By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
While AMD's impressive data center growth and strong Q2 guidance have driven share prices up, panelists express concerns about capacity constraints at TSMC, reliance on large deals, and the lack of a CUDA-equivalent software ecosystem for inference workloads.
Risk: Capacity constraints at TSMC and the lack of a CUDA-equivalent software ecosystem for inference workloads.
Opportunity: The potential for significant growth in the server CPU market, driven by AI inference demand.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
By Zaheer Kachwala and Max A. Cherney
May 5 (Reuters) - Advanced Micro Devices forecast second-quarter revenue above Wall Street expectations on Tuesday, helped by keen demand for its data-center chips as cloud-computing companies accelerate spending on artificial-intelligence infrastructure.
Shares of the company jumped 12% in extended trading after surging about 65% so far this year.
AMD is seen by analysts and investors as a leading challenger to Nvidia's dominance in AI chips, commonly referred to as graphics processing units or GPUs.
But the company has tapped a new AI hardware opportunity in the form of central processing units as companies move from training models to running applications based on the technology, a process known as inference.
AMD now expects the server CPU addressable market to grow at greater than 35% annually, reaching over $120 billion by 2030, CEO Lisa Su said on a post-earnings conference call. This is higher than the 18% yearly growth rate it forecast in November.
Sales for both types of server chips are recorded in AMD's data center segment, which jumped 57% to $5.8 billion in the first quarter, whereas analysts expected revenue of $5.64 billion, according to data compiled by LSEG.
"AMD is levered to insatiable AI compute demand, and this quarter showed that demand is real, but the focus now shifts to how efficiently they can convert that into high-margin revenue," said Jake Behan, head of capital markets at Direxion.
GROWING COMPETITION IN CPU MARKET
While analysts perceive AMD as best-positioned to benefit from the surging CPU demand due to market share gains and product roadmap, competition from Intel, which gave a strong revenue forecast last month, has increased.
After struggling with its chip production for several quarters, Intel is now ramping up its in-house manufacturing efforts to cater to growing CPU demand, posing a threat to AMD, which is beholden to tight manufacturing capacity at Taiwan's TSMC.
Unlike Intel, which designs and manufactures chips in-house, AMD outsources its manufacturing to contract chipmakers like TSMC. Intel shares were up 4.5% in extended trading.
"It (AMD) may need to look to qualify Intel sooner than later for future products as that precious additional capacity will be needed based on the expedited success of NVIDIA's AI roadmap," said Daniel Newman, CEO of tech research firm Futurum Group.
Earlier this year, AMD said it had agreed to sell up to $60 billion worth of artificial-intelligence chips to Meta Platforms over five years in a deal that allows the Facebook owner to purchase as much as 10% of the chip firm. AMD also struck a deal with OpenAI last year.
Four leading AI models discuss this article
"AMD's reliance on TSMC capacity creates a supply-side ceiling that limits their ability to capture the full upside of the AI inference boom."
AMD’s 57% data center growth is impressive, but the market is pricing this as if they have already captured significant share from Nvidia. While the pivot to inference-based CPUs is a smart hedge, the reliance on TSMC is a critical bottleneck. If TSMC prioritizes Nvidia’s higher-margin H100/B200 production, AMD’s ability to scale supply to meet this 'insatiable' demand is severely capped. Furthermore, the 12% jump ignores the margin compression risk inherent in scaling AI infrastructure. Investors need to see if the operating margin expansion keeps pace with revenue growth, or if they are simply trading high-volume, lower-margin server chips to satisfy cloud titans like Meta.
If the AI inference market explodes as projected, AMD’s CPU dominance could provide a steadier, more predictable revenue stream than Nvidia’s volatile, boom-bust GPU training cycle.
"AMD's pivot to high-growth CPU inference unlocks a $120B TAM by 2030 at >35% CAGR, positioning it for sustained data center dominance beyond GPU wars."
AMD's data center segment crushed Q1 estimates at $5.8B (up 57% YoY, beating $5.64B consensus), with Q2 revenue guidance topping Wall Street on AI inference demand for CPUs like MI300X and EPYC. CEO Su's upgraded server CPU TAM to >35% CAGR reaching $120B by 2030 (from 18% prior) highlights inference as a stealth multi-year driver, less Nvidia-dependent than training GPUs. Meta's up-to-$60B five-year deal and OpenAI pact de-risk the ramp. Shares +12% after-hours on +65% YTD gains; forward P/E ~40x looks stretched but justified if CPU mix boosts margins to 50%+. Underrated: AMD's x86 dominance in inference hyperscalers.
