AI models are having their iPhone moment. What’s Next?

Lately there has been a lot of talk of how the foundational models are quickly becoming like every other iPhone release. They are ho-hum, till the next one comes around. But it is not the right analogy. I have a more boring, and more accurate, analogy that will explain the growth so far, and how it will evolve.

I have been fortunate enough to have been involved with the last five cycles of technology. As a result, I have been able to see patterns in the history of technology. It doesn’t matter what the technology is – we go from shock and awe to ho-hum, go-to-work. A technology eventually becomes invisible to us.

Remember when broadband came around? That was in the late 1990s, and it was magical. I had a DSL connection in my East Village apartment. By 2020 I was part of the gigabit society. Speed had faded from the foreground. We just kept consuming the internet, its joys and jolts, without thinking about speeds.

Same will happen to AI. But let me give you some more examples.

The Clock Speed Wars

If you were around in the 1990s, you lived through the megahertz and then the gigahertz wars. For roughly three decades, from the mid-1970s through the early 2000s, the personal computer industry was all about how fast the processor ran. Intel and AMD were in a clock speed race. The Pentium 4 launched in November 2000 at 1.3 GHz and eventually topped out at 3.8 GHz. Faster was better.

The Pentium 4 era became notorious not for its speed but for its heat. Chips that once drew 15 to 20 watts were routinely hitting 70, 80, even over 100 watts, generating more thermal waste than useful work. The industry regrouped. It stopped talking about clock speed and started talking about performance per watt. Multi-core architectures replaced the single fast core. Apple’s M1 chip in late 2020 made performance per watt the story. By 2021, ARM itself declared that performance per watt was the new Moore’s Law. The question was what the chip could do while drawing almost no power, producing almost no heat, inside a device thin enough to forget you were carrying it.

The chip became invisible. It stopped being the product and became part of the product. Nobody buying a MacBook Air today asks about the M-series clock speed. They feel the battery last all day and the machine stay cool and they do not think about the processor at all. That is exactly the point.

It took roughly 30 years from the first commercial microprocessors to the clock-speed plateau in 2005. Another 15 years to the M1 reframing. The total arc from speeds and feeds to invisible efficiency ran about 45 years.

The Smartphone Upgrade Treadmill

The smartphone followed a compressed version of the same arc, running faster because the industry had already learned some lessons and because the market was way larger, and it moved much quicker than slower PC buying cycles.

The original iPhone, announced in 2007, was a genuinely new way of doing things. Nothing before it did what it did. The next three or four years were a sprint – cameras improved dramatically, screens improved dramatically, LTE replaced 3G, form factors settled, app stores created ecosystems that generated their own gravity. Annual upgrades were rational. Each year’s phone was genuinely different from the one before. It was so exciting. I still remember getting the iPhone 4 and being just simply astounded by how much it could do.

The plateau came faster than anyone expected. By the early 2010s, the core smartphone experience was essentially perfected in functional terms. The differential between one year’s model and the next shrank from obvious to marginal. Upgrade cycles that had been annual stretched to two years, then three, then beyond. Half my family is still using the iPhone 15 or iPhone 16, with no intention to upgrade. It’s typical. Industry data shows nearly a quarter of users stretching to three or four years between replacements.

Samsung introduced a 100 megapixel phone. It might have been pixels, but it didn’t solve any real problem most users had. OnePlus made faster charging a headline feature. Again good, but not amazing enough to spend more dollars. The features that remained genuinely useful – battery life, camera quality, storage – were also the least exciting to announce. The things that mattered were becoming infrastructure. The smartphone did not become worse. It became good enough that its goodness stopped generating conversation. The upgrade became an assumption rather than a desire.

This is what my friend Christian Lindholm, who once worked for Nokia and then for the design firm Fjord, calls his “of-course principle of design.”

Great design means that one look and the end user reacts by knowing what to do with a knob or a button, without as much as even thinking about it. Of course this knob is what turns the volume up, or brings up the home screen. This of course factor is at the heart of every great design – from the iPhone to the Braun alarm radio.

It is the same for underlying infrastructure technology. In comparison to 45 years for the PC, we first noticed the perceived plateau in roughly 7 years from the release of the first iPhone, somewhere around 2013 to 2015. The total arc from cataclysmic change to ho-hum ran, give or take, a decade.

The Wrong Analogy

In both previous cases, the underlying technology actually did plateau in measurable ways – clock speeds stopped climbing, annual phone improvements shrank. The ho-hum feeling tracked a real slowdown in capability improvement.

That is not what is happening with AI. Capability is not plateauing. The curve is still accelerating. GPT-4 to GPT-5 is not a shrinking increment. Reasoning models, multimodal capabilities, the proliferation of open-weight models that commoditize what were closed advantages – these are real jumps.

The ho-hum that is coming, the one already arriving at the edges, is something different. It is not the slowing of capability. It is the migration of AI from topic to infrastructure. It will go into the background. From the thing you think about to the thing that makes everything else work. Consider when the iPhone 18 marries Gemini into its operating system. (Of course, knowing Apple they will bungle it up.)

