Google’s custom-silicon deal with Marvell Technology is striking for its potential size. But the more consequential part of the agreement may be its breadth.
The deal extends well beyond the tensor processing units, or TPUs, that Google has spent more than a decade developing as an alternative to general-purpose accelerators. Marvell says the relationship encompasses AI inference accelerators, storage controllers, network interface controllers, memory-interface controllers, and near-memory compute.
That suggests Google is extending the logic behind the TPU deeper into the AI data center. If specialized silicon can improve the economics of AI compute, why stop at the accelerator?
Marvell has so far been the company publicly spelling out the scope of the relationship. Google didn’t issue a separate announcement, leaving Marvell’s securities filings and subsequent earnings commentary as the principal public sources of detail about the agreement.
View All The deal also includes an unusual warrant allowing Google to acquire nearly 59 million Marvell shares at $206.58 each. Most of those shares vest progressively as Google purchases qualifying custom products, effectively tying the equity award to as much as $120 billion in cumulative revenue through Marvell’s fiscal 2033. The $120 billion figure is therefore not a guaranteed Google purchase commitment, but the maximum revenue level associated with full vesting of the warrant.
Marvell CEO Matt Murphy said during the company’s fiscal Q2 2027 earnings call yesterday that the warrant covers programs already underway, new design wins, and potential future programs. Revenue covered by the agreement through fiscal 2028 was already included in Marvell’s previous forecasts, he said, with the potentially larger incremental impact beginning in fiscal 2029.
More silicon around the TPU
According to Brendan Burke, research director for semiconductors, supply chain, and emerging tech at Futurum Group, the agreement isn’t about Marvell taking the TPU away from Broadcom. It’s about Google creating additional custom-silicon programs around it.
“This deal expands the pie,” he told EE Times.
Burke expects Broadcom to remain involved in the core TPU program, while Google adds specialized chips elsewhere in the system. The success of the TPU, he said, is encouraging Google to apply custom silicon to inference, memory, storage, and networking, potentially giving it a range of proprietary data center silicon that begins to resemble the breadth of portfolios offered by companies such as Nvidia and AMD.
That expansion reflects a broader change in AI architecture. Memory and data movement, once treated largely as supporting functions around the processor, are becoming increasingly important to overall system performance.
Murphy made a similar argument on Marvell’s earnings call. He said the company has maintained for years that customization would eventually spread to “every hop in the network,” rather than remain confined to accelerators. Five years later, he said, that’s increasingly what Marvell is seeing.
The shift is particularly important as AI moves from massive training runs toward inference workloads where enormous volumes of tokens must be generated economically.
Why hyperscalers can specialize
For hyperscalers, the economics of custom silicon are different from those faced by smaller cloud providers, according to Carmen Li, CEO of Silicon Data (which provides pricing, benchmarking, and market data for compute) and Compute Exchange ( a marketplace for buying and selling compute capacity). A neocloud selling bare-metal accelerator capacity may have customers explicitly asking for an Nvidia H100 or B200, she told EE Times. Google, AWS, or Microsoft, by contrast, often sells a service or workload outcome. A customer may care about latency, throughput, or the price of running an inference workload without knowing—or caring—which processor sits underneath it.
“They can use whatever they want to,” Li said.
That gives hyperscalers more freedom to shift appropriate workloads onto their own silicon when doing so improves economics. If Google can provide the required service level with a TPU or another specialized ASIC at lower cost, Li said, there may be little reason for the customer to insist on a general-purpose GPU.
But that doesn’t mean every workload belongs on an ASIC. “It all depends on the workflow,” Li said.
A relatively stable workload can be easier to optimize around because the hyperscaler can forecast its requirements and design silicon accordingly. Workloads that are changing rapidly, such as emerging multimodal applications, carry greater risk because the specialized hardware may prove less adaptable.
That means hyperscalers are likely to maintain a mix: custom accelerators for predictable workloads and general-purpose GPUs where flexibility remains valuable.
Inference makes the calculation more compelling because recurring workloads can potentially be optimized around cost, power, and performance. Murphy said AI infrastructure companies are increasingly focused on the cost and performance associated with each token generated as the industry moves into what he called a “monetization era.”
Google designed, partner engineered
The growing use of custom silicon raises another question: If Google already has one of the semiconductor industry’s most sophisticated internal design teams, why does it need Marvell or Broadcom?
Burke draws a distinction between defining what a chip should do and turning that architecture into a manufacturable system.
Google specializes in compute architecture, including processor cores and interconnect, he said. Merchant semiconductor partners contribute expertise in areas such as SerDes, physical interconnect, advanced packaging, HBM integration, and foundry execution—technologies that benefit from experience across multiple high-volume customers.
A Google custom chip can therefore be “Google designed, but partner engineered,” Burke said. Google defines the function and architecture, while the semiconductor partner helps supply physical subsystems, package the system, and manage the foundry relationship.
Marvell’s appeal is that it has assembled an unusually broad collection of technologies around those increasingly critical parts of AI infrastructure. Its portfolio spans custom compute, switching, SerDes, optical connectivity, memory controllers, CXL, storage, and advanced packaging.
Murphy repeatedly emphasized that breadth during the earnings call. In scale-up networking, for example, he argued that hyperscalers increasingly want partners capable of delivering an end-to-end roadmap encompassing switches, copper interconnect, near-packaged optics, and eventually co-packaged optics.
“Point solutions at this juncture, we believe, are not going to get it done,” Murphy said.
More leverage for Google
Adding Marvell also gives Google another source of negotiating power.
Burke said Broadcom retains considerable pricing power in its relationships with customers. Although the Marvell programs appear largely complementary rather than a direct tradeoff with Broadcom, a credible second design partner gives Google more bargaining leverage when negotiating future programs and potentially allows it to retain more of the economics of its proprietary chips.
Marvell, meanwhile, gains access to an enormous potential revenue stream, but the custom-silicon portion of its business could become increasingly dependent on Google if the relationship approaches its maximum scale.
The arrangement illustrates a broader shift in bargaining power as hyperscalers become more sophisticated semiconductor customers. Rather than selecting one custom-chip supplier, they can increasingly distribute programs among several partners while keeping the architecture and software under their own control.
That doesn’t necessarily mean Nvidia loses.
Asked whether custom accelerators would take meaningful economic share from Nvidia GPUs or whether both markets could continue expanding, Li said overall demand remains the crucial variable.
“Both will grow,” she said. “Both will grow tremendously.”
The bigger change may therefore be happening less in the contest between GPUs and ASICs than in the definition of custom silicon itself.
Google began with specialized processors for machine learning. The Marvell agreement indicates that specialization is now spreading outward—into memory, storage, networking, and the links that move data between processors.
The TPU may have been the beginning. Increasingly, the rest of the AI data center is becoming custom too.
Read also:
[Qualcomm Bets Open-Source AI Software Can Break Nvidia’s Lock-In](https://www.eetimes.com/qualcomm-bets-open-source-ai-software-can-break-nvidias-lock-in/)
[IBM Makes Quantum Cryogenics Modular, but Scaling Problems Remain](https://www.eetimes.com/ibm-makes-quantum-cryogenics-modular-but-scaling-problems-remain/)
[AI Hardware’s Next Frontier Is Integration](https://www.eetimes.com/ai-hardwares-next-frontier-is-integration/)