AI models are ushering in their very own "iPhone moment" — what happens next?

God Translation Bureau is a compilation team under 36Kr, focusing on fields such as technology, business, the workplace, and life, and mainly introducing new technologies, new ideas, and new trends from abroad.

Editor's note: The exponential growth of AI capabilities is retreating into the background. When benchmark scores are no longer the core story, it will become an invisible infrastructure like optical networks. This article is from a compilation.

Recently, there has been a lot of discussion about how quickly foundation models have become like the release of each new generation of iPhones. Before a new model comes out, everything seems ordinary and uninteresting. But this analogy is not appropriate. I have a more boring but more accurate analogy that can explain the development logic so far and its future evolution trend.

I'm lucky enough to have witnessed the change of the past five technological cycles. Because of this, I can discern the patterns in the history of technological development. No matter what the technology is, we always go through a process from "being amazed" to "finding it ordinary and just a tool for work." A technology will eventually become something we take for granted.

Remember when broadband first emerged? It was in the late 1990s, and it was like magic. I installed a DSL connection in my apartment in the East Village at that time. By 2020, I had become a part of the gigabit network society. Speed has faded from people's focus. We just keep consuming the Internet, experiencing the joy and impact it brings, without even thinking about the network speed.

Artificial intelligence will go through the same process. However, let me give you a few more examples.

The Clock Speed War

If you lived through the 1990s, you witnessed the war between megahertz (MHz) and then gigahertz (GHz). From the mid - 1970s to the early 2000s, for about thirty years, the personal computer industry was completely competing around the operating speed of processors. Intel and AMD were engaged in a clock speed race. The Pentium 4 was launched in November 2000 with a starting clock speed of 1.3 GHz and eventually reached a peak of 3.8 GHz. In that era, the faster, the better.

The Pentium 4 era later became infamous, not because of its speed, but because of its heat generation. Chips that once consumed only 15 to 20 watts of power started to consume 70, 80, or even over 100 watts, generating more waste heat than useful work. So, the entire industry readjusted its direction. People stopped just talking about clock speed and began to focus on "performance per watt" (energy efficiency ratio). Multi - core architectures replaced single high - speed cores. The M1 chip launched by Apple at the end of 2020 made performance per watt the core selling point. By 2021, ARM itself announced that performance per watt was the new Moore's Law. The core question became: How can a chip work in a device so thin that you forget it's there, consuming almost no power and generating almost no heat?

The chip has become "invisible." It is no longer the product itself but a part of the product. Today, when people buy a MacBook Air, no one asks about the clock speed of the M - series chips. They can feel that the battery lasts all day and the machine stays cool without even thinking about the processor. This is exactly the key point.

It took about 30 years from the birth of the first commercial microprocessor to 2005 when the clock speed hit the ceiling. It took another 15 years until the M1 chip redefined the rules of the game. From blindly pursuing "parameters and speed" to achieving "invisible efficiency," the entire span took about 45 years.

The Smartphone Upgrade Treadmill

Smartphones have followed a compressed version of this trajectory and developed at a faster pace because the entire industry has learned from previous lessons, and the market scale is much larger, with an operation speed far exceeding the slower personal computer purchase cycle.

The original iPhone released in 2007 brought a brand - new way of doing things. Before that, nothing could do what it did. The next three or four years were a sprint - the camera improved significantly, the screen quality took a leap, LTE replaced 3G, the phone's form factor was basically set, and the App Store created an ecosystem with its own gravitational pull. Annual upgrades were reasonable. Each new phone was truly different from the previous generation. It was so exciting. I still remember when I got the iPhone 4; I was completely shocked by its powerful functions.

However, the plateau period came faster than anyone expected. By the early 2010s, the core experience of smartphones had basically reached perfection in terms of functionality. The gap between annual models shrank from obvious to negligible. The once - annual upgrade cycle has been extended to two years, three years, or even longer. Half of my family members are still using the iPhone 15 or iPhone 16 and have absolutely no intention of upgrading. This is quite common. Industry data shows that nearly a quarter of users extend their phone - replacement cycle to three or four years.

