Overview

Broadcom has emerged as the dominant provider of custom AI application-specific integrated circuits (ASICs) for the world’s largest hyperscalers and AI labs, reporting $10.8 billion in AI semiconductor revenue in fiscal Q2 2026 (ended May 3, 2026) – a 143% year-over-year increase [1][2]. The company’s six core custom chip customers include Google (Alphabet), Meta Platforms, Anthropic, OpenAI, and two unnamed clients [3][4]. However, the sustainability of this dominance is being questioned as customers pursue in-house silicon, alternative design partners, and as new competitors – from Marvell Technology to emerging startups – enter the market. This report examines the depth of Broadcom’s current relationships, the diversification strategies of its key customers, the competitive landscape, and the broader trend toward vertical integration in AI hardware, to assess how durable Broadcom’s custom AI chip business truly is.

Broadcom’s Current Custom Chip Relationships

Google (Alphabet) – The Anchor Customer

Broadcom’s relationship with Google is the most significant and the most scrutinized. The two companies have a long-term agreement covering multiple generations of Tensor Processing Units (TPUs) and AI networking, with the contract extending through 2031 [5][6]. Citi analysts estimate that Google alone accounts for approximately 35–40% of Broadcom’s total sales, making it by far the largest customer [7]. Broadcom has helped Google bring 14 of its most advanced chip designs to market over the past 12 years [5].

On the Q2 FY2026 earnings call, CEO Hock Tan confirmed the “long-term Google TPU agreement” and noted that Alphabet’s $80 billion equity capital raise in June 2026 – which included a $10 billion private placement from Berkshire Hathaway – directly benefits Broadcom as the leading ASIC provider for Google’s TPUs [8][9]. JPMorgan analysts have stated that the TPU v9 remains on track for a 2028 debut with Broadcom, and they believe the market underestimates Broadcom’s 18-month-plus design lead over competitors [5].

Meta Platforms – A Growing Partnership

Meta is one of Broadcom’s six core AI customers, and the relationship has deepened considerably. On the Q2 FY2026 earnings call, Tan detailed a plan to deploy 3 gigawatts (GW) of compute capacity with Meta through the end of 2028, with an initial 1 GW order beginning delivery in the second half of 2026 [3][10]. Meta’s 2026 capital expenditures are guided at $115–$135 billion, underscoring the scale of its infrastructure buildout [11].

Beyond chip supply, Broadcom and Meta are collaborating on a $125 million “Semiconductor Hub” at the UCLA Samueli School of Engineering, alongside Applied Materials, GlobalFoundries, and Synopsys [12]. The five-year initiative aims to accelerate research and workforce development for AI-powered chip technologies. This joint investment signals a strategic alignment that goes beyond a simple vendor relationship.

Anthropic – A Rapidly Expanding Account

Anthropic placed a $10 billion AI chip order with Broadcom in December 2025, and the company has a contract for 5 GW of next-generation TPU-based compute starting in 2027 [3][13]. Broadcom’s AI special purpose vehicle (SPV) with Apollo Global Management and Blackstone is designed to provide debt financing for these large-scale chip sales, with more than 20 GW of computing capacity expected through 2028 [10][14].

Anthropic is nearing its first-ever quarterly profit, projecting Q2 2026 revenue of at least $10.9 billion and an operating profit of $559 million [15]. Its annualized revenue of ~$45 billion is 35% higher than OpenAI’s $33 billion ARR [15]. As Anthropic scales, its reliance on Broadcom for compute capacity is increasing, not decreasing.

OpenAI and Other Customers

OpenAI has a contractual commitment for 1.3 GW of compute in 2027, part of a larger 10 GW agreement by 2029 [3][10]. Broadcom has begun delivering chips to OpenAI and is on track for production later in 2026 [10]. Two additional unnamed customers placed $6 billion in purchase orders during Q2 FY2026 [3].

Revenue Concentration and Financial Context

Broadcom’s AI semiconductor revenue represented 48.7% of total revenue in Q2 FY2026 ($10.8B out of $22.19B), and the company guided AI revenue to $16 billion in Q3 FY2026 (54.4% of total) [1][2]. For full fiscal 2026, Broadcom expects AI semiconductor revenue of $56 billion, and for fiscal 2027, the target is in excess of $100 billion [1][3]. The company booked over $30 billion in AI semiconductor orders during Q2 alone, far exceeding the $10.8 billion recognized [3].

