[Definitive Guide] What is Custom Silicon (ASIC)? [Broadcom, Marvell, NVIDIA: Who is the Winner?]

Everyone is desperately trying to buy NVIDIA GPUs, waiting months for the delivery of the latest servers, and the global stock market fluctuates with every earnings announcement from the company. That era of "AI investment = NVIDIA GPUs" is coming to an end.

As of 2026, what companies are looking for are "cheaper, more efficient, and more power-saving chips."

To meet this urgent demand, hyperscalers (giant IT companies) like Google, Amazon, and Microsoft have set their sights on breaking away from NVIDIA dependency and have simultaneously steered toward in-house production of their own dedicated chips, "custom ASICs."

And behind the scenes of this historic shift in hegemony, there are entities currently reaping the most enormous profits. They are the masterminds of chip design, ASIC manufacturers "Broadcom" and "Marvell."

Market data clearly tells the story of this radical change.

Shipments of custom ASICs are expected to record explosive growth of 44.6% year-on-year in 2026. This is a staggering pace, about three times the growth rate of NVIDIA GPUs (16.1%). As a result, the share of ASIC-equipped AI servers is projected to reach 27.8%, hitting its highest level since 2023.

What is even more shocking is what comes next. It is estimated that a "reversal phenomenon" where ASIC shipments overtake general-purpose GPUs will occur in 2028, and the AI accelerator market is on track to balloon to a scale of $600 billion (approximately 90 trillion yen) by 2033.

The phase of searching for the "next NVIDIA" has already begun.

In this article, we will thoroughly delve into the capabilities of the protagonists of this historic turning point, "Broadcom (AVGO)" and "Marvell (MRVL)," from an investor's perspective, while comparing them with the absolute champion, "NVIDIA (NVDA)."

Now, let's look at the "true future" of the AI infrastructure market.

Chapter 1: What is an ASIC—The shift from "general-purpose" to "dedicated"

1-1. The fundamental difference between GPUs and ASICs

To understand why the transition to ASICs is happening now, I will explain the difference between GPUs and ASICs using the analogy of "cooking tools."

• NVIDIA GPU (General-purpose GPU) = Chef's Knife

  • It is an extremely excellent general-purpose tool that can cut any ingredient. It can flexibly handle all kinds of workloads, from complex AI training to inference, advanced image processing, and video generation.

  • In the early stages of research and development, or in environments where algorithms evolve daily, this "use it for anything" nature becomes the strongest weapon.

• Custom ASIC = Sashimi Knife

  • It is a dedicated tool optimized to the extreme for the specific purpose of "slicing sashimi."

  • In terms of the speed of slicing sashimi, the beauty of the cross-section, and the energy efficiency of the task, no chef's knife can compete. However, it cannot cut a pumpkin (perform other calculations).

Technically defined, ASIC (Application-Specific Integrated Circuit) refers to a dedicated integrated circuit designed solely for a specific application or a specific AI workload.

| 比較軸					| NVIDIA GPU（汎用）			| カスタムASIC（専用）					|
|-----------------------|---------------------------|-----------------------------------|
| 汎用性と柔軟性			| 高い（あらゆるAIモデルに対応）	| 低い（特定のモデルやタスクに固定）		|
| 電力効率（特定タスク時）	| 中程度3〜5倍				| 高い								|
| TCO（総保有コスト）削減	| 基準値						| 30〜50%削減（超大規模運用時）			|
| 開発コスト・期間			| 不要（購入してすぐ使える）		| 莫大（数百億円規模・2〜3年を要する）	|
| 最も適した導入フェーズ	| 研究・実験・学習など			| 本番環境での推論・特定モデルの超大量処理	|

According to Broadcom's recent reports, custom AI ASICs can achieve a 30-50% reduction in TCO (Total Cost of Ownership) compared to NVIDIA GPUs for specific AI workloads.

By sacrificing versatility, you can gain overwhelming economic efficiency and power savings.

1-2. Why are hyperscalers shifting to ASICs all at once now?

In fact, it was 2015 when Google first introduced its AI-specific chips (TPUs) into its data centers. This was two years before NVIDIA announced its groundbreaking Volta architecture.

However, at that time, only Google could justify the enormous development costs, and it was nothing more than a "luxury for a few large corporations."

There are three clear reasons why this has become an industry-wide trend in 2026.

1. AI consumption has shifted from "experimentation" to "production-scale mass processing (inference)"

   • Generative AI services like ChatGPT, Gemini, and Copilot are currently processing billions of queries per day

   • In such production inference environments where "the same specific calculation is repeated billions of times," the flexible versatility of GPUs is no longer necessary

   • Instead, the power efficiency of dedicated chips that can reduce massive electricity costs becomes a life-or-death issue

2. The massive scale of cost reduction through economies of scale

   • The number of GPUs operated by Google, AWS, Meta, Microsoft, and others has already reached the hundreds of thousands

   • If infrastructure costs per unit (including power, cooling, and rack space) are reduced by 30-50%, the total savings for the entire company reach the scale of hundreds of billions to trillions of yen annually

   • They have reached a scale where the enormous initial development costs of ASICs can be recovered in just a few months

3. The Management Imperative of Breaking Free from "NVIDIA Lock-in"

   • Currently, NVIDIA's share of the AI GPU market is in a near-monopolistic state, totaling approximately 80-90% (over 90% for training-specific, and 60-75% for inference).

   • As long as this overwhelming market share continues, hyperscalers will have no bargaining power.

   • Incorporating "in-house proprietary chips" into the foundation of their infrastructure is a management strategy itself to break free from excessive dependence on NVIDIA and to control profit margins.

Chapter 2: The Custom ASIC Supply Chain—Who Makes Money and Where?

This is the most essential and important perspective in investment.

When people see news that "Google is using custom ASICs (TPUs)," many tend to misunderstand it as "Google is designing and manufacturing chips from scratch in-house."

