Why Broadcom and Marvell Are The Real Winners Of The AI Chip War

Every week there's a new headline about Big Tech building chips to challenge Nvidia. Amazon has quietly built what CEO Andy Jassy now calls a $50 billion chip business. Google's TPU Ironwood powers over a billion daily AI queries. Microsoft's Maia 200 is already running GPT-5.2 across Azure Copilot. Meta is four chip generations into its own processor roadmap. These developments frame this as existential pressure on Nvidia, as if the interesting question is whether Jensen Huang loses. It isn't. Two companies are sitting at a quiet toll booth in the middle of all this custom chip activity, collecting revenue from every hyperscaler regardless of who builds the best silicon. Their names are Broadcom and Marvell, and the structural position they occupy is the actual trade.

Let's break down what the training-versus-inference split means for chip economics, why a near-total duopoly in ASIC co-design means Broadcom and Marvell profit from the hyperscaler chip race without taking competitive risk, and what the investor case looks like when the numbers are laid out clearly.

Why Nvidia May Not Be the Only AI Chip Stock to Watch

Custom ASIC server shipments are projected to reach 27.8% of the total AI server market in 2026, the highest share since 2023, according to TrendForce. Those shipments are growing at 44.6% year-over-year, nearly three times the 16.1% growth rate projected for GPU-based AI servers in the same period. The raw numbers suggest Nvidia faces real competition.

But there's a detail the competitive narrative skips: the hyperscalers aren't building these chips in isolation. They define the product requirements and own the output. The actual silicon co-design work (architecture, IP, packaging, interconnect) runs through Broadcom and Marvell. Those two companies together control an estimated 95% of the custom AI ASIC co-design market, according to Tom's Hardware analysis.

So regardless of whether Amazon, Google, Microsoft, or Meta wins the custom chip arms race, Broadcom and Marvell get paid. Every time.

Think of it like Indus Towers in India. Jio, Airtel, and Vi fight hard for subscribers, but they all need cell towers. Indus Towers earns rent from every operator, indifferent to who wins the market share war. Broadcom and Marvell occupy a nearly identical structural position. The hyperscalers' competitive spending is their revenue.

AI Training vs Inference: Why Custom Chips Are Gaining Ground

To understand why custom chips are winning specific battles without replacing Nvidia entirely, you need to understand one distinction:

Training is when you build an AI model from scratch. The workload is exploratory, computationally intense, and changes as model architectures evolve. Nvidia's GPUs, with the CUDA software ecosystem and flexible architecture, dominate here. The H100, B200, and upcoming Vera Rubin chips are built for this kind of unpredictable, wide-ranging compute. No custom ASIC makes economic sense for training, because model architectures shift too fast to justify an 18-24 month chip design cycle.

Inference is every time someone gets a response from an AI product: a Copilot suggestion, a Gemini answer, a ChatGPT output. These workloads are stable, repetitive, and running at a scale where cost-per-token is the only metric that matters.

Inference now accounts for approximately two-thirds of all AI compute cycles, up from roughly one-third in 2023, as per Deloitte's TMT Predictions. At that volume, even a modest cost advantage on custom silicon compounds into billions in annual savings. Midjourney moved its image-generation workloads from Nvidia A100/H100 clusters to Google's TPU v6e. Monthly compute costs fell from $2.1 million to under $700,000 (a 65% reduction, or approximately $17 million saved per year from one migration decision).

Sources: Deloitte TMT Predictions 2026; TrendForce; Midjourney engineering disclosures

Inference is the structural crack in Nvidia's position. Training remains largely Nvidia's territory. But inference is where the money is, and that's where it's going.

Broadcom and Marvell: The Picks-and-Shovels Play in AI Chips

Here is where the investment case gets specific.

Broadcom co-designs Google's TPU chips. It co-designs chips for Meta's MTIA program. It has six confirmed hyperscaler customers including OpenAI and Anthropic. In April 2026, Broadcom and Google formalized a supply agreement running through 2031, committing Broadcom to develop and supply custom TPUs across Google's future chip generations. A five-year supply commitment is standard in aerospace and defense. In commercial semiconductors, it is rare enough to be treated as a structural signal.

Sources: Broadcom 8-K SEC filings, March 4 and June 3, 2026; Broadcom Q2 FY2026 earnings call transcript

CEO Hock Tan on the June 3 earnings call said: "Demand for XPUs and networking is simply insatiable." Q2 bookings for AI semiconductors were over $30 billion against $10.8 billion shipped (a 3-to-1 demand-to-supply ratio). The company's gross margin stands at around 77% of revenue.

Marvell (MRVL) is the number two player in this duopoly. The company co-designs Amazon's Trainium and Inferentia processors and Microsoft's Maia accelerator. Marvell has disclosed active design engagement across more than 50 new opportunities with over 10 customers, and projects up to $11 billion in revenue for 2026.

Every dollar Google, Amazon, Meta, or Microsoft spends on custom chip development is flowing through one of these two companies. The hyperscalers' escalating capital expenditure plans are Broadcom and Marvell's forward pipeline.

