AMD’s stock has surged ahead of earnings, but this analyst sees a problem ahead
AMD’s shares have sprinted into earnings on a wave of AI optimism, powered by enthusiasm for its MI300 accelerator family and ongoing share gains in data center CPUs. The setup looks strong: hyperscalers are spending, AI budgets remain elevated, and investors want a credible No. 2 to Nvidia. Yet one veteran semiconductor analyst sees a less comfortable path from here. The concern isn’t about whether AMD can sell accelerators—it’s about whether the next leg of growth can meet the expectations now embedded in the stock.
The core of the bear case is simple: when a narrative gets this good, execution risk rises and the market’s tolerance for any wobble falls. In AI hardware, those wobbles often come from bottlenecks that sit outside the chip itself—packaging, memory, software, and customer behavior.
The bottleneck has moved upstream
Even if AMD continues to win accelerator slots, the near-term ceiling is defined by supply of advanced packaging and high-bandwidth memory (HBM) rather than by demand. TSMC’s CoWoS capacity has been expanding, but allocation across the industry is tight and typically favors entrenched incumbents with firm long-term commitments. HBM—especially the latest HBM3/3E grades—remains a gating factor as well. If AMD can’t secure enough packaged capacity and memory at the right volumes and timelines, shipments can lag orders, elongating revenue ramps.
That creates two risks for a stock running into numbers:
– The expectations trap: With the multiple elevated, investors may be pricing in a smooth, steep accelerator ramp. Even modest supply slippage or staged customer qualifications can turn a “beat and raise” quarter into a “good, but not good enough” print.
– Margin math: Accelerators are margin-sensitive to packaging and HBM costs. Early in a product cycle, yields and learning curves can weigh on gross margin. If supply is constrained, AMD has less ability to optimize mix and absorb costs, making corporate gross margin progression choppier than bulls expect.
Software is better—but the moat hasn’t moved
Performance per watt matters, but so does stack maturity. AMD’s ROCm environment has improved meaningfully, and the company has made strides in enabling mainstream frameworks and key model libraries. Nonetheless, CUDA’s decade-plus head start still exerts gravity. Porting and performance-tuning at scale takes time and talent, and hyperscalers often standardize on toolchains that minimize friction.
The analyst’s point isn’t that AMD can’t close the gap—it’s that consensus sometimes assumes software hurdles shrink faster than procurement and deployment realities allow. In practice, migrations roll out in waves, often starting with specific workloads, internal teams, or regions. That cadence can flatten the near-term revenue curve even when the long-term trajectory is intact.
A crowded second lane
The “second source” seat next to Nvidia is more contested than headlines imply:
– Custom silicon: AWS (Trainium/Inferentia), Google (TPU), Microsoft (Maia) and Meta (MTIA) are carving out internal lanes. Every watt that moves in-house reduces the incremental TAM for merchant GPUs.
– Other merchant players: Intel’s Gaudi line, specialized inference silicon, and emerging entrants in networking and AI appliances are intensifying price and design-win competition, particularly in inference and at the edge.
– Arm in servers: In CPUs, AMD’s share gains could meet a stiffer headwind as Arm-based instances proliferate at cloud scale, while Intel’s next-gen Xeon lineups target power/performance per core—pressuring AMD’s price umbrella.
The net: AMD can win substantial deals and still face a more elastic pricing environment and staggered adoption patterns, complicating the “straight up and to the right” revenue narrative.
Mix and cyclicality still matter
Even in an AI upcycle, legacy segments tug on model outcomes:
– Semi-custom/console is late-cycle and volatile.
– Client PC demand has stabilized but remains sensitive to macro and replacement timing.
– Data center CPUs are a bright spot, but hyperscaler digestion can appear with little warning after big capex bursts.
Those moving parts affect not just top-line steadiness but also the blended margin profile. If accelerators grow from a lower starting margin and semi-custom normalizes down, consolidated gross margin can drift rather than march, which is dangerous when investors are primed for linear improvement.
What this analyst will watch on earnings day
– AI accelerator trajectory: Absolute revenue, backlog color, and—crucially—visibility into 2H supply of CoWoS and HBM. Watch mentions of vendor diversification and long-term capacity agreements.
– Software and ecosystem: Concrete updates on ROCm adoption, turnkey solutions, and performance parity on popular LLMs and inference benchmarks. Named enterprise deployments carry more weight than lab demos.
– Pricing and margins: Any commentary on accelerator ASPs, introductory discounts, or services/support bundling. Clues on where gross margin trendlines sit as yields improve.
– Data center CPU momentum: EPYC share gains versus both Intel’s new platforms and Arm alternatives. Look for signals on attach rates with AI workloads and networking/storage partners.
– Non-AI segments: The slope of semi-custom declines and the cadence of PC CPU refreshes—important for cash flow consistency.
– 12–18 month guideposts: Rather than a single-year target, investors need a path through next year’s product transitions, packaging capacity adds, and HBM node shifts.
The upside rebuttal
To be fair, the bull case has force. If AMD secures more advanced packaging slots and HBM, continues ROCm maturation, and converts multiple hyperscaler anchors into scaled deployments, the accelerator business can compound faster than skeptics expect. Each successful at-scale deployment reduces perceived switching costs for the next one. In CPUs, a sustained power/performance lead could preserve a pricing umbrella even as competitors sharpen their offers. And if AI capex remains robust into next year, the tide can lift more boats than models currently contemplate.
Bottom line
The stock’s pre-earnings surge says investors want AMD to be the definitive alternative in AI silicon. The analyst’s caution is that the next phase hinges less on raw chip performance and more on the plumbing that turns silicon into shipped, supported systems at scale—advanced packaging, HBM, and a software stack customers trust for mission-critical workloads. Those are solvable problems, but they operate on timelines and dependencies the market often underestimates. With the bar now high, even good execution may not clear it cleanly every quarter.
For investors, the tell will be management’s specificity on supply, software, and sustainable margins. If AMD can de-risk those three, the multiple can hold—or even expand. If not, the story likely remains intact, but the path may be bumpier than the pre-earnings rally implies.
This article is for information only and is not investment advice.
