Marvell’s Google AI chip deal could unlock $120bn as Broadcom faces new TPU rival

Marvell Technology, Inc. (NASDAQ: MRVL) has expanded its relationship with Google LLC through a broad custom silicon agreement covering artificial intelligence inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute associated with Google’s Tensor Processing Unit ecosystem. The agreement is backed by warrants allowing Google to purchase up to approximately 58.97 million Marvell shares at $206.58 each, with most of the potential equity vesting linked to future commercial purchases rather than being granted immediately. If Google reaches the highest purchase thresholds contemplated by the structure, the relationship could generate roughly $120 billion of revenue for Marvell through fiscal 2033, a scale capable of transforming a semiconductor company that generated $8.20 billion of total revenue in fiscal 2026. The agreement also introduces a significant new competitive variable for Broadcom Inc. (NASDAQ: AVGO), although Broadcom separately holds a long-term Google agreement covering future TPU generations and related AI rack components through 2031, making supplier diversification a more accurate interpretation than immediate displacement. Marvell shares closed at $237.27 on August 19, up 9.9% during the session, approximately 9.3% over five trading sessions and 21.7% over one month as investors reassessed the company’s position in hyperscale custom silicon.

Why does Google’s Marvell agreement matter if Broadcom already has a TPU contract running through 2031?

The immediate temptation is to describe Marvell as taking Google’s Tensor Processing Unit business away from Broadcom, but the disclosed agreements do not support such a simple conclusion. Broadcom announced in April that it had entered into a long-term agreement to develop and supply custom TPUs for future Google generations, alongside a supply assurance agreement for networking and other components used in next-generation AI racks through as late as 2031. That arrangement remains current, meaning Google has not publicly terminated its longstanding Broadcom relationship merely because Marvell now has a larger role.

The more important development is that Google appears to be diversifying the semiconductor architecture surrounding its artificial intelligence platform. Modern AI infrastructure contains far more custom silicon than the central accelerator itself, with specialised processors controlling storage, networking, memory movement, inference and increasingly the movement of data between different memory tiers. Marvell’s agreement spans several of those categories and explicitly includes artificial intelligence inference accelerators, giving it access to a broader opportunity than supplying peripheral connectivity components alone.

Diversification also reduces Google’s supplier concentration. The amount of computing required for Gemini, Google Cloud and external TPU customers is becoming too large for Google to depend unnecessarily on a single semiconductor design partner when capable alternatives exist. Maintaining relationships with both Broadcom and Marvell can increase engineering flexibility, improve supply resilience and give Google greater commercial leverage as spending reaches tens of billions of dollars.

For Marvell, merely becoming a second source around Google’s TPU ecosystem can still be transformative. The custom silicon market is expanding quickly enough that winning part of a growing programme may be worth substantially more than winning all of a smaller historical programme. This is why the market reaction should not automatically be interpreted as a zero-sum transfer of existing Broadcom revenue to Marvell.

Broadcom shares nevertheless fell 4.6% to $362.48 on August 19 because investors immediately reduced the probability that Broadcom will capture as much of Google’s future custom silicon wallet as previously assumed. The reaction reflects future share-of-wallet uncertainty rather than evidence that current contracts have disappeared.

How does the Google warrant turn chip purchases into potential ownership of Marvell?

The warrant structure is unusually important because it ties Google’s potential equity ownership directly to the commercial relationship. Marvell has granted Google the right to purchase up to approximately 58.97 million shares at an exercise price of $206.58, implying total exercise proceeds of about $12.18 billion if every warrant eventually becomes exercisable and Google chooses to exercise them.

Google is therefore not simply receiving a $12.2 billion stake for free. It would have to pay the stated exercise price for the shares, and most of the warrants become available only when commercial conditions are satisfied.

Approximately 1.3 million shares are scheduled to vest in equal quarterly instalments during the first year following execution. The much larger remaining portion is linked to discretionary purchases from Marvell beginning in the third quarter of fiscal 2027 and extending through the end of fiscal 2033, with vesting connected to defined revenue milestones.

This creates a powerful alignment mechanism. Google receives greater potential ownership as it buys more Marvell technology, while Marvell accepts possible dilution only when the associated commercial relationship becomes correspondingly larger.

