The AI infrastructure build-out is no longer a future story. It is happening right now, at a scale that is reshaping the entire semiconductor industry. In February 2026, NVIDIA reported quarterly revenue of $68.1 billion — a record — with its Data Center segment alone generating $62.3 billion in a single quarter, up 75% from the same period a year earlier. For fiscal year 2026, the company posted $215.9 billion in total revenue, up 65% year-over-year. These are not projections. They are filed results, sourced directly from NVIDIA’s 8-K filed with the Securities and Exchange Commission on February 25, 2026.
At the same time, Advanced Micro Devices reported full-year 2025 revenue of $34.6 billion, up 34%, with its Data Center segment — encompassing EPYC server CPUs and Instinct AI GPUs — reaching a record $16.6 billion for the year. The message from both companies is the same: enterprise and hyperscale customers are spending aggressively on AI compute, and the demand shows no signs of slowing.
What Is Driving the AI Infrastructure Boom
Three interlocking forces are sustaining the current capex supercycle in AI infrastructure.
Agentic AI and inference compute. The industry has shifted from training-dominated workloads toward inference — the continuous, real-time serving of AI models at scale. NVIDIA CEO Jensen Huang described it explicitly in the company’s Q4 FY2026 press release: “The agentic AI inflection point has arrived.” Inference workloads are always-on, require low latency, and scale with the number of users and queries — making them a structurally durable driver of GPU demand in ways that one-time model training runs are not.
Hyperscaler capex commitments. The four largest cloud providers — Amazon Web Services, Microsoft Azure, Google Cloud, and Meta — have each signaled record capital expenditure plans for 2026. NVIDIA’s own press release noted that AWS, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure will be among the first to deploy instances based on the company’s next-generation Vera Rubin platform. Meta has also committed to a multiyear, multigenerational strategic partnership with NVIDIA, including large-scale deployments of Blackwell and Rubin GPUs. CoreWeave, a specialist AI cloud provider, is partnering with NVIDIA to build more than 5 gigawatts of AI factories by 2030.
Platform transitions. The industry is in the middle of a major architectural refresh cycle — from Hopper to Blackwell, and from Blackwell to Rubin. Each generation brings meaningful improvements in inference economics. NVIDIA’s internal benchmark data cited in the Q4 FY2026 filing shows Blackwell Ultra delivers up to 50x better performance and 35x lower cost for agentic AI workloads compared to Hopper. These step-change improvements accelerate replacement cycles and sustain upgrade demand.
The Stocks Investors Are Watching
NVIDIA (NVDA) remains the dominant position in AI infrastructure. Its Q4 FY2026 results — $68.1 billion quarterly revenue, $62.3 billion from Data Center, $1.62 non-GAAP EPS — exceeded expectations, and Q1 FY2027 guidance of approximately $78.0 billion signals continued acceleration. The company has also returned $41.1 billion to shareholders during fiscal 2026 through buybacks and dividends, with $58.5 billion remaining under its repurchase authorization. The primary risk is geopolitical: NVIDIA explicitly noted it is not assuming any Data Center compute revenue from China in its Q1 FY2027 guidance, reflecting ongoing export control restrictions.
Advanced Micro Devices (AMD) is executing a credible second position in AI accelerators. AMD’s Data Center segment revenue grew 39% year-over-year in Q4 2025 to a record $5.4 billion, driven by EPYC server processor share gains and the ramp of Instinct MI300X GPU shipments. For Q1 2026, AMD guided revenue of approximately $9.8 billion, representing 32% year-over-year growth. AMD also faces China export headwinds — its MI308 shipments to China are restricted, and the company absorbed approximately $440 million in inventory charges in 2025 as a result. The Helios rack-scale platform, previewed at CES 2026, and partnerships with HPE, Cisco, and TCS represent meaningful long-term catalysts.
Broadcom (AVGO) is increasingly central to the custom silicon story. Hyperscalers building proprietary AI accelerators — Google’s TPUs, Meta’s MTIA chips — rely on Broadcom for custom ASIC design and high-bandwidth networking. Broadcom’s AI revenue has grown substantially as hyperscale customers look to optimize cost-per-token by running inference on purpose-built silicon rather than off-the-shelf GPUs. This is a long-cycle business with deep customer lock-in.
Marvell Technology (MRVL) occupies a similar custom silicon niche, supplying optical interconnects and custom compute for AI clusters. Marvell has guided for significant AI-driven revenue acceleration in its fiscal 2026.
Taiwan Semiconductor Manufacturing (TSM) is the foundational layer beneath all of the above. TSMC manufactures chips for NVIDIA, AMD, Broadcom, Apple, and virtually every major fabless semiconductor company. Its 3nm and 2nm advanced nodes are the manufacturing platform enabling next-generation AI accelerators. TSMC’s annual report filed with the SEC (20-F, filed April 2025, accession number 0001193125-25-083423) details the company’s capacity expansion plans in Taiwan, Japan, and the United States. TSMC’s Arizona fabs are receiving multi-billion-dollar subsidies under the CHIPS and Science Act, with the goal of bringing leading-edge U.S. production online by the late 2020s.
Macro Context: Capex Cycle and Rate Sensitivity
The current AI infrastructure investment cycle is unusual in that it has proven relatively insensitive to interest rate fluctuations. Unlike traditional corporate capex, AI infrastructure spending is being treated by hyperscalers as a strategic necessity rather than a discretionary investment. This does not mean rate sensitivity is zero — highly leveraged AI cloud companies and smaller data center REITs are more exposed — but the core semiconductor demand from large-cap hyperscalers has remained robust even as borrowing costs have stayed elevated relative to 2020–2021 levels.
The larger macro risk to watch is trade policy. U.S. export controls on advanced semiconductors to China have already materially affected NVIDIA’s revenue mix and AMD’s Instinct GPU shipments. Any escalation — additional countries added to restricted lists, tighter definitions of controlled technology — could reduce addressable market size meaningfully. Conversely, any relaxation of controls would represent a positive catalyst for both companies.
What to Watch Going Forward
- NVIDIA Q1 FY2027 earnings (expected May 2026): Does the $78 billion revenue guidance hold? Watch Data Center margin trajectory as Rubin transitions into production.
- AMD Q1 2026 results (expected late April/early May 2026): Can Data Center revenue sustain the growth rate with China headwinds? Monitor Instinct GPU shipment volumes and EPYC market share data.
- Hyperscaler capex disclosures: Each earnings call from AWS, Azure, Google Cloud, and Meta will include updated capital expenditure guidance. These numbers are the leading indicator for AI chip demand 12–18 months forward.
- Export control policy: Any updates from the Bureau of Industry and Security (BIS) on AI chip export rules will move semiconductor stocks immediately.
- TSMC capacity reports: CoWoS advanced packaging availability remains a potential bottleneck for NVIDIA’s highest-end GPU production. Watch for any supply constraint signals.
Disclosure: This article is for informational purposes only and does not constitute investment advice. The author does not hold positions in any of the securities mentioned. Past performance of stocks discussed is not indicative of future results. Investors should conduct their own due diligence and consult a qualified financial advisor before making investment decisions. All financial figures cited are sourced from public regulatory filings as noted.