Broadcom Google Chip Ties Face an Earnings Test as Marvell Moves In
Broadcom enters its September 2 earnings report with a clear conflict around its Google business, despite record AI semiconductor growth. The Broadcom Google partnership now carries both enormous revenue potential and a new competitive threat. Google has signed a long-term supply agreement with Broadcom, yet it is also expanding custom-chip work with Marvell.
That tension matters because Broadcom already guided for fiscal third-quarter AI semiconductor revenue of $16 billion, more than triple the prior-year level. Investors are therefore looking beyond a routine earnings beat. They want evidence that Broadcom can protect its role inside Google’s TPU roadmap while expanding programs with other major AI companies.
Marvell is the immediate pressure point. Its latest Google agreement does not establish that Broadcom has lost a TPU generation. However, it shows that Google wants more than one credible custom-silicon partner. The earnings call must clarify whether this is ordinary supplier diversification or an early shift in Broadcom’s competitive position.
Broadcom Earnings Put the Google Relationship Under a Microscope
Broadcom’s results will test whether extraordinary AI growth can outweigh new doubts about its position inside Google’s chip program.
Broadcom scheduled its fiscal third-quarter report for September 2 after the US market closes. Its earnings calendar lists a conference call at 5 p.m. Eastern time.
The timing is unusually important. Broadcom’s previous quarter established a demanding baseline for revenue, margins, and future commitments. It also increased the market’s dependence on management commentary about a small group of large AI customers.
Broadcom reported fiscal second-quarter revenue of $22.2 billion, up 48 percent from the prior year. Semiconductor revenue reached $15 billion, while infrastructure software contributed $7.2 billion.
AI semiconductor revenue reached $10.8 billion, representing 49 percent of total company revenue. That figure grew 143 percent from the prior year and exceeded management’s earlier forecast.
The company’s quarterly results also guided for third-quarter revenue of approximately $29.4 billion. Broadcom forecast $20.5 billion from semiconductors and $8.9 billion from infrastructure software.
Within semiconductors, management projected $16 billion in AI revenue, up more than 200 percent from the year-earlier quarter. That forecast makes AI execution the central question in the upcoming report.
Broadcom’s AI category includes custom accelerators, often called XPUs, and networking components used to connect large computing clusters. An XPU is a processor designed around a customer’s specific workloads rather than broad, general-purpose computing.
Networking accounted for almost 40 percent of Broadcom’s second-quarter AI revenue. That mix matters because Broadcom sells more than processor design services. It also supplies switches, interconnect technology, and other components required to operate thousands of accelerators together.
Bookings added another strong signal. Broadcom said AI semiconductor bookings exceeded $30 billion during the second quarter, compared with $10.8 billion shipped.
Inventory rose to $4.3 billion, and days of inventory increased from 68 to 86. Management said it was securing supply for expected AI growth during the second half of 2026.
Those figures suggest that the third quarter is not only about investor expectations. Broadcom has already committed working capital, engineering resources, and manufacturing capacity to the projected expansion.
Still, the numbers alone will not settle the debate. The market has already seen Broadcom deliver rapid growth while holding its longer-term forecast steady.
Management expects $56 billion in fiscal 2026 AI semiconductor revenue and more than $100 billion in fiscal 2027. The upcoming report must show how signed programs convert into production revenue without weakening execution or margins.
The original market discussion%2BOpinions%2Bon%2BUpcoming%2BEarnings%2Band%2BAI%2BChip%2BDevelopments) highlighted social attention around that guidance. It also noted Broadcom’s second-quarter revenue and generally favorable analyst positioning.
That discussion captures the immediate sentiment, but the underlying issue is larger than one quarter. Broadcom must demonstrate that its custom-chip pipeline remains defensible as customers seek more negotiating leverage.
Why Broadcom Google Chip Development Matters Now
The Broadcom Google relationship sits at the center of Broadcom’s claim that custom AI accelerators can become a durable business rather than a temporary spending surge.
Google designs Tensor Processing Units, or TPUs, for machine learning workloads across its internal services and cloud platform. Broadcom helps convert those architectures into production-ready silicon and supplies associated networking technology.