However, the actual value chain is as follows:

[Custom ASIC Supply Chain and Capital Flow]

1. Hyperscalers(Google, Meta, Amazon, OpenAI, etc.)

   • Determine the [specifications] for "wanting to process these types of calculations at this speed to run AI models."

2. Design Partners(Broadcom / Marvell)

   • Convert those required specifications into "advanced silicon blueprints (physical design)" that can actually be manufactured, and incorporate their own communication IP, etc. This is where the highest added value is created.

3. Foundry (TSMC)

   • Based on those precise blueprints, they perform the [physical manufacturing] using cutting-edge processes like N2 and A16.

   • Related Article: [Complete Preservation Edition] Understanding TSMC: What are 2nm and A16? The Full Picture of the AI Chip Manufacturing Monopoly

4. System Integrator (Dell / Super Micro / HP)

   • They incorporate the finished chips into massive AI server racks and cooling systems, then deliver them to data centers.

   • Related Article: [Complete Preservation Edition] What are AI Server Racks? Dell, Super Micro, HPE [For Companies Incorporating NVIDIA GPUs]

Hyperscalers know 'what kind of dish they want to cook,' but they lack the physical design know-how to 'burn that into nano-level silicon that runs with minimal power and heat,' as well as the advanced IP (intellectual property) required to make chips communicate with each other at the speed of light.

According to JPMorgan's estimates, in the high-end AI ASIC design market, Broadcom holds approximately 80-85% share, and Marvell holds approximately 10-12%, creating a de facto duopoly.

Behind the news that 'Google has developed its own chips,' the fact that these two companies actually hold the majority of the value-added in design is an important point that investors often overlook.

Chapter 3: NVIDIA (NVDA) — The GPU King's Adaptation and Counterattack to Custom ASICs

Will the era of 'NVIDIA' end with the rise of ASICs? In conclusion, the king has already made its next move, and its position remains rock solid.

3-1. Latest Earnings Highlights (Q1 FY2027: May 20, 2026)

| 財務指標				| Q1 FY2027 実績	| 備考											|
|-----------------------|---------------|-----------------------------------------------|
| 売上高					| $81.62B		|（前年同期比 +85%）成長率は鈍化しつつも絶対額は驚異的	|
| データセンター売上比率	| 約92%			| 完全なAIインフラ企業としての地位を確立（売上高$75.2B）|
| 粗利益率				| 75%			| ハードウェア企業として歴史上類を見ない高水準			|
| AI GPU市場シェア		| 約80〜90%		| 学習用途で90%超、推論で60〜75%を維持				|
| CUDAエコシステム開発者	| 400万人以上	| 20年以上の蓄積による強力な堀（モート）				|

3-2. NVIDIA's Answer to the 'Custom ASIC Era' — The Impact of 'NVLink Fusion'

CEO Jensen Huang brilliantly defied the simple prediction that NVIDIA's sales would fall if ASICs became widespread. On March 31, 2026, NVIDIA announced a groundbreaking mechanism called the 'NVLink Fusion Platform'.

This allows customers (hyperscalers) to build entire servers using NVIDIA's ultra-high-speed communication network, the 'NVLink ecosystem,' when creating their own custom ASICs. Specifically, it enables the seamless integration of custom chips designed by companies like Marvell with NVIDIA's Vera CPU, ConnectX NIC, BlueField DPU, and Spectrum-X switches to create AI clusters.

Industry insiders are calling this the 'Ecosystem Tax on Custom Chips'.

Data centers that adopt NVLink Fusion will inevitably incorporate NVIDIA-made components (either CPUs, switches, or network equipment). In other words, even if the star of computing shifts from NVIDIA's GPUs to other companies' ASICs, they have created a structure where NVIDIA is guaranteed to generate revenue from the network and control parts of the entire data center. It is nothing short of brilliant to have incorporated even the rise of ASICs, which were thought to be competitors, into their own revenue ecosystem.

Behind CEO Huang's boast at GTC 2026 that there is '1 trillion dollars in committed orders by 2027,' lies this precise infrastructure dominance strategy.

3-3. [NVDA 5+1 Framework Analysis]

NVDA Overall Rating: ★★★★★ (5/5)

• By layering the new 'ecosystem tax' mechanism of NVLink Fusion onto the 20-year-old software moat of CUDA, they have prepared a formation to defend and expand their revenue base even amidst the structural change of the rise of ASICs. The outcome of China export restrictions and trends in the inference market share are the biggest points to watch over the next 12-18 months.

Chapter 4: Broadcom (AVGO) — The True Ruler of the 'Custom Silicon Empire'

In the custom ASIC era, the company that will receive the most direct and certain benefits is 'Broadcom.' Under the genius management of CEO Hock Tan, the company has solidified its position as the 'emperor' of the semiconductor industry.

4-1. Latest Earnings Highlights (Q2 FY2026: June 3, 2026)

| 財務指標				| Q2 FY2026 実績	| 備考													|
|-----------------------|---------------|-------------------------------------------------------|
| 売上高					| $22.19B		|（前年同期比 +48%）AI半導体およびソフトウェアが過去最高		|
| データセンター売上比率	| 約49%			|（前年同期比+143%）AI半導体が全体の牽引役（$10.8B）			|
| 粗利益率				| 77.1%			| ソフトウェア統合後も高い水準を維持（EBITDAマージンは69%）		|
| カスタムAI ASICシェア	| 約55〜70%		| Google（TPU）、Meta等向けカスタムシリコン設計で圧倒的シェア	|

In the Q2 earnings report, the stock price adjusted temporarily as it fell slightly short of some investors' excessive expectations (a significant upward surprise in guidance), but the sales growth of AI chips alone is hitting explosive figures of +140% year-on-year.

More importantly, CEO Hock Tan declared that "there is a clear path to over $100B in revenue from AI chips alone by 2027." The massive $73B (approximately 11 trillion yen) order backlog guarantees solid profitability through mid-2027.

