Beyond NVIDIA: 4 Custom AI Chip Designers Gaining Global Market Share

For a time, investing in AI seemed straightforward: buy the company selling the tools. In this case, the tool was a GPU, an expensive one, with even pricier software to go with it.

That approach paid off. NVIDIA became the clear winner by providing the hardware everyone needed, especially when demand was high, and alternatives were limited. When people rush to buy computing power, the top supplier usually benefits.

Eventually, the biggest cloud companies stopped asking how to get more GPUs and started wondering why they were paying for this setup in the long term.

That’s where custom chips come into play.

The next stage of AI infrastructure isn’t about getting rid of GPUs completely. Instead, it’s about moving repetitive, predictable, and costly tasks to custom chips that handle them better. Training advanced models still relies on GPUs, but tasks like inference, recommendations, and certain cloud services are shifting to specialised chips.

This shift is important because large cloud companies operate at such a big scale that even small improvements add up fast. Cutting power use, latency, or costs by a little can save billions. In AI, saving 20% isn’t just a technical win; it can change the whole business model.

General-purpose GPUs are great because they’re flexible. They can handle many different tasks, and when things are changing quickly, that flexibility is very valuable.

But flexibility comes at a cost.

GPUs need to handle many tasks, so they have extra features and complexity. This is helpful when workloads are unpredictable, but less appealing when a cloud provider knows exactly what it needs the chip to do at a large scale every day.

A custom chip can be designed specifically for that job.

This usually leads to less wasted chip space, better power efficiency, improved use, and lower overall costs. It also gives companies more control, since they don’t have to wait in line with everyone else to buy the same chips from the same supplier.

This part often gets overlooked. Custom chips aren’t just about better performance; they also give companies more leverage.

If the AI boom is like a land rush, then having your own custom chips is like owning the factory instead of just buying tools from a store.

Many companies are part of this shift, but the four most important right now are Google, AWS, Broadcom, and Marvell.

Two of these are large cloud providers making their own chips to improve their business. The other two are chip companies helping make these systems work at scale.

They each play a different role, which makes this trend interesting.

Google has been working on custom AI chips for longer than most people realise. Its TPU programme is not a side project. It is a serious attempt to reshape the economics of machine learning inside Google and across Google Cloud.

The architectural idea is simple enough to explain and hard enough to execute. TPUs are built around systolic arrays, which are highly specialised for the matrix operations that dominate many AI workloads. Instead of trying to be broadly programmable like a GPU, the chip focuses on moving data through a tightly structured compute fabric very efficiently.

This makes TPUs less flexible. It also makes them very good at the jobs they are built for.

That trade-off is the whole point. Google controls not only the chip, but also much of the surrounding software stack through TensorFlow, JAX and compiler tooling. That means it can tune the entire path from model to machine. In AI infrastructure, owning the chip is useful. Owning the way workloads are shaped to fit the chip is better.

What Google has really built is not just a processor. It is a system. And systems tend to be stickier than components.

If a customer optimises around TPU instances, Google is no longer merely renting compute. It is selling an environment.

Amazon’s chip strategy is broader and, in some ways, more commercially obvious.

AWS is building a family of infrastructure silicon rather than one flagship product. Graviton handles general-purpose compute, Nitro offloads networking and security, Inferentia targets inference, and Trainium is designed for training. The names sound like they were generated by a branding committee locked in a room with a whiteboard and too much coffee, but the strategy is sensible.

Amazon wants to own more of the economics of AWS.

That matters because the cloud is a margin business disguised as an innovation business. Customers want performance and convenience. Providers want efficiency at scale. If Amazon can serve more AI demand on infrastructure it controls more directly, it can improve its cost structure and reduce its dependence on third-party suppliers.

This is especially relevant for inference. Training attracts headlines because it is dramatic and expensive. Inference is where usage compounds. Once a model is in production, it has to keep answering questions, processing requests, ranking ads, recommending products and doing whatever useful thing justified building it in the first place. That is where cost efficiency becomes ruthless.

A custom chip does not need to be universally loved. It needs to be cheaper for the workloads that run all day.

Amazon understands this very well.

Broadcom is the sort of business the market periodically rediscovers and then acts surprised by, even though its role is sitting in plain sight.

The company is increasingly central to custom AI silicon, not because consumers know the brand, but because hyperscalers need a partner that can help design, package and connect these systems at an industrial scale. Broadcom is showing up behind the scenes where technical ambition meets manufacturing reality.