This concentration creates both opportunity and risk. While the revenue growth is extraordinary, the dependence on a handful of hyperscaler customers – particularly Google – makes Broadcom vulnerable to any shift in those customers’ sourcing strategies.

Customer Diversification Threats

Google’s In-House and Alternative Partner Moves

Google represents the most immediate diversification risk to Broadcom. The clearest signal came in mid-2026 with reports that Google’s 10th-generation TPU, codenamed “Icefish,” is being developed with MediaTek rather than Broadcom, with mass production targeted as early as 2028 [16][17]. The main compute die will be manufactured by TSMC using a 1.4-nanometer process, while Samsung will produce a memory-facing I/O die using its 2-nanometer process [16][18]. Previous TPU generations were designed with Broadcom.

Macquarie downgraded Broadcom stock to neutral from outperform on June 4, 2026, citing Google’s move to develop AI chips in-house and expecting Broadcom’s market share to “decline meaningfully” in 2027 [19]. However, JPMorgan has pushed back, noting that the TPU v9 (with Broadcom) is on track for 2028 and that Broadcom’s design lead remains substantial [5].

Google is also taking steps to commercialize its TPUs externally. During the Q1 2026 earnings call, CEO Sundar Pichai announced that Google will begin selling TPUs to select customers, with deals already signed with Anthropic and Meta [20]. Google is partnering with Blackstone on a joint venture to build a neocloud compute-as-a-service company, deploying 500 MW of TPU capacity by 2027 [20][21]. This external sales push could actually benefit Broadcom if it increases total TPU volume, but it also signals Google’s growing confidence in its own silicon capabilities.

Additionally, Google plans to use Intel to produce more than 3 million specialized TPUs by 2028, after months of testing Intel’s technology [22]. While this is a manufacturing diversification rather than a design diversification, it reduces Google’s dependence on TSMC and, indirectly, on Broadcom’s design ecosystem.

Meta – Stable but Not Immune

Meta’s in-house silicon efforts center on the Meta Training and Inference Accelerator (MTIA) family, which the company has been developing for years. However, there is no public evidence that Meta is reducing its reliance on Broadcom. The 3 GW deployment plan through 2028 is locked in, and Meta has not announced any alternative custom ASIC partner such as Marvell [23]. The $125 million UCLA research hub further cements the relationship.

Meta’s AI restructuring in May 2026, which included laying off approximately 8,000 employees and reassigning 7,000 staff to AI-related teams, is focused on accelerating AI development rather than changing chip suppliers [11]. Meta is also among the select customers that will receive Google’s TPUs, but this is complementary to, not a replacement for, its Broadcom-based compute [20].

Anthropic – Multi-Sourcing Without Reducing Broadcom

Anthropic is aggressively diversifying its compute sources across multiple providers, but this diversification is additive rather than substitutive. The company has a 10-year deal worth over $100 billion with Amazon Web Services to use Trainium chips, is in discussions with Microsoft to adopt Maia 200 chips, and has agreed to pay SpaceX $1.25 billion per month through May 2029 for NVIDIA GPU-based compute [15][24][25]. However, its Broadcom commitment – 5 GW of TPU-based compute starting in 2027 – remains a core part of its infrastructure plan [3].

Anthropic CEO Dario Amodei has stated that the company has had “difficulties with compute” due to surging demand, which suggests that Broadcom’s capacity is needed, not redundant [24]. Anthropic is not developing its own silicon; as an AI lab, it focuses on models, not chip design.

Summary of Diversification Risk by Customer

Customer	Broadcom Exposure	Diversification Threat	Timeline
Google	35–40% of Broadcom revenue	High – Icefish moving to MediaTek, Intel manufacturing, external TPU sales	2028+
Meta	Significant but smaller	Low – 3 GW commitment, joint research hub, no alternative partner	Through 2028
Anthropic	Growing	Low – increasing reliance, no own silicon, multi-sourcing is additive	2027+
OpenAI	Early stage	Low – long-term contract, no alternative ASIC partner announced	2027–2029

Competitive Threats from Marvell Technology

Marvell’s Custom ASIC Position

Marvell Technology is Broadcom’s most direct competitor in the custom AI ASIC market. Marvell has two confirmed hyperscaler programs: Amazon’s Trainium chips and Microsoft’s Maia 100/200 accelerators [26][27]. Marvell’s custom silicon business is projected to exceed $10 billion annually by fiscal 2029 [28].

