Jensen’s move achieves three goals at once: supporting Marvell to curb Broadcom’s market dominance; tying Marvell into his own ecosystem to collect tolls; and simultaneously leveraging public endorsement of Marvell to pressure Broadcom’s stock price, securing public support for his strategic narrative.
Article author and source: Wall Street Journal
Overnight, the game has changed.
Last night, Broadcom released earnings that in any normal year would have sparked market euphoria: AI semiconductor revenue of $10.8 billion in Q2, up 143% year-over-year; and a Q3 guidance of $16 billion in AI chip revenue, expected to grow over 200% year-over-year. Yet, the stock plunged more than 13.7% after hours—quite the awkward turn.
Two days prior, on the stage of the Taipei International Computer Show (Computex), NVIDIA CEO Jensen Huang stepped up, stood beside Marvell CEO Matt Murphy, and told the audience the now widely cited phrase:
The next trillion-dollar company, ladies and gentlemen.
Before the words were even out, Marvell’s stock surged over 20% in pre-market trading, then rose another 37% over the next two trading days. Juxtaposing these two scenarios reveals the most profound unsaid truth in today’s AI chip industry: financial figures are no longer the only narrative—it’s the realignment of the competitive landscape that truly tells the story.
A暴跌 was misinterpreted—Broadcom’s earnings, by any traditional standard, were outstanding. Q2 total revenue reached $22.19 billion, a 48% year-over-year increase and a record high; adjusted EPS was $2.44, up 54% year-over-year, marking the 13th consecutive quarter of AI revenue growth; annual AI chip revenue is expected to reach $56 billion, with reaffirmed guidance that it will surpass $100 billion by 2027.
So why did it drop? The market’s answer is simple and harsh:
Expectation gap under conditions of high valuation.
Total revenue of $22.19 billion was slightly below the analyst consensus estimate of $22.27 billion; the infrastructure software segment contributed $7.18 billion, below the expected $7.32 billion; most critically, some of the most aggressive buy-side institutions internally forecasted Q3 AI chip revenue above $16 billion. Broadcom has risen approximately 40% year-to-date, significantly outperforming the Nasdaq’s 16% gain over the same period, and just reached a historical high of $481.57 ahead of the earnings report. With such a fully priced valuation:
In the market's eyes, "not winning enough" is no different from "losing."
Amid this plunge, there lies a more intriguing signal:
Broadcom has not raised its 2027 AI revenue guidance, which remains at the broad statement of "over $100 billion." In the context of Marvell's strong rise, Broadcom's silence itself has become a silent signal.
To truly understand the game against Marvell, one must first grasp where Broadcom’s moat in the AI chip market comes from and why it is so strong.
Broadcom is not a GPU company, but it is the undisputed leader in the custom AI accelerator (XPU/ASIC) market, currently holding approximately 70% of the global share and firmly established as the market leader. Its business model centers on deep co-design partnerships with hyperscale tech companies: Google’s TPU (Tensor Processing Unit) series has been co-designed by Broadcom since its first generation in 2014, and their agreement has been officially extended through 2031 as of April 2026.
According to Mizuho analysts' estimates, Broadcom's AI revenue from its relationships with Google and Anthropic is projected to reach approximately $21 billion in 2026 and rise to $42 billion in 2027.
This appears to be an unbreakable pattern.
However, Marvell's emergence fundamentally undermined its certainty.
Marvell is not a new player, but its transformation over the past two years has been astonishing. In fiscal year 2026 (ending January 2026), Marvell’s data center revenue reached $6.1 billion, accounting for 74% of total revenue—a 42% year-over-year increase. More importantly, Marvell has provided custom chip design services to Amazon (Trainium series AI accelerators), Microsoft (Maia AI accelerators), Meta (data processing units), and Google (Axion ARM CPUs), with 18 active custom silicon projects in hand.
In other words, Marvell is quietly becoming the second—or even parallel—choice for all hyperscale customers seeking to escape Broadcom’s monopoly.
NVIDIA’s calculation: Jensen Huang’s $2 billion investment in Marvell and the NVLink Fusion strategic partnership represent the most strategically imaginative move in recent semiconductor investment history.
To understand the logic behind this move, it’s first necessary to grasp the structural threats NVIDIA currently faces (far from comprehensive):
With the explosive growth in inference demands from large models like GPT, major hyperscale tech companies in North America are procuring custom chips at an unprecedented scale to reduce their reliance on NVIDIA GPUs. Google has TPU, Amazon has Trainium, Microsoft has Maia, and Meta has MTIA—behind these chips are the design capabilities of Marvell, Broadcom, and MediaTek.
A cruel paradox thus emerges:
Marvell's largest customers are precisely the companies most eager to replace NVIDIA.