This collaboration gives Google an alternative to relying entirely on merchant GPUs. It also gives Broadcom exposure to AI computing demand without requiring it to market a general-purpose accelerator under its own brand.
The relationship gained formal protection in April 2026. Broadcom disclosed that it had signed a long-term agreement to develop and supply custom TPUs for future Google generations.
The same regulatory filing included a supply assurance agreement covering networking and other components. That agreement extends through 2031 for Google’s next-generation AI racks.
This is stronger evidence than informal supply-chain reports. It confirms that Google expects Broadcom to participate in multiple future TPU generations and related infrastructure.
However, the agreement does not guarantee fixed revenue every quarter. It also does not establish that Broadcom will receive every custom-chip program Google develops.
Broadcom’s filing contains an important caution around another part of the arrangement. Anthropic plans to access approximately 3.5 gigawatts of next-generation, TPU-based computing capacity through Broadcom beginning in 2027.
The filing says that usage depends on Anthropic’s continued commercial success. Broadcom, Google, and Anthropic were also discussing operational and financial partners when the disclosure was filed.
Gigawatts measure power capacity rather than chip revenue. The unit helps describe the physical scale of an AI deployment, but it does not reveal the final processor mix or Broadcom’s economic share.
That distinction is critical. Large capacity commitments signal demand, yet revenue depends on construction schedules, chip yields, customer adoption, financing, and system configuration.
Google’s product roadmap provides additional context. In April, the company introduced TPU 8t for training and TPU 8i for inference, which runs already-trained models.
Google says TPU 8t can scale to 9,600 processors and two petabytes of shared, high-bandwidth memory in one superpod. A superpod links many accelerators into a single computing system.
The company also says TPU 8t delivers up to 2.7 times better training performance per dollar than Ironwood. Google’s TPU technical overview attributes that gain to architecture, memory, networking, and workload specialization.
TPU 8i targets inference, including applications that support many simultaneous AI agents. Google says the design uses additional on-chip memory and specialized communication hardware to reduce serving costs.
Those are company claims, not independent benchmark conclusions. They nevertheless show why the Broadcom Google program is commercially significant.
Google is separating training and inference into distinct systems. Each system requires processor design, memory integration, interconnects, packaging, and manufacturing coordination.
That complexity creates more potential content for Broadcom in each deployed rack. It also creates openings for other suppliers that can handle part of the system.
Broadcom’s advantage comes from combining custom silicon with Ethernet networking and high-speed SerDes technology. SerDes circuits convert data between serial and parallel formats across fast chip connections.
Yet Google’s incentive is different from Broadcom’s. Google wants high performance, dependable supply, lower operating costs, and bargaining leverage across multiple vendors.
That difference explains why a long-term agreement can coexist with supplier diversification. The contract protects Broadcom’s participation, while Google retains reasons to cultivate credible alternatives.
Marvell Turns Supplier Diversification Into a Competitive Test
Marvell’s Google agreement changes the debate from whether Broadcom has demand to whether Broadcom can preserve its share of that demand.
Google recently expanded its custom semiconductor relationship with Marvell. The agreement includes incentives tied to future custom-product revenue over several years.
Reports describing the arrangement connected the potential vesting conditions to a large cumulative revenue opportunity. Those conditions should not be treated as guaranteed sales.
The structure instead shows Google rewarding Marvell for scaling a broader relationship. The covered product areas reportedly include custom processors, networking, and memory-related components.
That overlap puts Marvell closer to Broadcom’s strategic territory. Both companies can help hyperscalers design application-specific chips and connect them inside large data centers.
Marvell’s win does not prove it is replacing Broadcom on Google TPUs. Broadcom’s agreement through 2031 remains in force, and no authoritative disclosure has identified a canceled Broadcom program.
The more defensible conclusion is that Google is reducing supplier concentration. It can assign different chips, functions, or generations to different partners while maintaining established programs.
That strategy gives Google several benefits. It reduces execution risk, adds manufacturing options, and strengthens its position during commercial negotiations.
