Buy Crypto Markets Spot FuturesBTC Earn Event Centre Rewards Hub

Broadcom reported fiscal Q3 2026 revenue of $29.59 billion, up 86% year over year, while AI semiconductor revenue surged 221% to $16.7 billion. The company expects AI semiconductor revenue to acceleraBroadcom reported fiscal Q3 2026 revenue of $29.59 billion, up 86% year over year, while AI semiconductor revenue surged 221% to $16.7 billion. The company expects AI semiconductor revenue to accelera

Broadcom Earnings: AI Chip Revenue Triples as the Custom Silicon Race With Marvell Heats Up

#US Stocks

AI$0.01944+1.25%

CHIP$0.05415+30.51%

TRADE$0.03138-2.18%

Author: James Mitchell

2026/09/03 15:29

Key Takeaways

Broadcom’s Q3 AI semiconductor revenue surged 221% to $16.7 billion, confirming strong demand for custom AI accelerators and networking. Marvell’s expanding Google partnership, however, shows that the market is becoming more competitive as attention shifts from design wins to revenue execution.

Broadcom reported fiscal Q3 2026 revenue of $29.59 billion, up 86% year over year, while AI semiconductor revenue surged 221% to $16.7 billion. The company expects AI semiconductor revenue to accelerate again to approximately $21.7 billion in Q4, reinforcing the rapid expansion of custom AI accelerators and networking inside hyperscale data centers.

The results arrive only weeks after Marvell disclosed an expanded custom silicon partnership with Google covering products attached to the TPU ecosystem. Taken together, the two developments point to a new phase in the custom AI chip market: demand remains exceptionally strong, but the investment debate is shifting toward competition, deployment timing and how quickly large design programs translate into recognized revenue.

What to Know

Broadcom’s fiscal Q3 revenue rose 86% year over year to $29.59 billion, while non-GAAP EPS reached $3.32.
AI semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially.
Broadcom expects Q4 AI semiconductor revenue of approximately $21.7 billion, representing 236% year-over-year growth.
Broadcom and Google have a long-term agreement covering future generations of Google TPUs, plus networking and other components for next-generation AI racks through up to 2031.
Marvell separately expanded its Google partnership across AI inference accelerators, network interface controllers, memory interface controllers, storage controllers and near-memory compute attached to the TPU ecosystem.
Google’s Marvell warrant contains 240 purchase-linked tranches, with each tranche tied to $500 million of Custom Products revenue. The implied $120 billion figure is therefore a maximum cumulative purchase threshold for full performance-based vesting, not a committed $120 billion Google order.
The key question is increasingly not whether hyperscalers want custom AI silicon, but which suppliers can convert large design programs into sustained revenue while capturing more semiconductor content around each AI system.

Broadcom’s AI Business Is Reaching a New Scale

Broadcom’s latest earnings provide one of the clearest signals yet that custom AI accelerators are becoming a major part of the AI infrastructure buildout. Fiscal Q3 revenue reached $29.59 billion, up 86% from $15.95 billion a year earlier, while Semiconductor Solutions revenue more than doubled to $20.84 billion. The strongest growth came from AI, where semiconductor revenue reached $16.7 billion, up 221% year over year and 54% sequentially.

The trajectory is still accelerating. CEO Hock Tan said demand for both custom AI accelerators and networking remains very strong, and Broadcom expects AI semiconductor revenue to reach approximately $21.7 billion in Q4, up 236% year over year. That would represent another roughly $5 billion of sequential AI revenue growth in a single quarter. Broadcom also guided for approximately $34.8 billion in total Q4 revenue, representing 93% year-over-year growth. Broadcom’s official Q3 FY2026 results

The scale matters because Broadcom occupies a different position in the AI semiconductor market from Nvidia. Nvidia primarily provides broadly programmable accelerated computing through its GPU platform, while Broadcom works with large cloud and technology companies to develop custom accelerators optimized around specific workloads and also supplies networking components required to connect increasingly large AI clusters. The two models are not mutually exclusive; hyperscalers are increasingly building infrastructure around a combination of general-purpose GPUs, custom accelerators and specialized networking.