Nvidia CEO Jensen Huang declared at Computex 2026 that Marvell “is going to be the next trillion-dollar company,” specifically citing its optical interconnect technology as essential for disaggregated AI data centers [29]. Nvidia committed a $2 billion investment into Marvell, further validating its technology [29].

Technology and Financial Comparison

Marvell’s core strength lies in high-speed networking, optical interconnects, and custom ASICs – the “connective tissue” that enables efficient data flow between GPUs [30]. This is complementary to Broadcom’s strength in both ASICs and networking, but Marvell’s optical leadership is a genuine differentiator.

Financially, Marvell is much smaller. In Q1 FY2027 (ended April 2026), Marvell reported revenue of $2.42 billion (up 28% YoY), with data center representing 76% of sales at $1.83 billion [28]. This compares to Broadcom’s $10.8 billion in AI semiconductor revenue alone. Marvell’s GAAP net income cratered 81% year-over-year to $34.5 million, and the stock trades at a forward P/E of 65 [30].

A critical risk for Marvell is that 76% of its revenue comes from customers who are “actively building their own in-house silicon to replace Marvell,” a risk factor the company itself discloses in SEC filings [30]. This means Marvell faces the same in-housing threat as Broadcom, but with less diversification.

Which Customers Is Marvell Targeting?

Marvell’s confirmed custom ASIC customers are Amazon and Microsoft – companies that are not Broadcom’s core custom chip clients. Broadcom’s six core customers are Google, Meta, Anthropic, OpenAI, and two unnamed; Amazon and Microsoft are notably absent from that list. This suggests that Marvell and Broadcom are not yet directly competing for the same design wins, but the lines could blur as both companies expand.

Marvell’s interconnect business is growing quickly, with the company projecting 70% growth this year and total revenue expected to climb 40% to nearly $11.5 billion [28]. However, the custom ASIC revenue is still a fraction of Broadcom’s.

Startup Threats (Architect Labs and Others)

Architect Labs – Too Early to Matter

Architect Labs emerged from stealth on June 18, 2026, just one day before this report. The Palo Alto-based startup has raised $24 million in seed funding from Kindred Ventures, TQ Ventures, Race Capital, and angel investors from Nvidia, Google, and OpenAI [31]. The company is building an AI system to design custom chips, aiming to create a “designless semiconductor industry” where anyone with a workload can access world-class chip design.

Co-founders Ebrahim Hussain (former Apple and Tesla chip designer) and Aaditya Subedi (Harvard AI researcher) have credible backgrounds, but the company has no announced chip products, no customers, and no tapeout timeline [31]. The $24 million seed round is minuscule compared to the hundreds of millions needed for a single chip tapeout.

Architect Labs represents a long-term thesis that AI-driven design tools could lower the barrier to custom silicon, potentially expanding the total addressable market. However, as a near-term competitive threat to Broadcom, it is negligible. The company is years away from production silicon, and its model – democratizing chip design – is more likely to create new entrants than to displace existing hyperscaler relationships.

Other Emerging Startups

Several other startups are targeting the AI chip market, but most focus on inference rather than the training ASICs that are Broadcom’s core business:

• d-Matrix: Microsoft-backed startup shipping its Corsair inference accelerator in June 2026. The chip uses on-chip SRAM and claims 10x faster inference for smaller workloads. d-Matrix is partnered with Broadcom on its SquadRack system, making it a collaborator rather than a competitor [32].
• SambaNova Systems: Over $1.5 billion in funding, targeting inference with its SN50 chips. The company claims superior performance for serving AI models, but its market is inference, not hyperscaler training ASICs [33].
• Groq: After a $20 billion “not-acquisition” deal with Nvidia in December 2025, Groq is pivoting to an inference neocloud business. It is no longer an independent chip competitor [34].
• XCENA: Raised $135 million for its MX1 memory-side compute chip, which uses CXL to process data within memory modules. Mass production at Samsung is scheduled for end of 2026, but the product addresses memory bottlenecks, not custom ASIC design [35].
• Lightmatter: Developing photonic interconnects for AI data centers, with $850 million raised from Google, Fidelity, and others. Photonics could disrupt the networking portion of Broadcom’s business, but Lightmatter is a component supplier, not an ASIC competitor [36].

