AMD's latest GPU is delivering Moonshot AI's newest trillion-parameter model at significantly lower inference costs than Nvidia's competing chip, according to new benchmark results that could reshape enterprise AI deployment decisions. The MI355X (AMD's Instinct accelerator) ran the 2.8-trillion-parameter Kimi K3 model at less than half the hourly cloud rental cost of Nvidia's B300, even though Nvidia retained an edge in raw processing speed.

What Makes Kimi K3 So Demanding?

Kimi K3, released by Moonshot AI on July 16, represents a significant leap in model scale and capability. The model uses a mixture-of-experts architecture, a design pattern that activates only the most relevant portions of the network for each task, allowing it to handle massive parameter counts more efficiently. With 2.8 trillion total parameters and a context window of 1 million tokens, the model can process roughly 800,000 words at once, making it suitable for complex document analysis and extended reasoning tasks.

The sheer size of Kimi K3 creates a hardware bottleneck. The model requires 288 gigabytes of high-bandwidth memory per GPU, a specification that limits practical deployment to only the newest accelerators from both AMD and Nvidia. This requirement effectively eliminates older hardware from the equation, creating a direct head-to-head comparison between the companies' latest offerings.

How Do the Costs Actually Compare?

Cloud providers currently offer MI355X instances from about $2.50 per GPU hour, compared with roughly $6 for the B300, according to results from SemiAnalysis's InferenceX platform, an open-source vendor-neutral tool that continuously measures AI inference performance. This price difference translates directly to token generation costs. Although Nvidia's B300 delivered superior throughput in some tests, AMD's lower rental price allowed the MI355X to produce tokens at a lower overall cost per token.

The cost advantage extends beyond Kimi K3. Earlier InferenceX tests showed a similar pattern across other large models. On the smaller Kimi K2.6 model, Nvidia's B300 delivered 29 percent more throughput per GPU, while the MI355X produced tokens at 23 percent lower cost. When running MiniMax M3, AMD's chip was also 24 percent cheaper per token. Signal65 reported in July that the MI355X generated between 1.4 and 2.1 times more tokens per dollar than Nvidia's B200 across several models.

Why This Matters for Enterprise AI Deployment

• Cost Efficiency at Scale: For large data center operators running inference workloads continuously, the difference between $2.50 and $6 per GPU hour compounds rapidly. A single GPU running for a month could cost $1,800 on AMD versus $4,320 on Nvidia, a difference of $2,520 per GPU per month.
• Ecosystem Lock-in Pressure: Nvidia continues to lead in absolute throughput and benefits from widespread adoption of its CUDA software ecosystem, which makes it the default choice for many organizations. AMD's lower inference costs provide a compelling alternative for cost-conscious cloud providers and enterprises.
• Model Size Democratization: As increasingly large open-weight models demand more memory and computing capacity, AMD's pricing advantage could make trillion-parameter models more accessible to organizations with tighter budgets.

The results arrived shortly before AMD's second-quarter earnings report on August 4, with analysts expecting revenue of about $11.2 billion, representing growth of 46 percent from the same period the previous year. AMD's data center division generated $5.8 billion in revenue during the first quarter, making inference cost efficiency a strategically important competitive lever.

The Broader Context: Continual Learning and Model Optimization

While the AMD-Nvidia cost comparison dominates recent headlines, the AI industry is simultaneously grappling with a deeper technical challenge: how to continuously improve models after their initial training phase. Mind Lab, a Chinese AI company founded in October 2025, has been advancing continual learning techniques that could reshape how models like Kimi K3 are optimized over time.

In December 2025, Mind Lab conducted end-to-end reinforcement learning training on Kimi K2, a trillion-parameter mixture-of-experts model, using only 64 Nvidia H800 GPUs and achieving results close to full-parameter training while consuming about 10 percent of the GPU resources required by conventional methods. This represents a significant efficiency breakthrough, though the technique has not yet been widely adopted across the industry.

The company's approach uses LoRA (low-rank adaptation) adapters, specialized modules that allow models to be fine-tuned for specific tasks without retraining the entire network. Mind Lab released the full version of Macaron-V1 on July 21, achieving state-of-the-art results in six of 12 benchmark tests. The flagship version, called Venti, is a 748-billion-parameter model that required only four billion additional parameters in LoRA adapters to exceed the performance of its base model.

These advances in post-training and continual learning are becoming increasingly important as the industry recognizes that raw model size alone no longer determines capability. Turing Award winner Richard Sutton, widely regarded as the father of reinforcement learning, stated at this year's World Artificial Intelligence Conference that the central path for the next generation of AI would be driven by experience, while the static labeled-data paradigm had reached its ceiling.

For enterprises evaluating AI infrastructure investments, the convergence of these trends creates a complex decision landscape. AMD's cost advantage on inference makes it attractive for running large models at scale, while advances in continual learning suggest that the models themselves will become more efficient and adaptable over time. The competition between AMD and Nvidia on cost, combined with ongoing improvements in model optimization techniques, is likely to accelerate the adoption of trillion-parameter models across industries.