Broadcom’s latest numbers therefore do more than confirm another strong quarter. They suggest that custom compute is becoming a rapidly scaling second lane of AI infrastructure spending rather than a niche alternative to GPUs.

Google Shows Why Custom AI Chips Are Becoming a Bigger Market

Google provides one of the clearest examples of how this architecture is developing. In April 2026, Broadcom disclosed that it had entered into a long-term agreement with Google to develop and supply custom Tensor Processing Units, or TPUs, for future generations. The companies also entered into a Supply Assurance Agreement covering networking and other components used in Google’s next-generation AI racks through up to 2031. Broadcom’s April 2026 SEC filing

That agreement shows why the economics of custom AI infrastructure extend well beyond the accelerator itself. At hyperscale, thousands of processors must communicate with each other, access memory and move enormous amounts of data across increasingly large clusters. As a result, networking and connectivity silicon become part of the performance equation rather than merely supporting infrastructure.

Broadcom is positioned on both sides of that expansion. It can benefit from the custom accelerator and from parts of the infrastructure required to connect those accelerators, which makes the company’s AI exposure broader than a simple custom-chip story. This fits into the larger shift discussed in MEXC’s analysis of why AI is becoming a capital expenditure cycle: spending is spreading from individual processors into data centers, networking, memory, power and the physical systems needed to deploy AI at scale.

Google, however, is also demonstrating why that opportunity is likely to become increasingly competitive.

Marvell’s Google Deal Changes the Competitive Picture

On August 19, Marvell disclosed an expanded commercial relationship with Google covering a much broader set of products than the phrase “AI chip deal” might suggest. According to Marvell’s Form 8-K, the partnership spans custom silicon programs attached to the TPU ecosystem, including AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute. Marvell’s Google agreement in its SEC filing

That breadth matters because it shows where competition is moving. Broadcom and Marvell are not simply competing over a single accelerator design; both have opportunities to capture semiconductor content surrounding increasingly complex hyperscale AI systems. As AI clusters expand, interfaces, networking, storage and memory connectivity can become strategically important businesses alongside custom compute itself.

The commercial structure also needs to be interpreted carefully. Google received a warrant to purchase up to 58,970,907 Marvell shares at an exercise price of $206.58. A relatively small portion of those shares vests over time, while most of the warrant is tied to purchases by Google and its affiliates. The performance-based portion consists of 240 equal tranches, with one tranche vesting for every $500 million in Custom Products revenue generated during the specified period through Marvell’s fiscal 2033.

Multiplying 240 tranches by $500 million produces the widely cited $120 billion figure, but that does not mean Google placed a guaranteed $120 billion order with Marvell. It represents the cumulative Custom Products revenue level required for all 240 purchase-linked warrant tranches to vest. This distinction is important because the deal is much better understood as a performance-linked framework around a potentially very large commercial relationship than as a fixed $120 billion contract.

For additional context on the original market reaction and deal mechanics, MEXC previously examined why MRVL jumped after the expanded Google AI chip agreement.

Broadcom vs. Marvell Is Not Simply a Fight Over Who Makes Google’s TPU

The two Google agreements make it tempting to frame Broadcom and Marvell as competitors fighting over the same TPU contract, but the primary-source disclosures support a more nuanced interpretation. Broadcom’s April filing explicitly covers future generations of Google TPUs as well as networking and other components through up to 2031, while Marvell’s August filing describes a broader collection of custom silicon programs that attach to the TPU ecosystem.

Nothing in those disclosures establishes that Marvell has simply replaced Broadcom at Google. A more plausible reading is that Google is building an increasingly large AI silicon ecosystem in which multiple semiconductor partners can serve different products, functions and generations. As the amount of silicon required around each AI cluster expands, the relevant competitive question becomes less about one supplier “winning Google” and more about how much semiconductor content each supplier can capture around Google’s infrastructure.

