When the conversation turns to AI chips, Wall Street still reflexively reaches for Nvidia, Intel and Broadcom. But two very different companies are now writing a more intriguing story — one that could reshape where profit pools form in the AI infrastructure boom. And neither looks anything like a conventional semiconductor play.
Amazon, the e-commerce and cloud colossus, has built a custom-silicon operation that now runs at an annualized revenue pace of more than $20 billion. Marvell Technology, a veteran in networking silicon, has just locked in a $2 billion strategic investment from Nvidia to become the indispensable connective tissue inside AI data centers. Amazon is using its cloud to invent a chip empire; Marvell is riding the networking layer to a windfall. Together, they underline a powerful shift: the spoils of AI hardware are migrating from those who simply design chips to those who control system-wide infrastructure.
Amazon has never worn the “chip company” label. That’s getting harder to justify. In its latest fiscal first-quarter results, the company disclosed that its custom silicon portfolio — including the Graviton general-purpose CPU, the Trainium AI training and inference accelerator, and the Nitro networking and storage virtualization chip — had reached an annualized revenue run rate north of $20 billion. That represents a sequential jump of nearly 40% and a year-over-year growth rate in the triple digits.
Even that eye-popping figure, CEO Andy Jassy warned, “somewhat masks the true scale.” During the earnings call, Jassy laid out an arresting thought experiment: if Amazon’s chip business were a standalone entity and sold its output to AWS and third-party customers in the manner of any leading chipmaker, its annual revenue would run to $50 billion. “By our estimates, our custom silicon business is now one of the top three in data center chips globally,” he said.
In a matter of a few years, Amazon has moved from being one of the chip industry’s biggest customers to a rival that sits at the same table as its largest incumbents. The Trainium franchise alone carries more than $225 billion in revenue commitments. Anchor clients Anthropic and OpenAI have respectively contracted for about 5 gigawatts and 2 gigawatts of Trainium compute capacity.
Trainium2 is effectively sold out. Trainium3 only started shipping in 2026 and delivers 30% to 40% better price-performance than its predecessor — and it is already almost fully booked. Even Trainium4, still roughly 18 months from volume shipments, has racked up substantial reservations. Every workload that migrates from a third-party GPU to Trainium lowers the compute cost paid by AWS. Jassy estimates that Trainium will save Amazon “tens of billions of dollars in capital expenditure” annually, while contributing a meaningful operating-margin advantage in inference.
If Amazon is pursuing vertical integration from cloud to silicon, Marvell Technology has taken a less conspicuous path — one that is equally essential.
Marvell designs the high-speed Ethernet switch chips that shuttle enormous datasets between AI server clusters with ultra-low latency. Its network interface cards and data processing units, or DPUs, offload tasks such as encryption and load balancing from central processors. That hardware never trains a large language model directly, but its impact is brutal when it fails: a single malfunctioning switch or a congested link can idle entire rows of GPUs, burning developers’ time and billions in sunk capital.
That critical role has now attracted the ultimate partner. Nvidia recently injected $2 billion of strategic capital into Marvell, with the two companies set to jointly accelerate next-generation Ethernet switch chips and DPUs purpose-built for Nvidia’s AI platforms. The deal hands Marvell a direct ticket into an ecosystem where hyperscale AI players are already pouring hundreds of billions of dollars.
This year, the five largest hyperscalers are expected to spend a combined $720 billion on AI capex. But the shape of that spending is shifting. Training workloads are still Nvidia’s fortress; inference, however, demands chips that are lower−power, massively deployable and more cost−efficient. Marvell’slow−power inference engines and custom architectures are built precisely for that moment. And with a market capitalization dwarfed by Nvidia’s nearly $5 trillion valuation, Marvell offers significantly more headroom for earnings beats and multiple expansion. In a market saturated with obvious AI winners, Marvell remains a sleeper.
On the surface, the two stories seem unrelated — one is a cloud giant that started making its own chips to slash costs, the other is a networking specialist that made itself indispensable to the GPU king. But the underlying wager is identical: AI compute value is no longer concentrated inside a single GPU; it is diffusing across the entire compute fabric, from servers to switches to the fiber that links them.
Companies that can deliver cloud scale, connection efficiency or system-level optimization — beyond what a pure-play chip designer can offer — are now grabbing the next set of tickets in the AI infrastructure buildout. For investors uneasy about chasing pure chip names at stretched multiples, Amazon and Marvell offer an alternative lens. Their latest earnings report, and the strategic check from Nvidia, simply make that logic impossible to ignore.
