Nvidia's Grip on AI and Gaming GPUs Is Loosening Becuase of Its Own Custom Silicon Chips

Last quarter, NVIDIA had a record revenue. Data centers continue to sell more and more. The short-term outlook on the company's situation is not shaky. The question customers are asking is changing, and it's the question. Historically, it was easy: Which GPU is the fastest?

Now it is about tokens per dollar, tokens per watt, and tokens per rack. That's a change in AI hardware pricing, GPU availability, VRAM considerations, and whether you need to buy a new graphics card. The hottest new competition won't be from AMD or Intel. It is being made by Nvidia's largest customers, from custom silicon.

Distinguish Between Training and Inference

For years, Nvidia's competitive edge was built on its full stack: the raw performance of its GPUs, its CUDA and networking capabilities, its AI software, its developer tools, and its supply chain muscle, keeping the rest of the world on its toes. All that remains is the same. However, AI workloads are moving from training to inference, and that's where it matters.

Training workloads are constantly evolving, and general-purpose GPUs excel at them. Inference is different. It's when a model is used for tasks after training, such as answering questions, writing code, creating images, or running an AI agent. It is repetitive; it operates at a very large scale, and at this scale, small savings in power consumption, cooling, memory bandwidth, and latency become significant in monetary terms.

The Custom Silicon From Hyperscalers

That's precisely why hyperscalers have been trying to create their own chips. Google has Ironwood. AWS has Trainium. Meta is promoting MTIA. Microsoft has Maia. Broadcom, on the other hand, is already seeing significant AI business from other companies' custom chips and networking gear. This is not the case to this degree as it was two years ago, and it has significantly altered the competitive landscape.

According to WCCFTech, a recent channel check indicated that Nvidia's market share in the inference space may drop to approximately 50 percent by 2028 as other companies, such as AMD, TPUs, Trainium, Maia, and more, continue to improve. It is worth noting that this is analyst research and not Nvidia's own guidance, so we should regard it as a projection rather than a confirmed fact.

But it's not just that there's a lot of disagreement about the near-term trend; some recent reporting suggests that Nvidia's share of inference workloads is actually holding steady or even rising, rather than falling. Nevertheless, the long-term forecasts represent a genuine shift in the approach to decision-making. It's no longer a question of whether this can outperform NVIDIA on all benchmarks.

It's "is it good enough, and will it be cheaper to operate. While NVIDIA still has the healthiest ecosystem in the industry, inference is where everyone else is truly getting a chance to take a piece of the pie – and that hasn't shown up in the numbers quite yet.

AMD's Instinct MI350 and Memory Argument

AMD is not to be overlooked here. Instinct MI350 series is designed for AI training and inference and will ship with 288GB of HBM3E memory and 8TB/s of bandwidth. This is important because memory capacity and speed can also be limiting factors for large AI models, as can computing. The one area where AMD is still weak is software.

While the journey of ROCm over the last couple of years has been impressive, it doesn't come close to Nvidia's developer ecosystem in terms of maturity and depth. The real trouble occurs when buyers are primarily concerned with price, memory size, and the availability of an alternative offering from one vendor.

If you take it down to gaming gear, it's a different story. For desktop graphics cards, Nvidia is poised to be virtually unbeatable. According to KitGuru, Nvidia's share of the add-in board market dropped to about 94% in Q42025, while AMD accounted for 5% and Intel for approximately 1%. The term competition is not in the context of a competition.

AMD and Intel Can Compete in VRAM Use

That's where things begin to fall apart. NVIDIA's RTX 5060 offers Blackwell architecture, DLSS 4, and 5th-generation tensor cores, but the standard version will also come with 8GB of VRAM. To some 1080p gamers, that's just right. If you are playing any of the newer titles, 8GB is starting to feel a little short.

If you are playing any of the newer titles, 8GB is now starting to feel a little too short. It's a point where AMD and Intel can attack. AMD's RX 7600 XT has a 16GB variant with up to 320 GB/s of memory bandwidth. With 12GB of VRAM, Intel's Arc B580 entered the Budget Battle, and reviewers have responded positively, recognizing it as a solid value pick despite ongoing questions about driver reliability.

Google's work attempts to capture some of the inference workloads.

While Nvidia still reigns supreme in ray tracing, creator applications, AI software, and upscaling quality, this is the ultimate selling point that should prompt consumers to pause before automatically choosing NVIDIA. All this is no indication of any dramatic collapse. It seems like erosion has occurred in small sections.

Meta has its own dedicated in-house AI chips. This is where AMD is finding its advantage in memory-heavy deployments, where its hardware actually is competing with its rivals. Intel targets budget gamers with VRAM being a more critical factor than brand loyalty. So, is Nvidia's supremacy coming to an end? Not yet, in the case of AI training.

Not yet in gaming GPUs. However, in AI inference, cost per token calculations, custom silicon, and budget graphics cards in particular, the cracks are quite visible now. In buyers' eyes, it's a positive sign. The more competitive, the better the price, VRAM sizes, performance, and the more sensible the upgrading. NVIDIA is still the clear leader. However, for the first time in years, the throne is less secure than it was.