Using Marvell’s expected second-quarter basic share count of approximately 899 million as an illustrative baseline, the full 58.97 million warrants would increase the share base by roughly 6.6%. After issuance, Google would hold approximately 6.2% of the enlarged share count before considering other changes in shares outstanding.

That would be meaningful dilution for existing investors, but the economics differ sharply from issuing 59 million shares merely to attract a customer. Full exercise would also bring approximately $12.18 billion of cash into Marvell at the $206.58 exercise price and would occur only after an exceptionally large commercial relationship had developed.

Reuters estimated that Google could become Marvell’s fifth-largest shareholder if the maximum warrant position were exercised. Such ownership would further align the companies beyond an ordinary supplier relationship, although Google would remain a minority investor rather than gaining control.

The structure resembles a growing pattern across artificial intelligence infrastructure in which large technology customers receive financial participation in suppliers whose capacity they help create. The attraction is obvious because a hyperscaler directing billions of dollars toward one semiconductor company can participate in some of the equity value created by that spending.

The risk is that increasingly circular relationships make the artificial intelligence supply chain harder for investors to analyse. Revenue generated from a customer that also becomes a shareholder is economically real when products are delivered and paid for, but the alignment can blur the traditional distinction between customer, strategic investor and financing partner.

How large could $120 billion of Google-linked revenue become relative to Marvell’s existing business?

The potential commercial scale is enormous relative to Marvell’s current financial base. Marvell generated record fiscal 2026 revenue of approximately $8.20 billion, meaning the roughly $120 billion maximum opportunity cited around the Google agreement is equivalent to more than 14 times an entire year of current company revenue.

That $120 billion should not be treated as contracted backlog or guaranteed sales. The outcome depends on Google making the purchases required under the commercial framework, Marvell executing the relevant designs successfully and the associated products reaching production at the contemplated scale.

Even an illustrative calculation shows why the market reacted strongly. If $120 billion were distributed evenly across seven fiscal years, the average would exceed $17 billion annually, more than twice Marvell’s total fiscal 2026 revenue. Actual revenue would almost certainly be uneven because semiconductor programmes move through design, qualification and volume production cycles, but the comparison demonstrates how transformative full achievement could be.

Marvell does not need to reach the maximum for the agreement to matter materially. Capturing only one-quarter of the theoretical $120 billion would represent approximately $30 billion of cumulative revenue, still almost four times the company’s fiscal 2026 sales.

The economics will ultimately depend on gross margin as much as revenue. Custom silicon can involve substantial engineering work and typically carries different margins from merchant semiconductor products because hyperscale customers negotiate aggressively and contribute to product specifications. A $10 billion programme generating weak returns is less valuable than a smaller programme earning strong margins.

Marvell’s broader intellectual property portfolio can improve those economics because the company does not begin each design from zero. It can reuse technologies spanning high-speed SerDes, Arm compute, storage, security, silicon photonics, advanced packaging, chiplets and custom high-bandwidth memory while adapting them to a specific customer architecture.

This ability to combine existing intellectual property with customer-specific designs is central to the custom ASIC model. Development expenditure can be spread across multiple large programmes while individual hyperscalers obtain processors designed more precisely for their workloads.

The Google agreement therefore creates two opportunities simultaneously. It can generate direct custom silicon revenue while giving Marvell stronger validation when pursuing similar programmes with other hyperscalers.

Why is Google broadening custom AI silicon beyond the central TPU accelerator?

Artificial intelligence infrastructure increasingly resembles a specialised computing factory rather than a collection of identical processors. The central accelerator receives most of the attention, but enormous quantities of silicon are required to feed those accelerators with data, move information between servers, manage storage and keep increasingly complex memory hierarchies productive.

Marvell’s agreement reflects this architectural fragmentation. The disclosed programmes include artificial intelligence inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute, allowing Marvell to participate across several bottlenecks created as AI clusters grow.

Inference is particularly important because the industry is shifting from primarily training models toward running those models for billions of user and software-agent requests. Inference workloads can reward specialised processors designed around particular models, latency requirements or power constraints rather than relying exclusively on general-purpose graphics processing units.

