Nvidia’s moat is getting attacked from multiple sides

Key takeaways

• Cerebras attacks Nvidia on inference speed, while Google pushes deeper into custom chips with Marvell.

• Hyperscalers increasingly want chips designed around specific workloads rather than one processor doing everything.

• Nvidia still has powerful software and networking advantages, but the AI chip market is becoming more specialised.

Nvidia has spent the artificial intelligence boom selling the industry's favourite all-purpose engine. Now, some of its biggest customers want engines built specifically for them, while smaller rivals are redesigning the engine altogether.

Two developments this week capture that shift. Cerebras unveiled new hardware on 18 August designed to make artificial intelligence (AI) inference dramatically faster. One day later, Google expanded its custom-chip relationship with Marvell Technology, potentially creating up to 120 billion USD of revenue for Marvell through fiscal 2033 if performance targets are reached.

Investors noticed. On 19 August, Marvell closed at 237.27 USD, up almost 10%. The bigger story, though, is not one day's share-price move. It is that Nvidia increasingly has to defend its position from two directions at once.

The GPU is becoming one lane, not the whole motorway

Nvidia's graphics processing units (GPUs) became the AI standard because they are powerful and flexible across many workloads. That matters especially for training advanced models, where requirements change quickly.

Inference is different. Once a model is trained, it repeatedly answers queries, making specialised chips potentially more efficient. Google has its tensor processing unit (TPU), Amazon has Trainium, while Meta is developing its Meta Training and Inference Accelerator (MTIA). These chips trade some flexibility for better performance, power efficiency or cost on specific tasks.

The market is therefore splitting. Nvidia offers programmable GPUs, hyperscalers such as Google and Amazon build custom processors, while specialists such as Cerebras rethink the architecture entirely. Marvell and Broadcom sit behind many custom designs, providing chip, networking and connectivity expertise.

Different jobs are getting different tools.

Cerebras attacks the architecture

Cerebras pushes specialisation much further.

Instead of linking many conventional chips, it builds processors roughly the size of a dinner plate. Its new CS-4 system uses three, keeping computing and memory closer together and reducing costly data movement.

Cerebras says the system uses 50% fewer components than its previous design and targets inference, where faster responses can improve both user experience and data-centre economics.

That does not make Cerebras an automatic Nvidia replacement. Nvidia still benefits from scale, software, reliability and a broad ecosystem. But Cerebras raises a more important question: does every AI workload need a GPU?

Google wants a chip made for the job

Google represents another form of pressure.

Hyperscalers operating enormous data centres increasingly have enough scale to design chips around their own workloads. Google's expanded Marvell agreement covers processors, storage and networking around its TPUs.

This looks more like supplier diversification than Broadcom being pushed aside. Broadcom separately agreed to develop future Google custom AI chips through 2031.

The lines between competitor and partner are also blurring. Marvell helps customers develop custom silicon while working with Nvidia, including through Nvidia's NVLink ecosystem.

Nvidia's defence may therefore be broader than protecting GPU market share. Even when customers use another processor, Nvidia can still try to remain part of the surrounding system.

Where the attack can stall

Specialised chips only make economic sense at sufficient scale. Designing them is expensive, manufacturing can be difficult, and software must work reliably. A benchmark advantage matters little if customers cannot deploy it easily.

Nvidia can also defend its position by improving inference performance while keeping customers inside its software, networking and development ecosystem. The early warning sign for investors is therefore not another impressive chip launch. It is whether workloads actually move.

The opposite risk also matters. AI computing demand could grow quickly enough for Nvidia, custom-chip designers and specialists to expand simultaneously. Market share can fall while revenue still rises if the market underneath it grows even faster.

Investor playbook

• Watch deployment, not demonstrations. Custom chips matter more when hyperscalers put them into production across large, recurring workloads.

• Separate training from inference. Nvidia may remain stronger in one part of AI even as alternatives gain ground elsewhere.

• Track the surrounding infrastructure. Networking, memory and connectivity can gain importance as data centres mix several types of processors.

• Treat custom-chip suppliers carefully. Large design wins create opportunity, but dependence on a handful of enormous customers also creates concentration risk.

A moat can narrow without disappearing

The important question is no longer whether somebody can build a faster AI chip than Nvidia. Somebody usually can, for a particular workload under particular conditions. The harder question is whether alternatives can match Nvidia's combination of performance, software, networking, supply and ease of deployment at scale.

That is why Cerebras and Google's Marvell deal matter together. One attacks the architecture. The other attacks the assumption that hyperscalers need to buy a standard processor in the first place. Neither means Nvidia's moat disappears. But AI computing is becoming more specialised, and that makes the moat more complicated to defend. The castle still looks formidable. There are simply more roads around it.