4-2. The 6 Mega-Customers and the Realism of the "$100B Goal"

There are currently 6 customers for the custom AI ASICs (which the company calls XPUs) that Broadcom develops. The lineup consists entirely of frontrunners leading global AI development.

| 顧客企業	| コミットメントと現状							| 規模感・影響度					|
|-----------|-------------------------------------------|-------------------------------|
| Google	| TPU v7 Ironwood、v8に向けた長期契約			| 最大にして最古参の超重要顧客		|
| Anthropic	| 2026年に1GW、2027年に3GW以上へインフラ拡大		| 3年間で規模を3倍に急拡大中		|
| Meta		| 「MTIA」の複数世代展開、2029年までの長期契約	| 複数ギガワット規模の強固な基盤		|
| OpenAI	| 第1世代XPUを2027年に展開開始、1GW以上の規模	| $100億を超える超大型合意			|
| 未発表A	| 業界アナリストの推測：Apple / ByteDance		| （未確認ながら巨大なポテンシャル）	|

According to estimates by analysts at Mizuho Securities, the contract with Anthropic alone, which is rapidly expanding its infrastructure, could bring Broadcom $21 billion in AI revenue in 2026 and $42 billion in 2027.

Furthermore, Broadcom has partnered with investment funds Apollo and Blackstone to establish an "AI XPU Platform." They are proceeding with a plan to deploy an astronomical computing capacity exceeding 20GW by 2028, and the first phase, a $35 billion financing, is already being led by Apollo.

4-3. [AVGO 5+1 Framework Analysis]

AVGO Overall Rating: ★★★★☆ (4/5)

• While it is the most direct beneficiary in the custom ASIC era, it is a stock where one must closely monitor the balance between high valuation and customer concentration risk.

• The "visualized demand" of a $73 billion backlog is a strong support, but it is essential to verify the quality of the contract details (whether it is for chips alone or higher value-added rack-scale solutions) on a quarterly basis.

Chapter 5: Marvell Technology (MRVL) — The Rise of the "Hidden Protagonist" in the Custom ASIC Era

In the era of AI hardware and custom ASICs, the company showing the most dramatic stock price performance is "Marvell Technology".

5-1. Latest Earnings Highlights (Q2 FY2026: June 3, 2026)

| 財務指標				| Q1 FY2027 実績	| 備考													|
|-----------------------|---------------|-------------------------------------------------------|
| 売上高					| $1.41B		|（前年同期比 +18%）AI向けデータセンター事業の急拡大が寄与		|
| データセンター売上比率	| 約48%			| データセンター向けが$678Mとなり前年同期比+42%の大幅増		|
| 粗利益率				| 52.8%			| 堅実な利益率を維持（カスタムASICおよびネットワーク製品）		|
| カスタムAI ASICシェア	| 約10〜15%		| Broadcomに次ぐカスタムASIC市場の主要なチャレンジャー（第2位）	|

In the Q1 earnings report, explosive growth in the data center segment (custom ASICs, high-speed optics, etc.) drove the company to achieve record-high revenue ($2.42B, +28% year-on-year).

While clearing the market's bullish expectations, CEO Matt Murphy emphasized that "AI-related bookings are expanding at an extraordinary pace," and against that backdrop, announced a significant upward revision to the full-year (FY2027 and FY2028) earnings outlook.

Winning large-scale projects for 800G/1.6T switches for cutting-edge AI clusters and next-generation custom silicon strongly guarantees solid medium- to long-term profitability.

5-2. Broadcom and Marvell's Brilliant "Division of Territory"

Both companies specialize in "custom ASIC design services", but if you look carefully at their customer bases, you can see that they are not engaged in a bloody direct competition, but rather have brilliantly divided their territories.

| クラウド・AI企業	| Broadcomの顧客		| Marvellの顧客	|
|-------------------|-------------------|---------------|
| Google			| ◎（TPU v7/v8等）	| —				|
| Meta				| ◎（MTIA等）		| —				|
| Anthropic / OpenAI| ◎					| -				|
| Amazon（AWS）		| —					| ◎（Trainium）	|
| Microsoft（Azure）	| —					| ◎（Maia）		|

While Broadcom has a firm grip on the frontier AI labs developing cutting-edge AI models (Google, Meta, Anthropic, OpenAI), Marvell has secured the world's two largest cloud infrastructure providers (Amazon AWS, Microsoft Azure) as solid partners.

5-3. The Unique Position Propelling Marvell: "Optical Interconnect"

Marvell's strength is not just in chip design.

It lies at the intersection of three technologies: "custom silicon × optical interconnects × data center networking".

As data centers that process AI grow larger, the network wall (communication congestion)—the question of "how to make tens of thousands of chips communicate with each other without latency and with minimal power"—becomes a greater bottleneck than the raw computing power of individual GPUs or ASICs.

Since acquiring Inphi for $10 billion in 2021, Marvell has led the industry in "optical DSP (digital signal processing)" technology, which converts electrical signals into optical signals for high-speed transmission. Improving computing power and expanding communication bandwidth are two sides of the same coin, and Marvell is in a unique position to provide solutions for both to AWS and Microsoft.

Furthermore, AWS and Meta are strongly promoting "Ethernet" as the network standard for next-generation AI clusters, which is a massive tailwind for Marvell's high-speed Ethernet switch portfolio.

5-4. [MRVL 5+1 Framework Analysis]

MRVL Overall Rating: ★★★★☆ (4/5)

• While often overshadowed by Broadcom, it is a "hidden protagonist" that secures the world's largest cloud operators, AWS and Microsoft. Its high volatility is a double-edged sword, but for investors who can tolerate the risks of Amazon dependency and price competition from new entrants (such as MediaTek), it offers appeal as a satellite position in a portfolio.

• Note that while the star rating is on par with Broadcom, it possesses a different quality of appeal. Through its participation in NVLink Fusion, the structure of "generating revenue from the expansion of ASIC demand regardless of NVIDIA's success or failure" can be evaluated as having relatively lower-dispersed ASIC exposure compared to the success/failure risks of the UALink strategy borne by Broadcom, or the concentration risk on capital-dependent customers like Anthropic and OpenAI.