Samsung launched a phone with 100 million pixels. The pixel count did increase, but it didn't solve any real problems faced by most users. OnePlus made faster charging speed its core selling point. This is also good, but it's not exciting enough to make people willing to pay for it. And those truly useful functions - battery life, photo quality, storage space - are the most boring and unexciting when they are released. Those truly important things are becoming infrastructure. Smartphones haven't gotten worse. They've just become good enough that their excellence no longer sparks conversations. Upgrading to a new phone has become a matter of course rather than a strong desire.

This is what my friend Christian Lindholm (who worked at Nokia and later joined the design firm Fjord) calls the "Principle of 'Obvious' Design."

Great design means that end - users can subconsciously know how to operate a knob or a button just by looking at it, without even thinking. It goes without saying that this knob is for turning up the volume, or this button is for returning to the home screen. This "obvious" factor is at the core of all great designs - from the iPhone to Braun's alarm clock radio.

The same is true for underlying infrastructure technologies. Compared with the long 45 - year cycle of personal computers, we first noticed this perceptual plateau period about 7 years after the release of the original iPhone (around 2013 - 2015). From a disruptive change to being ordinary, the entire cycle is about a decade.

The Wrong Analogy

In the previous two cases, the underlying technologies did reach a plateau in terms of measurable indicators - the clock speed stopped increasing, and the annual improvements in phones were negligible. That sense of ordinariness corresponds to a real slowdown in the improvement of technological capabilities.

But this is not what is currently happening with artificial intelligence. The capabilities of AI have not reached a plateau. Its development curve is still accelerating. The leap from GPT - 4 to GPT - 5 is not a tiny, diminished increment. Inference models, multimodal capabilities, and the proliferation of open - source weight models (which have turned what were once closed - source exclusive advantages into mass - market commodities) - these are all real leaps.

The upcoming sense of ordinariness (which is already emerging in some marginal areas) is of a completely different nature. It's not a slowdown in the growth of technological capabilities, but a migration of artificial intelligence from a "hot topic" to "infrastructure." It will retreat into the background. It will change from "something you're thinking about" to "the cornerstone that makes everything else work." Just think about what will happen when the iPhone 18 deeply integrates Gemini into its operating system. (Of course, based on what I know about Apple, they might mess it up.)

This is the commodification of infrastructure. There is a much better historical analogy for this.

The Unnoticed Light

I was lucky enough to witness the dawn of optical networks. As a young journalist at that time, I reported on it with sincere excitement. George Gilder was preaching the "telcosm" to us young, attentive guys. The future seemed to lie in those glass fibers thinner than human hair.

In the mid - 1990s, the Internet backbone that carried intercontinental data ran at a speed of only 45 megabits per second (Mbps). Home users had to wait several minutes to download a photo via dial - up Internet. Bandwidth limitations were everywhere.

To understand what changed all this, you need to understand how light travels in optical fibers. A single glass fiber can only carry a limited amount of data at a single wavelength - you can think of it as a single lane on a highway. The insight of wavelength - division multiplexing (WDM) technology is that you can send multiple signals through the same fiber simultaneously, each using a different color of light, just like a prism decomposes white light into a spectrum. Each color carries its own independent data stream. One fiber thus becomes multiple. And dense wavelength - division multiplexing (DWDM) takes it a step further, stuffing dozens or even hundreds of closely spaced wavelengths into a single fiber. Early systems could only carry a few channels, while DWDM can ultimately support 96 channels simultaneously, each with an independent wavelength, each carrying a full - flow data stream.

This technology never had a press conference or any media reviews. DWDM was commercially deployed in the mid - 1990s and started doing something extraordinary yet completely silently: what seemed like limited capacity became almost unlimited in practical applications.

In the early 2000s, the capacity of a single - wavelength channel climbed from gigabits per second to 100 gigabits per second. Modern DWDM systems can carry 51.2 terabits per second (Tbps) of data in a single pair of fibers. The optical fibers deployed in the 1980s are now transmitting signals 645 times faster than 20 years ago, without laying a single new cable underground. It is estimated that the theoretical capacity of a single standard fiber exceeds 600 terabits per second, which means that the current deployment only uses about one - sixty - thousandth of the carrying capacity of this glass fiber.

