Why Did Marvell's Stock Surge? A Deep Dive into Google's Custom AI Chips and the $20 Billion Revenue Target
Overview
Marvell Technology raised its fiscal 2028 revenue target to approximately $20 billion at its investor day on October 6, above the prior market consensus. The stock closed at $287.01 that day, up 5.81%, with intraday gains at one point approaching 10%, bringing the company's market capitalization back to about $251.7 billion, according to data from StockAnalysis's MRVL quote page. On the same day, Broadcom shares rose about 4% in sympathy.
What truly drove the re-rating of expectations was not just that single figure. The company also raised its fiscal 2029 custom chip business revenue target from $10 billion to more than $12 billion, and provided a fiscal 2031 revenue range of $70 billion to $90 billion. Against analysts' prior forecast of approximately $46.85 billion for fiscal 2031, this set of targets implies that management believes orders from hyperscalers' in-house AI chips are far larger than what sell-side models had assumed. Google is the most critical customer in this logic chain, and the agreement the two companies signed at the end of July has already tied order scale directly to equity through a warrant.
Key Takeaways
The magnitude of the target jump far exceeds the expectation gap. According to Reuters, the $20 billion fiscal 2028 target is above the Wall Street consensus of $18.2 billion, while the fiscal 2031 range of $70 billion to $90 billion has a midpoint roughly 1.7 times analysts' forecasts.
The custom chip target was raised by 20%. Based on Investing.com's summary of the investor day, the fiscal 2029 custom chip revenue target was raised from $10 billion to more than $12 billion, with a fiscal 2031 earnings per share target set at more than $30.
Interconnect is larger than custom chips. In the midpoint breakdown of $80 billion for fiscal 2031, interconnect accounts for approximately $37.5 billion, custom compute about $30 billion, and switching and storage about $10 billion. This shows that Marvell's AI narrative is not just about ASICs—connectivity and optical interconnect are the largest single segment.
The $12 billion figure in the Google agreement has a clear source. According to Marvell's 8-K filing with the U.S. Securities and Exchange Commission, the warrant vests in tranches of one for every $500 million in custom product revenue, for a total of 240 tranches, corresponding exactly to the sales potential of up to approximately $120 billion by fiscal 2033 that Reuters mentioned.
The scale gap with Broadcom remains enormous. According to Broadcom's third quarter earnings release filed with regulators, its quarterly AI semiconductor revenue has already reached $16.7 billion, higher than the full-year fiscal 2029 custom chip business target Marvell has set.
Customer concentration is the other side of this narrative. In its fiscal 2026 annual report, Marvell disclosed that a single direct customer contributed 14% of net revenue, and a single distributor contributed 37%, with the two together accounting for more than half.
The Numbers from the Investor Day That Actually Changed Expectations
The target jump from fiscal 2028 to fiscal 2031
Marvell's investor day was held in New York, in accordance with the conference notice published in early August. The target framework presented at the event included several layers: fiscal 2028 revenue of approximately $20 billion, of which data center business accounts for about $18 billion; fiscal 2029 custom chip revenue of more than $12 billion; and fiscal 2031 revenue of $70 billion to $90 billion, with earnings per share above $30. CEO Matt Murphy also raised the company's total addressable market for 2030 to approximately $400 billion.
These figures need to be understood against the current scale. According to Marvell's second quarter earnings filing with regulators, revenue for the quarter ended August 1, 2026 was $2.739 billion, a record high, up 37% year-over-year, of which data center business was $2.172 billion, accounting for 79%, up 46% year-over-year. Third quarter guidance is $3.15 billion, plus or minus 5%. On an annualized basis, the current run rate is roughly around $12 billion. To reach $20 billion in fiscal 2028, it would need to sustain nearly double-digit growth over the next two years, while the fiscal 2031 midpoint target would mean quadrupling from that base.
Targets alone do not constitute performance, but they change the valuation anchor. The market had previously priced Marvell based on a fiscal 2031 revenue assumption of approximately $46.85 billion. Management's midpoint is $80 billion, and the earnings per share expectation was raised from about $19 to more than $30. The immediate stock price reaction was essentially the market's first round of discounting this new anchor.
