This is the $Advanced Micro Devices(AMD)$ investment thesis.
The bulk of the costs to run AI models comes from inference and training.
$NVIDIA(NVDA)$ dominates training due to its first-mover advantage, NVLink, and CUDA.
Currently, most models run inference tasks on the same chips used for training, which is why the costs are so similar.
But this is about to change.
As the gains from pre-training begin to flatline, training workloads will decline while inference demand will skyrocket.
Inference requires as much memory as possible within a single server.
$Advanced Micro Devices(AMD)$ excels here.
Its expertise in chiplet design allows it to connect multiple memory dies to the chip core through its proprietary interconnect, Infinity Fabric.
This modular design enables AMD to use cheaper manufacturing nodes for simple memory dies and more advanced nodes for chip cores, reducing overall costs.
Thus, as the bulk of AI workloads shifts from training to inference, chip demand will likely shift from $NVIDIA(NVDA)$ to $Advanced Micro Devices(AMD)$ , maximizing performance while lowering costs for the AI labs.
Long $Advanced Micro Devices(AMD)$ .
Bar chart titled Exhibit 5 AI economics per million tokens estimated for 2026 tokens. Vertical axis labeled Per Million Tokens in dollars from 0.00 to 5.00 in increments of 1.00. Horizontal axis categories Revenue Inference Cost Training Cost Other Opex Operating Profit. Blue bar for Revenue at 4.29. Blue bar for Inference Cost at 3.36. Red bar for Training Cost at 3.58. Blue bar for Other Opex at 1.12. Blue bar for Operating Profit at 4.18. Source text Goldman Sachs Global Investment Research at bottom.
For whom haven't open CBA can know more from below:
🏦 Open a CBA today and enjoy privileges of up to SGD 20,000 in trading limit with 0 commission. Trade SG, HK, US stocks as well as ETFs unlimitedly!
Find out more here：
Comments