Folks, if you're still trading Nvidia, Broadcom, and AMD as three ordinary tech stocks, I have to tell you directly: you may have already missed the core of this AI

Folks, if you're still trading Nvidia, Broadcom, and AMD as three ordinary tech stocks, I have to tell you directly: you may have already missed the core structural restructuring of this AI round.

The truly important question is no longer 'which company's stock price rises faster' but: in the next three years, who will hold control of AI? Is it the company that simply sells chips? The company that provides high-speed connection pipes? Or those who can define the operating rules of the entire data center system, making AI run like never-stopping super factories?

If you think through this question, you'll see the investment logic has completely changed. As of late May 2026, AI is no longer a chatbot, image generator, or software toy for writing code; it's becoming 24/7 never-stop industrial super factories. This power structure reshaping will determine where global capital flows over the next decade and how much real dividend we ordinary US investors can capture, rather than just chasing short-term market fluctuations.

Today's video, in the most grounded and understandable way, from an ordinary investor's perspective, will thoroughly dissect the real positions, interrelationships, and the entire AI infrastructure power structure of Nvidia, Broadcom, and AMD. After listening, you'll have a clear 'AI power map': who defines the rules, who earns stable cash flow, who provides cost-effective flexibility as a backup, and how ordinary people can intelligently position to avoid the biggest risks and capture truly long-term structural opportunities.

Over the past two years, those who understood early the underlying logic of 'AI shifting from software to infrastructure' have significantly outperformed the S&P 500. Now is the critical window before the second wave of AI industrialization.

1. Nvidia: From Chip Company to AI System Defined

If you still see Nvidia purely as a GPU manufacturer, honestly, that view is a bit outdated. It now does far more than sell chips; it's comprehensively defining how the entire AI system should run efficiently, how to substantially reduce costs, and how to achieve large-scale deployment.

You can imagine it as three roles perfectly combined: AI computing standard setter, data center architect, and AI factory system general contractor.

In the past, when enterprises and cloud providers bought servers, they typically separately purchased CPUs, GPUs, network cards, storage devices, then had engineering teams piece them together. This was inefficient, costly, had compatibility issues, and was troublesome to maintain. Now the situation has fundamentally changed. More and more large customers, like Microsoft, Meta, Amazon, Google Cloud, and Oracle, directly purchase complete AI rack solutions from Nvidia. These include not only top GPUs but also bundled CPUs, high-speed network interconnects, liquid cooling systems, software optimization stacks, and standardized rack designs. At this point, Nvidia earns not just hardware profit per chip but the entire system-level premium and long-term ecosystem service revenue.

The competitive dimension has also completely changed. In the past, everyone compared whose GPU peak performance was faster; now they compare who can make each AI output cheaper, more power-efficient, and more effective. The core metric is single token cost, which you can think of as 'oil price' or 'electricity price' in the AI world. Whoever can continuously drive down this cost can infinitely expand scale and capture more market share; whoever's cost remains high will be gradually squeezed out.

Nvidia's repeated emphasis at Computex 2026 and subsequent GTC conferences has shifted from simply 'faster' to 'cheaper' and 'end-to-end system efficiency'. The next-generation Rubin architecture platform aims to significantly reduce inference single token cost. According to official disclosures and industry feedback, compared to Blackwell, Rubin is expected to achieve up to 10x inference token cost reduction, and when training MoE models, required GPU count can be reduced to a quarter. Rubin uses 3nm advanced process, supports HBM4 high-bandwidth memory, single GPU performance up to 50 PFLOPS (FP4 precision), and has begun large-scale deployment through major cloud providers like AWS, Google Cloud, and Microsoft Azure in the second half of 2026.

Using a very life-like analogy: previously AI was like a cooking competition – who makes better food; now AI competes on who can make millions of high-quality meals a day at the lowest cost and still be profitable. At this point, the real winner is no longer the individual skilled chef but the one who controls the entire supply chain design, factory layout, and operational efficiency.

Nvidia's complete system solution includes Rubin GPU, Vera CPU, high-speed NVLink 6 interconnect, Spectrum-6 network switch, whole-rack liquid cooling, BlueField-4 smart network card and storage processing unit, and advanced CoWoS packaging. These components are designed from the start as a tightly coordinated unified system with one core goal: make AI run cheaper, more power-efficient, easier to scale, and lower maintenance costs.