Market reaction and the linkage with Broadcom
The intraday price action that day was not one-directional. According to Investing.com's record, the stock fell 3% at the start of the event, then rose as much as 10%, and closed up 5.90%. StockAnalysis's closing data showed $287.01, up 5.81%, with a 52-week range of $70.69 to $329.88. The width of that range itself shows that over the past year the market has undergone a dramatic reassessment of the company's positioning.
More noteworthy is the spillover effect. Reuters reported that Broadcom rose about 4% that day. The two companies are not competing for the same order, but they share the same assumption: hyperscalers will continue to hand more and more AI compute to custom chips rather than general-purpose GPUs. Marvell quantified this assumption into a five-year target, and the market promptly applied the same logic to Broadcom.
Why Custom Chips Suddenly Became the Main Story
The division of labor between ASICs and general-purpose GPUs
Custom chips in the industry usually refer to application-specific integrated circuits, or ASICs, jointly developed by chip design companies and cloud providers, serving only specific workloads for specific customers. Such chips sacrifice generality in exchange for higher performance per watt and lower unit cost on the target task. Google's Tensor Processing Unit is among the most mature examples. According to Google Cloud's official release notes, the TPU7x in the seventh-generation Ironwood family entered preview on November 24, 2025, and became generally available on March 31, 2026, targeting training and inference for large language models, mixture-of-experts models, and diffusion models.
What needs to be clarified is division of labor, not replacement. The value of general-purpose GPUs lies in adaptability—model architectures, frameworks, and operators change every few months, while an ASIC typically takes one to two years from project initiation to mass production. For training tasks with frequent model iterations, general-purpose GPUs remain the lowest-risk choice; for inference tasks whose architectures have stabilized and require long-term, large-scale repetitive execution, the cost advantage of custom chips can be fully realized. This is also why Google continues to purchase Nvidia products while expanding its in-house chip deployment.
Inference workloads forced a recalculation of the cost equation
The core variable driving this round of customization is the scale of inference demand. Nvidia reported revenue of $96.2 billion in its second quarter earnings release for the period ended July 26, 2026, of which data center business was $89 billion, up 117% year-over-year. CEO Jensen Huang described the current phase in the release as compute equaling revenue, with demand still accelerating. When inference call volumes reach this magnitude, every percentage point of unit cost difference is amplified into billions of dollars in spending differences, and the return on investment for in-house chips follows.
The direction of capital expenditure corroborates this. On Alphabet's second quarter earnings call, CFO Anat Ashkenazi said the company was raising its full-year 2026 capital expenditure guidance from $180 billion to $190 billion to $195 billion to $205 billion, with roughly 60% going to servers and 40% to data centers and networking equipment, as detailed in Investing.com's transcript of the call. Management also noted that revenue from external sales of TPU systems has entered the cloud business backlog, but most will be recognized in 2027. For chip suppliers, this is both a source of order visibility and the reason targets are set for fiscal 2028 rather than the present.
What Exactly Is Bound Between Google and Marvell
The warrant ties revenue to equity
What is truly worth reading closely is that 8-K filing. According to the disclosure, Marvell and Google entered into a commercial collaboration on July 29, 2026, covering the development of custom semiconductor products, including AI inference accelerators, storage controllers, and network interface controllers. As part of the agreement, Marvell granted Google a warrant for 58,970,907 shares of common stock at an exercise price of $206.58, expiring on August 18, 2033.
The vesting method is divided into two parts. Of these, 1,360,867 shares vest on a time basis, in quarterly installments over the first year. The remainder is tied to revenue, divided into 240 equal tranches, from the third quarter of fiscal 2027 to fiscal 2033, with one tranche vesting for every $500 million increase in custom product revenue.
Where the $120 billion figure comes from
A simple calculation of the vesting terms makes the source of the so-called sales potential clear. 240 tranches multiplied by the $500 million revenue threshold per tranche equals exactly $120 billion, which fully matches Reuters's statement that, if all performance conditions are met, the collaboration could generate up to approximately $120 billion in sales by fiscal 2033. In other words, this figure is not an order amount but the cumulative revenue scale required for the warrant to fully vest—a ceiling assumption, not a commitment.