That distinction changes how Broadcom and Marvell should be compared. Broadcom already operates at enormous scale in custom accelerators and networking, giving it an established base of deployed revenue and customer relationships. Marvell is smaller but is extending its position across custom silicon, connectivity and interfaces, creating the potential to capture a larger share of future hyperscaler infrastructure programs.

The arrival of additional suppliers can therefore be both validating and competitive. Marvell’s expansion into the TPU ecosystem provides evidence that the custom silicon opportunity is becoming large enough to support a broader supplier base, but it also means Broadcom cannot assume that every incremental dollar of hyperscaler custom-chip spending will automatically flow to the incumbent.

Marvell’s Latest Earnings Show the Other Side of the Custom Silicon Boom

Marvell’s fiscal Q2 2027 results provide another useful piece of the same industry picture. The company reported record revenue of $2.739 billion, up 37% year over year, while Data Center revenue growth accelerated to 46%. CEO Matt Murphy said AI-related bookings remained exceptionally robust and that Marvell expected a significant acceleration in its Custom business beginning in the second half of fiscal 2027. The company also raised its fiscal 2027 and fiscal 2028 revenue outlooks. Marvell’s official fiscal Q2 2027 results

Those results independently reinforce the same broad signal coming from Broadcom: hyperscaler demand for custom AI silicon and connectivity remains strong. The difference is the maturity of the revenue base. Broadcom is already generating tens of billions of dollars in quarterly AI semiconductor revenue, whereas Marvell is starting from a much smaller base while adding large custom programs that could materially reshape its future business if they move successfully from development into volume production.

That makes revenue timing especially important for Marvell. A major design win establishes an opportunity, but production schedules, customer deployment and actual purchases determine when and how much of that opportunity reaches the income statement. This was already the central issue identified ahead of the quarter in MEXC’s Marvell earnings preview on custom AI chips and 1.6T optics: the most important question was not simply whether Marvell could win additional AI programs, but how those programs would convert into recognized revenue.

Broadcom and Marvell therefore present two different stages of the same custom silicon cycle. Broadcom has already reached scale and must demonstrate that it can sustain and defend it; Marvell has accumulated increasingly important opportunities and must demonstrate that they can become scale.

The Custom AI Chip Trade Is Moving From Design Wins to Execution

This may be the most important conclusion from the recent Broadcom and Marvell developments. The first phase of the custom AI silicon story was largely about validation: investors needed evidence that hyperscalers would seriously invest in proprietary accelerators even though Nvidia already offered a powerful and flexible GPU ecosystem. That question is becoming easier to answer as Google continues developing TPUs, Broadcom reports AI semiconductor growth above 200%, and Marvell reports strong AI bookings alongside an expanding custom pipeline.

The next phase is more demanding because the market now has to distinguish between design wins and financial execution. A large program needs to move through development, manufacturing, data-center deployment and customer purchasing before it becomes meaningful reported revenue. Investors also have to evaluate how much semiconductor content a supplier captures around each accelerator, whether manufacturing capacity can support the intended ramp and how aggressively hyperscalers diversify their suppliers.

This changes the investment debate around both companies. Broadcom must demonstrate that it can maintain large existing customer relationships while continuing to expand custom accelerator and networking revenue. Marvell must prove that its growing collection of hyperscaler programs can translate into material revenue at the pace implied by market expectations.

Put differently, Broadcom needs to defend scale; Marvell needs to convert opportunity into scale. That is a more useful way to understand the next stage of the custom AI chip race than simply counting design-win announcements.

Nvidia Still Matters, but Custom Silicon Is Becoming a Second Major AI Compute Lane

None of this means custom silicon is replacing Nvidia. That interpretation would oversimplify how hyperscalers are actually building AI infrastructure. Nvidia remains the dominant platform for broadly programmable accelerated computing, supported by its GPU architecture, networking products and software ecosystem, while custom accelerators address a different economic problem.

For companies operating AI infrastructure at enormous scale, silicon optimized around specific internal workloads can offer advantages in performance, power efficiency and total system economics. The result is increasingly a coexistence model in which hyperscalers can deploy Nvidia GPUs for flexible general-purpose accelerated computing alongside custom XPUs and ASICs designed for specific workloads.

