A few weeks ago, we attended Computex, one of the world's largest technology trade shows. We've seen some pretty boring shows over the past 20 years or so, but this one was possibly the most boring yet, which will probably surprise exactly no one. The climate simply wasn't right for a fun show. DRAM pricing is still cooked, NAND pricing is equally cooked, and as a result, GPU pricing is... pretty horrible.
It's not exactly a fun time to be a PC gamer, especially if you need a new gaming PC or a platform upgrade. Now simply isn't the time for any of that. The question everyone around here has been asking is: When will the right time come? When will this nightmare be over?
It's an interesting question that no one can really answer right now, but there seem to be two distinct opinions about how all of this will play out.
There's the "you'll own nothing and be happy" crowd, and then there are those of us who are simply hoping to ride this out and return to something resembling normality. To be clear, we don't think things will ever go back to exactly how they were five years ago or so, and that's what I want to explore today.
Nvidia's Financial Flywheel
By "more normal," we mean that products such as DDR5 memory, or DRAM in general, will eventually become cheap again, as will SSDs. In fact, with production ramping up and competition from China increasing, we could find ourselves in a situation where DRAM and NAND pricing completely collapses once it becomes clear that these AI data centers aren't actually going to be built in time.
I suspect that day is looming sooner than some realize, but maybe I'm just coping.
Still, you have to stop and question reality sometimes, because much of what is happening right now simply doesn't make sense. A recent report revealed that Nvidia sold $5.4 billion worth of GPUs to a shell company called Valor, which is being used to transfer GB200 GPUs to Elon Musk's AI company, xAI.
Nvidia is under immense pressure from Wall Street to continuously post blockbuster revenue figures. Because Valor is technically an independent entity, Nvidia can book that $5.4 billion as immediate revenue from a completed sale, moving the physical inventory entirely off its books.
To make the deal happen, Nvidia didn't just sell the chips. It also helped fund the buyer...
To make the deal happen, Nvidia didn't just sell the chips. It also helped fund the buyer. Nvidia invested roughly $1.9 billion of its own cash in Valor as a limited partner. This creates a financial flywheel: Nvidia injects equity into the shell company, which then uses that cash to buy Nvidia chips. Nvidia gets its cash back while recording massive revenue, all while securing xAI as a long-term, locked-in customer.
This structure effectively shifts much of the immense financial risk associated with unproven AI infrastructure away from the primary technology companies. A healthy company, or at least one that truly believes in AI as strongly as Nvidia claims to, would not be doing this, in my opinion. Rather, this is what you do when you know that what you're propping up is BS and you're trying to make as much money as possible before it all falls apart.
So, if that is indeed the case, why has Nvidia done this? The obvious answer is money, but is this pure greed, or was it born out of desperation? I tend to believe it's more the latter, and I'll explain why.
How GeForce Ran Out of Road on Gaming Performance
For years, it has become increasingly obvious that Nvidia can no longer achieve the kinds of performance gains it once delivered. Back in the early 2000s, Nvidia was able to release a new GeForce generation roughly every 12 months, and those updates almost always delivered substantial performance gains. As a gamer, it was basically impossible to keep up.
As exciting as this period was, it wasn't the best of times for Nvidia. In the decade between 2000 and 2010, Nvidia was a tiny fraction of the size it is today. Instead of fueling an artificial intelligence gold rush, the company was fighting a fierce battle on the PC market, primarily selling graphics cards to PC gamers and tech enthusiasts.
During this era, Nvidia's annual revenue grew from less than $1 billion to somewhere between $3 billion and $4 billion. At the time, that kind of growth would have been exciting for the GPU maker. The company was a major player in an emerging market, and the constant rollout of new products meant gamers were always scrambling to hand over their cash for the latest gear.
The good times would continue for another decade, but as 2020 approached, Nvidia faced a problem, and the company knew about it before we did. The gaming industry was larger than ever and continuing to grow. Nvidia had slowly become the dominating player in the GPU market, so on the surface, everything looked good.
The problem, however, was the product. Making it better had become both increasingly difficult and increasingly expensive.
As we said, in the early days, you could expect a new GeForce generation roughly every 12 months, and the gains were massive. In fact, it was fairly common to see a generational performance uplift of around 50% right up to the GeForce 10 series, based on the Pascal architecture. The GTX 1070, for example, was around 60% faster than the GTX 970.
Although gamers didn't realize it at the time, the GeForce RTX 20 series was the first sign of trouble for Nvidia.