Therefore, Jensen’s response was not to block this trend, but to integrate it into his own ecosystem, as the architecture of the NVLink Fusion platform conceals clever design and profound innovation:
Each NVLink Fusion platform must include at least one NVIDIA component—whether it’s the Vera CPU, ConnectX network card, BlueField DPU, or Spectrum-X switch. NVIDIA also controls the licensing of NVLink IP. This means that even if hyperscale customers commission Marvell to design “de-NVIDIAized” custom chips, as long as these chips are deployed within the NVLink Fusion architecture, NVIDIA still generates revenue per rack.
This is a toll booth, a tax on "custom ASICs."
NVIDIA, through Marvell, secured its share of the profits in the era of custom chips. Meanwhile, Marvell’s technical capabilities have been integrated into NVIDIA’s ecosystem, enabling hyperscale customers to deploy customized AI accelerators while retaining NVIDIA’s software stack and supply chain support—enhancing the appeal of the NVLink Fusion platform.
Jensen's move is a three-in-one strategy:
- Support Marvell to counter Broadcom's market share monopoly;
- Meanwhile, tie in Marvell into your own ecosystem to collect toll fees;
- It also leveraged public endorsement of Marvell to suppress Broadcom’s stock price, securing public support for its strategic narrative.
So it must be acknowledged that, by supporting Marvell, Jensen demonstrated his unparalleled strategic and execution capabilities.
Google’s “multi-vendor strategy”: The first crack in Broadcom’s relationship Broadcom’s partnership with Google is a business alliance built on deep technical integration. Broadcom has played a central role in designing the chip architecture, IP, interconnect technologies, and packaging for Google’s TPUs, serving as a key enabler of TPU silicon realization, while Google retains control over the chip architecture and software stack. Together, the two companies have co-designed seven generations of TPUs, reflecting a deeply entrenched collaboration.
However, a quiet evolution in Google's internal strategy is underway.
Google’s multi-vendor strategy is beginning to take shape. On the Ironwood TPU (7th generation), Google partnered with MediaTek to co-design a cost-optimized version for inference, codenamed “Zebrafish,” aiming to be 20–30% cheaper than Broadcom’s solution. For the next-generation TPU v8 training chip (codenamed “Sunfish”), Broadcom continues to handle the training chip, while MediaTek focuses on the inference chip, giving Google leverage to foster competition between the two suppliers.
More importantly, Google has been negotiating with Marvell on a new AI inference chip; if the partnership is finalized, Marvell will become Google’s third design partner in its AI chip ecosystem.
This means:
Broadcom's relationship with Google is quietly evolving from an "exclusive partnership" to one of "primary suppliers." Broadcom remains the core and dominant player in Google's custom chip ecosystem and cannot be replaced in the short term, but its indispensability is beginning to show clear signs of weakening. This is the first tangible sign of the competitive pressure from Marvell, driven by Jensen, on Broadcom's fundamental position.
The hidden battle over interconnect standards: Beneath the open competition in chip manufacturing, a covert battle over interconnect standards is unfolding simultaneously, with Broadcom and NVIDIA standing on opposite sides.
NVIDIA's flagship offering is NVLink: a high-performance, fully proprietary interconnect standard with a 4-5 year first-mover advantage, deeply embedded in global AI infrastructure through the CUDA ecosystem.
Led by AMD and Intel, with Broadcom formerly involved, the UALink (Ultra Accelerator Link) consortium aims to establish an open industry standard for scalable interconnect solutions for non-NVIDIA chips. The UALink 2.0 specification was officially released on April 7, 2026, introducing new architectures such as In-Network Compute, which theoretically reduces distributed training time by up to 30%.
However, Broadcom later quietly exited the UALink alliance and shifted to developing its own interconnection technology. This move is intriguing and reveals another layer of Broadcom’s strategic thinking:
Chart your own independent path between open standards and private ecosystems.
The advantage of NVLink lies in its peak performance and deep integration with the software ecosystem; the value of UALink lies in avoiding vendor lock-in and providing larger customers with greater choice. Neither pathway will determine a winner in the short term, but ownership of the interconnect standard will determine the future power structure of AI data center architectures—that is, market share.
Photonics, the next battlefield. If custom chips are the current main arena, then silicon photonics and optical interconnects are emerging as the critical and more strategic battleground that will determine the outcome of the next phase.
The driving force is the physical laws themselves:
As AI systems scale from single racks to distributed clusters spanning multiple racks and data centers, the bandwidth limits and power costs of traditional copper cables are no longer sustainable. Optical interconnects offer more than twice the energy efficiency of copper cables, along with longer transmission distances and higher bandwidth density.
Marvel has made highly aggressive moves in this field.
In December 2025, Marvell completed its acquisition of photonic interconnect startup Celestial AI for up to $5.5 billion, gaining access to its “Photonic Fabric” technology platform. This technology enables optical connections between chips from any point to any point, forming the foundational basis for building “light-driven computing fabrics.”
Marvell CEO Murphy explicitly stated at Computex that the physical limits of copper cabling are approaching the inside of the rack, and co-packaged optics (CPO) is the only solution.