Amazon Trainium Shows How Big the Custom AI Chip Market Can Become

In his April 9, 2026 shareholder letter, Andy Jassy made a disclosure that deserves more attention than it received.

Amazon's custom silicon portfolio crossed $20 billion in annual revenue run rate in Q1 2026, growing at triple-digit rates year-over-year. The portfolio covers Graviton (CPU), Trainium (AI accelerator), and Nitro (infrastructure chip). Nearly 40% sequential growth in a single quarter.

Sources: Andy Jassy 2026 shareholder letter, April 9, 2026; Amazon Q1 2026 earnings call

Jassy's calculation: if Amazon sold these chips to external customers the way Nvidia does, the business would represent approximately $50 billion in annual run rate, placing it among the top three data center chip companies globally.

The forward-looking detail is the more interesting one. Trainium4, not expected for broad availability until late 2027, already has substantial pre-orders. Companies like Anthropic and OpenAI are treating custom silicon deployment as a multi-year capital commitment on the same planning horizon they use for data center construction.

Meanwhile, Bedrock, AWS's AI model platform, now runs most of its inference on Trainium. Almost 80% of Fortune 100 companies are using Bedrock.

Custom AI ASIC Market Growth: The Numbers Investors Should Know

Sources: TrendForce Global AI Server Forecast, January 2026; Bloomberg Intelligence AI Accelerator Chips 2026 Outlook; Deloitte Technology, Media & Telecommunications Predictions 2026

The trajectory is not ambiguous. Custom ASICs are growing at nearly three times the rate of GPU-based AI servers. By 2030, TrendForce projects roughly 40% of AI server shipments will run on custom chips. That isn't displacement. Nvidia still powers the majority, especially for training. But it is a fundamental restructuring of where compute dollars flow and who intermediates that flow.

Key Risks for Broadcom and Marvell in the AI Chip Race

1. Nvidia closes the inference cost gap. Nvidia's NIM (Nvidia Inference Microservice) software stack and the Vera Rubin architecture are specifically targeting inference economics. If Nvidia delivers a generational improvement in cost-per-token for inference in the 2026-2027 product cycle, the 65% TCO advantage of custom ASICs shrinks, and the migration incentive with it. This is the most credible near-term risk to the thesis.

2. Hyperscalers build in-house design capability. Google is furthest along. Its in-house TPU engineering team is mature enough that reducing Broadcom reliance is at least theoretically possible over a multi-year timeline. If any major hyperscaler internalizes chip design work that currently flows through Broadcom or Marvell, the 95% market share figure starts to erode. The risk is gradual, not sudden as chip design expertise takes years to build, but it is real.

3. Model architecture shifts break ASIC assumptions. Custom ASICs are optimized for today's transformer-heavy inference patterns. A substantial shift toward new architectures like state-space models, novel mixture-of-experts configurations, or post-transformer paradigms, could shorten the useful life of chips designed for current workloads below the 3-5 year horizon their economics assume. The 18-24 month design cycle becomes a liability if the target moves.

4. TSMC capacity stays a shared bottleneck. Every chip in this story runs on TSMC's advanced nodes. CoWoS advanced packaging capacity is scaling to 120,000-130,000 wafers per month in 2026, but Nvidia holds roughly 60% of current allocation. A TSMC capacity crunch delays custom ASIC timelines without discriminating between programs, it's a blunt instrument that affects Broadcom's customers and Nvidia simultaneously.

5. Customer concentration risk for Broadcom. Almost half of Broadcom's Q2 FY2026 revenue came from AI semiconductors. A large portion of that flows from a small number of hyperscale customers. If a single major relationship slows materially, quarterly revenue trajectory changes sharply. Broadcom's Q3 guidance of $16 billion, while 200%+ YoY growth, landed below the sell-side consensus of $17.2 billion and the stock fell approximately 14% on the day. The market is pricing in continued acceleration; any stumble in that narrative gets repriced quickly.

How Should Investors View This

The chip war headlines will run for years. The "will Nvidia survive" narrative generates clicks, and it will keep doing so. But the investor's job is to find the structural beneficiary, not the most dramatic story.

Three points are worth holding onto.

1. Broadcom and Marvell are not competing against Nvidia. They are selling to the companies that are competing against Nvidia. This is categorically different risk exposure. Their revenue grows with custom chip adoption whether or not those chips ever displace Nvidia's training dominance.
2. The inference shift is durable. Two-thirds of AI compute is already inference, and that share grows as deployed AI product usage scales faster than new training runs. The Midjourney case isn't an outlier, it is what the math looks like for any company processing billions of inference requests at production scale.
3. Amazon's $50 billion chip signal matters beyond Amazon. If one cloud company has quietly built a chip business that would rank in the global top three, the assumption that AI chip economics are static is already obsolete. Every hyperscaler is running the same calculation, and Broadcom and Marvell are on the other side of every one of those conversations.

The most interesting trade in the AI chip war isn't picking a winner. It's identifying who collects the toll regardless of who wins.