Flagship GeForce GPU Comparison - Gen on Gen
Let's start by looking at 10 generations of flagship Nvidia GPUs, excluding dual-GPU models and Titan-class products. The GTX 285 offered around a 60% performance increase over the previous-generation flagship, the old 9800 GTX+. It cost just $360 and was rated for a TDP of only 215W.
The following year, Nvidia initially fumbled with the GTX 400 series but quickly replaced it with the much better 500 series, giving us the GTX 580. Another roughly 200W product, it delivered a massive 70% performance uplift.
After that, we had to wait a year and a half for the GTX 680. Although it was only 40% faster than the 580, it was still one of the most technologically significant updates Nvidia had ever pulled off. Moving from a mature 40nm process to a cutting-edge 28nm process allowed Nvidia to shift its entire engineering focus from raw computing power at all costs to extreme performance-per-watt efficiency. As a result, the GTX 680 dropped to just 170W.
Then, in 2013, the GeForce 700 series arrived, led by the GTX 780 Ti. At $700, it was 40% more expensive than the 680, but it was also around 60% faster. In fact, the $330 GTX 770 was able to match or even slightly beat the GTX 680. The 770 was essentially a tuned 680 offered at a substantial discount.
A few years later, the GeForce 900 series arrived, and the 980 Ti offered another 40% performance increase over the 780 Ti. It also packed twice as much VRAM, jumping from 3GB to 6GB.
That brings us to most people's favorite GeForce generation, including Nvidia CEO Jensen Huang's. Looking back, Jensen probably views the Pascal architecture and GeForce 10 series so fondly because they largely sold themselves. It was something he could genuinely be proud of and enjoy alongside the rest of us, without having to spin endless BS just to generate interest.
Anyway, for me personally, this was the last generation that felt genuinely exciting. The GTX 1080 Ti was a massive 70% faster than the 980 Ti. The series also offered more VRAM and was generally more power-efficient.
Power efficiency, or simply power consumption, is something we should keep in mind as we move on. The RTX 2080 Ti was released just one year after the 1080 Ti, but it offered only a 30% performance increase, the weakest generational uplift we had seen from Nvidia in a very long time. Power consumption also climbed to 270W, the highest of any Nvidia flagship GPU at that point.
For the first time, with the GeForce RTX 20 series, the GPU's generational price increase was larger than the performance uplift .
Still, this doesn't look too bad on paper because we're ignoring the more than 40% increase in price. For the first time, the price increase was larger than the performance uplift.
Nvidia tried to sweep this issue under the rug with promised technologies such as ray tracing and upscaling. We all know how well that played out for the GeForce 20 series. Not nearly as well as advertised, let's just say that. The result was Nvidia's most disappointing GeForce generation since the initial Fermi release.
Looking at the graph, it appears Nvidia managed to recover with Ampere, or the RTX 30 series, as the 3090 Ti offered a 50% increase over the 2080 Ti. Not too bad after two years.
But look a little closer and the cracks begin to appear. The 3090 Ti increased power consumption by ~80% over the 2080 Ti. In other words, Nvidia was very much brute-forcing that 50% performance increase. Not only that, but the MSRP also doubled from $1,000 to $2,000.
Two years later, in 2022, Nvidia did somewhat better with the RTX 4090. It offered another 60% performance increase while slightly reducing power consumption, and it even cost a little less at $1,600.
After that, it took Nvidia another three years to release the RTX 5090, which was just 30% faster but came with a staggering TDP of 575W and consumed 590W in our testing.
MSRP vs Gains: The Flagship Price Trajectory
Still, as we said, this picture doesn't look too bad for Nvidia, and that's because flagship GPUs mask what is really going on. But before we move on, here is a look at the MSRP of these flagship GPUs adjusted for inflation. This means the $360 GTX 285 would cost $558 in today's money.
What we can clearly see is that, in the earlier years, gamers regularly received substantial performance gains at relatively stable and predictable prices.
But as we move from left to right across the graph, the opposite begins to happen. Large performance gains become less common as higher prices become the norm.
What a $500 GPU Actually Got You, Year by Year
Now, if we look at GPUs priced around $500, or as close as we can get to that price point after adjusting for inflation, we find a very different picture.
The GTX 570 was 55% faster than the previous-generation flagship while costing slightly less, making it an excellent gen-on-gen uplift that took only a year to achieve.
