In the context of semiconductor design, Chris Bergey, Arm's Senior VP and GM of the Client Line of Business, referred to the "waterfall effect" as a way to describe a long-standing industry practice: today's flagship chip technology eventually cascades down to tomorrow's mid-range and budget devices. Chipmakers often amortize their massive development costs by reusing last year's high-end system-on-chip (SoC) in lower-priced products or selling slightly detuned versions for cheaper devices. Of course, consumers have historically benefited from this waterfall of innovation. Advanced features like faster processors, better graphics, and new connectivity standards all eventually reach affordable gadgets as the cutting-edge moves forward.
Yet this "waterfall effect" is currently under strain, according to Bergey. In an interview, Bergey explained to us that the surging complexity and costs at the top end are widening the gap between premium and mainstream chips. The result is a potential breaking point where simply repackaging last year's silicon for cheaper devices "may not work" anymore.
Premium chips are more expensive than ever
Prices have been rising for a long time
The cost to build leading-edge chips has skyrocketed, which threatens to disrupt the economics of trickle-down tech that allows many of us to buy high-end hardware from previous years, or see the benefits of high-end features eventually making their way to budget-oriented solutions. Developing and manufacturing a flagship processor on the latest process node now costs "hundreds of millions of dollars", according to Bergey. For example, Qualcomm's latest Snapdragon 8 Elite Gen 5 is rumored to cost upwards of $200 per chip, and those costs have been rising every single year. To put that in perspective, a single premium smartphone SoC can cost as much as an entire budget device.
One big reason for this increase in cost is that cutting-edge fabrication has become far more expensive. Foundries like TSMC charge much more for newer, smaller transistor nodes. Apple's A-series chips, which form the basis of Apple's M-series, are a good way to illustrate that point. Back in 2013, the Apple A7 (built on a 28nm process) was manufactured on wafers that cost about $5,000 each, whereas by 2024, the A17/A18 Pro (using a 3nm-class TSMC process) saw wafer prices around $18,000. That's a three-and-a-half-times jump in just a decade, which has seen the cost per millimetre of wafer jump dramatically.
However, the true breaking point has been the onset of artificial intelligence and the need for significantly more functionality in today's chips. These days, modern smartphone and PC SoCs aren't just a CPU and GPU, as they also integrate dedicated AI accelerators (NPUs), advanced image processors for cameras, 5G modems, and other specialized IP blocks. Bergey notes that artificial intelligence workloads in particular are stretching chip designs: you now have to accommodate "three computing elements – CPU, GPU, and NPU – and those are all growing" to meet AI demands. It seems as if each generation aims for double-digit performance gains, often by adding transistors and custom accelerators, and this essentially means that die sizes are kept consistently large. Again, going back to Apple, the company's A-series went from about 1 billion transistors on the A7 to 20 billion on the A18 Pro, and they're definitely not the only ones.
These chips are not shrinking much despite moving to advanced nodes, as those advanced nodes aren't offering the density increase from generation to generation that we used to see in the past. On this point, Bergey puts it simply: "You get the performance [uplift]; you're not necessarily getting shrinkage." On top of that, designers often trade the theoretical density gains for adding more on-chip capabilities or maintaining higher performance at a given die area. As a result, a flagship SoC might stay between 100mm squared and 300mm squared (Apple's M1 landed at 118.91mm squared, for example), but on a process node that costs far more per millimetre squared, resulting in a much higher unit cost. And because leading-edge nodes like 3nm use expensive techniques (extreme ultraviolet lithography, new materials) and have yield challenges, the manufacturing cost per chip stays high. In short, bleeding-edge performance now comes at a steep price, and it's rising with each generation.
It's not just the processor silicon, either. Memory costs in devices are also climbing as a side-effect of AI and software. Premium smartphones and laptops now often ship with 12 to 16 GB of RAM (plus larger storage) to accommodate AI models and hefty operating systems. "The other thing that AI really stresses is memory bandwidth, but even especially memory size," Bergey tells me, noting that where once the OS and apps drove memory needs, it's now on-device AI models that dictate those needs based on their high consumption of RAM. More or faster memory chips, and the power to run them, mean a higher bill-of-materials costs for device makers, adding to the pressure that raises prices at the high end.
As for why these advancements are coming to devices in the name of on-device AI, Bergey's answer here is particularly insightful:
It's interesting, because I have a lot of people that are like, you know, they're skeptical. Like, "Hey, tokens are so cheap in the cloud, why would, why do we need a new on device AI?" The part that makes me quite assured that it's going to happen is that I can tell you, most of those players that own the cloud that are doing that are like, we need to get this stuff on devices, because, you know, as we know, we're in the early days of AI. And there's a lot of... you know, the cost economics are cloudy at best right now, relative to what things are priced at versus what they cost.
In this instance, a great example of the economics of pricing comes from the API costs of using AI through the likes of OpenAI, Google, or Anthropic. While we don't know how much it costs OpenAI to execute a single request through its API, we do know that the company hasn't been able to profit from its advancements thus far.
