· 8 min read
Big Tech does not want to buy every AI chip from the same shelf.
Nvidia remains the center of the AI-computing boom, but cloud companies are also designing custom silicon for specific workloads. Marvell’s rising forecasts show how valuable that second market is becoming.
Fig. — One workload, many possible chips.
The easiest way to imagine the AI-chip market is as a race between Nvidia and everyone else.
The real market is becoming more complicated.
Cloud companies still buy enormous numbers of general-purpose AI accelerators, but they are also designing chips for workloads they understand extremely well. That creates a growing market for custom silicon.
Why general-purpose GPUs became dominant
GPUs are flexible. A company can use the same broad computing platform for training different models, running experiments and serving many types of workloads.
That flexibility is valuable when the software is changing quickly.
Nvidia also built a deep software ecosystem around its hardware, which makes switching costs much larger than the price of the chip alone.
Why custom chips still make sense
Once a workload becomes predictable and massive, flexibility can become less important than efficiency.
A cloud company may know exactly which operations it performs billions of times every day. Designing silicon around that pattern can improve performance per watt, reduce cost or reduce dependence on a single supplier.
Custom does not mean completely in-house. Companies can define the architecture while partners such as Marvell help with design, networking, packaging or production.
Marvell is a useful signal
Reuters reported that Marvell raised its fiscal 2028 revenue forecast to about $20 billion as demand for custom data-center chips increased.
The company projected $12 billion in custom-chip revenue for fiscal 2029, up from a previous target of $10 billion.
Those forecasts do not prove every custom-chip project will succeed. They do show that hyperscalers are spending enough on alternatives to create a very large business around them.
Custom chips do not automatically replace Nvidia
A company can use both.
General-purpose accelerators are useful for fast-moving research and workloads that need flexibility. Custom silicon can be attractive for high-volume tasks where the company can optimize aggressively.
The result is a layered market rather than a winner-takes-all replacement cycle.
The hidden issue is supply
AI demand has made advanced wafers, memory, packaging and networking strategic constraints.
Reuters reported that AMD is planning several years ahead to secure more advanced manufacturing and memory capacity as it expands AI-chip supply for 2027.
That means chip strategy is partly about resilience. More architectures and more suppliers can give cloud companies additional paths when one part of the supply chain is tight.
What to watch
The next battle is not only benchmark performance.
Watch cost per useful unit of work, software compatibility, memory bandwidth, interconnect speed, power efficiency and how quickly a company can bring a new design into production.
The most valuable AI chip may not be the fastest chip in isolation. It may be the chip that fits the workload, power budget and supply chain best.