The GTX 670 was a little more expensive but still delivered a solid 45% uplift after roughly 18 months. The GTX 770 was a letdown, only roughly matching the previous-generation flagship and making it just 20% faster than the GTX 670. However, it was almost 20% cheaper after adjusting for inflation, so it still represented a decent improvement in value.
The GTX 970 was a gem. Ignoring the small issue with its partitioned VRAM, it was 40% faster than the 770 for the same money. Then we arrived at Pascal, which once again delivered an amazing uplift. The GTX 1070 was 60% faster than the 970 while costing just 14% more.
Two years later, we arrived in 2019 and were presented with the worst mid-range offering Nvidia had ever released: the RTX 2060. Thanks to a price increase, it lands very close to $500 after adjusting for inflation, at $465 in today's money, while offering just a 10% performance increase over the GTX 1070. In fact, 10% is generous, as the uplift was closer to 6%. At launch, it didn't even have any ray tracing titles to hang its hat on.
The RTX 3060 Ti was a bright spot, arriving just a year later. The RTX 30 series was largely spoiled by cryptocurrency mining, though that isn't especially relevant to what we're examining here. The 3060 Ti was 60% faster than the 2060 and arguably too good, as it effectively killed Nvidia's next generation.
The RTX 4060 Ti was a disgrace. The 16GB model launched with a $500 MSRP, which is equivalent to roughly $546 in today's money. For that, you received a performance increase of just 5%.
Most recently, we got the RTX 5070, which is supposed to cost $550 and, at least for a time, actually did. Compared with the RTX 4060 Ti, it delivers a 50% performance uplift.
A clear pattern is emerging from this data. It is becoming increasingly difficult for Nvidia to deliver meaningful performance gains on a regular basis. Not only is the gap between product releases growing, but the gap between notable performance improvements has become even wider.
The Same Picture at Face Value: Mid-Range MSRPs Over Time
Now, for those who wish to claim we're cherry-picking the data to make Nvidia look bad, or something crazy like that, here is a simple MSRP comparison, again targeting the $500 price point. As you can see, despite changing some of the GPUs, the same pattern emerges. In fact, this comparison is a little more favorable to the older-generation GPUs.
This line graph clearly shows the pattern. From 2010 to 2017, we saw fairly frequent GPU releases at similar prices, delivering similarly strong performance gains. There was a small dip in the middle, but nothing overly concerning.
For example, if you bought a GTX 680, you only had to wait a few years to receive around a 60% performance increase with the GTX 980, without spending more than you did originally. Then, two years later, the GTX 1080 delivered another 60% performance increase, once again for roughly the same money.
Performance Comparison: 5-Year Generation Blocks
Beyond that, however, we start to see massive generational declines in performance gains, meaning gamers have to wait much longer for worthwhile improvements at similar price points.
This is my favorite graph for demonstrating the issue Nvidia and its customers have run into.
From 2010 to 2015, Nvidia's ~$500 GPU offerings improved by a whopping 140%, as seen when moving from the GTX 580 to the GTX 980. That represents an entirely different gaming experience, unlocking a new level of graphical fidelity in just five years.
But Nvidia did it again, and actually improved by even more, from 2015 to 2020. Although the jump from the GTX 980 to the RTX 3070 makes the 3070 look extremely impressive, as we just saw, a large portion of those gains came from the Pascal generation.
Also, keep in mind that if we avoided VRAM limitations, as the GTX 980 had only 4GB of VRAM, the performance uplift would probably be closer to 150%. That is still a sizable improvement, but because we're moving from 4GB to 8GB in this comparison, the uplift isn't solely the result of increased compute performance.
Whichever way you slice it, the result is much the same. Nvidia was more than doubling performance during this period, which brings us to the 2020-to-2025 comparison.
If this doesn't clearly illustrate the treacherous waters Nvidia was heading into, we don't know what will. A performance increase of just 50% from 2020's RTX 3070 to 2025's RTX 5070 is shockingly poor.
A New Normal: Where Does This Leave PC Gamers?
What this tells me is that even if the AI bubble does pop, and at this point I believe it's more a matter of when than if, we're never going back to the way things were. We can't. Those days are over, and there's nothing anyone can do about it, including Nvidia.
If anything, the situation is likely to get worse for both gamers and Nvidia. Nvidia relies on new process nodes to improve its products. If those nodes take longer to develop and become more expensive, we're going to see longer development cycles and more expensive products, making it extremely difficult for Nvidia to offer gamers the kinds of improvements they have grown accustomed to.