It may also accelerate internal innovation. Competing partners can present different packaging, interconnect, and system designs for a rapidly changing set of AI workloads.
For Broadcom, this is a share question rather than an immediate demand collapse. Google’s overall AI infrastructure budget can grow while Broadcom receives a smaller portion than investors previously expected.
That possibility explains why Broadcom’s stock reacted negatively when the Marvell arrangement became public. Markets had treated Google’s TPU program as a particularly defensible element of Broadcom’s AI forecast.
Competition changes the quality of that forecast. Revenue backed by exclusive or highly concentrated relationships deserves different assumptions from revenue contested across several suppliers.
Analysts quoted before earnings have generally maintained positive views on Broadcom. Their optimism centers on the signed Google agreement, new customer programs, and the enormous scale of hyperscaler spending.
The skeptical view focuses on future bargaining power. More qualified suppliers can pressure design fees, component share, and the duration of individual awards.
Google’s approach also reflects a broader shift in AI infrastructure. Hyperscalers increasingly want processors optimized for their own models, cloud services, and power constraints.
Nvidia remains central because its GPUs support a broad software environment and many different workloads. Custom accelerators offer another route for large, predictable workloads where buyers can justify years of engineering.
Broadcom and Marvell help customers take that route. Their competition therefore sits inside a larger contest between merchant platforms and customer-specific systems.
The two routes are not mutually exclusive. Google offers its own TPUs while continuing to provide Nvidia GPUs through Google Cloud.
That mixed strategy reduces dependence on any single architecture. It also allows customers to select hardware based on model, software, availability, and cost requirements.
Broadcom’s earnings call should therefore avoid framing the market as a simple replacement battle. A better test is whether Broadcom’s content per system continues rising across a mixed accelerator environment.
Networking is central to that test. GPU and XPU clusters both require high-speed connections, switches, optical components, and rack-scale coordination.
If Broadcom loses some processor work but expands networking content, its total opportunity can remain substantial. If competitors take both compute and networking positions, the pressure becomes more serious.
The upcoming report needs enough detail to distinguish those outcomes. Aggregate AI revenue growth will not explain which customers, product categories, and deployments are driving the increase.
The Real Earnings Test Is Conversion, Not Bookings
Broadcom must show that large orders and capacity commitments are becoming shipped systems without creating an unsustainable gap between expectations and execution.
Second-quarter AI bookings above $30 billion indicate demand far beyond the products Broadcom shipped during that period. Bookings can improve visibility, but they are not identical to recognized revenue.
Custom chips follow long development and deployment cycles. Designs require verification, manufacturing, packaging, system integration, and customer qualification before volume revenue appears.
A delay at any stage can move revenue between quarters. Data center power, construction, cooling, and financing can also constrain deployment after chips are ready.
Broadcom’s increased inventory reflects confidence in the second-half ramp. It also raises the cost of a schedule change if customers alter deployments or supply assumptions.
The September report should clarify whether Broadcom achieved its projected $16 billion in third-quarter AI semiconductor revenue. It should also explain the balance between XPUs and networking.
That balance influences both growth quality and margins. Broadcom projected consolidated gross margin near 74 percent, below the second quarter’s 77.1 percent.
Management attributed the decline to a larger semiconductor mix relative to higher-margin infrastructure software. It did not characterize the change as a structural deterioration in semiconductor economics.
Investors still need to examine the claim carefully. Faster hardware growth can lift revenue while reducing the company-wide gross-margin percentage.
Operating leverage can offset that mix effect. Broadcom forecast an operating margin near 67 percent, roughly stable from the previous quarter.
Another issue is customer concentration. Broadcom names six core custom-chip customers, but a limited number of hyperscalers drive much of the planned growth.
Google is the most established program. Meta and OpenAI represent later ramps, while Anthropic’s access model depends partly on continued commercial performance.
Broadcom said it had delivered silicon for OpenAI and expected production late in 2026. It also described a larger deployment commitment beginning in 2027.
Meta’s program is expected to begin deliveries in the second half of 2027. Two additional customers were expected to start shipments in late 2026 and accelerate during 2027.