• The cost paid for this is higher volatility due to its smaller corporate scale.

Chapter 6: Deep Dive into Competitive Ecosystems and Market Cycles

6-1. Other Market Competitors

• AMD (MI300X/MI325X)

  • Aims to replace GPUs in both training and inference, but the ROCm software ecosystem still faces a structural gap compared to CUDA.

  • Data center revenue is growing at a high rate of +57% year-over-year (Q1 2026), but the absolute scale remains only a fraction of NVIDIA's.

• Intel (Gaudi 3)

  • Positioned for niche inference cost optimization. Potentially structurally lagging behind.

  • Large-scale turnaround strategy is currently limited.

• Google-MediaTek Alliance

  • A New Challenger to Broadcom's High-End Oligopoly

  • MediaTek is rapidly expanding its business, targeting $2 billion in AI ASIC revenue in Q4 2026 alone

6-2. Current Status of Market Dynamics and Cycles

AI infrastructure investment is expected to reach a total of $660 billion to $690 billion in 2026 for the six hyperscalers (major players including Google, Amazon, Microsoft, Meta, OpenAI, and Anthropic), with estimates suggesting that 75% of this is directed toward AI-related infrastructure.

This supports a shift in demand from the "experimental phase" to production-scale processing (inference) and suggests that near-term AI capital expenditure is more structural than cyclical in nature.

Meanwhile, physical constraints such as TSMC's manufacturing capacity and the strain on power and cooling infrastructure are variables that must continue to be monitored as bottlenecks defining the pace of growth.

If a phase of slowing hyperscaler investment were to occur (e.g., the rise of skepticism regarding AI return on investment or a worsening macroeconomic environment), NVIDIA, Broadcom, and Marvell would likely all experience significant stock price volatility, which must be recognized as a cyclical risk .

Chapter 7: A Complete Comparison of the Three Companies—How Should We Investors Strategize?

Based on the analysis so far, I will summarize the positioning and investment appeal of the three companies in a matrix.

| 評価軸			| NVDA								| AVGO									| MRVL										|
|---------------|-----------------------------------|---------------------------------------|-------------------------------------------|
| ビジネスモデル	| プラットフォーマー(HW+SW+ネットワーク)	| インフラテック複合体（半導体+VMware）		| 設計＋光インターコネクト＋ネットワーク 			|
| 構造的モート	| CUDA＋NVLink Fusion				| ASIC設計シェア80%超＋VMwareロックイン		| AWS/Azure顧客基盤＋光DSP技術					|
| NVIDIAとの関係	| エコシステムの中心。ASICからも収益化	| UALinkでNVLink(NVIDIA)に対抗			| NVLink Fusion公式パートナー (NVIDIAとも協業）	|
| 顧客の質		| 分散（学習市場全体）＋地政学的集中リスク	| Google/Meta＋Anthropic/OpenAI			| Amazon/Microsoft (2026年設備投資で業界最大級)	|
| 財務の質		| 非GAAP粗利益率75%、FCFマージン約60%	| 営業利益率67.3%、FCFマージン40%超			| 非GAAP粗利益率58〜59%、成長投資フェーズ		|
| 成長ドライバー	| 市場独占＋Physical AI＋NVLink 収益化	| $730億バックログ＋AI XPUプラットフォーム	| FY2028目標（売上150億ドル）＋CXL技術			|
| バリュエーション	| 時価総額5.02兆ドル、PER 31.76倍		| 時価総額1.78兆ドル、PER 62.32倍			| 時価総額約1,500億ドル台、PER 60.45倍			|
| 主なリスク		| 対中輸出規制、推論市場でのシェア低下	| 顧客集中（上位3社で7割）、循環的資金リスク	| Amazon依存、MediaTek等新規参入、ボラティリティ	|
| 総合評価		| ★★★★★（5/5）						| ★★★★☆（4/5）							| ★★★★☆（4/5、質的にはより低分散）				|

Thinking about the Custom ASIC Era (A Warning Against Zero-Sum Thinking)

The simplification that "if ASICs grow, GPUs will be sold off" is incorrect.

Even if NVIDIA's market share (%) declines slightly, the speed at which the overall AI investment pie is expanding exceeds that, so NVIDIA's revenue (in absolute terms) continues to hit record highs.

In the 2026-2030 AI infrastructure market, the view that "ASICs do not replace GPUs, but rather expand the entire ecosystem " is more in line with reality.

Chapter 7: Concrete Investment Strategies Utilizing the New NISA

For Japanese investors, it is important how to combine these stocks within the "New NISA (Growth Investment Quota)," which allows for maximizing tax-free benefits.

Pattern A: The Standard, Simple Strategy (Core with QQQM)

Use the "QQQM" ETF, which tracks the Nasdaq-100 Index, as the core of your portfolio. QQQM already includes NVIDIA (approx. 11%) and Broadcom (approx. 6%) as top holdings. This alone automatically covers the majority of AI hardware.

On top of that, purchase "Marvell (MRVL)," a mid-cap growth stock not included in QQQM, as an individual stock in your satellite quota (about 5-10% of investment funds). This combination is a highly sophisticated method that minimizes management effort while still aiming for the upside of custom ASICs.

Pattern B: AI Infrastructure Concentration & High-Dividend Hybrid Strategy (AVGO + MRVL + QQQM)

For those who want to enjoy the thrill of individual stock investing, I strongly recommend purchasing Broadcom (AVGO) directly within the growth investment quota. In addition to its undervaluation at a P/E of 25x, its 1.5% dividend yield acts as a powerful cushion for long-term holding.

By holding NVDA indirectly through QQQM and using the remaining funds to invest in MRVL, which is expected to show high growth, you can achieve both stable cash flow and explosive growth.