No one writes articles to praise it, and there's no need to. Because DWDM works, no one notices the bandwidth problem anymore. YouTube was born, Netflix became possible, and Zoom online meetings during the pandemic were also made possible. The capacity has been growing silently for twenty years, supporting all the applications above it, without requiring anyone's tribute or gratitude.

It's not like the clock speed trajectory of personal computers - where capabilities slow down and the topic shifts; nor is it like the trajectory of smartphones - where the category is already good enough and no longer stimulates the desire to buy. The curve of optical networks is: technological capabilities continue to grow - and are still growing - while the topic has completely detached from it because this growth has been integrated into everything and no longer requires anyone's deliberate attention.

Where Is Artificial Intelligence Heading?

The capabilities of artificial intelligence will continue to climb. There is no reason to think that the research curve will flatten. Models will become more powerful, more efficient, and more specialized. The open - source weight ecosystem will narrow the gap between cutting - edge technologies and general commodities. Inference costs will continue to decline (as they have been), dropping by an order of magnitude approximately every one or two years. Its raw capabilities, like the raw bandwidth of DWDM, will continue their silent exponential growth.

What will stop growing is the discussion about it. The breathless and sensational reports about each new model release now feel different from those in 2022. The release speed is getting faster, the benchmark scores are constantly rising, but the sense of surprise is gradually diminishing.

When GPT - 3 appeared, it felt like a miracle, a new iPhone moment. When GPT - 4 arrived, it felt like a major upgrade, like the launch of the M1 chip. Now, as the fifth - and sixth - generation models start to circulate, the questions people ask have changed. It's not because the models' capabilities have weakened, but because the technological capabilities themselves are no longer the core story. The real story lies in what this capability is wrapped in.

AI will become like the former DWDM: the underlying layer that you don't see but that makes everything above it work. It will be in the camera that determines how a photo is exposed; in the chip that manages the power consumption of a laptop; in the hospital monitor that detects the early deterioration of a patient's condition; in the contract that has been reviewed before you read it.

The models won't disappear. They just won't be the unit of discussion anymore. No one talks about DWDM when starting a video call. When people use products powered by foundation models, no one will talk about the foundation models themselves.

This is not good news for foundation model labs that want to keep coming up with new metrics and the valuations of companies like Nvidia. But the industry is not stupid. Just look at the stock trend of Ciena from its IPO to the first ten years, and you can almost predict the movement curve of the stocks of these labs, even if you can't accurately predict the scale.

It took about a decade for the iPhone to transform from a "disruptive breakthrough" to "infrastructure." The clock speed story of personal computers took longer because of different interests and different ways of industry operation. But artificial intelligence is developing faster than both of them. The infrastructure - based architecture already exists in enterprise software, developer tools, and hardware roadmaps. The transformation on the consumer side usually lags behind by a few years. My prediction is: by 2028, people's question will no longer be "which AI is being used," but "what can this thing do that it couldn't do before." The models will have gone underground by then.

The companies that build foundation models are not necessarily the ones that will define the AI era. DWDM was built by carriers such as Nortel and Lucent, most of which are now forgotten or merged. And the Internet layer above them - Google, Amazon, Netflix - has captured the value released by the optical network infrastructure. Infrastructure provides possibilities, but it cannot determine who will be the final winner.

Commodification is already in progress. Open - source weight models are narrowing the advantages that closed - source cutting - edge models once had. Inference costs are dropping so fast that model capabilities themselves are no longer a moat that can be defended. The real moat will be those specific applications that make the underlying capabilities both "indispensable" and "invisible."

Don't stare at the benchmark scoreboard. Pay attention to how the benchmark scores disappear from the conversation. When we stop asking which model scores the highest in the inference test and start asking why our software seems smarter without any changes, then this transformation is complete.

The optical fibers have been buried underground. We just don't know what will run on them yet.

Translator: boxi.