Of course, every coin has a risk side. These risks are real: current valuation already discounts much future growth; large customers like Microsoft and Amazon are pushing their own chips to reduce dependence; AMD and other custom ASIC solutions are gradually siphoning some mid-low-end demand. But we must emphasize: in the short term, no company can fully replace Nvidia's dominant position in system-level integration, software ecosystem, and full-stack optimization. Its CUDA development platform and full-stack optimization capability remain one of the strongest moats in the industry.

2. Broadcom: The Invisible Pipe King of AI Infrastructure, Severely Underestimated Stable Cash Flow Machine

If Nvidia is the shining star on center stage, Broadcom is the most critical yet most easily overlooked 'infrastructure pipe king' behind the scenes.

Many ordinary investors find Broadcom not 'sexy', not like Nvidia making headlines often, but professional institutional investors love such companies best – they don't rely on storytelling to attract attention but earn real cash flow and high gross margins.

Broadcom's core position in AI can be summarized in one sentence: AI data center connectors + custom chip basic supplier. It is the key bridge that enables thousands of servers to truly form a super intelligent cluster.

It mainly does three big things, each critically important:

First, AI network infrastructure. AI computing is no longer a single server computing alone, but thousands or even hundreds of thousands of servers must interconnect at high speed and work in real time. How to achieve low-latency, high-bandwidth communication between servers? Rely on Broadcom's Tomahawk series and Jericho series high-speed Ethernet switch chips, network controllers, and SerDes physical layer technology. Simply put: without Broadcom's high-speed network infrastructure, many large AI training and inference clusters simply cannot run efficiently. In 2026, as AI cluster scales continue to expand, Broadcom's revenue in this segment grew very steadily.

Second, custom ASIC chip trend. This may be Broadcom's most underestimated growth engine currently. Tech giants like Google, Meta, Amazon, OpenAI, Microsoft are all developing their own custom chips (XPU), but they typically don't handle production themselves; they partner with Broadcom for co-design and tape-out. In fiscal Q1 2026, Broadcom AI semiconductor revenue reached $8.4 billion, up 106% YoY, with custom AI accelerators contributing heavily. It has signed multi-generation deep cooperation agreements with Meta, supporting multi-gigawatt deployment by 2029, and added important new customers like OpenAI. Current AI chip backlog orders reach $73 billion, with management targeting AI chip revenue over $100 billion by 2027.

Third, software cash flow base. Through successful integration of VMware, Broadcom has significantly strengthened enterprise cloud and AI infrastructure software capabilities, bringing very stable subscription revenue. VMware Cloud Foundation latest version 9.1 is specifically optimized for production-grade AI workloads, providing Broadcom with a stable cash flow source resistant to economic cycles.

Extended insight: According to industry analysis reports, hyperscalers spending $1 on custom AI accelerators typically spend an additional $0.4-0.6 on network interconnection and infrastructure, a large portion supplied by Broadcom. Its growth rate is not as explosive as Nvidia, but extremely stable, high gross margin, ample cash flow, typical slow bull high-quality asset. In current high interest rate environment, companies offering high visibility, certain cash flow are especially favored by long-term institutions.

3. AMD: AI Computing Cost-Effectiveness Challenger, Important Backup Providing Flexibility

AMD's position in the overall AI power structure is very clear – it is not the rule setter but a strong alternative provider and cost-effectiveness challenger. Its existence keeps the market healthily competitive, preventing monopoly by a single supplier.

Core business mainly three: MI350 series GPU, data center EPYC CPU, and AI accelerator integration solutions. It mainly uses higher cost-effectiveness and more open ROCm ecosystem to challenge Nvidia's closed CUDA system.

Opportunities mainly in three scenarios: cost-sensitive small and medium cloud providers, AI inference-dominated market, and CPU+GPU integrated architecture scenarios. MI350X has huge improvement over previous generation, equipped with up to 288GB HBM3e memory, memory bandwidth up to 8 TB/s, providing better per-dollar performance in some long-context inference scenarios. It offers higher cost-effectiveness in some training and inference tasks, especially suitable for customers who don't need top ecosystem but pursue cost optimization.

But challenges are equally clear: CUDA ecosystem developer stickiness and maturity gap still large; system-level full-stack integration relatively weak; large customers still gradually building confidence in full trust and large-scale deployment. In 2026, AMD's overall share in AI accelerator market is around 5-7%, far below Nvidia's over 80%.

So reality is direct: AMD is now a 'backup-type core asset'. It provides important flexibility options and price competition pressure, but in the short term, it cannot comprehensively shake Nvidia's dominance. Its explosive potential will largely depend on the depth of custom chip wave and large-scale commercial expansion of AI inference market.