Pattern C: The All-in-One, Maximum Diversification Strategy (Utilizing SMH)

If you want to avoid the risks of individual stocks, utilize the US semiconductor ETF 'SMH (VanEck Semiconductor ETF)'. Although the expense ratio is 0.35%, slightly higher than QQQM, it allows for a single, diversified investment across the entire custom ASIC value chain, including NVDA, AVGO, TSMC (manufacturing), and MU (memory).

Summary: How to Win in the Era of AI Hardware and Custom ASICs

The greatest technological innovation in human history, AI, has entered the era of 'custom ASIC chips' for infrastructure optimization.

• As hyperscalers drive demand for inference-specialized custom ASIC chips, the presence of these three companies is growing.

• The king, 'NVIDIA (NVDA)', will continue to be in demand for inference as well as training. Furthermore, in the custom ASIC field, they are building a new revenue pillar through their 'NVLink ecosystem' to stay competitive.

• 'Broadcom (AVGO)' is the prince of custom ASICs. By capturing demand in the inference sector, their backlog from Anthropic and OpenAI seems like a treasure trove. This is a stock that is sure to capture the ASIC market.

• 'Marvell (MRVL)' may look relatively small, but that is the flip side of opportunity. If they can overcome volatility and competition, they can achieve significant returns. Their ability to grow cooperatively with NVIDIA is a major appeal.

Without being swayed by superficial news or stock price volatility, let us identify the true cash-generating power of companies and their positions in the value chain to build a solid portfolio.

Thank you for reading until the end.

Bonus: '5+1' Framework Analysis of Each Company

As a bonus, I have summarized a detailed '5+1' framework analysis for 'Nvidia', 'Broadcom', and 'Marvell'.

NVIDIA (NVDA): 5+1 Framework Analysis

[Nvidia] Factor 1 | Structural Moat, Classification: Hybrid of Type C (Ecosystem Lock-in) + Type B (Market Leadership)

• NVIDIA's greatest strength is not the hardware performance itself, but the 'CUDA' software development environment, which has been accumulated over more than 20 years. With over 4 million developers building models on CUDA, major machine learning frameworks are fundamentally optimized for CUDA first. The cost for companies to switch to another company's chip is not just monetary, but weighs heavily in terms of the 'time axis of development delays', which is the essential depth of this moat.

• Furthermore, the 'NVLink Fusion' platform announced on March 31, 2026, evolved the nature of this moat one step further. This establishes an 'ecosystem tax' mechanism where even custom ASICs designed by competitors must incorporate NVIDIA's peripheral components such as Vera CPUs, ConnectX NICs, BlueField DPUs, and Spectrum-X switches. As evidenced by Marvell's participation as an official partner, NVIDIA has succeeded in incorporating even the ASIC camp, which should be their 'enemy', into their own revenue structure.

• Depth of Moat: Structural and self-reinforcing. The developer ecosystem of CUDA is difficult to replicate even with years of investment by new entrants, and NVLink Fusion is worthy of praise for extending this moat to the 'outside of the computing chip'.

• Trends of the last 2-3 years: While maintaining over 90% share in the training market, the workload shift to ASICs in the inference market has caused their share to drop to 60-75%. It is more accurate to say that the moat is 'being reconstructed while shifting the battlefield' rather than 'expanding'.

[Nvidia] Factor 2 | Competitive Durability

• Accumulation of Tacit Knowledge: The experience of over 20 years of silicon design and software co-optimization that supports the generational evolution of GPU architecture (Hopper -> Blackwell -> Vera Rubin) cannot be easily imitated.

• Regulation and Geopolitical Protection (and Risks): Export controls to China are not a "protection" for NVIDIA, but a clear headwind. With the H20 restrictions in April 2025, NVIDIA recorded a $4.5 billion inventory write-down, and its response has been inconsistent since then. Although the BIS (Bureau of Industry and Security) conditionally lifted the ban on H200 exports to China in January 2026 (with a 50% cap on shipments relative to US shipments, a 25% government share, and mandatory third-party verification), actual shipments remain stagnant due to national security reviews, and as of February 2026, actual shipments to China remain near zero. Access to the Chinese market (estimated by NVIDIA at approximately $50 billion annually) is the biggest swing factor for future performance.

• Network Effects and Ecosystem Density: While the CUDA ecosystem continues to expand, there is a structural vulnerability in inference applications where reliance on CUDA is not as strong as in training, making it relatively easy to switch through software rewrites.

• Capital Wall x Time Wall: Google (TPU), Amazon (Trainium), Microsoft (Maia), and Meta (MTIA) have all already executed multi-generational ASIC investments, effectively breaking through the "time wall." Therefore, the focus of competition has shifted to the remaining "versatility in the training market" and "peripheral monetization via NVLink Fusion."

• Triggers for Downgrading Durability Score: (1) If shipments to China remain effectively zero after the second half of 2026, (2) if market share in the inference market falls significantly below 60%, (3) if adoption of NVLink Fusion remains sluggish. It is necessary to check the trends in data center revenue sub-segments (Hyperscale vs ACIE) in every quarterly earnings report.

[Nvidia] Factor 3 | Pricing Power and Quality of Earnings

• The non-GAAP gross margin for FY2027 Q1 is 75.0% (guidance), maintaining a historically unprecedented high level for a hardware company. Free cash flow reached $48.6 billion in Q1 alone (up from $35 billion in the previous quarter), with no divergence from reported earnings. GAAP net income is $58.3 billion, and the profit margin relative to $81.6 billion in revenue is at an extremely high level.

• The source of pricing power is clearly structural (CUDA moat) and not temporary due to supply-demand tightness. However, the loss of sales opportunities due to export controls to China (Q2 guidance assumes zero data center revenue from China) is manifesting more as a "risk of growth rate plateauing" than as a risk to gross margin itself.

• Shareholder returns are extremely aggressive, with the quarterly dividend increased 25-fold from 1 cent to 25 cents, and a new $80 billion share buyback program established. This is a sign that management has high confidence in its current cash-generating ability.