New strategies in the face of AI
Chip building is a different game these days
Companies have started to shift their strategies in order to continue offering technology in multiple price segments. Rather than relying on things like the binning process (where a CPU is categorized after production, based on its performance and capabilities), companies have started to design specifically for new segments that don't quite meet the high-end, but are better than the classic mid-range. The Arm C1-Premium core is a great example of those, coming in at roughly 35% smaller in area than Arm's top-end C1-Ultra core, yet it achieves similar performance on many key benchmarks. This kind of innovation is one way to address this issue, as it allows a chip focused on the mid-to-high-end market to be built more cost-efficiently without making use of last year's high-end core, and without cutting down the high-end core of today.
Even at the very low end, we see technology steps rather than a smooth cascade. For example, in cellular connectivity, 5G modems are finally reaching entry-level devices, but OEMs didn't achieve this by slapping last year's expensive 5G chip into a $100 phone. Instead, Bergey tells me, that "no one is offering a 5G chipset at the same price as the 4G. It is probably at about a 15% premium, but the carriers want the 5G." He mentions how consumers may finally see that value, and the original cost of that technology has still come "way, way down" since then. He likens it to meeting in the middle, where the cost probably comes down by about 70% or so, and the consumer either sees the value proposition for that smaller price premium, or simply doesn't have another choice as the industry has moved on to that technology as a default by that stage.
The current wave of AI is a double-edged sword in this regard. On the one hand, AI features are a big part of what's driving much of the added cost and complexity in modern chips. On the other hand, AI could also drive new value and revenue that justify those costs. Companies are betting that consumers will see on-device AI capabilities, such as intelligent voice assistants, generative AI in apps, advanced camera processing, and more, as worth the premium. In Bergey's view, the device overall is becoming more important to people's lives, essentially absorbing functionality that might have once lived in the cloud or on a PC. If smartphones and personal devices truly become "AI companions" that assist in everyday tasks, users might be willing to spend more on hardware that delivers those benefits.
There’s also the question of new business models around AI. Some device makers can subsidize higher hardware costs with software and services income, such as tying in app store ecosystems, subscription services, or advertising platforms, that generate revenue. With AI, there may be opportunities to monetize new services (AI-driven apps, personalized AI assistants, and more), which could offset slim margins on the hardware itself. When I asked about what happens if consumers are essentially priced out of high-end hardware, given the rising costs, Bergey alluded to that exact solution. He noted that companies are examining how AI "could be disruptive to some of the monetization models" on mobile and looking to take advantage of the new use cases it enables. In short, if on-device AI makes a device far more useful or sticky for a user, there may be ways to recoup the investment beyond just the asking price.
As already mentioned, too, the cost of running AI in the cloud is a big reason for companies to try and shift those processing burdens to consumers instead. If those tasks can be handled locally on the device's NPU or GPU, companies can save massively on cloud infrastructure in the long run, and this offloading can be sold back to the consumer as a performance improvement, too. As already stated, the economics around AI are "cloudy": offloading inference is a great way to gain a clearer picture.
Tech is at a turning point
The old playbook won't work anymore
The industry is approaching a pivotal point where the old playbook of steadily cascading technology to cheaper tiers needs an overhaul. There's "tension" in the system right now, Bergey tells me, and in his view, it's primarily AI that's driven this sudden change..
Some of the technology costs could be exceeding some of the supply demand curve or the price sensitivity curve, right? I think that is a question that exists. Now, again, I am worried I'm being myopic, right? I mean, it's not that many years ago that people said nobody would ever buy a phone for $1,000 and Apple came out with $1,000 phone and now we think that's normal. So I am worried that I'm gonna regret these things, but I think that this continues to be a dance, right? Of, you know, the technology costs absolutely are going up, but what has always happened is the utility is so high that people will pay it, right? Or we'll find a technical solution. And I think that will happen, but I just would say that, you know, there's just a lot of tension in the system right now. Because of AI. And because AI puts pressure on logic, puts pressure on systems, puts pressure on memory. I think it feels like it's causing more tension than maybe what we've had in the past.
In the big picture, the breaking of the old waterfall is not so much a failure as an evolution of the industry's approach. Innovation isn't slowing down; if anything, it's happening in more places simultaneously. Bergey told me that we're entering an "exciting innovation cycle," largely driven by AI and novel computing needs. High-end devices will continue to stretch capabilities (and budgets) in ways that would have seemed unbelievable a decade ago, from AI co-processors to desktop-class performance on a mobile chip. Competition and economies of scale will then push vendors to find ways to bring those advancements downmarket in a viable form.
For tech enthusiasts and consumers alike, the outcome could still be a positive one. With that said, you might see companies reuse older flagship chips more and more, an industry trend that's been happening for a while now, in an effort to still provide high-end experiences to consumers. Bergey believes the reality is that devices will become "more important and more valuable" to us, even if they come at a higher cost. The waterfall may not flow as freely as before, but it's being redirected and reshaped, and nobody really knows where it's going just yet.