Google has already demonstrated the economics of custom acceleration through several generations of TPUs. Designing processors around its own workloads can reduce dependence on Nvidia Corporation and potentially lower computing cost when utilisation is sufficiently high to justify custom development.

The opportunity now extends beyond Google’s internal systems. Google Cloud increasingly offers TPU capacity to external developers and large artificial intelligence companies, turning proprietary silicon into a commercial cloud differentiator rather than an internal cost optimisation alone.

Anthropic provides one of the clearest examples. Broadcom, Google and Anthropic disclosed in April that Anthropic is expected to access approximately 3.5 gigawatts of next-generation TPU-based capacity beginning in 2027 as part of a wider multi-gigawatt commitment.

As external demand grows, Google needs a much larger semiconductor supply chain. One design partner may remain strategically important while additional suppliers handle complementary processors, alternative architectures or later generations.

Marvell’s portfolio is particularly relevant because the company already supplies technologies surrounding the accelerator. Its optical interconnects, Ethernet switching, storage controllers and custom ASIC expertise allow it to address the movement of data as well as the computation itself.

That distinction matters because AI performance increasingly depends on the entire system. An accelerator capable of extraordinary arithmetic performance remains underutilised when networking, memory or storage cannot deliver information fast enough.

The Google agreement therefore positions Marvell around one of the central semiconductor themes of the next several years: the value in artificial intelligence is spreading from the accelerator toward the infrastructure that keeps the accelerator busy.

Does Marvell’s Celestial AI acquisition become more strategically important after the Google deal?

Marvell completed its acquisition of Celestial AI in February, bringing the Photonic Fabric optical interconnect platform into its current portfolio. The technology uses optical connectivity to address scale-up communication between processors and memory across increasingly large artificial intelligence systems.

The acquisition initially looked like a strategic bet on an emerging architecture. The Google agreement makes that bet more interesting because hyperscalers increasingly need custom silicon, memory connectivity and optical interconnects to be designed as parts of one system rather than purchased independently.

AI clusters are becoming physically larger. Accelerators that once communicated primarily within one server increasingly need extremely high-bandwidth links across racks and eventually across multiple racks operating as one computing system.

Electrical interconnects become harder to scale as distance and bandwidth increase because power consumption and signal integrity deteriorate. Optical technologies can move information farther with greater bandwidth efficiency, making photonics increasingly important as clusters expand.

Marvell now owns a much broader connectivity stack following Celestial AI and XConn Technologies. The company has scale-out Ethernet switching, optical DSPs, silicon photonics, CXL technologies, advanced packaging and custom ASIC capabilities in addition to the acquired Photonic Fabric architecture.

This creates an opportunity to design custom silicon around the entire data movement problem rather than compete only for the central accelerator die.

Google is an attractive customer for such capabilities because it controls both the software workloads and physical data centre architecture. A hyperscaler capable of redesigning processors, memory and networking together can extract more benefit from specialised silicon than an enterprise buying isolated components.

The Celestial AI acquisition also increases execution risk. Marvell must integrate acquired engineering teams and convert photonics roadmaps into production products while simultaneously supporting rapidly expanding custom programmes.

The company’s capital allocation has become more aggressive as artificial intelligence opportunities expand. Investors need evidence that acquisitions such as Celestial AI and XConn contribute design wins and future revenue rather than simply increasing goodwill, intangible assets and operating expenses.

The Google relationship provides encouraging validation, although the disclosed agreement does not establish that every acquired technology will necessarily be deployed in Google systems. It demonstrates that Marvell now has access to a customer environment where those capabilities could become commercially relevant.

Why does the Marvell agreement matter for Nvidia even if Google is not trying to replace GPUs everywhere?

The rise of custom accelerators represents one of the most important long-term competitive risks facing Nvidia, but the relationship is more nuanced than a direct replacement cycle. Hyperscalers continue purchasing enormous numbers of Nvidia GPUs while simultaneously designing custom processors for workloads where they believe specialised architectures provide better economics.

Google was an early example through its TPU programme. Amazon Web Services developed Trainium and Inferentia, Microsoft developed Maia, and Meta Platforms has expanded its internal accelerator efforts as infrastructure spending increases.

The economic incentive strengthens with scale. Designing a custom processor can require hundreds of millions or billions of dollars of engineering and software investment, but a hyperscaler deploying millions of accelerators can spread that development cost across an enormous installed base.

