Artificial Intelligence Infrastructure Investment Panorama: GPU, Memory, Network - Which of the Three Tracks Will Dominate?

In 2026, the construction of global artificial intelligence infrastructure is advancing at an unprecedented pace. Morgan Stanley predicts that by 2028, nearly $3 trillion in AI-related infrastructure investment will flow through the global economy, and that over 80% of the spending will still be ahead. In 2026 alone, the combined capital expenditures of leading global technology companies on AI infrastructure have already exceeded $600 billion. Omdia further forecasts that by 2030, global data center cumulative investment will approach $1.6 trillion.

The scale of this round of capital expenditure is rare in the history of technology. Hyperscale tech companies are expected to spend a total of $660 billion to nearly $700 billion in 2026. AI infrastructure has evolved from “nice-to-have” technical investment into strategic spending that determines the competitive landscape for enterprises. The AI Factory market has crossed an irreversible tipping point and is evolving into an entirely new form of industrial organization—characterized by ultra-high capital intensity, clear geopolitical attributes, and complex engineering-and-technology barriers.

For investors, understanding the industrial-chain structure of artificial intelligence (AI) infrastructure and the flow of capital is the prerequisite for grasping this technology investment cycle. Starting from the three core hardware tracks—GPU, Memory, and Networking—combined with the latest market data and industry logic, this article breaks down the investment value and key targets in each track.

GPU: The “Engine” of Compute Infrastructure

GPU is the most core computing unit in AI infrastructure and, at present, the largest share of capital expenditures. Research and Markets data shows that the global AI infrastructure market size will grow from $71.88 billion in 2025 to $90.91 billion in 2026, with a compound annual growth rate of 26.5%. This figure is expected to further increase to $226.95 billion by 2030. GPUs and accelerator systems dominate this growth.

In terms of market performance, the stock price trends of leaders in the GPU track confirm capital’s pursuit of compute infrastructure. In the early hours of June 30 Beijing time, the three major U.S. stock indexes all closed higher: the Nasdaq Composite Index rose 2.07% to 25,820.14. NVIDIA (NVDA) closed at $194.97, up 1.27%, with a total market cap of approximately $4.72 trillion. AMD (Advanced Micro Devices) closed at $539.49, up 3.43%, with a total market cap of approximately $879.7 billion. The Philadelphia Semiconductor Index rose about 3.83% that day, and the year-to-date cumulative gain had already reached 93.55%.

The investment logic in the GPU track rests on two structural factors. First, the compute demand for large-model training and inference continues to rise—ranging from the expansion of model parameters to the scaling of inference deployments, the compute consumption curve has not yet shown an inflection point. Second, the entry barriers on the supply side are extremely high, including multiple moats in architecture design, process technology, and software ecosystems (such as CUDA), which enables leading players to maintain strong pricing power in the foreseeable future.

Among the notable targets, NVIDIA remains the absolute leader in AI compute, and its product roadmap and breadth of customer coverage still serve as an industry benchmark. AMD continues to catch up in both data-center CPU and GPU directions, with a gain of 141.3% year-to-date. Cantor Fitzgerald recently raised its target price for AMD to $700. In addition, as a core supplier of semiconductor manufacturing equipment, Applied Materials (AMAT) surged 10.82% on June 29 to close at $694.64, reflecting the market’s continuing expectations for expanded chip production capacity.

Memory: “Locked-In” Capacity and Pricing Power

If GPU is the “brain” of AI computing, then high-bandwidth memory (HBM) is the “nerve fibers” that keep the brain running at high speed. During AI training and inference, memory bandwidth directly determines whether compute units can be adequately fed with data—this is known as the “memory wall” bottleneck.

High-bandwidth memory demand is rising rapidly as training and inference models continue to expand. The market generally observes that major production capacity has been locked in by large customers in advance for 2026 and even 2027, leaving extremely limited near-term supply elasticity. This supply-demand imbalance gives memory suppliers stronger bargaining power in pricing, order visibility, and profitability.