None of these startups directly threaten Broadcom’s custom ASIC relationships with hyperscalers. The inference-focused startups are more of a threat to Nvidia’s GPU inference dominance than to Broadcom’s model of designing training accelerators for specific customers.

In-House Silicon Efforts and Their Impact on Broadcom

The Hyperscaler Vertical Integration Trend

The broader trend of hyperscalers building their own silicon is the most significant long-term risk to Broadcom’s business model. Amazon, Google, Microsoft, and Meta all have active in-house chip programs, and the trajectory is toward greater self-sufficiency.

Amazon is exploring selling its Trainium chips externally. AWS AI chief Peter DeSantis confirmed early-stage talks to allow external use of Trainium, and CEO Andy Jassy stated in his April 2026 shareholder letter that if the chip business operated standalone, it could reach an annual run rate of approximately $50 billion – comparable to Intel’s revenue [27][37]. Amazon’s current Trainium capacity is fully booked, and even next year’s Trainium4 chips are already allocated [27].

Microsoft unveiled seven in-house MAI models at Build 2026, co-designed with the Maia 200 inference accelerator. Microsoft’s AI business surpassed an annual revenue run rate of $37 billion, up 123% year-over-year [38]. While Microsoft still depends on Nvidia for training compute, the company is clearly building toward greater self-sufficiency.

Google is the most advanced, with its TPU program now spanning 14 generations and expanding to external sales. The Icefish shift to MediaTek is a concrete example of Google reducing dependence on Broadcom for design.

Meta continues to develop its MTIA family, but has not yet reached a scale that would allow it to replace Broadcom. The 3 GW commitment suggests Meta still needs Broadcom’s design expertise and capacity.

Is Broadcom a Bridge or a Long-Term Partner?

The critical question is whether Broadcom’s custom chip business is a temporary bridge to full in-house development for its customers, or a sustainable long-term partnership. The evidence suggests a mixed answer.

For Google, the relationship appears to be transitioning. The long-term agreement through 2031 provides a multi-year runway, but the Icefish program with MediaTek indicates that Google is actively building alternative design capabilities. JPMorgan’s view that Broadcom has an 18-month design lead suggests that Google will continue to rely on Broadcom for cutting-edge designs while developing in-house capability for later generations [5].

For Meta, the relationship appears more durable. Meta’s in-house MTIA chips are not at a scale to replace Broadcom, and the joint research hub signals a collaborative rather than competitive dynamic. Meta’s AI infrastructure needs are growing so rapidly that it needs multiple sources of supply.

For Anthropic and OpenAI, Broadcom is a long-term partner because these AI labs do not have the resources or expertise to develop their own silicon. They will remain dependent on external ASIC providers for the foreseeable future.

The Structural Advantage of Scale

Broadcom’s scale is a powerful moat. The company booked over $30 billion in AI semiconductor orders in a single quarter, and its $100 billion+ FY2027 target implies a production volume that few competitors can match [3]. The AI SPV with Apollo and Blackstone provides financing capacity that startups and even some hyperscalers cannot replicate [14].

CEO Hock Tan has noted that Broadcom’s visibility now extends into fiscal 2028, with “substantial growth” expected from fiscal 2027 [19]. The company’s ability to secure wafer and HBM supply for 2026 and 2027, and to work on securing supply for 2028 and 2029, gives it a supply chain advantage that new entrants will struggle to match [4].

Sustainability Assessment

Key Risks

1. Google’s Icefish shift to MediaTek (2028+): This is the most concrete and highest-severity risk. If Google moves a significant portion of its TPU design work to MediaTek, Broadcom could lose a substantial share of its largest customer’s business. Macquarie expects Broadcom’s market share to “decline meaningfully” in 2027 [19]. However, the TPU v9 (with Broadcom) is still on track for 2028, and the contract runs through 2031, so the erosion will be gradual.