Custom silicon can also be designed around the hyperscaler’s software stack, networking architecture and power limits. This potentially improves performance per dollar for predictable workloads such as inference.

Nvidia retains major advantages through CUDA, rapid product cycles, broad model compatibility and the ability to supply complete systems. Customers that require flexibility across rapidly changing workloads may continue preferring general-purpose GPU infrastructure even when a custom accelerator is cheaper for a particular task.

The result is likely to be a heterogeneous market rather than wholesale substitution.

That environment is attractive for Marvell because it does not require one customer architecture to defeat Nvidia globally. Marvell can help hyperscalers design specialised processors while simultaneously supplying networking and connectivity technology into GPU-based systems.

Its March collaboration with Nvidia around NVLink Fusion illustrates that dual positioning. Marvell can participate when customers build systems around Nvidia and when they develop custom accelerators intended partly to reduce reliance on Nvidia.

This makes Marvell less dependent on predicting which individual accelerator architecture dominates. The company can monetise the movement, storage and processing of data across several ecosystems.

The Google agreement strengthens the custom silicon side of that strategy without eliminating Marvell’s opportunity inside Nvidia infrastructure.

How serious is the competitive threat to Broadcom after investors erased more than 4% from $AVGO?

Broadcom remains the benchmark against which Marvell’s custom silicon ambitions are measured. It has long-standing hyperscale relationships, enormous engineering resources and substantial intellectual property across networking, connectivity and custom accelerators.

Its April Google disclosure provides particularly strong protection against claims of immediate displacement. Broadcom has a long-term agreement to develop and supply future generations of Google TPUs and a separate supply assurance agreement covering networking and other next-generation AI rack components through up to 2031.

The question is therefore not whether Broadcom remains a Google supplier. It clearly does.

The question is how much of Google’s rapidly expanding custom silicon spending Broadcom will capture relative to Marvell.

If Google previously relied overwhelmingly on Broadcom for TPU development and now allocates an increasing number of accelerators and attach processors to Marvell, Broadcom can continue growing while still losing potential market share.

That distinction explains the August 19 stock reaction. Broadcom fell 4.6% to $362.48 even though no disclosed contract was cancelled, while Marvell jumped 9.9%.

Investors were repricing future opportunity rather than current revenue.

The size of Google’s artificial intelligence investment means both companies may ultimately win. Morningstar characterised the development as evidence of an expanding opportunity rather than straightforward displacement, an interpretation consistent with the coexistence of the two disclosed agreements.

Competition could also benefit Google commercially. Maintaining credible alternatives gives the hyperscaler leverage when negotiating engineering fees, wafer allocation, production pricing and future roadmap commitments.

Marvell must now prove that it can execute at Broadcom’s level. Winning a design is only the beginning of a custom silicon programme because hyperscalers expect chips to reach production on schedule, hit demanding power and performance targets and scale reliably across enormous deployment volumes.

A significant failure could cost not only one generation but future programmes.

Broadcom’s incumbency therefore remains meaningful. Marvell has gained an important position, but the next several years will determine whether that position becomes a durable second franchise or eventually rivals Broadcom’s custom AI scale.

Do Marvell’s current financial results support the much larger company implied by the Google opportunity?

Marvell entered the agreement with improving fundamentals rather than depending entirely on future artificial intelligence promises. First-quarter fiscal 2027 revenue reached a record $2.418 billion, increasing 28% year over year and exceeding the midpoint of management’s guidance.

Data centre revenue increased 27%, supported by artificial intelligence demand across electro-optics, custom silicon, storage and switching. The company also reported record operating cash flow of approximately $639 million.

Management guided second-quarter revenue to approximately $2.7 billion, plus or minus 5%, implying around 35% year-over-year growth at the midpoint. Marvell has said revenue growth should accelerate through fiscal 2027 as artificial intelligence bookings translate into shipments.

The company is scheduled to report second-quarter results on August 27, making that release the next immediate test of whether the broader AI acceleration is progressing as expected.

Marvell has already raised its outlook for fiscal 2027 and fiscal 2028 because of stronger-than-expected demand across 800G and 1.6T optical products, Ethernet switching, data centre interconnect, custom XPU and XPU-attach technologies.