Broadcom’s earnings show how quickly that second lane is scaling, while Marvell’s expanding Google relationship shows that the lane is becoming more competitive. The larger semiconductor story is therefore not that spending is simply moving from Nvidia to Broadcom or from Broadcom to Marvell. Instead, the AI infrastructure stack itself is becoming more specialized, distributing value across general-purpose compute, custom accelerators, networking, optical connectivity, memory interfaces and other components required to connect increasingly large clusters.

That expansion creates more addressable semiconductor content, but it also increases the importance of execution. As hyperscalers add more suppliers and more specialized architectures, simply having exposure to “AI” becomes less informative than understanding exactly where a company sits in the infrastructure stack and how directly that position converts into revenue.

What Broadcom and Marvell Need to Prove Next

For Broadcom, the most immediate checkpoint is its own Q4 guidance. The company expects AI semiconductor revenue to rise from $16.7 billion in Q3 to approximately $21.7 billion in Q4, and delivering that sequential acceleration would provide further evidence that hyperscaler custom accelerator and networking deployments remain on track. Google is another important verification point because Broadcom’s April filing establishes a long-term relationship spanning future TPU generations and next-generation AI racks through up to 2031.

Marvell faces a different test. Management has already reported robust AI bookings and expects a significant acceleration in its Custom business during the second half of fiscal 2027, so the next stage is conversion. Investors will need to watch how quickly Google and other hyperscaler programs move into production, how much recognized revenue they generate and whether Marvell’s connectivity businesses expand alongside custom compute.

Together, these checkpoints should gradually answer the question now facing the custom silicon market: which suppliers can turn increasingly ambitious hyperscaler AI infrastructure plans into large, repeatable semiconductor revenue?

Explore Broadcom and Marvell on MEXC

Broadcom and Marvell represent two different exposures to the expanding custom AI infrastructure market, and users can follow both companies through Real U.S. Stocks on MEXC. The broader MEXC U.S. Stocks market page provides access to major U.S. equities and market data, while dedicated pages are available for Broadcom (AVGO) and Marvell Technology (MRVL).

MEXC also currently lists Stock Futures linked to both names for users seeking directional exposure: AVGO Stock Futures and MRVL Stock Futures. Product availability and terms may vary by region, and leveraged products involve additional risk.

Conclusion

Broadcom’s fiscal Q3 results confirm that custom AI silicon has moved well beyond an experimental side business. AI semiconductor revenue reached $16.7 billion, up 221% year over year, and management expects another sharp acceleration to approximately $21.7 billion next quarter. At the same time, Marvell’s expanding Google relationship shows that the opportunity around hyperscale custom compute is becoming broad enough to support competition across accelerators, networking, memory interfaces and other surrounding silicon.

That combination marks an important transition for the sector. The custom AI chip story no longer depends mainly on proving that hyperscalers want proprietary silicon; the evidence for demand is increasingly visible in Broadcom’s revenue, Marvell’s bookings and Google’s expanding supplier relationships. The harder question now is execution: how rapidly design programs move into production, how much semiconductor content each supplier captures and how defensible those positions remain as hyperscalers diversify their infrastructure.

The custom silicon race is therefore moving from design wins to revenue conversion — and Broadcom versus Marvell is becoming one of the clearest places to watch that shift unfold.

Market Opportunity

Gensyn Price(AI)

----

USD

Gensyn (AI) Live Price Chart

The articles shared on this page are sourced from public platforms and are provided for reference only. They do not represent the position or views of MEXC. All rights belong to James Mitchell. If you believe any content infringes upon the rights of a third party, please contact [email protected] for prompt removal. MEXC does not guarantee the accuracy, completeness, or timeliness of any content and is not responsible for any actions taken based on the information provided. The content does not constitute financial, legal, or other professional advice, nor should it be interpreted as a recommendation or endorsement by MEXC. For expert insights and in-depth analysis, visit MEXC Learn.