TL;DR · 30-second read
The Short Version
Broadcom, a company that designs specialized computer chips, lost about $191 billion in stock market value in a single day. That is roughly one-eighth of the whole company. The drop followed a news report that Google had scaled back its plans for chips Broadcom builds for it.
These chips are made for Google alone and run its artificial intelligence services. If Google wants fewer, Broadcom cannot simply sell them to someone else.
The lesson: some of the artificial intelligence boom’s biggest winners depend on decisions made inside a few giant tech companies.
Broadcom shares fell 12%, erasing about $191 billion in market value in a single session, after a report that Google had cut its plans for Broadcom-built AI chips, 24/7 Wall St. reported. Barron’s described the report as Google cutting AI chip shipments. Neither company has disclosed the size, timing or chip generation involved.
The reported change touches Broadcom’s most visible custom-silicon relationship: an agreement disclosed in April 2026 under which Broadcom develops and supplies future generations of Google’s tensor processing units (TPUs), Google’s in-house AI chips, along with networking and other components for Google’s next-generation AI racks, through up to 2031.
Executive Summary
The reported cut matters most for what the market’s reaction revealed. Broadcom’s custom-silicon business is built on long-term design relationships with hyperscalers, the few companies that run cloud data centers at enormous scale, and its Google agreement gives that business a tangible commercial basis. A one-day loss of about $191 billion, on a report that quantified nothing, shows how much of Broadcom’s valuation investors attach to one customer’s deployment roadmap.
The episode also shows a structural feature of custom AI chips. Long-dated agreements secure a supplier’s design role, but volumes follow the customer’s capital spending and model plans. Marvell, which signed its own Google agreement in August 2026 with warrants vesting per $500 million of custom revenue, works under the same logic. So do the memory and networking suppliers whose sales depend on how many of those chips get built.
Why One Customer’s Reported Change Moved $191 Billion
A custom AI chip is known in the industry as an ASIC, short for application-specific integrated circuit. It is designed around a single customer’s workloads, software and data center racks. That is its strength: it can run a hyperscaler’s stable, high-volume jobs more cheaply and with less energy than a general-purpose processor. It is also why its revenue is tied so closely to one buyer. A Google tensor processing unit is built for Google’s systems. If Google deploys fewer of a given generation, Broadcom cannot redirect that production to another cloud operator.
That relationship has anchored Broadcom’s custom-silicon story. In April 2026 the company disclosed a long-term agreement to develop and supply future TPU generations, plus networking and other components for Google’s next-generation AI racks, through up to 2031. Investors had reasonably treated that as a multi-year revenue stream. A 12% drop worth about $191 billion implies a market value of roughly $1.6 trillion beforehand. The decline is the market repricing that stream on a report about Google’s plans, before any detail of size or timing emerged.
The selloff makes the operating reality visible: in custom silicon, the supplier’s revenue curve is the customer’s deployment roadmap. Google decides which generation it deploys, when, and in what volume. Customer concentration sits alongside expensive design programs and pricing pressure from increasingly sophisticated cloud buyers among the recognized risks of Broadcom’s model. The share move shows how heavily investors weight that risk once a single program is in question.
A Long-Term Agreement Is Not a Volume Guarantee
The Broadcom–Google agreement establishes what Broadcom will design and supply. Broadcom has not publicly disclosed minimum purchase volumes under it. Multi-year custom-chip agreements generally secure the design role across product generations. The number of chips actually ordered follows the customer’s capital spending, model plans and alternatives. A report that the volume side has changed can therefore move the stock even if the agreement itself stays in force.
Marvell’s Google arrangement writes that distinction into the contract. Its August 2026 agreement covers custom silicon tied to the TPU ecosystem, including inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory computing, and it includes a purchase-linked warrant. A warrant is a right to buy a company’s shares at a preset price. Here, most of the shares vest in tranches tied to cumulative custom-product revenue, one tranche for each $500 million. The structure pays out only as revenue is realized, and by design it does not guarantee that the contemplated revenue will materialize.
