The volume of AI bubble skepticism reached a peak in recent weeks. Critics point to the same issue: massive compute investments have been deployed, but is anyone actually using them? Has demand been overextended? AMD's earnings report states a more straightforward fact: customers with identifiable names are buying its AI chips with real money, and purchasing more than they did last quarter.
Data Center: Doubled Year-over-Year, Still Accelerating Quarter-over-Quarter
AMD's revenue for Q3 2026 was $13.377 billion, up 45% year-over-year and 16% quarter-over-quarter, setting a new record for a single quarter. Data center revenue accounted for $8.546 billion, up 97% year-over-year and 27% quarter-over-quarter, increasing its share of total revenue from 47% a year ago to 64%.
A more critical signal comes from the customer list. AMD disclosed that Anthropic, Meta, Microsoft, OpenAI, and Oracle Cloud Infrastructure are running AI models and workloads on the first Helios rack-scale solutions. Helios features Instinct MI450 series GPUs and sixth-generation EPYC processors, directly competing with NVIDIA's Grace Blackwell and Vera Rubin systems. In other words, in the field of AI accelerators dominated by NVIDIA for years, AMD has secured not just letters of intent or testing partnerships, but live workloads.
Lisa Su's wording in the earnings report is noteworthy: she expects server revenue to double year-over-year in 2027, and data center AI business growth to outpace the company's forecast made 90 days ago. When a chipmaker proactively raises its long-term outlook, it typically indicates that forward-looking demand signals from customers are stronger than previously expected.
The Bullseye of Bubble Theory Was Avoided
The core anxiety behind current AI bubble theories can be summarized by two questions: who is paying for this, and do those payers have the capacity to keep paying?
BBC recently focused on new cloud providers represented by CoreWeave, which expand through borrowing against long-term lease contracts and GPU collateral. CoreWeave alone carries debt as high as $35.1 billion. The business models of such companies have not yet undergone full-cycle verification; if financing environments tighten or compute rental prices fall, the fragility of the entire chain will be exposed. Meanwhile, although Alphabet's third-quarter report showed doubled cloud revenue, free cash flow turned negative, and long-term debt rose significantly, triggering market concerns about AI infrastructure draining cash flows.
In this context, AMD's earnings report stands exactly opposite to this anxiety. AMD sells compute hardware; it does not need to prove whether specific AI applications will eventually become profitable, only that someone is still paying for compute. The Q4 revenue guidance of $17 billion (approximately 66% year-over-year and 27% quarter-over-quarter growth at the midpoint) essentially says: the payers are still here, and they will pay even more next quarter.
However, this does not mean bubble theory has been refuted. Among AMD's customers, OpenAI and Anthropic themselves are heavy spenders; to what extent their compute procurement relies on continuous financing remains a question unanswered by the earnings report. Meta and Microsoft's capital expenditure discipline will also face stricter scrutiny in the coming quarters. AMD's earnings prove that people are buying now, not that buyers will always afford to buy.
Agentic AI Is Changing the Fate of CPUs
Focusing solely on GPUs would overlook a subtle narrative pivot in this AMD earnings report.
In the earnings call, Lisa Su mentioned a judgment: CPU adoption is just beginning, and we expect demand to surge sharply as agents are more widely deployed and handle more complex tasks. This logic deserves elaboration. Agentic AI applications, such as systems like Meta Muse that can run independently for hours, drive virtual machines, and execute multi-step tasks, rely far more heavily on server CPUs than traditional conversational AI. Every distribution, scheduling, and isolation of agent tasks requires support at the CPU level.
Demand for EPYC server processors remained strong in the third quarter, cited by AMD as the primary driver of its data center growth, ranking ahead of Instinct GPUs. This means AMD's AI narrative is not simply betting on GPU replacement of NVIDIA, but also securing a position on the line where CPUs regain importance. For observers concerned about the homogenization of AI compute demand, this represents a structural change worth noting.
Summary
Can the AI narrative sustain itself? AMD's earnings suggest yes, but the manner of sustainability is changing. The core of the AI narrative from 2024 to 2025 was "whoever has GPUs wins," with NVIDIA as the sole protagonist. By late 2026, the narrative began to diverge: agentic AI restored relevance to CPUs, and rack-scale systems shifted competition from single chips to holistic solutions. AMD's earnings temporarily quieted the voices of AI bubble skeptics. It proved that demand persists, customers are real, and profits are being generated. However, it did not—and cannot—answer the ultimate question: when compute supply climbs at such a steep slope, can the application ecosystem digesting this compute keep up? What Lisa Su can do is continue to deliver numbers worthy of the phrase "demand is accelerating" in every upcoming quarter.