Could NVIDIA Get Dethroned? New AI Chips Are Faster and Cheaper Than GPUs

The AI computing market is seeing a shift: the old way of measuring power through “GPU-hours” is gradually giving way to a more practical metric — the cost of generating text. Against this backdrop, alternative solutions are starting to look a lot more attractive than traditional NVIDIA GPUs. According to data shared by a representative of the infrastructure company Nebius, 90–95% of corporate demand today comes not from training models, but from inference — meaning running already-built AI. This has also changed pricing strategies: companies are increasingly calculating costs not based on hardware runtime, but on the price per million tokens processed.

In this system, chips from the startup Groq look significantly more cost-effective — estimates put their usage at roughly 5–10 cents per million tokens, while NVIDIA’s Blackwell-based solutions (B100, B200, B300) run about 25 cents. On top of that, Groq doesn’t just win on price — its chips can deliver up to 800 tokens per second compared to around 450 for NVIDIA GPUs, giving them a roughly 70–80% speed boost in text generation tasks.

That said, it’s still too early to talk about a full-on GPU replacement. NVIDIA maintains a key advantage in model training — that remains the most resource-heavy phase, and specialized solutions like Groq aren’t competing there just yet. Essentially, the market is starting to split into two segments: training stays with GPUs, while inference is gradually moving to more narrowly specialized chips. To put things in perspective: renting NVIDIA GPUs remains expensive even under these new metrics. For example, an H100 on the spot market goes for about $2.95 per hour, an H200 — $3.50, and a Blackwell B200 — up to $6.50. Interestingly, at the end of 2025, NVIDIA and Groq signed a non-exclusive licensing agreement for inference technologies. At the time, that deal became the largest in the startup’s history and indirectly confirmed that Groq’s approach has potential — potential even the market leaders acknowledge.

This situation matters for gamers too — the growing demand for AI compute directly affects graphics card prices and availability. If part of that workload shifts to specialized accelerators like Groq, the pressure on the GPU market could ease, which could help stabilize gaming graphics card prices and improve their availability.

For now, the industry is moving toward a split of roles: general-purpose GPUs remain the backbone, but niche solutions are slowly eating away at some of their tasks. This makes the market more competitive and could impact both the enterprise sector and everyday users.

What do you think — will the rise of alternative AI chips lead to lower graphics card prices, or will GPUs remain the go‑to tool for both AI and gaming? Share your thoughts in the comments.