It is infrastructure commoditization. And there is a much better historical analogy for it.

The Light Nobody Sees

I was lucky to watch the dawn of optical networking. As a young reporter I wrote about it with genuine excitement. George Gilder was preaching telcosm to us young punks paying attention. The future seemed to live in strands of glass thinner than a human hair.

In the mid-1990s, the internet backbone that carried data across continents ran at 45 megabits per second. Home users waited minutes to download a photograph over dial-up. The constraint was everywhere.

To understand what changed it, you need to understand how light moves through fiber. A single strand of glass can carry only so much data at one wavelength – think of it as one lane of a highway. Wavelength Division Multiplexing, or WDM, was the insight that you could send multiple signals down that same strand simultaneously, each on a different color of light, the way a prism splits white light into its spectrum. Each color carries its own independent stream of data. One fiber becomes many. Dense Wavelength Division Multiplexing – DWDM – pushed this further, packing dozens, then scores, of tightly spaced wavelengths onto a single fiber. Where earlier systems carried a handful of channels, DWDM eventually supported 96 simultaneous channels, each at its own wavelength, each carrying its own full stream of traffic.

There were no launch events for this. There were no reviews. DWDM entered commercial deployment in the mid-1990s and began doing something remarkable, and entirely invisible: capacity that seemed finite became functionally unlimited.

In the early 2000s, channel capacity climbed from gigabits per second to 100 gigabits per second per wavelength. Modern DWDM systems can carry 51.2 terabits per second down a single fiber pair. Fiber deployed in the 1980s is now running signals 645 times faster than it was 20 years ago, with no new cable in the ground. One estimate puts the theoretical capacity of a single standard fiber strand at over 600 terabits per second, meaning current deployments use roughly 1/60,000th of what the glass can carry.

Nobody wrote about this. Nobody had to. Because DWDM worked, nobody noticed the bandwidth problem. YouTube became possible. Netflix became possible. Zoom calls during a pandemic became possible. The capacity was simply there, having grown silently for twenty years, enabling everything above it without requiring acknowledgment from anyone.

Not the PC clock-speed arc, where capability slowed and conversation shifted. Not the smartphone arc, where the category became good enough and stopped generating desire. The optical networking curve, where capability kept growing – is still growing – while the conversation moved entirely away from it, because the growth had become embedded in everything and required no one to pay attention.

Where AI Goes From Here

AI capability will keep climbing. There is no reason to expect the research curve to flatten. Models will become more capable, more efficient, more specialized. The open-weight ecosystem will compress the gap between frontier and commodity. Inference costs will continue to fall, as they have been falling, by an order of magnitude roughly every year or two. The raw capability, like the raw bandwidth of DWDM, will continue its silent exponential.

What will stop growing is the conversation about it. The breathless coverage of each new model announcement has a different texture than it did in 2022. The releases come faster, the benchmarks climb, but the surprise is attenuating.

When GPT-3 appeared, it felt like a visitation. A new iPhone moment. When GPT-4 arrived, it felt like a significant upgrade. Like the arrival of the M1. Now, as fifth and sixth generations circulate, the question people ask has changed. Not because the models are less capable. Because capability is no longer the story. The story is what the capability is inside.

AI will become what DWDM became: the layer you cannot see that makes everything above it work. It will be inside the camera that decides how to expose the photograph. Inside the chip that manages the laptop’s power. Inside the hospital monitor watching for early deterioration. Inside the contract that was reviewed before you read it.

The models will not disappear. They will stop being the unit of conversation. No one talks about DWDM when they open a video call. No one will talk about the foundation model when they use what the foundation model made possible.

This is not good for the valuations of foundation labs and the ilk of Nvidia, who want to come up with new metrics. But the industry isn’t stupid. Just follow the Ciena stock from IPO to first 10 years, and you can almost predict the curvature, if not the scale, of the stocks of these labs.

The iPhone took roughly a decade to shift from rupture to infrastructure. The PC clock-speed story took longer because the stakes were different and the industry moved differently. But AI is moving faster than either. The infrastructure framing is already present in enterprise software, in developer tools, in hardware roadmaps. The consumer shift tends to lag by a few years. My estimate: by 2028, the question will no longer be “which AI” but “what does this do that it did not do before.” The model will have gone underground.

The companies building foundation models are not necessarily the companies that will define the AI era. DWDM was built by Nortel, Lucent, carriers now mostly forgotten or absorbed. The internet layer above them – Google, Amazon, Netflix – captured the value that the optical infrastructure enabled. Infrastructure enables; it does not determine who wins.

Commoditization is already underway. Open-weight models are compressing the advantage that closed frontier models once held. The cost of inference has fallen so fast that capability is no longer a defensible edge. The edge will be the particular applications that make the underlying capability feel indispensable and invisible at once.

Watch not the benchmark. Watch the disappearance of the benchmark from the conversation. When we stop asking which model scores highest on reasoning tests and start asking why our software feels smarter without us having changed anything, the transition will have happened.

The optical fiber is already in the ground. We just don’t know yet what runs on it.