bottleneckScore 75/100Research
AMD claims Helios delivers 10-15% more performance and 30% lower tokens-per-dollar vs. NVIDIA; pricing pressure from TSMC supply constraints
“How much of that is because TSMC is charging more to just deliver chips? ... we actually win on both sides. Um our Helios systems have more performance um than our competition. And so, you know, let's call it anywhere from, you know, 10 to 15% plus more performance. And we also have lower uh tokens uh per dollar. So, you know, up to a 30% benefit for our customers in terms of tokens per dollar.”
Why it matters
Lisa Su explicitly attributes pricing pressure to TSMC's constrained capacity and higher costs, while claiming significant efficiency gains; this suggests TSMC's advanced packaging/CoWoS bottleneck is translating into both cost pressure and a potential performance differentiation opportunity for AMD.
Investment implication
If AMD's token-per-dollar claims hold in customer deployments, it could drive share gains vs. NVIDIA among hyperscalers cost-sensitive on inference workloads; TSMC's capacity constraints are likely to remain a pricing lever through 2027, benefiting AMD's competitive positioning on cost-per-token.