The Google agreement potentially extends that growth runway well beyond the current forecast period.

Investors should nevertheless distinguish design wins from recognised revenue. Custom semiconductor programmes can take several years from architecture definition through tape-out, qualification and mass production.

Revenue attached to the most ambitious Google programmes may therefore arrive significantly later than the headline agreement.

Marvell also incurs development expenditure before volume production begins. Operating expenses are rising as the company integrates acquisitions and increases engineering investment around advanced process nodes and artificial intelligence products.

Gross margin will provide another important signal. Marvell guided fiscal second-quarter GAAP gross margin between 52.1% and 53.1% and non-GAAP gross margin between 58.25% and 59.25%.

A rapidly expanding custom business can increase absolute profit while affecting the product mix and consolidated margin percentage. Investors should focus on gross profit dollars and operating leverage rather than assuming every dollar of future Google revenue carries the same margin as Marvell’s highest-value merchant products.

Is Google’s warrant ultimately shareholder-friendly despite potential 6% dilution?

The warrant creates a genuine dilution risk because nearly 59 million new shares would be substantial relative to Marvell’s existing share base. Based on an illustrative 899 million basic shares, full issuance would increase the count by approximately 6.6%, leaving Google with about 6.2% of the enlarged total before considering future repurchases or other issuance.

Evaluating that dilution in isolation would miss the commercial structure.

Most warrants vest only when Google purchases substantial volumes of Marvell products. Existing shareholders therefore surrender ownership percentage only when Marvell has generated corresponding revenue from one of the world’s largest technology companies.

Full exercise would also require Google to pay approximately $12.18 billion to Marvell, assuming the $206.58 exercise price applies across the maximum warrant amount. That would create significant additional corporate liquidity.

The exercise price sits below Marvell’s August 19 close of $237.27, meaning the warrants already possess economic value if vested and exercisable at current prices.

Google’s incentive increases when Marvell shares appreciate. At the same time, Marvell shareholders benefit when the commercial relationship driving warrant vesting contributes enough earnings and strategic value to support that appreciation.

The structure becomes unattractive only if dilution occurs without sufficient economic return from the customer relationship.

That appears less likely because vesting is linked primarily to purchases rather than merely the passage of time, although investors should review the detailed warrant provisions as future filings clarify the exact relationship between revenue milestones and share vesting.

Strategically, the warrant gives Google a reason to care about Marvell’s broader success. A hyperscaler that becomes a meaningful shareholder has an additional economic incentive to expand a supplier whose technology is central to its infrastructure.

The arrangement may also complicate Marvell’s relationships with other hyperscale customers. Amazon, Microsoft or Meta may prefer suppliers perceived as commercially neutral rather than financially aligned with Google.

Marvell will need to demonstrate that Google’s potential ownership does not compromise confidentiality or roadmap independence across competing customers.

Custom silicon companies routinely maintain strict information barriers because their customers compete directly. The possibility of Google becoming a major shareholder raises the importance of those governance protections.

Why did Marvell shares jump nearly 10% when much of the Google revenue remains years away?

Marvell closed at $237.27 on August 19, up 9.85% during the session. The stock had closed at $217.08 on August 12, producing an approximately 9.3% five-session gain even after the semiconductor sector experienced substantial volatility during the intervening period.

Compared with the July 20 close of $194.94, Marvell was up approximately 21.7% over one month. The stock has traded within a 52-week range of roughly $61.44 to $329.88, leaving the August 19 close about 28% below the high while standing almost 286% above the low.

The rally reflects the market assigning a higher probability that Marvell can become one of the major suppliers of custom AI infrastructure rather than remaining primarily a connectivity semiconductor company.

A Google relationship is especially valuable as validation because hyperscale custom chips require engineering expertise, manufacturing execution and intellectual property across multiple advanced semiconductor disciplines. Google expanding the relationship signals confidence that Marvell can support programmes at enormous scale.

Investor enthusiasm is also magnified by operating leverage. Marvell’s existing revenue is only a fraction of the maximum Google-linked opportunity, meaning incremental programme wins can change long-term earnings estimates materially.