[Nvidia] Factor 4 | Capital Efficiency and Quality of Earnings

• Because it maintains a fabless model, capital intensity is low, and ROIC is at the industry's highest level. The FCF margin for FY2027 Q1 reached approximately 60% (FCF $48.6 billion / Revenue $81.6 billion), and there are no signs of significant divergence between reported earnings and cash flow or accruals (accrual-based profit inflation). The balance sheet is also healthy, and financial risk derived from leverage is limited.

• On the other hand, it has been pointed out that the quarter-over-quarter growth rate of "Hyperscale" (approximately $38 billion, about 50% of data center revenue) as a breakdown of the data center business is relatively slowing, which suggests the possibility that hyperscalers' in-house ASIC investments are beginning to substitute for a portion of compute demand.

[Nvidia] Factor 5 | TAM and Scaling Power

• NVIDIA's growth drivers have shifted to "continued monopoly in the training market," a new horizontal expansion called Physical AI (robotics and autonomous driving), and the development of a new market worth approximately $200 billion called "Agentic AI" using the Vera CPU. CEO Huang announced at GTC 2026 that there are "$1 trillion in committed orders by 2027," which is an ambitious goal that includes an indirect monetization strategy through NVLink Fusion.

• Risk of Commoditization Trap: In the inference market, the economics of ASICs (30-50% reduction in TCO) continue to erode the value proposition of GPU versatility, and some analysts predict that NVIDIA's inference share will fall to 20-30% by 2028. This point is the variable that should be watched most closely as NVIDIA's weakness.

[Nvidia] Factor 6 | Valuation Asymmetry

• As of August 1, 2026, NVIDIA's stock price is $207.40, its market capitalization is approximately $5.02 trillion, and its trailing P/E ratio is 31.76x. Revenue has reached $253.49 billion in the last 12 months.

• Bull Case: While the P/E ratio does not feel expensive in absolute terms, when looking at the PEG ratio that accounts for growth rates, the expectations priced in by the market are by no means excessive. Considering the diversification of revenue sources through the CUDA moat and NVLink Fusion, a P/E ratio in the 31x range can be interpreted as reasonable, or even slightly undervalued, for a "core infrastructure company for AI."

• Bear Case: The complete evaporation of Chinese sales may not be fully priced into the stock price as a long-term risk. If a return to the Chinese market (estimated by NVIDIA at approximately $50 billion annually) is realized, this is a clear upside factor against consensus. Conversely, if restrictions on China are further strengthened, it will manifest as a downside risk.

• Evaluation of Asymmetry: While the market has not yet paid a sufficient premium for the high quality of earnings represented by 'monopolistic status in the training market + ecosystem tax via NVLink Fusion,' the structural risk of declining share in the inference market cannot be ignored. Overall, the valuation is in the 'fair to slightly undervalued' range, and no extreme overvaluation is observed.

Broadcom (AVGO): 5+1 Framework Analysis

[Broadcom] Factor 1 | Structural Moat, Classification: Type B (Market Leadership) close to Type A (Bottleneck Monopoly)

• Broadcom is not merely a semiconductor company, but a conglomerate of an 'AI Semiconductor Business' (custom XPU design services + AI networking chips) and an 'Infrastructure Software Business' (VMware).

  • Moat 1: In the custom ASIC design services market, it is the de facto gatekeeper of the high-end segment, holding an overwhelming share of 80-85% according to JPMorgan's estimates. It is considered to be more than a year ahead of competitors in advanced packaging technology called '3.5D XDSiP,' and is in a position to secure priority access to TSMC's manufacturing lines.

  • Moat 2: The very fact that it has secured all 'six core customers'—Google, Meta, Anthropic, and OpenAI—who are at the forefront of generative AI development as design partners is the most eloquent proof of its technical prowess and delivery capabilities.

  • Moat 3: VMware's virtualization infrastructure is deeply embedded in large corporate IT infrastructure, making it extremely difficult to switch due to physical and cost-related reasons. High-margin subscription revenue exceeding 90% supports the massive R&D investments of the semiconductor division.

• Strategic Move: Countering at the Standard Specification Level via UALink

  • Broadcom is leading the open interconnect standard 'UALink' together with AMD, Intel, Cisco, Google, HPE, Meta, and Microsoft. This is a direct counter-axis to NVIDIA's NVLink (and its extension, NVLink Fusion), signifying that Broadcom is staking its claim in the battle for hegemony over industry interconnect standards themselves, rather than just remaining a chip design service provider.

  • However, it should be noted that UALink has been pointed out to face 'tragedy of the commons' type challenges due to the pace of specification development and differing interests among member companies, and it must be discounted that it has not yet achieved the same centripetal force as NVLink Fusion.

[Broadcom] Factor 2 | Competitive Durability

• The biggest concern regarding Broadcom's durability is not the thickness of the moat itself, but the 'customer concentration risk.' In the Q2 FY2026 earnings call announced on June 3, 2026, Google, Anthropic, Meta, and OpenAI were named as core customers, and CEO Tan reiterated the suggestion of 'over $10 billion in full-cycle AI semiconductor revenue by 2028.'

• However, in the Q2 earnings report, it became clear that the contract with Anthropic was for 'chip-only sales' rather than the 'rack-scale' (high-value-added contracts including network equipment other than chips, etc.) that the market had expected, which led to a decline in the stock price. This is a correction of expectations regarding the substantive economic value of the contract, not a denial of the growth story itself, but investors need to pay close attention to the details of the contract.

• Regulatory and Geopolitical Protection: While it shares supply chain risks with NVIDIA in terms of TSMC dependency, the impact of direct export controls to China is limited (as all customers are US hyperscalers).

• Network Effects: The backlog has reached $73 billion, and Q2 2026 quarterly bookings alone hit $30 billion (against $10.8 billion in shipments), signaling continued strong demand. The CEO has stated that visibility extends through 2028.

