Advanced Micro Devices Inc "Raising Estimates, PT to $730 Following AI Day"

Research evidence excerpt

Advanced Micro Devices Inc "Raising Estimates, PT to $730 Following AI Day"

0B From To % ch Cons.

C2025 baseline) - a meaningful step up vs. the co's forecast from last year of > Q1 1.37 1.37 0 1.37

$500B by C2028. Importantly, AMD continues to expect GPUs to comprise the Q2E 1.65 1.65 0 1.61

vast majority of the TAM, reflecting the value of GPU programmability/flexibility as Q3E 1.86 1.91 3 1.86

models and workloads continue to evolve rapidly across both AI training and Q4E 2.80 2.89 3 2.64

inference. Related to the transition from training-driven compute demand to 12/26E 7.68 7.82 2 7.45

inference-driven demand, the company pointed to C2026 as the first year in which 12/27E 14.63 18.65 27 13.46

~60% of existing AI compute capacity is to be utilized for inference workloads 12/28E 21.11 27.57 31 18.53

rather than training (vs 50/50 in 2025) - which we find particularly notable

Timothy Arcuri

considering that AMD still forecasts training compute/FLOPS to increase by ~5x/ Analyst

year through C2030. timothy.arcuri@ubs.com

+1-415-352 5676

Helios officially launched and is ramping into 4Q. The company officially

unveiled its Helios rack-scale platform, with specs and performance largely in line Natalia Winkler, CFA

with expectations (e.g., 432GB HBM4 per GPU, 23.3TB/s of memory bandwidth, Analyst

and peak compute performance of up to ~34x vs. MI355X at mid-to-high natalia.winkler@ubs.com

+1-415-352 4626

concurrency levels). Customer shipments are expected to begin in September

(C3Q26-end), with management emphasizing a strong C4Q26 ramp (and for this Gianmarco Vella

reason, we forecast AMD DC GPU revenue +37% in 2H vs 1H). AMD also Associate Analyst

suggested Helios delivers up to 30% more tokens-per-dollar for inference gianmarco.vella@ubs.com

+1-415-352 4555