According to an October 3 statement by Alexey Navolokin, AMD’s Asia-Pacific Regional Director, AI’s shift from single prompts to continuous operational workflows is causing token consumption to surge as systems must constantly reason, iterate, and call tools. Every operation on cloud infrastructure incurs usage-based charges, accumulating into massive recurring expenses.

Rather than introducing a specific product to solve this pain point, AMD presented a strategic direction: treating agentic AI as a distributed infrastructure challenge. Therefore, instead of continually expanding cloud budgets, the fundamental solution is to allocate and share workloads across multiple environments.

AMD’s Numbers and Accompanying Conditions

The average workload baseline established by AMD is the most notable highlight: 5.7 million input tokens and 574,000 output tokens per user daily. When applying this benchmark to a fleet of 500 AMD AI PCs deployed in a hybrid model (50% on-premises and 50% cloud), enterprises could reduce total costs over three years by roughly 40% to 60% compared to running entirely on the cloud, depending on the cloud model chosen.

For desktop deployments, AMD presented specific unit economics for the AMD AI PRO R9700 lineup. This machine is expected to handle around 18 million tokens per day with an estimated electricity cost of $64.80 per month. Under the company’s calculation model, the three-year cost for this setup caps at $6,533, a stark contrast to the $81,108 required when relying entirely on cloud services.

However, all the figures above are based on AMD’s own analysis and assumptions; actual costs will vary depending on workloads, cloud models, hardware utilization rates, electricity prices, and deployment models. It should be emphasized that the 40%–60% savings figure is currently only a projection provided by the manufacturer, without independent third-party verification.

Regarding the market context, AMD cited a survey from Anthropic’s State of AI Agents 2026 report, indicating that 57% of surveyed enterprises have integrated AI agents into multi-step workflows. This is also partner-cited data rather than findings from AMD’s own research.

The “Workload Right-Sizing” Strategy

Navolokin proposed the “workload right-sizing” approach to optimize infrastructure for specific types of tasks. Specifically, tasks that run continuously or require low latency should be deployed locally or at the edge. Conversely, massive workloads requiring flexible scalability are better suited for cloud computing.

On-premises processing offers clear cost benefits: upfront hardware investment eliminates ongoing token query fees incurred in the cloud. Furthermore, equipping local machines with sufficient memory and compute capacity enables smooth execution of models supporting tool calling and multi-context workflows, simultaneously safeguarding internal data and minimizing latency.

Notably, Navolokin emphasized that an employee running multiple AI agents simultaneously creates substantial concurrency demands across compute, memory, orchestration, and data retrieval—far exceeding the consumption of single-agent AI assistants. Consequently, infrastructure strategy requires a holistic perspective, harmoniously balancing CPUs, GPUs, memory, network infrastructure, and software rather than relying solely on GPU performance.

The Equation for Self-Hosters

While this argument is not new, the figures AMD presented bring greater clarity to the economic equation. Individuals or organizations paying recurring API fees for repetitive tasks can directly compare this data against hardware investment costs. For those already running local models, this perspective reinforces the rationale behind their choice. The only difference is that AMD focuses on enterprise-scale scenarios rather than personal use.

However, the figure of 18 million tokens per day warrants caution. This is merely a manufacturer assumption tailored to a specific hardware configuration and does not reflect a universal limit for other hardware systems. The core merit of the solution lies in the principle of allocation: shifting high-frequency, low-latency tasks closer to the user while keeping massive workloads in the cloud. The key remaining question is the realism of AMD’s assumed 50/50 ratio, as real-world enterprise workload patterns will vary considerably.