For anyone modeling these suppliers, the metrics that matter are the ones the selloff turned on: design wins, volume ramps, margins, and whether each new product generation is renewed. The headline length of an agreement matters far less.
Marvell Sits on the Same Fault Line
A reported pullback from Broadcom could easily be read as a gain for Marvell, which competes for content across Google’s AI systems. Nothing disclosed so far supports that reading. Neither company has said whether any reduced Broadcom volume moves to another supplier, shifts to a later chip generation, or drops out of Google’s plans altogether.
Marvell is clearly preparing for scale. In an 8-K filed November 12, it disclosed that its board had grown from seven to ten directors. Lam Research CFO Douglas Bettinger, who was CFO of Avago Technologies (now Broadcom) from 2008 to 2013, joined the audit committee. Walmart Chief People Officer Donna Morris and Intuit Chairman and CEO Sasan Goodarzi joined the compensation committee. Announcing the appointments, Chairman and CEO Matt Murphy said “Marvell is entering an extraordinary period of growth as demand for AI infrastructure continues to accelerate.”
The same concentration logic applies to that growth. Marvell’s warrant tranches vest only as Google-linked custom revenue accrues, so its exposure to a single hyperscaler’s roadmap is written into the contract. Broadcom’s $191 billion day shows how the market prices that kind of exposure when the roadmap comes into doubt.
Downstream: Memory and Networking Feel Roadmap Shifts Differently
Custom chip volumes also matter to memory makers. High-bandwidth memory (HBM) is stacked DRAM placed beside the processor so it can be fed data quickly. An illustrative 2025 model from The Information Network gives 7.2 million merchant GPUs eight HBM stacks each and 1.5 million custom ASICs three stacks each. That works out to 57.6 million stacks for GPUs versus 4.5 million for ASICs. On those assumptions, a change in custom chip volume moves memory demand far less per unit than a change in GPU volume.
That gap may narrow. In the same model’s sensitivity cases, doubling ASIC units to 3 million raises total HBM stack demand by 7.2% at three stacks per chip, 12.1% at four and 21.7% at six. These are calculations, not forecasts. They still show that as custom accelerators carry more memory, a hyperscaler’s choices about custom chip volume matter more to memory suppliers.
Networking is the other exposure. Broadcom’s April agreement bundled networking and rack components with the accelerators, and its switching business benefits when ASIC deployments increase the bandwidth needed between chips. How far the reported change travels through Broadcom’s business depends on whether it touches only the accelerators or the rack-level networking too. Meanwhile, Nvidia GPUs continue to handle research, new model development and demanding inference, and GPUs and custom ASICs coexist in the same data centers. A shift in one custom program is not, on its own, evidence of less AI computing overall.
Background
Hyperscalers, the handful of companies such as Google that operate cloud data centers at enormous scale, began developing their own AI processors years ago because general-purpose chips become expensive at their volumes. Google’s TPU is a prominent example. Merchant GPUs from Nvidia and AMD are off-the-shelf processors any customer can buy. They stay dominant in many large-model deployments because they are flexible and supported by mature software ecosystems.
Broadcom is a leading supplier of custom AI processors and of the high-speed networking chips that connect thousands of them, and those two businesses reinforce each other. Marvell Technology competes for content across the AI system, including accelerators, storage and network controllers, and connectivity. It signed a Google agreement in August 2026 covering custom silicon tied to the TPU ecosystem. Source: Data Center AI Chips Could Change the Winners: Broadcom and Marvell, 24/7 Wall St.’s analysis of how hyperscalers’ custom AI chips reshape revenue for Broadcom, Marvell and memory suppliers. Related coverage: Broadcom Stock Plunges 12% and Loses $191 Billion After a Report That Google Cut Its Plans for Broadcom AI Chips (24/7 Wall St.) and Broadcom Stock Falls on Report Google is Cutting AI Chip Shipments (Barron’s). Primary sources: Marvell Technology Form 8-K filed November 12, 2026; Marvell Appoints Three New Independent Directors to its Board (Exhibit 99.1 to 8-K, November 12, 2026).Sources