NVIDIA’s collaboration with Marvell also identifies silicon photonics and optical interconnects as core areas of partnership. Amazon AWS has even issued warrants to Marvell for its photonic fabric products, supporting its innovations in optical-scale interconnects.
Whoever first moves optical interconnects from the lab to large-scale mass production will control the pricing of the infrastructure for the next-generation AI factory. Broadcom will certainly not be absent from this competition, but Marvell’s depth and speed of investment in this direction have already given it a first-mover advantage.
Three scenarios for the evolution of the market landscape: Scenario One: Broadcom retains its core position, while Marvell fills the gaps (this scenario has the highest probability, within 2–3 years)
This is the most likely near-term direction.
Broadcom’s short-term revenue dominance is nearly unshakable, thanks to its long-term agreement with Google through 2031, custom chip projects with OpenAI and Meta, and AI chip backlog orders totaling up to $7.3 billion—it remains the king. Marvell captures the additional demand from hyperscale customers that Broadcom cannot fully address, as well as the market share diluted by Broadcom’s multi-supplier strategy.
Both coexist, but Broadcom's relative premium will gradually narrow.
Scenario Two: The Interconnected Ecosystem Determines the Landscape (Medium Term, 3–5 Years)
If NVLink Fusion gains deep adoption from more hyperscale customers, Marvell will establish an asymmetric advantage among emerging AI infrastructure clients through its customized chip capabilities embedded within NVIDIA’s ecosystem. If Broadcom cannot offer a similar ecosystem integration solution, its growth will become increasingly dependent on the continuation of its relationship with Google alone.
This scenario has no impact on Broadcom's software business, but the valuation logic for its chip business will undergo a qualitative change.
Scenario Three: Photonics Reshuffles (Long-term, 5+ years)
If Marvell-Celestial’s photonic fabric technology achieves the first large-scale commercial deployment (Marvell’s current market expectation is for billion-dollar revenue by 2028–2029), the entire AI data center interconnect market will undergo an architectural overhaul. In this scenario, the company that first establishes dominance in optical interconnect standards will become the next “tax collector” in AI infrastructure, following CUDA.
Marvell's enormous ambition is clearly evident here.
The domino effect of this battle—where NVIDIA, using Marvell, has challenged Broadcom—extends far beyond the market value fluctuations of just these two companies.
TSMC: The biggest mutual beneficiary and the biggest bottleneck. Broadcom, Marvell, NVIDIA, Google—all players rely on TSMC’s advanced process and CoWoS packaging capacity. TSMC’s 3nm and 2nm capacity will remain in persistent short supply, long-termly sought after by the market, becoming a structural bottleneck for the entire industry.
So, TSMC is still extremely, extremely, extremely good!
Arm Holdings: The Hidden Winner. Arm has joined the NVLink Fusion ecosystem, allowing its licensees to natively integrate NVLink connectivity into their chips. Google’s Axion CPU, Amazon’s Graviton, and Microsoft’s Azure Cobalt are all based on Arm architecture, meaning Arm is playing an increasingly indispensable role in the custom chip strategies of its largest customers.
MediaTek: An Unexpected Beneficiary. By participating in the design of Google's TPU v8 inference chips, MediaTek has successfully entered the high-end AI data center market, injecting new imaginative possibilities into its long-standing narrative as a consumer electronics company.
HBM memory (SK hynix, Samsung, Micron): Custom ASIC expansion is driving increased demand. Google’s Ironwood TPU is equipped with 192 GB of HBM3E memory, and next-generation XPUs will only continue to increase demand for HBM. Memory manufacturers are among the most stable beneficiaries in this landscape, as they don’t need to take sides but gain more buyers.
This war has no end—returning to the most fundamental question at the start of the article: Has Marvell’s competitive impact already appeared in Broadcom’s financial reports?
The answer is:
The outline is beginning to emerge, but no critical damage has been done yet.
Broadcom's Q3 AI chip guidance of $16 billion is below even the most aggressive expectations, reflecting both Broadcom's traditionally conservative guidance style and market adjustments pricing in potential share dilution. Broadcom's core business—including Google's TPU, Meta's MTIA, and OpenAI's custom chips—remains solid over the next two years.
But the narrative of Broadcom's "absolute monopoly" has ended, and the pricing logic of being "one of the primary suppliers" is now taking its place—a shift that has systemic implications for a tech stock that has enjoyed high premiums.
Jensen Huang played this move with exceptional precision:
He did not compete directly with Broadcom; instead, he redefined the power structure of the AI chip era by supporting Marvell, building the NVLink Fusion ecosystem, and making that public declaration on the Computex stage in Taipei. In Jensen’s narrative, the future is not about “who makes the best GPU,” but about “who builds the infrastructure ecosystem that large-scale customers cannot leave”—that is Jensen’s core interest.
Hock Tan, the Broadcom leader known for acquisitions and financial discipline, is facing the most complex and challenging competitive landscape of his career: he must simultaneously contend with market share erosion from Marvell, Google’s multi-vendor strategy, and narrative erosion from NVIDIA’s ecosystem.