The market reaction should nevertheless be kept in perspective. Marvell does not suddenly possess $120 billion of guaranteed backlog, and investors paying a higher valuation today are assuming successful execution across products whose revenue may not emerge for several years.

Sentiment is therefore strongly constructive but execution-sensitive.

The August 27 earnings report becomes the next near-term catalyst. Investors will want updated fiscal 2027 and fiscal 2028 guidance, commentary on custom silicon design wins and greater clarity around the timing of revenue associated with hyperscale customers.

The market will also watch whether management changes capital allocation after issuing such a large warrant. A potential future $12 billion cash inflow is economically relevant only if the warrants vest and are exercised, meaning Marvell cannot finance current operations as though that money already exists.

What are the key takeaways from Marvell’s Google custom AI silicon agreement?

• Marvell has expanded its Google relationship across AI inference accelerators, storage controllers, network interface controllers, memory interface controllers and near-memory compute associated with the TPU ecosystem.
• Google received warrants to purchase up to approximately 58.97 million Marvell shares at an exercise price of $206.58.
• Full exercise would require approximately $12.18 billion of payments to Marvell rather than representing a free $12.2 billion equity grant.
• Most of the warrant position is tied to future commercial purchases through fiscal 2033, aligning potential dilution with growth in Google-linked revenue.
• The maximum commercial opportunity has been estimated at roughly $120 billion through fiscal 2033, but that amount is conditional and should not be treated as guaranteed backlog.
• The $120 billion maximum is more than 14 times Marvell’s $8.20 billion fiscal 2026 revenue, demonstrating how significantly the relationship could change the company if Google reaches the upper purchase thresholds.
• Full warrant exercise could give Google roughly 6% of an enlarged Marvell share base using the company’s current expected basic share count as an illustrative denominator.
• Broadcom remains a major Google supplier under a separate agreement covering future TPU generations and AI rack components through up to 2031, making diversification more likely than immediate displacement.
• Marvell reported record Q1 fiscal 2027 revenue of $2.418 billion, up 28%, and guided Q2 revenue to approximately $2.7 billion before the new Google agreement was disclosed.
• Marvell shares closed at $237.27 on August 19, up 9.9% for the day, approximately 9.3% over five trading sessions and 21.7% over one month, while Broadcom fell 4.6%.

Can Marvell turn Google’s custom silicon opportunity into a second hyperscale franchise beside Broadcom?

The Google agreement changes the perception of Marvell’s artificial intelligence business because it connects several technologies the company has spent years assembling into a single hyperscale commercial opportunity. Custom ASIC expertise, optical connectivity, Ethernet switching, storage controllers, advanced packaging and the acquired Celestial AI photonics platform increasingly address different parts of the same problem: moving and processing enormous amounts of data efficiently inside AI infrastructure.

Google’s willingness to connect potential equity ownership with future purchases adds unusual strategic weight to that relationship. The warrant structure effectively says that when Google buys enough Marvell silicon to create substantial economic value, Google can participate in some of the equity value that relationship helps generate. Marvell accepts dilution, but only alongside potentially transformative revenue and more than $12 billion of possible exercise proceeds.

The agreement is not evidence that Broadcom has lost Google. Broadcom’s TPU and rack component contract remains in place through up to 2031, and the amount of artificial intelligence infrastructure Google intends to deploy may be large enough to support significant growth for both suppliers. The change is that Broadcom can no longer be assumed to own the overwhelming majority of every future custom silicon opportunity around the Google TPU ecosystem.

For Marvell, that distinction is enormous. The company does not need to replace Broadcom outright or challenge Nvidia across every accelerator workload. It needs to establish itself as a repeatable hyperscale design partner capable of winning several large pieces of each AI system and carrying those programmes into volume production.

The $120 billion maximum opportunity provides a compelling headline, but the more important evidence will emerge more gradually through tape-outs, production ramps, customer concentration, gross margins and cash generation. Semiconductor agreements create value only when designs become working chips that customers deploy at scale.

Marvell now has something much more valuable than a theoretical addressable market. It has an agreement with one of the world’s largest AI infrastructure buyers and an incentive structure designed to expand alongside that buyer’s spending. Whether it becomes the defining contract in Marvell’s transformation will depend on execution, but the August 19 market reaction shows investors already believe the probability has increased substantially.