From market data, the momentum of the memory track has also been validated. Micron Technology (MU) closed at $1,145.28 on June 29, up 1.14%. SK Hynix, another core player in the HBM market, together with Micron and Samsung Electronics, forms the “iron triangle” of global high-bandwidth memory supply. Samsung Electronics’ weight in AI-infrastructure-related investment portfolios is also not to be underestimated.

The investment logic of the memory track differs from GPU’s: it is not simply a contest of technology leadership, but rather a competition over the speed of capacity expansion and the depth of customer lock-in. Because HBM has a complex manufacturing process and a long yield ramp cycle, the manufacturer that first achieves stable, large-scale mass production will gain a significant first-mover advantage. In addition, with the explosion of AI inference scenarios—expected to see inference compute demand exceed training demand—requirements for memory capacity and bandwidth will further increase.

Networking: AI’s “Nervous System” and the Next Bottleneck

In the networking track, a consensus is taking shape: the larger the AI cluster scale, the more likely network bandwidth will become the new bottleneck. In a report released in May, Bank of America estimated that the AI networking market will reach $316 billion by 2030, higher than its prior forecast of $240 billion.

The logic behind this judgment is as follows: AI training clusters are evolving from the thousand-card level to the ten-thousand-card level—and even to the hundred-thousand-card level. At this scale, communication efficiency between GPUs directly determines the effective utilization rate of total compute power. The so-called “zombie GPU” effect in the industry—where expensive GPUs sit idle due to I/O waiting—is becoming one of the most troublesome problems for hyperscale customers. Evaluation metrics are shifting from simply the number of floating-point operations (FLOPS) to time-to-first-token (TTFT) and vector retrieval speed.

During the 2026 Summer Davos, Lan Shangli, Global Senior Vice President at Ericsson, proposed that the first wave of AI investment has flowed into chips and data centers, but the winners in the next stage may be telecom operators laying fiber-optic cables and building base stations. He compared networking to the nervous system of “physical AI”—large language models are the brain, robots and drones are the body, and the network is responsible for enabling the brain to drive the body.

On the side of networking equipment, Broadcom (AVGO) is a name you cannot bypass. As a core supplier of AI networking chips (such as switch ASICs), Broadcom benefits deeply from upgraded demand for interconnect bandwidth within data centers. Although its stock price has recently pulled back, institutions such as Jefferies still maintain a “Strong Buy” rating, with an average target price of about $513.58. On June 29, Broadcom closed at $372.45, up 2.04%.

In addition, Cisco Systems, as a traditional networking equipment giant, is also actively transforming to meet the new needs of AI data centers. It closed up 3.45% on June 29 to $117.70. Dell, as an AI server system integrator, closed up 3.78% to $414.61.

Horizontal Comparison of the Three Tracks and an Investment Perspective

From the perspective of their positions in the industrial chain, there are significant differences among GPU, Memory (Memory), and Networking (Networking):

GPU track sits at the very top of the value chain, enjoying the highest gross margins and technology premium, but it also faces the highest valuation levels and market expectations. NVIDIA’s current price-to-earnings ratio (TTM) is about 29.86x. Given its growth rate, this valuation is not extreme among tech giants, but any slowdown in demand growth could trigger a valuation repricing.

Memory track has even more pronounced cyclical characteristics. Tight supply for HBM may temporarily mask the cyclical fluctuations of traditional DRAM and NAND, but investors still need to watch for changes in the supply-demand relationship after large-scale capacity is released. The current pattern of capacity being locked through 2026–2027 provides relatively clear mid-term earnings visibility for this track.

Networking track currently receives less market attention than GPU and memory, but this may precisely imply a larger expectation gap. Bank of America’s forecast of a $316 billion market size in 2030 suggests that the networking track’s compound growth rate over the next few years may exceed current market consensus expectations.

From the risk perspective, the three tracks share risks including: a marginal slowdown in AI capital expenditure, geopolitical disruptions to the supply chain, and the impact of technology route changes (such as new paradigms like in-memory computing and optical interconnects) on the existing industrial landscape. In research on more than 200 enterprises, Omdia identified four major core challenges: ROI and time-to-IPO, digital sovereignty, the AI talent gap, and systemic engineering complexity. These challenges will affect the investment return cycles of each track to varying degrees.

How to Invest in AI Infrastructure on Gate?