Whereas 10 years ago you could realistically upgrade your graphics card every 12 months, or certainly every 18 to 24 months, noteworthy upgrades over the past five years have been pushed out to something closer to a five-year cycle. In another 5 years, we could be looking at 7 or 8-year upgrade cycles for gamers.
So there's the time component, which will see Nvidia having to entice gamers to buy their products long after they've freshly come out of the oven. Then there are the economics of semiconductor manufacturing to contend with.
For nearly two decades before 2019, wafer prices remained relatively flat or even declined as manufacturing processes matured. However, the introduction of EUV, or extreme ultraviolet lithography, combined with rising supply chain complexity and TSMC's immense pricing power, has caused the cost of cutting-edge wafers to increase by roughly 400% over the past decade.
TSMC Flagship 300mm Wafer Cost History
| Process Node | Est. Cost Per Wafer (USD) | Key Industry Driving Factor | |
|---|---|---|---|
| 2016 | 10nm (FinFET) | ~$6,000 | Apple transitions smartphone chips to early 10nm layout. |
| 2017 | 10nm / 7nm Risk | ~$6,500 | Transition year as TSMC preps its highly successful 7nm node. |
| 2018 | 7nm (N7 - Deep UV) | ~$9,300 | Massive scaling jump. Shaders and mobile CPUs adopt N7 layout natively. |
| 2019 | 7nm+ (N7+ / Early EUV) | ~$10,000 | TSMC introduces EUV machines for the first time on critical layers. |
| 2020 | 5nm (N5 - Full EUV) | ~$16,900 | The first price explosion. Multi-patterning EUV tools push costs up by roughly 70%. |
| 2021 | 5nm (Mature Node) | ~$17,000 | Global chip shortages maximize fab utilization rates across the board. |
| 2022 | 4nm (N4 Evolution) | ~$18,000 | Post-pandemic supply crunch: TSMC enforces a broad 10% structural hike. |
| 2023 | 4nm / 3nm Early (N3B) | ~$18,500 | Early 3nm variants struggle with high defect densities and yield challenges. |
| 2024 | 3nm (N3E Mainstream) | ~$20,000 | Crucial baseline node for next-gen AI accelerators and flagship mobile silicon. |
| 2025 | 3nm Upgraded (N3P) | ~$21,000 | High-demand premium; TSMC levies an additional 4% to 5% price adjustment. |
| 2026 | 2nm (N2 - Gate-All-Around) | ~$29,000 | The second price explosion. Transition away from FinFET to nanosheet design. |
Nvidia wasn't willing to absorb those costs through lower margins. In fact, it wanted to increase its margins, so gamers simply got less. That's why we've gone from a 256-bit-wide memory bus and 8GB of VRAM with the GTX 1070 in 2016 to a 192-bit-wide memory bus and just 12GB of VRAM with the RTX 5070 nearly 10 years later. Of course, Nvidia could reduce its margins, but given the technological slowdown we've illustrated in this article, even that isn't a viable long-term strategy. At best, it would buy the company a little time.
This is why Nvidia started leaning so heavily on software features such as DLSS upscaling and frame generation. These technologies don't require making the die larger, and in a sense, they allow Nvidia to simulate performance gains over previous-gen hardware.
So that explains the GeForce side of the equation. But for Nvidia, selling directly to consumers was never as desirable as conducting business-to-business transactions. The company always knew that the server and professional markets would pay far more for its hardware than gamers would. It simply needed a way to fully exploit that market, and AI was their way in.
How all of this will play out remains to be seen. Whatever happens, I believe Nvidia knows that returning its focus to PC gaming isn't a viable long-term strategy for the business, given the technological challenges involved. The company simply no longer views the market as profitable enough.
Of course, there will be future GeForce GPUs, and I'm sure many of them will be very impressive. Hell, the GeForce 60 series could be amazing. But the point, or rather the problem for Nvidia, is that it can no longer release new generations regularly, and offering gains substantial enough to convince gamers to upgrade every generation has become far less common.
PC gaming remains a multibillion-dollar industry. So if not Nvidia, someone else will eventually fill the void, given enough time. Realistically, Nvidia will continue serving the gaming market. But when opportunities such as cryptocurrency mining and AI data centers come knocking, the company will capitalize on them, just as it has before. When those markets inevitably crash, Nvidia will once again become a gaming company.
So, see you again soon, Nvidia. Hopefully, you can find legitimate reasons for gamers to upgrade more than once a decade. Fingers crossed.