2. Hyperscaler in-housing acceleration: Amazon’s exploration of external Trainium sales and Microsoft’s growing in-house capabilities could reduce the total addressable market for custom ASIC design services. If hyperscalers become self-sufficient in chip design, Broadcom’s role could shift from design partner to pure-play manufacturer, which would compress margins.

3. Marvell’s growth in custom ASICs: Marvell’s $10 billion+ custom silicon target, combined with Nvidia’s endorsement and investment, makes it a credible #2 player. If Marvell wins additional hyperscaler accounts (e.g., Google or Meta), it could directly erode Broadcom’s market share.

4. Customer concentration: With 35–40% of revenue from Google alone, any disruption to that relationship would have an outsized impact on Broadcom’s financials.

5. Margin pressure: Broadcom’s gross margin declined to 77.1% in Q2 FY2026, down 230 basis points year-over-year, due to product mix shifts toward lower-margin ASICs and TPUs [2]. CFO Kirsten Spears noted that “ASICs and TPUs have lower margins, but AI networking has rich margins, which helps offset some of the pressure” [4]. As ASIC revenue grows faster than networking, margins could continue to compress.

Mitigating Factors

1. Multi-year contracts and deep entrenchment: Broadcom has long-term agreements with all four major AI customers (Google through 2031, Meta through 2028, Anthropic and OpenAI through 2027–2029). These contracts provide revenue visibility and switching costs for customers.

2. Design lead and execution track record: JPMorgan estimates Broadcom has an 18-month design lead over competitors, and the company has brought 14 of Google’s most advanced designs to market [5]. This expertise is not easily replicated.

3. Diversification across six core customers: While Google is the largest, Broadcom has five other core AI customers, including two unnamed ones that placed $6 billion in orders. This reduces single-customer risk.

4. Networking as a complementary moat: Broadcom’s AI networking business provides rich margins and is essential for connecting AI accelerators. Even if customers design their own ASICs, they will still need Broadcom’s networking silicon.

5. Supply chain and financing advantages: Broadcom’s ability to secure wafer and HBM supply, and its AI SPV with Apollo and Blackstone, provide capacity and financing that competitors cannot match.

6. Total addressable market growth: AI infrastructure spending is expected to exceed $700 billion in 2026, up from ~$400 billion in 2025 [10]. Even if Broadcom loses some market share, the absolute revenue opportunity is growing rapidly.

Timeline Assessment

• 2026–2027: Broadcom’s dominance is intact. AI revenue is growing 143% YoY, and the $100 billion+ FY2027 target is supported by $30 billion in quarterly bookings. The Google Icefish risk is not yet material.
• 2028–2029: The Icefish program with MediaTek could begin to impact Broadcom’s Google revenue. Marvell’s custom silicon business may reach $10 billion, but this is still a fraction of Broadcom’s scale. Hyperscaler in-housing will be more advanced, but Broadcom’s networking business and long-term contracts provide a buffer.
• 2030+: The sustainability of Broadcom’s custom ASIC business depends on whether it can continue to offer design advantages that hyperscalers cannot replicate internally. If Google, Amazon, and Microsoft achieve full self-sufficiency in chip design, Broadcom’s ASIC revenue could plateau or decline. However, the networking business and the emergence of new AI customers (e.g., from the startup ecosystem) could offset this.

Conclusion

Broadcom’s custom AI chip business is sustainable in the medium term (through 2028) but faces structural headwinds in the long term. The company’s deep relationships, design expertise, scale, and networking moat provide significant competitive advantages. The most serious threat is Google’s gradual shift toward alternative design partners, but this is a multi-year process, not an immediate disruption. Marvell is a credible #2 but remains much smaller and faces its own in-housing risks. Startups like Architect Labs are not near-term threats. The broader in-housing trend will eventually reduce the addressable market for custom ASIC design services, but Broadcom’s networking business and its ability to serve a diversified set of AI customers – including AI labs that will never build their own silicon – provide a foundation for continued growth. The key risk to monitor is the pace at which hyperscalers achieve self-sufficiency in chip design, and whether Broadcom can maintain its design lead as customers gain more internal expertise.