The implications of this structure for investors are two-way. On one hand, it binds customer interests to the supplier's stock price, reducing the probability of orders being diverted; on the other hand, the higher the revenue, the more shareholders are diluted. Based on the October 6 closing price of $287.01 and the exercise price of $206.58, the intrinsic value of approximately 240,000 shares per tranche is close to $19 million, equivalent to about 4% of the $500 million in revenue needed to trigger that tranche. This cost does not appear in gross margin, but it genuinely exists in the denominator of earnings per share.
Where Marvell, Broadcom, and Nvidia Each Stand
The scale gap with Broadcom
On the custom chip track, Broadcom remains far ahead. According to its third quarter earnings release for the period ended August 2, 2026, quarterly revenue was $29.6 billion, up 86% year-over-year, of which AI semiconductor revenue was $16.7 billion, up 221% year-over-year and 54% quarter-over-quarter. The company's fourth quarter guidance is total revenue of $34.8 billion and AI semiconductor revenue of $21.7 billion. CEO Hock Tan said in the release that demand for custom AI accelerators and networking products remains strong.
Placing these figures alongside Marvell's targets makes the magnitude difference immediately clear: Broadcom's AI semiconductor revenue in a single quarter already exceeds the $12 billion target Marvell set for its full-year fiscal 2029 custom chip business. Therefore, the investor day target framework is less a challenge to Broadcom than a declaration that Marvell has transformed from a supplier to a single customer into a second source covering multiple hyperscalers.
Is Nvidia facing replacement or diversion?
Understanding custom chips as a replacement for general-purpose GPUs is the most common misreading of this theme. Nvidia's single-quarter data center revenue of $89 billion is higher than Marvell's fiscal 2031 full-year revenue midpoint of $80 billion, and the latter also includes non-AI compute businesses such as interconnect, switching, and storage. Even if Marvell and Broadcom's targets are fully realized, custom chips are taking a portion of incremental budgets, not the main body of the existing market.
A more realistic description is diversion. The backbone of training clusters is still composed of general-purpose GPUs, while custom chips are taking an increasingly large share on the inference side and for specific model architectures. And between these two types of compute, connectivity is a shared necessity. Marvell placing its largest fiscal 2031 revenue segment in interconnect rather than custom compute is itself a statement of this judgment: regardless of who provides the compute, data must flow between racks and between clusters, and demand for optical interconnect and network interfaces grows in the same direction as total compute, largely independent of who wins the chip orders.
The AI chip story has never been about a single stock. Track price changes across the upstream and downstream in the U.S. equity sector as well
Risks, Scenarios, and Follow-Up Watch Points
Customer concentration and warrant dilution
In its fiscal 2026 annual report, Marvell disclosed that in the fiscal year ended January 31, 2026, one direct customer contributed 14% of net revenue and one distributor contributed 37%. The company explicitly stated in its risk factors that the risk of relying on a small number of customers for most of its revenue is increasing, with major customer concentration continuing to rise, while sales are highly concentrated in the data center end market, and the gain or loss of key customer design wins would have a material impact. For a company building a five-year target on orders from a handful of hyperscalers, the weight of this risk factor is above normal levels.
Warrant dilution is the second layer. The revenue-based vesting design means that the more successful the business, the more certain the dilution. The third layer is execution risk: the cycle from design win to mass production for custom chips is measured in years, and any delay in process nodes, packaging capacity, or the optical interconnect supply chain would push back the target timeline. The fourth layer is the assumption itself: the $400 billion total addressable market and the fiscal 2031 target range are both built on the premise that AI capital expenditure will continue to grow rapidly, and that premise depends on whether cloud providers' return on investment materializes.
Three scenarios
In the scenario where targets are met, custom products from Google and other hyperscalers ramp up as planned, warrant tranches vest steadily, the $20 billion fiscal 2028 figure becomes a verifiable interim checkpoint, and the market gradually shifts its valuation anchor from current expectations toward management's targets.
In the scenario where the pace is delayed, orders materialize but mass production timing shifts back by one to two quarters, changing the shape of the revenue curve while keeping the endpoint unchanged. The damage to the stock price in this case is typically concentrated in a single quarter's guidance
