The AI Compute Market Isn't One Market — It's Two, and the Bottleneck Decides Who Wins

The market is "starving for compute." That is the phrase Piper Sandler used this week when it initiated coverage of five semiconductor stocks — Nvidia, Broadcom, AMD, Arm Holdings, and Marvell — with its highest Overweight rating across the board. The analyst projects the AI compute market will reach $2.2 trillion by 2030 and calls Broadcom the "ASIC compute king".

The headline is right about the direction but wrong about the mechanism. The constraint holding up this market is not demand. It is supply — specifically, who controls the packaging lines at TSMC, and how much room the bottleneck leaves for everyone besides Nvidia. The five stocks Piper Sandler rated equally are not five equal beneficiaries of one market. They are participants in two markets that are growing at different rates, with different margins, and different paths to the constraint.

The Constraint Has Moved

Every AI accelerator — Nvidia's Blackwell, AMD's MI300 series, Broadcom's custom chips for Google and Amazon — must pass through TSMC's CoWoS advanced packaging before it can be shipped. TSMC's own leadership has said CoWoS capacity is approximately three times short of demand. Lead times run 52 to 78 weeks. The foundry wafers are booked through 2028. High-bandwidth memory, which must be packaged alongside the logic dies, is similarly allocated. SK Hynix's 2026 HBM output is already allocated.

The bottleneck is no longer whether a company can design a competitive chip. It is whether it can get one packaged.

And the packaging slots are not distributed evenly. Nvidia controls an estimated 60% of TSMC's CoWoS allocation. AMD holds roughly 11%. Broadcom's hyperscaler ASIC customers account for about 20%. Everyone else shares the remainder.

This allocation share is not a function of performance. AMD's chips are competitive on price and inference workloads. But 11% of the packaging pipeline is a production ceiling, not a demand problem. That is why AMD holds only 6% to 8% of AI accelerator revenue despite having products that many engineers consider equivalent for specific use cases.

Two Markets, Not One

The AI compute market has split into two businesses with different growth rates, different economics, and different competitive dynamics.

The merchant GPU market — where Nvidia and AMD sell standardized chips — is growing at approximately 16.1% year-over-year in merchant GPU shipments. Nvidia dominates this market, but the total addressable pool is constrained by the CoWoS slots it can secure and the power available in data centers.

The custom ASIC market — where Broadcom designs purpose-built chips for Google, Amazon, Microsoft, and Meta — is growing at approximately 44.6% year-over-year. Broadcom holds an estimated 60% share of this segment by 2027. The ASIC backlog is reported at $73 billion. Hyperscalers are prioritizing custom silicon to bypass merchant GPU allocation queues and to match their specific inference workloads.

These are not substitute products. They are adjacent markets. Nvidia dominates training; ASICs are increasingly handling inference. The hyperscalers are not choosing between Nvidia and Broadcom. They are buying both, for different stages of the AI pipeline.

The implication for investors is that you are not buying into one "compute" market. You are buying into a specific position within a specific supply allocation.

Nvidia Is Not the Only One, but It Is Different

Piper Sandler assigned the same Overweight rating to all five names. But the five companies have fundamentally different profit structures.

Nvidia reported $96.2 billion in revenue for the second quarter of fiscal 2027, with 75% gross margins and 64% operating margins on a trailing twelve-month basis. Its return on invested capital is 87%. Free cash flow over the past twelve months is $127 billion. It generates more cash than most countries collect in taxes.

Compare that with the rest of the group:

• AMD generates 12% operating margins and 8% ROIC, with $8.4 billion in trailing free cash flow.
• Marvell produces 17% operating margins and 6% ROIC, with $1.7 billion in trailing free cash flow.
• Arm Holdings has 18% operating margins and 8% ROIC, with $1.5 billion in trailing free cash flow.

Nvidia does not just have market share. It has margin depth that no competitor can match at this stage. A 64% operating margin on $385 billion in annual revenue produces cash that compounds the advantage. The other four companies combined generate a fraction of that cash flow.

That is not an argument against the other four. It is an argument that Nvidia is in a different category than the other four, even within the "compute" bucket.

The Valuation Spread Is the Real Story

If all five companies are buying into the same "starving for compute" market, why do they trade at wildly different multiples?

Company	P/E (TTM)	Operating Margin	ROIC
Nvidia	27x	64%	87%
Broadcom	45x	N/A	N/A
AMD	128x	12%	8%
Marvell	75x	17%	6%
Arm	260x	18%	8%

Nvidia trades at 27 times trailing earnings. AMD trades at 128 times. Arm trades at 260 times.

These spreads are not arbitrary. They reflect the gap between current profit quality and future profit expectations. Nvidia is already earning at margins most semiconductor companies can only imagine. AMD and Arm are pricing in a future where they capture significantly more of the AI compute market than they currently do — despite holding a fraction of the packaging capacity and a fraction of the profit rate.

Arm is the most extreme case. It earned $7.6 billion in revenue over the past twelve months, growing 25% year-over-year, but shrank 13% quarter-over-quarter in the most recent period. At 260 times earnings, the market is pricing in years of continued high growth, a dominant position in AI chip architecture, and zero execution risk. The company holds no manufacturing capacity, controls no packaging slots, and licenses its designs to the same foundries that allocate the bottleneck.

What Actually Changes the Thesis

The "starving for compute" narrative treats supply as an abstract shortage. In reality, it is a specific allocation at a specific facility, controlled by one company — TSMC — that decides who gets what.

The structural question for all five of these stocks is not whether demand for compute will grow. It is whether the packaging bottleneck relaxes faster than demand expands, and whether the allocation shifts away from Nvidia.

If TSMC brings CoWoS capacity online as planned — the company has been scaling from roughly 35,000 wafers per month in late 2024 toward 130,000 — the constraint eases. AMD gets more slots. Broadcom gets more slots. The market looks more like a broad growth story and less like an allocation game.

If the constraint persists, Nvidia's 60% share of packaging capacity is a durable moat. AMD remains structurally capped. Broadcom's ASIC growth becomes the faster-growing alternative, but it depends on hyperscaler relationships that can shift. Arm's premium valuation requires execution that the company has yet to demonstrate at this scale.

The key issue is not whether the AI compute market is growing. It is. The more important question is whether you are buying the company that controls the constraint, or the companies that are waiting in line for it.