AMD chief executive Lisa Su is facing fresh doubts as delayed AI chips and Nvidia’s CUDA moat squeeze the company.
Su has been one of the chip industry’s most respected leaders since taking charge of AMD in 2014. She dragged the company from near irrelevance into a serious CPU rival to Intel, helped by the Zen architecture and sharp execution.
That turnaround made AMD a credible force in consumer PCs, servers and data centre chips, while investors enjoyed a long share price climb.
Recent quarters look less shiny. AMD’s data centre revenue is still growing, but it has not matched the AI hopes baked into the stock. The MI300 accelerator series arrived later than expected, giving Nvidia more time to tighten its grip on AI training and inference. As a result hyperscalers have already sunk money, code and staff time into Nvidia’s CUDA software ecosystem.
AMD’s rival ROCm platform has improved, but it still trails CUDA in maturity, support and developer adoption. The result is a switching problem. Customers may like an alternative, but moving big AI workloads away from CUDA remains a faff.
Production issues have made matters worse. Yield problems and supply limits around the MI300X have restricted AMD’s ability to chase booming AI hardware demand.
AMD keeps talking up its roadmap, but constrained shipments have left customers and investors grumbling. Nvidia is estimated to control about 80 to 90 per cent of the high-performance AI chip market.
That lead comes from its hardware, software stack and years of developer optimisation, rather than one clever GPU launch.
Su’s strategy has been to pitch AMD as a more open and cheaper alternative to Nvidia. That argument lands with some researchers and buyers wary of lock-in, but enterprise customers usually want proven tooling and support.
Outside AI, AMD faces pressure in older hunting grounds. Intel has shown signs of recovery with newer processors, leaving Chipzilla better placed to claw back some ground in CPUs.
The PC market remains weak as corporate refresh cycles slow and consumer demand stays patchy. AMD still has strong console chip deals with Sony and Microsoft, but those contracts are lower-margin than the data centre prizes investors fancy.
The CPU business has hit bumps as well. Zen 4 delivered decent gains, while Zen 5 reviews were more mixed in some workloads. That has given Chipzilla room to fight harder in servers, where AMD had previously made painful inroads.
Financially, the strain is visible. AMD’s gross margins remain healthier than in its darker days, but spending on AI research and development is rising fast. Operating expenses have climbed as Su pushes faster product cycles and broader platform work.
The cocaine nose jobs of Wall Street have grown more cautious, with some analysts trimming forecasts because of the slower AI ramp. Many still like AMD’s long-term potential, but the old assumption that Su would always beat expectations has taken a knock.
TSMC remains both a strength and a vulnerability for AMD. Access to leading-edge manufacturing has helped AMD compete, but it must fight Nvidia, Apple and Broadcom for limited capacity.
Advanced CoWoS packaging, which matters for AI accelerators, remains a bottleneck and hurts smaller players more than giants with larger volume commitments.
AMD has tried to ease that pressure by widening its manufacturing options, including work with Samsung’s foundry. That helps diversify supply, but qualifying chips across foundries is complicated and any slip can mean delays or defects.
The AI boom still gives AMD a real opening. If demand keeps exploding, even second place in accelerators could be highly profitable. The MI300 line has a pitch around performance per watt, which could matter as data centres hit power and cooling limits.
Some cloud providers have begun deploying MI300 systems and have praised the raw compute. Others remain wary of software optimisation problems and long-term support, which is where Nvidia’s head start still bites.