4. Real Power Structure and Interdependency of the Three Companies

Putting Nvidia, Broadcom, and AMD together, the entire AI industry can be clearly divided into three interdependent layers of power structure, not a simple three-kingdom competition:

First layer: System definition layer – Nvidia, responsible for defining how AI runs, architectural standards, and software ecosystem. It is the rule-maker and ecosystem dominator of the entire game.
Second layer: Infrastructure layer – Broadcom, responsible for efficient AI connection, network transmission, and custom underlying chips. It is the 'pipes and bridges' enabling the entire system to actually run and coordinate smoothly.
Third layer: Alternative competition layer – AMD, providing price competition, cost-effectiveness choices, and market flexibility. It prevents monopoly risk from being too high while offering customers backup options, keeping market vitality.

These three layers are not zero-sum but a highly interdependent ecosystem. Nvidia dominates standard setting, Broadcom provides key physical support and custom capability, AMD maintains market competition pressure. This layered power structure is a typical feature of AI entering industrialization phase, making the entire industry healthier and more sustainable, also giving investors good opportunities for diversified positioning and risk hedging.

5. Three Most Critical Variables and Biggest Inflection Points in the Next Three Years

What will truly determine market direction in the future is not one quarter's earnings of a company but three deep structural issues:

First, power constraint. Power consumption of AI servers is continuously and rapidly rising. According to Goldman Sachs' latest forecast, US data center power demand will rise from 31 GW in 2025 to 41 GW in 2026, possibly soaring to 66 GW by 2027. Global data center total electricity consumption may reach 945 TWh by 2030, equivalent to Japan's annual total consumption, with AI accelerator server electricity growth rate about 30% per year. Imagine tens of thousands of high-power Rubin architecture racks running at full power simultaneously – this has escalated from a simple data center issue to a grid stability and national energy system issue. Liquid cooling technology, nuclear restart, long-term natural gas contracts, transformer mass production, etc., will become new investment hotspots.

Second, custom chip trend. Cloud providers continue pushing custom chips to reduce reliance on a single supplier, but this trend actually makes Broadcom earn more in custom design and network infrastructure.

Third, enterprise investment return rate problem. Can AI projects continuously bring real, quantifiable profit growth for enterprises, or become a huge cost black hole? This will be the biggest test point for market sentiment shift in 2027-2028.

What is the most critical inflection point next? The smooth large-scale rollout of Rubin architecture and actual resolution of power bottleneck. If power supply can keep up and system costs are controllable, AI capital expenditure cycle will easily extend to 2028 or beyond; if severe power shortage or enterprise ROI decline appears, the market may experience periodic sharp adjustments and capital reallocation.

6. How Should We Ordinary US Investors Respond to This AI Power Structure?

I'll compress the most practical advice into three points for easy direct implementation:

1. Don't all-in on a single stock. Nvidia is still the core high-growth target (market share over 80%), suitable as core holding in a portfolio, but must strictly diversify risk, not pin all hopes on one company.

2. Focus on three main lines:

   • Nvidia's system definition capability (high growth but volatile, suitable for long-term core position);
   • Broadcom's stable infrastructure cash flow (slow bull, deep moat, suitable for balanced allocation);
   • AMD's cost-effectiveness flexible opportunity (explosive potential but higher uncertainty, suitable as satellite position).
     At the same time, look more at the underlying level, focusing on infrastructure tracks that truly support long-term AI factory operation: power, nuclear, liquid cooling, data center equipment, transformers, etc.

3. Long-term perspective + strict risk management. Use Dollar Cost Averaging (DCA) to gradually build positions, focusing on company fundamentals, cash flow, and moat depth, not short-term stock price emotional fluctuations. Remember, AI is a decade-level infrastructure super cycle. Pay special attention to inflationary pressure from rising electricity costs, regulatory changes (some states have begun to impose new regulations on data center electricity and taxes), and geopolitical supply chain risks. Consider relevant industry ETFs or index funds to further diversify single company risk.

My final judgment: This AI cycle is far from over; it's just entering the most exciting and imaginative industrialization acceleration phase. The real profit logic is to invest in the power structure of the entire AI infrastructure, not chase short-term concept speculation and emotional fluctuations.

Folks, are you now investing in three isolated chip companies, or investing in the power distribution of the entire AI industrial empire? Do you value short-term explosive power more, or long-term monopoly ability and stable cash flow? Have you realized that the players who truly control wealth flow in the long run may be those who control system standards, data flow connections, and energy supply?