For investors looking to participate in AI infrastructure investment opportunities, the Gate platform offers diversified entry routes.

Gate has listed more than 12,500 stock products, including U.S. stocks, Hong Kong stocks, and Korean stocks. The platform now fully supports 7×24 hours trading of U.S. stocks, Hong Kong stocks, and Korean stocks—covering the pre-market, intraday, after-hours, overnight sessions, and weekend market-closure periods. This means investors do not need to be constrained by the opening hours of traditional exchanges, and can adjust their positions more flexibly based on market dynamics.

On AI-infrastructure-related stock products, Gate covers many of the core companies mentioned in this article: NVIDIA (NVDA), AMD (AMD), Micron Technology (MU), Broadcom (AVGO), Applied Materials (AMAT), Cisco (CSCO), Dell (DELL), and more. Investors can use Gate’s stock trading module to complete the allocation and rebalancing of these targets in one place.

Conclusion

In 2026, AI infrastructure has evolved from conceptual storytelling into a capital-expenditure competition backed by real money. The tens of billions of dollars—actually, hundreds of billions—invested annually by hyperscale tech companies are weaving GPUs, high-bandwidth memory, and high-speed networks into a global compute infrastructure network.

The GPU track benefits from the highest technical barriers and the most direct mapping of compute demand, making it the direction with the strongest current certainty. The memory track, supported by a supply-demand pattern with capacity locked in, offers the clearest mid-term earnings visibility. The networking track, due to insufficient market recognition, may contain the largest expectation-gap opportunity.

The investment cadence and risk-reward characteristics of the three differ. Investors can make differentiated allocations according to their own risk preferences and investment horizons. Gate’s 7×24 hour stock trading and extensive product coverage provide flexible and efficient execution tools for this allocation.

The construction cycle of AI infrastructure is far from over. As Jensen Huang said at NVIDIA’s 2026 annual shareholder meeting, AI infrastructure is the largest-scale infrastructure project in human history. In this multi-year wave of capital expenditures, understanding the structure and rhythm of the industrial chain may deliver more long-term value returns than chasing short-term hot spots.

FAQ

Q1: Which sub-sectors are mainly covered by AI infrastructure investment?

It mainly includes three core hardware tracks: GPU (graphics processing unit, responsible for AI compute acceleration), high-bandwidth memory (HBM, solving the “memory wall” bottleneck), and data center networking (solving interconnect communication problems for large-scale clusters). In addition, it includes supporting areas such as data center cooling, power systems, and a software orchestration layer.

Q2: Why is Networking (Networking) considered the next big trend for AI investment?

As AI training clusters expand from the thousand-card scale to the ten-thousand- and hundred-thousand-card scales, communication efficiency between GPUs becomes a key bottleneck for effective compute utilization. Bank of America predicts that the AI networking market size will reach $316 billion by 2030. Networking is likened to the nervous system of “physical AI,” which is the infrastructure enabling intelligence to move from data centers to the real world.

Q3: Can AI-infrastructure-related U.S. stock targets be traded on Gate?

Yes. Gate has listed more than 12,500 stock products, covering the U.S., Hong Kong, and Korean markets, including core companies such as NVIDIA (NVDA), AMD (AMD), Micron Technology (MU), Broadcom (AVGO), and more. The platform supports 7×24 hour trading and covers pre-market, intraday, after-hours, overnight, and weekend sessions.

Q4: What are the main risks faced by AI infrastructure investment today?

Main risks include: demand pullback driven by slower AI capital expenditure growth, geopolitical disruptions to the chip supply chain, the impact of technology route changes (such as new paradigms like in-memory computing and optical interconnects) on the existing landscape, and correction pressure caused by overly high valuations in certain tracks. Investors need to allocate in line with their own risk tolerance.

Q5: What is the expected market size for AI infrastructure in 2026?

Research and Markets data shows that the global AI infrastructure market size is expected to grow from $71.88 billion in 2025 to $90.91 billion in 2026, with a year-on-year growth rate of 26.5%. Another institution forecasts it will reach $465 billion by 2033. In 2026 alone, global leading technology companies’ combined capital expenditures on AI infrastructure have already exceeded $600 billion.