• Capital Wall x Time Wall: With a clear follower like Marvell and new entrants such as the Google-MediaTek alliance, firms like TrendForce have pointed out the possibility that Broadcom's nearly 70% market share could be gradually eroded through 2028.

The Hidden Variable of Customer Quality—Circular Financing Risk

• Broadcom's six core customers are not monolithic. While Google and Meta fund their capital expenditures with abundant free cash flow generated from their core businesses (search advertising and social media advertising), the situation for Anthropic and OpenAI is different.

• Amazon has invested over $80 billion in total in Anthropic, Google has invested over $40 billion in Anthropic, and Microsoft has invested approximately $27 billion in OpenAI while securing a $250 billion Azure usage contract from them, creating a circular structure where 'investors are also customers' that is spreading across the industry. OpenAI expects to burn approximately $27 billion in cash in 2026 (projected to expand to $63 billion in 2027), while annual revenue remains at the $20 billion level, with the gap primarily filled by capital increases.

• It is reasonable to consider that the portion of Broadcom's backlog for Anthropic and OpenAI may be indirectly dependent on this circular financing, and that there is a difference in the 'quality of demand' compared to the portion for Google and Meta. Rather than feeling reassured by the $73 billion backlog figure alone, being aware of its breakdown (whether demand is funded by self-capital or by capital increases/strategic investments) is a crucial perspective for evaluating AVGO's durability.

[Broadcom] Factor 3 | Pricing Power and Quality of Earnings

• Revenue for Q2 FY2026 (announced June 3, 2026) was $22.19 billion (+48% YoY), with AI semiconductor revenue reaching $10.8 billion, a staggering 143% increase YoY. Operating margin reached a record level of 67.3% (+200bp YoY), and adjusted EBITDA margin reached 69%.

• Meanwhile, gross margin itself is under slight downward pressure due to the expansion of the AI mix, which CFO Spears explained as 'not implying a structural change in the semiconductor business.' This is a point that should be continuously monitored in the gross margin trend for each earnings report.

[Broadcom] Factor 4 | Capital Efficiency and Quality of Revenue

• Even after the VMware integration, the free cash flow margin has maintained a high level of around 40%, boasting cash-generating power comparable to software companies. Shareholder returns (dividends + share buybacks) in the most recent quarter exceeded $10 billion, and the stock price has recorded a double-digit percentage increase year-to-date (see valuation below).

• The combination of semiconductors (capital-intensive with high demand volatility) and software (stable, high gross margin) provides a level of earnings stability not seen in other AI semiconductor companies.

[Broadcom] Factor 5 | TAM and Scalability

• Broadcom's growth drivers, in addition to the steady digestion of the $73 billion backlog, include the 'AI XPU Platform' established in partnership with Apollo and Blackstone. This plan aims to deploy over 20 gigawatts of computing power by 2028, with the first phase (approximately $35 billion in scale) already underway under Apollo's leadership. Both vertical integration (expanding from design services to infrastructure operation) and horizontal expansion (acquiring new customers) are functioning effectively.

• Risk of Commoditization Trap: Given the rise of the Google-MediaTek alliance and the fact that Marvell is handling Amazon Trainium, the high-end ASIC design market could also be exposed to price competition in the future. However, at present, Broadcom's advantage in advanced packaging technology mitigates this risk to a certain extent.

[Broadcom] Factor 6 | Valuation Asymmetry

• As of August 1, 2026, Broadcom's stock price is $374.45, its market capitalization is approximately $1.78 trillion, and its trailing P/E ratio is 62.32x (based on the last 12 months of revenue of $75.47 billion). The stock price fell 5.03% on this day during a market-wide correction.

• Bull Case: The $73 billion backlog and management's guidance of 'over $100 billion in AI semiconductors alone in 2027' are factors that justify the current high P/E ratio. According to JPMorgan's estimates, given Broadcom's backlog and booking pace, this target may even be conservative.

• Bear Case / Pricing for Perfection: A P/E ratio in the 60x range carries the risk that the stock price will overreact to even 'minor mistakes' in execution. In fact, despite beating market expectations for both revenue and profit in the Q2 earnings report, the stock price saw a correction due to disappointment regarding the quality of the guidance (whether it was rack-scale or chip-only). This is a symbolic example of the fragility of a valuation that has 'perfect execution' priced in.

• Asymmetric Risk of Customer Concentration: It is estimated that the top three companies (Google, Anthropic, and Meta) account for approximately 70% of AI revenue. If two of these companies were to slow their investment pace simultaneously, the goal of exceeding $10 billion by 2027 would collapse. This concentration risk is the most important variable in judging the merits of a high P/E ratio.

Marvell (MRVL): 5+1 Framework Analysis

[Marvell] Factor 1 | Structural Moat, Classification: Type D (Differentiation through Concentration Strategy) + Partial Type C (Ecosystem Lock-in)

• Marvell avoids direct, all-out confrontation with Broadcom and has achieved a clear division of labor in its customer base. While Broadcom secures the frontier generative AI labs (Google, Meta, Anthropic, OpenAI), Marvell has two of the world's largest cloud infrastructure providers, Amazon (AWS Trainium) and Microsoft (Azure Maia), as its core partners.

• Marvell's uniqueness lies in the fact that it stands at the intersection of three technical domains: 'custom silicon design,' 'optical interconnects,' and 'data center networking.' The optical DSP (digital signal processing) technology derived from Inphi, which it acquired for $10 billion in 2021, provides a solution to the biggest bottleneck accompanying the large-scale expansion of data centers: high-speed, low-power communication between AI chips. The tailwind for 'Ethernet,' which AWS and Meta are promoting as the network standard for next-generation AI clusters, is also a positive factor for Marvell's high-speed Ethernet switch product line.

• Furthermore, in the NVLink Fusion announcement in March 2026, Marvell was listed as an official chip design partner. This means they have obtained a 'passport' to slip their own custom chips into NVIDIA's massive ecosystem, rather than being in opposition to it.