These schedules create a bridge between current Google-led revenue and broader diversification. The bridge is promising, but several programs remain earlier than Google’s production history.
That leaves Broadcom exposed to two related risks. Google can diversify suppliers before newer customers reach scale, or those newer programs can ramp more slowly than expected.
Neither outcome requires AI demand to disappear. Even a strong market can disappoint when revenue arrives later or carries less content than investors assumed.
The wider spending environment remains enormous. Alphabet, Amazon, Meta, and Microsoft were expected to direct hundreds of billions toward data centers during 2026.
An AI spending review reported that the four companies planned up to $720 billion of combined spending, primarily for AI infrastructure.
That total supports the demand argument, but spending does not flow evenly to each supplier. Cloud companies can change accelerator mixes, defer projects, or negotiate lower component costs.
Revenue concentration also gives individual customers unusual influence. One architecture decision can affect several years of engineering work and manufacturing capacity.
Broadcom’s long-term agreements reduce some uncertainty. They cannot remove adoption, competitive, financing, or execution risks.
The company’s fiscal 2027 forecast above $100 billion remains the largest point of tension. Broadcom reiterated that level after its second quarter rather than raising it.
A restrained forecast can prove conservative. It can also indicate that management sees constraints not captured by headline bookings.
The coming earnings discussion should connect orders to specific production windows. It should also identify which assumptions support the 2027 target without disclosing customer-confidential details.
Broadcom’s Other AI Customers Reduce Risk, but Google Still Sets the Standard
Broadcom’s expansion beyond Google strengthens its long-term case, although those programs have not yet matched Google’s scale or production maturity.
Broadcom has described engagements with Google, Anthropic, OpenAI, Meta, and two other core customers. That roster gives the company exposure to several distinct AI business models.
Google operates consumer services, develops Gemini models, and sells cloud infrastructure. Its TPU program has also progressed through several hardware generations.
Anthropic develops frontier models and sells access through enterprise products and cloud partners. Its arrangement connects model demand with TPU-based compute supplied through Broadcom and Google.
OpenAI is pursuing custom silicon while continuing to use large quantities of merchant accelerators. Broadcom said production for their joint program remained scheduled for late 2026.
Meta is developing MTIA, its internal accelerator family for AI workloads. Broadcom expects its work with Meta to expand through multiple generations.
These customers want different system characteristics. Training large models prioritizes memory capacity, communication bandwidth, and cluster reliability.
Inference emphasizes response time, utilization, energy consumption, and serving cost. Recommendation systems can require another balance of memory access and specialized computation.
Broadcom can reuse intellectual property across customers, including interfaces and networking technology. However, each custom program also requires dedicated engineering and validation.
That model differs from selling one standard processor to many buyers. A custom design can deepen a relationship, but it can also increase dependence on a customer’s roadmap.
Google remains the best test of Broadcom’s execution because the partnership combines long history, multiple TPU generations, and large-scale deployment.
If Broadcom maintains its Google position while adding Meta and OpenAI, customer diversification becomes a genuine strength. Revenue growth would then rely on several independent deployment cycles.
If Google reallocates material work before those programs scale, diversification would arrive later than the competitive pressure. That timing difference matters more than the customer count alone.
The Broadcom Google agreement through 2031 provides meaningful visibility. Still, investors need details about the scope of work within each future generation.
A supplier can remain under contract while its share of a system changes. More components can also be split among processor, networking, memory, optics, and packaging vendors.
Broadcom argues that its combined silicon and networking capabilities create differentiated value. That position is credible because modern AI systems depend on communication as much as individual chip speed.
Large models distribute computation across thousands of processors. Slow or inefficient connections can leave expensive accelerators waiting for data.
Broadcom’s Tomahawk switches and SerDes technology target that bottleneck. Co-packaged optics, which places optical connections close to switch silicon, offers another path for improving bandwidth and power efficiency.
These products extend Broadcom’s opportunity beyond any single TPU award. They can support clusters built with custom XPUs or merchant GPUs.
Competition remains intense. Marvell sells custom silicon and networking products, while Nvidia combines accelerators with its own networking platform.