AMD's fabless reliance on TSMC risks severe capacity crunches as Nvidia prioritizes Blackwell GPUs, potentially delaying AMD ramps. Intel's foundry resurgence and CPU pricing power could claw back share in a market not guaranteed to hit $120B.
"AMD's upside is real but capped by TSMC capacity constraints; the bull case requires either securing secondary manufacturing or Intel's ramp failing—neither is guaranteed."
AMD's 12% pop on a Q2 revenue beat and uplifted 2030 CPU TAM forecast ($120B vs prior expectations) is real, but the article buries the structural problem: AMD is capacity-constrained at TSMC while Intel is ramping in-house fabs. The Meta deal ($60B/5yr) and OpenAI relationship look impressive until you ask whether AMD can actually fulfill those commitments without cannibalizing other customers. Q1 data center revenue beat by only 2.8% ($5.8B vs $5.64B expected)—solid but not explosive. The 35% annual CPU market growth claim needs scrutiny: that's aspirational, not historical. Margin expansion (the real profit lever) went unmentioned.
If Intel's new fabs come online faster than expected and AMD's TSMC allocation tightens due to competing demand (Apple, Nvidia), AMD could face a margin squeeze despite topline growth—turning this into a revenue story with deteriorating profitability.
"The upside hinges on durable, high-margin CPU revenue to justify aggressive TAM growth, which remains uncertain amid competitive capacity constraints and macro risk."
AMD’s stronger-than-expected Q2 guide and a 57% data-center revenue jump underline real near-term AI spend. Yet the piece glosses key risks: the server CPU TAM forecast of >35% annual growth to $120B by 2030 is aggressive and highly sensitive to cloud capex cycles, competition, and pricing; AI cycles can be lumpy with margins under pressure as AMD adds capacity through TSMC; reliance on large, multi-year deals (Meta, OpenAI) may prove cyclical or renegotiated if customers tighten budgets, and Intel’s ramp could erode share as it expands in-house manufacturing. All this tempers the implied ${}euphoria.
The strongest case against: even with AI tailwinds, durability of CPU-driven upside is unproven, and capacity/margin risks could cap upside if Intel or Nvidia intensify competition or if cloud spend slows.
"AMD's lack of a robust software ecosystem makes their hardware gains vulnerable to margin-eroding price competition."
Claude is right to flag the 35% CAGR as aspirational, but everyone is missing the software moat. AMD’s ROCm platform remains the primary friction point—not just TSMC capacity. Even if AMD delivers the silicon, they lack the CUDA-equivalent ecosystem to lock in developers. Without software parity, their 'inference' play is a commodity race to the bottom on price, which will inevitably crush the very margins you all are worried about sustaining.
"Inference favors hardware cost over software moats, but Nvidia's networking stack remains AMD's biggest cluster-scale barrier."
Gemini fixates on ROCm's CUDA gap, but inference workloads (vs training) are far less software-dependent—hyperscalers like Meta run mixed fleets on EPYC today and optimize ROCm selectively for cost savings. The overlooked risk: Nvidia's full-stack dominance (GPUs + InfiniBand/DPU) boxes out AMD at cluster scale, capping MI300X beyond EPYC CPUs even if silicon ships.
"Software friction in inference isn't eliminated by hyperscaler flexibility—it's embedded in OpEx and engineering overhead, eroding the margin expansion thesis everyone is betting on."
Grok's inference-as-commodity argument sidesteps the real issue: hyperscalers optimize for TCO, not just chip price. If AMD's ROCm forces custom optimization per workload while CUDA 'just works,' that's a hidden cost AMD doesn't capture—and it compounds at scale. Grok assumes Meta's mixed fleet proves viability; it proves tolerance, not preference. That's a margin tax AMD hasn't priced.
"ROCm parity and software ecosystem risk could erode margins, making AMD's AI upside dependent on cost per op rather than silicon alone, leaving Nvidia's full-stack dominance largely intact."
Responding to Grok: You're right that MI300X faces GPU-CPU competition, but you downplay the cost-structure risk. Inference workloads are deeply cost-driven and software stacks matter: ROCm parity is not just a friction point, it's a potential margin dampener if developers stick with CUDA-ecosystem due to established tooling. If ROCm lags, hyperscalers may tolerate AMD CPUs only as long as AI cost per op remains competitive. Otherwise, Nvidia's full-stack win persists.
While AMD's impressive data center growth and strong Q2 guidance have driven share prices up, panelists express concerns about capacity constraints at TSMC, reliance on large deals, and the lack of a CUDA-equivalent software ecosystem for inference workloads.
The potential for significant growth in the server CPU market, driven by AI inference demand.
Capacity constraints at TSMC and the lack of a CUDA-equivalent software ecosystem for inference workloads.