[Marvell] Factor 2 | Competitive Durability

• Accumulation of Tacit Knowledge: The first-mover advantage in optical DSP technology has been cultivated over more than five years since the Inphi acquisition, and it is difficult to replicate in a short period.

• Regulatory and Geopolitical Protection: Because the customers are US hyperscalers like AWS and Microsoft, the direct impact of export restrictions to China is limited.

• Network Effects: A plan has emerged for Amazon to sell its self-developed 'Trainium' chips to external companies beyond the AWS framework; if this is realized, the external proliferation of chips designed by Marvell could become a new pillar of royalty revenue.

• Wall of Capital x Wall of Time: Giants of smartphone chips like Qualcomm and MediaTek have begun entering the data center ASIC market, and MediaTek, in particular, is showing aggressive movement, aiming for $2 billion in AI ASIC revenue in Q4 2026 alone. The risk of price competition in design service fees is a more significant threat to Marvell, which is smaller in scale than Broadcom.

• Triggers for downgrading the durability score

  1. Moves by AWS to fully internalize Trainium design or switch to other design partners

  2. Price destruction by new entrants such as MediaTek

  3. Prolonged slump in the communications infrastructure segment (15-20% of sales)

[Marvell] Factor 3 | Pricing Power and Quality of Earnings

• Revenue for Q1 FY2027 (announced May 27, 2026) was $2.418 billion, up 28% year-over-year, with data center revenue reaching 76% of the total, recording 11% growth quarter-over-quarter. GAAP gross margin was 52.1%, non-GAAP gross margin was 58.9%, and management has provided guidance for non-GAAP gross margin of 58.25-59.25% from Q2 onwards.

• Compared to Broadcom's 77% range, the gross margin looks inferior, but this is due to differences in business mix (including networking and optical products) and scale, and does not indicate a lack of structural price competitiveness. CEO Murphy stated that 'AI-related bookings are expanding at an extraordinary pace' and has raised the full-year guidance.

[Marvell] Factor 4 | Capital Efficiency and Quality of Earnings

• According to statistics as of July 2026, Marvell's ROE is 16.03%, ROIC is 6.87%, and free cash flow (trailing 12 months) is $1.67 billion.

• Compared to Broadcom's FCF margin of over 40%, Marvell is still in a growth investment phase and should be evaluated as being in development in terms of capital efficiency. The debt-to-equity ratio is 0.29 and the current ratio is 3.28, so there are no major concerns regarding financial health itself.

[Marvell] Factor 5 | TAM and Scale Power

• Marvell's management has set a goal of $15 billion in revenue and over $5 in EPS by FY2028, nearly doubling from current levels. In the recent Q1 FY2027 earnings report, the full-year FY2027 revenue outlook was raised to approximately $11.5 billion, and a forecast was also provided that it would reach $16.5 billion by FY2028.

• Growth drivers include the start of mass production for Amazon Trainium4 (late 2026 onwards, 3x performance improvement over the previous generation), the expansion of Microsoft Azure Maia deployment, and CXL (Compute Express Link) technology derived from Xconn Technologies, which was acquired for $550 million in January 2026. CXL is expected to become a core technology for data center design from 2027 onwards as a next-generation memory-to-GPU interconnect standard.

• Horizontal expansion: They are also involved in the design of Google Axion (Arm-based CPU), and moves to reduce dependence on a single customer are also progressing.

• Note on Customer Quality

  • As mentioned above, Amazon and Microsoft fund their capital expenditures with the abundant cash flow generated from their core businesses (e-commerce, cloud, Office/Azure), which is a relative strength of Marvell when compared to Broadcom's new customers (Anthropic, OpenAI).

  • However, they are not completely unrelated; as mentioned, Amazon has invested over $80 billion in Anthropic, and part of AWS's growth (including consumption via Trainium/Bedrock) is, in some respects, indirectly benefiting from this circular funding supply.

  • Nevertheless, relative to the scale of their core businesses, AWS and Azure, this impact is limited, and it is reasonable to assess that the hierarchy of risk is one level different from the structure where Broadcom is directly dependent for a significant portion of its revenue on customers with high dependence on capital increases, such as Anthropic and OpenAI.

[Marvell] Factor 6 | Valuation Asymmetry

• Marvell's stock price has shown extremely high volatility entering 2026. After recording a significant rise of +198% at one point year-to-date, it experienced a sharp drop of about -33% in July, highlighting its extreme price movement (beta of 2.20).

• According to statistical data as of the end of July 2026, the market capitalization is approximately $153-181 billion (with variations across multiple sources, recently in the $150 billion range), the trailing P/E ratio is 60.45x, the forward P/E ratio is 38.54x, and the PEG ratio is 1.12.

• The consensus from 43 analysts is a "Strong Buy," with a bullish view prevailing that the 12-month target price is in the $250 range (an upside potential of around +50% from the recent stock price). Furthermore, its inclusion in the S&P 500 was decided in June 2026, which serves as a tailwind in terms of supply and demand due to capital inflows from index funds.

Case Scenarios

• Bull Case: Compared to Broadcom's market capitalization (approx. $1.78 trillion), Marvell's market cap is only about one-tenth of that, and this "asymmetry in corporate scale" is the greatest investment opportunity. The higher the probability of achieving the FY2028 target ($15 billion in revenue), the greater the upside potential for the stock price.

• Bear Case: The high multiple in the 60x P/E range has priced in growth expectations well in advance, and as the July plunge shows, it is a structure prone to large stock price fluctuations even with slight deterioration in sentiment. The high dependence on Amazon (a significant portion of data center revenue is via AWS) is a vulnerability that is as important as, or even more important than, Broadcom's customer concentration risk.

• Evaluation of Asymmetry: The high volatility itself falls into a case where the "range of uncertainty is wide (Unevaluable)." It is reasonable to keep the position size smaller than that of Broadcom, in a manner commensurate with this range of uncertainty.