AMD competes in accelerators and processors. Hyperscalers also maintain internal engineering teams that can absorb more design responsibility over time.
Broadcom therefore needs to defend both technological content and commercial relevance. A long customer list does not guarantee that each relationship produces attractive economics.
The next useful evidence will come from production milestones. Delivered silicon, qualified systems, customer deployments, and recognized revenue carry more weight than broad capacity announcements.
For technology buyers, this competition has practical consequences. More custom accelerators can widen infrastructure choices, but they can also fragment software and operational workflows.
Teams evaluating AI systems must compare more than benchmark results. They need to track model compatibility, capacity availability, networking, deployment schedules, and long-term support.
Maintaining a searchable record of technical claims can help teams evaluate those changes over time. A structured engineering knowledge base can connect announcements, benchmarks, and deployment notes without treating marketing claims as settled facts.
What to Watch After Broadcom Reports
Three signals will determine whether Broadcom’s Google exposure remains a durable advantage or becomes a concentration problem.
The first signal is third-quarter AI semiconductor revenue and its product mix. Broadcom previously forecast approximately $16 billion, up more than 200 percent from the prior year.
Meeting that target would confirm that the second-half production ramp is underway. Missing it would raise questions about deployment timing, manufacturing, or customer demand.
The internal mix matters just as much. Growth led by both XPUs and networking would support Broadcom’s claim that it gains content across complete AI systems.
Growth concentrated in one category would be less informative. It might reflect a temporary shipment schedule rather than a broader expansion.
The second signal is management’s description of the Google relationship after the Marvell agreement. Broadcom does not need to identify confidential chip assignments.
It does need to explain whether its 2031 agreement, expected content, and future TPU work remain consistent with previous assumptions.
A clear statement that multiple generations remain on schedule would strengthen the long-term case. Vague language or reduced visibility would increase concern about supplier diversification.
Investors should also listen for changes in terminology. “Strategic,” “substantial,” and “long term” sound reassuring, but they do not quantify program share.
Production timing, design milestones, and component categories provide better evidence. The most useful commentary will connect contractual commitments with actual deployment work.
The third signal is progress across OpenAI, Meta, Anthropic, and the two unnamed customers. Broadcom has presented these programs as the foundation for growth beyond Google.
OpenAI production was expected to begin late in 2026. Any confirmation of qualified silicon and initial shipments would strengthen the diversification argument.
Meta deliveries are scheduled later, beginning in the second half of 2027. Updates on orders or design milestones would help show whether that schedule remains intact.
Anthropic’s position requires separate treatment because its arrangement involves access to TPU-based compute capacity. Consumption depends partly on Anthropic’s continued commercial success.
Broadcom should therefore distinguish chip shipments, infrastructure access, customer commitments, and final utilization. Combining them into one capacity figure can obscure important economic differences.
These three signals also frame the central Broadcom earnings debate. The question is not whether AI infrastructure demand exists.
The question is whether Broadcom can convert that demand into durable revenue while customers add suppliers and negotiate harder.
Google’s custom-chip strategy makes the issue visible because it combines rapid technical progress with deliberate vendor diversification. Broadcom remains deeply involved, but Marvell now has a stronger route into adjacent work.
A strong report would show that the market is expanding faster than Broadcom’s share is being challenged. It would also provide evidence that newer programs are moving from design into production.
A weaker report would not necessarily invalidate custom silicon. It would indicate that forecasts moved ahead of deployment schedules, customer concentration, or supplier economics.
Readers should treat any immediate stock reaction with caution. One trading session can reflect positioning, expectations, or changes in guidance rather than the long-term value of the technology.
Broadcom Google chip development will remain strategically important after the earnings call. The relationship covers future TPUs and related networking through 2031, while Google continues expanding its supplier base.
The best next step is to compare management’s new statements with its previous commitments. Track the $16 billion AI revenue target, Google program language, and production milestones across newer customers.
Those facts will reveal whether Broadcom is broadening its AI business or relying on expectations that still need years of execution. What evidence would change your view: faster shipment growth, clearer Google commitments, or verified progress from the newer chip programs?
