition, ordinary people lose.

Even Old Memory Is Becoming Valuable Again

The stranger part is that the shortage is not limited to advanced products like DDR5 and HBM. Even old memory is becoming valuable again.

In June 2026, TrendForce reported that structural tightening in mature-node DRAM supply was forcing some consumer DRAM buyers to adopt legacy memory products to secure larger supply allocations. Demand for older products such as DDR2 and DDR3 has returned. TrendForce estimated that DDR2 contract prices would rise about 55% to 60% in the second quarter of 2026, followed by another 35% to 40% increase in the third quarter.

Even more striking, TrendForce said some OEMs and ODMs were downgrading memory specifications to control system costs. Some DDR4 designs were being replaced with DDR3 solutions, and certain DDR3-based products were being redesigned to use DDR2.

That sounds like regression, but it fits the current reality.

When new memory is too expensive and supply is too tight, vendors look for ways to reduce specifications. A user may see a “new device,” but inside it may have a more conservative configuration, less memory, slower storage, and longer wait times. A company may see a “usable server” or a “usable endpoint,” but it may no longer offer the same performance that the same price used to buy.

The technology industry appears to be moving forward, but cost pressure can pull many products backward.

That is a quiet form of degradation.

MacBook Prices Are Rising, and the Notebook Market Is Cooling

On July 1, 2026, TrendForce said Apple’s across-the-board MacBook price increases had reshaped market expectations. It forecast Apple notebook shipments of about 23.1 million units in 2026, while global notebook shipments were expected to decline 13.6%.

This is not just an Apple story. One line in TrendForce’s explanation matters: strong AI server demand continues to divert semiconductor supply-chain resources away from consumer electronics, while prices for memory, power management ICs, and key raw materials such as gold and copper remain elevated, keeping pressure on notebook manufacturing costs.

In other words, MacBook price increases are not merely about Apple wanting higher margins. The entire cost structure has changed. Apple simply has the brand power and ecosystem stickiness to pass more of that cost on to consumers.

As for the iPhone 18, what we have today is still mostly rumor and analyst speculation. It should not be treated as confirmed fact. But we do not need to bet on whether one specific iPhone model will rise in price. The industry data already shows the direction: Gartner expects smartphone prices to rise 13% because of memory costs, and TrendForce forecasts global smartphone production to fall about 16.2% year over year in 2026 to 1.051 billion units. TrendForce also warned that the decline could be even sharper if memory price increases remain elevated and brands are forced to raise retail prices repeatedly.

Phones, computers, and servers are all being affected by the same thing: memory and storage are no longer cheap.

Companies Are Not as Comfortable as They Look

There is a common misunderstanding: if ordinary users cannot afford upgrades, surely companies can.

Not necessarily.

In November 2025, Network World cited a Counterpoint Research report warning that DDR5 64GB RDIMM modules, widely used in enterprise data centers, could cost twice as much by the end of 2026 as they did in early 2025. The report also said enterprise procurement teams were already feeling price increases across servers, PCs, and smartphones, and that most companies have limited leverage in choosing memory suppliers unless they are hyperscalers or large AI data center buyers.

That means companies are not standing outside the AI boom. They are also being squeezed by AI data centers.

Meta offers a symbolic example. In July 2026, Network World reported that Meta developed a custom CXL chip called Vistara to reuse older DIMMs from decommissioned servers in newer machines. Meta said the performance of about 40% of its millions of servers was limited by insufficient memory, while RAM often lasts about twice as long as the rest of the machine.

That is revealing.

If even Meta is seriously studying how to reuse old memory, what will ordinary companies do when memory prices rise? Most likely: delay upgrades, lower specifications, reuse old parts, stretch budgets, and tolerate the slowness they can tolerate.

We used to say the cloud was powerful. Now we may need to add: even the cloud is worried about memory.

Software Is Also Becoming Less Careful About Performance

Hardware getting more expensive and replacement cycles getting longer are only half of the problem.

The other half is that software is becoming less restrained.

AI makes code faster to produce, but it does not automatically make code more efficient. In fact, it can push software production into a new kind of expansion: more features, more dependencies, more background services, more generated code, more shallow abstractions, and more implementations that are accepted simply because they run.

In the past, writing slow code took a meaningful amount of human time. Now AI can quickly generate a feature that looks complete and structurally plausible. But performance, memory use, edge cases, and long-term maintenance still need human review.

The problem is that humans may not actually review them.

When requirements move faster, iteration moves faster, and management expectations move faster, software naturally gets heavier. Meanwhile, users’ computers are not getting stronger at the same pace. In many cases, they are being replaced later because hardware is more expensive. So old computers have to run new software, low-memory devices have to run high-memory apps, and ordinary phones have to carry more AI features, background sync, and cloud service hooks.

What users feel is not “AI made this software smarter.” What they feel is: why does a normal app take so long to open?

This is another way AI can damage performance culture. It makes generation cheap, while making restraint expensive.

Compute Inflation Reaches Everyone

Many people think AI costs only live on cloud companies’ balance sheets. OpenAI, Anthropic, Google, Meta, and Microsoft buy GPUs, build data centers, and burn electricity. What does that have to do with me?

A lot.

Those costs do not stay inside data centers. They travel through the supply chain and reach everyone.

AI competes for HBM, so memory makers shift capacity toward HBM. HBM crowds out conventional DRAM, so PC and smartphone memory prices rise. Memory prices rise, so device prices rise. Device prices rise, so users delay upgrades. Upgrade cycles stretch, so older devices remain in use. Older devices run heavier software, so the user experience gets slower.

The chain is not complicated:

AI data centers get faster, ordinary devices get more expensive.

Ordinary devices get more expensive, replacement cycles get longer.

Replacement cycles get longer, the share of old devices rises.

The share of old devices rises, software feels slower.

At the same time, enterprise servers become more expensive, and companies either pass the cost to customers or compress internal IT budgets. In the end, users may pay more for devices, more for services, and may still work on slower machines inside companies.

This is what I would call compute inflation.

It is not just GPU prices. It is the rising cost of every device that computes, stores data, and consumes memory. PCs, phones, servers, SSDs, RAM sticks, cloud services, and AI subscriptions all begin to affect one another through the same cost pool.

A Few Places Get Faster. Many Places Get Slower.

AI is not worthless.

It can write code, summarize documents, generate content, assist search, and improve the productivity of some people. It will also push chips, servers, data centers, and energy systems forward.

But that does not mean the whole society’s experience becomes faster.

The opposite may happen. We may enter a strange phase in which a small number of AI data centers have extreme compute density, fast networks, huge memory pools, and enormous storage capacity, while ordinary users keep using older computers, smaller memory configurations, slower SSDs, and crowded phone storage. Software companies keep adding AI to products, operating systems keep getting heavier, web pages keep expanding, and apps keep adding background tasks.

The cloud gets faster. The local device gets slower.

The model gets smarter. The machine struggles harder.

The technology gets more advanced. The user hesitates longer before upgrading.

That may be the most ironic part of the AI era. It does not distribute “fast” evenly to everyone. It concentrates speed in a few expensive infrastructure layers, then spreads the cost across society.

So when I look at AI now, I do not only think of intelligent assistants, automatic coding, and a productivity revolution.

I also think of more expensive RAM, delayed upgrades, rising MacBook prices, unaffordable DDR5, old computers staying alive, downgraded configurations, bloated software, and the seconds users spend waiting after clicking an ordinary app.

Because of AI, the world is getting slower.

References