Turing, an AI infrastructure and cloud training company, has publicly confirmed it is routing approximately 10% of its AI training compute through AMD hardware, explicitly reducing its reliance on Nvidia. The disclosure is notable because training workloads — not just inference — have historically been the most CUDA-locked segment of the AI stack, where switching costs are highest and AMD's software ecosystem (ROCm) has lagged most visibly.

For AMD, the signal matters because it suggests real enterprise dollars are moving onto MI300-series hardware beyond the headline wins with a handful of hyperscalers. AMD grew revenue 34.3% YoY to $34.6B, but net margins remain thin at 12.5% versus Nvidia's 55.6%, and the market is essentially pricing AMD as a 'if ROCm matures' option. A sustained customer diversification narrative could accelerate multiple expansion.

For Nvidia, 10% of one operator's training budget is noise at the revenue level — the company generated $215.9B in revenue (+65.5% YoY) with 71.1% gross margins, reflecting near-monopoly pricing power. But the concern is directional: if cost-sensitive operators begin treating AMD as a credible 10–20% blending option, the pricing ceiling on Blackwell and future architectures faces secular pressure over a 12–24 month horizon.

The bull/bear tension here is asymmetric by ticker. AMD's upside is a re-rating story if adoption broadens; the risk is that ROCm friction keeps this a one-off headline. Nvidia's downside is not revenue today but multiple compression if the narrative of an unassailable moat softens — worth watching into the next round of hyperscaler capex commentary and AMD's next earnings for MI-series attach rate disclosures.