That is valuable territory.

The custom chip story is not just about drawing a clever design. It is about getting that design built, integrated and deployed into a network that actually works under real-world conditions. This is where Broadcom has become hard to ignore. It has meaningful exposure not only to custom ASIC design but also to the switching and networking layers that keep giant AI clusters functioning.

And this is where the narrative gets better.

Investors like to talk about compute because it is glamorous. More tera-this, more peta-that, more chips per rack, more whatever. But once clusters get large enough, the elegant bottleneck moves from raw compute to data movement. A very fast accelerator waiting on a congested network is just an expensive heater.

Broadcom sells much of the plumbing that prevents that from happening.

That plumbing may be less cinematic than a GPU launch. It may also be more durable.

Marvell plays a different role. It is smaller, more selective and often framed as the challenger rather than the incumbent. That framing is fair, but slightly understates the opportunity.

In AI infrastructure, you do not need to beat every rival everywhere. You need to win the right sockets in the right systems.

Marvell has been building a position in custom silicon, DPUs, optical connectivity and data movement. All of those matter more as AI systems become larger, denser and more distributed. If Broadcom is the giant with broad strategic leverage, Marvell is the company picking up valuable share among hyperscalers seeking alternatives, flexibility, and specialised engineering support.

Its DPU strategy is a good example. Offloading networking, storage, and security functions from the main accelerator sounds technical because it is. It is also financially intuitive. The less time a premium AI chip spends doing infrastructure chores, the more time it spends doing billable compute. That improves the economics of the whole system.

Marvell also has exposure to optical and interconnect technology, which matters because electrical scaling has limits and AI clusters are running into them. There is only so much data you can push around inefficiently before physics sends you the bill.

That makes Marvell relevant in a part of the stack that is getting more important, not less.

The cleanest way to think about this theme is that AI infrastructure is maturing from a chip market into a systems market.

In the early phase of a boom, the most important product is the one in shortest supply. For AI, that was the GPU. Fair enough.

As the market develops, buyers get more sophisticated. They stop asking for the scarce thing and start asking what architecture actually gives them the best economics at scale. That question leads naturally toward custom chips, proprietary networking, compiler optimisation, power efficiency, packaging and interconnect design.

In other words, it leads away from a single product and toward an integrated stack.

That does not make NVIDIA irrelevant. Far from it. NVIDIA still has the dominant software ecosystem, enormous developer mindshare and a product set that remains essential for frontier training and flexible workloads. But dominance in one layer does not prevent value from appearing in adjacent ones.

Sometimes the most profitable part of a gold rush is not the gold. It is the roads, the railways, and the people selling highly specialised equipment to miners who suddenly discover they need far more infrastructure than they expected.

AI is beginning to look like that.

The lazy version of this argument says custom chips are coming for NVIDIA. That is too neat.

A better way to see it is this: hyperscalers are trying to reserve GPUs for the workloads where GPUs are most valuable, while shifting repeatable production jobs onto cheaper, more optimised hardware where possible.

That can be bad for NVIDIA at the margin without being catastrophic in absolute terms.

It is entirely plausible that NVIDIA continues to grow while losing some share. When the market itself is expanding rapidly, those two things can coexist quite happily. Investors sometimes struggle with this because “share loss” sounds bad in isolation. It is less dramatic when the pie is getting much larger.

The bigger question is where the incremental economics settle.

If more AI workloads move onto proprietary silicon, then more value accrues to the hyperscalers controlling deployment, and to the semiconductor companies enabling that custom buildout. Broadcom fits that description well. Marvell increasingly does too. Google and Amazon benefit through better cloud economics and tighter customer lock-in.

That is why this theme matters.

The AI story is not becoming less attractive. It is becoming less singular.

The semiconductor hierarchy beneath AI is starting to fracture, but not in the way headlines usually imply.

This is not a simple story of one winner being toppled by four challengers. It is a story of the market getting more specific. General-purpose compute still matters. But once workloads become large enough and predictable enough, custom infrastructure starts to make financial sense. And when it makes financial sense at hyperscaler scale, it usually spreads.

That is what is happening here.

Google and AWS are proving that custom silicon can become a strategic asset inside the cloud. Broadcom and Marvell are proving that the companies enabling those systems can capture meaningful value as well.

The next phase of AI may still be powered by accelerators.

It just may not all look like a GPU anymore.