AMD and Google are gaining ground in the artificial intelligence chip market as businesses look for lower-cost alternatives to Nvidia’s expensive graphics processing units (GPUs). AMD’s Instinct accelerators are attracting customers looking to reduce computing costs, while Google’s Tensor Processing Units (TPUs) are gaining traction with companies seeking specialised hardware for AI workloads.

The shift reflects a gradual change in how businesses purchase AI computing. Rather than depending entirely on Nvidia, some customers are testing multiple chip suppliers to improve price efficiency, secure computing capacity and reduce dependence on a single vendor. Nvidia remains the dominant player, but improvements in AMD’s software ecosystem and Google’s expanding TPU business are creating more competition in a market that was once difficult for rivals to enter.

Key takeaways

  • AMD and Google are attracting more customers seeking alternatives to Nvidia’s AI chips.
  • AMD’s Instinct accelerators are gaining interest, particularly for AI inference workloads.
  • AMD’s ROCm software improvements have made it easier for developers to work with its processors.
  • One cloud provider reported that the time its employees spent helping customers resolve ROCm issues declined by 80% to 90% compared with the previous year.
  • Google is expanding access to its custom TPU chips beyond its own cloud infrastructure.
  • Customers are exploring Google’s TPUs because they can offer lower costs for selected AI workloads.
  • Google TPU 8 systems have attracted interest worth hundreds of millions of dollars in potential computing spending, according to a provider cited by The Information.
  • Nvidia continues to dominate the market, with its software ecosystem, manufacturing scale and product availability providing important advantages.
  • The emerging trend is multi-vendor AI infrastructure rather than an immediate replacement of Nvidia.

Why Are Customers Looking Beyond Nvidia?

Nvidia has built a powerful position in the AI chip market through its GPUs, networking technology and software ecosystem.

Its processors are widely used to train large AI models and operate AI applications. The company’s CUDA software platform has also become an important part of the development environment used by AI researchers and engineers.

This combination has created a strong competitive advantage.

However, Nvidia’s market position has also encouraged customers to look for alternatives.

The cost of operating AI infrastructure can be substantial. Businesses running large workloads need to consider the purchase price of processors, electricity consumption, data-centre expenses, software compatibility and ongoing maintenance.

When companies run AI applications at scale, even a modest reduction in the cost of computing can improve their operating economics.

The Information reported that some businesses are increasingly evaluating AMD and Google hardware as they seek more affordable ways to run AI models.

The shift is particularly visible in inference.

Inference refers to the computing process through which an already-trained AI model generates responses, produces content or performs tasks for users.

As AI applications move from experimentation into commercial deployment, inference workloads can become a significant recurring expense.

This creates an opportunity for chipmakers that can deliver suitable performance at a lower overall cost.

AMD’s Instinct Chips Are Becoming a More Serious Alternative

Advanced Micro Devices, commonly known as AMD, is one of Nvidia’s most important competitors in the AI accelerator market.

The company offers its Instinct series of accelerators for data centres and AI workloads.

For years, convincing customers to adopt AMD processors was difficult because Nvidia had a substantial advantage in software support and developer familiarity.

That situation is gradually changing.

According to The Information, companies that previously considered AMD hardware difficult to deploy are now taking its products more seriously.

One example is Superlinked, whose chief executive said the company had been testing AMD chips for several months and had become more interested in using them as smaller open-weight AI models gained popularity.

Open-weight models allow developers to access and use model parameters under the applicable licence conditions. They can be deployed on different infrastructure platforms, giving businesses more flexibility over where and how they run AI applications.

The availability of smaller models has created additional opportunities for lower-cost computing hardware.

For some workloads, businesses do not necessarily need the most expensive or highest-performance processor available.

They need hardware that can run the required model reliably while keeping operating expenses under control.

AMD’s ROCm software is improving

One of AMD’s biggest historical challenges has been its software ecosystem.

Nvidia’s CUDA platform has become widely used by developers, researchers and AI infrastructure providers.

This makes switching hardware more complicated than simply replacing one processor with another.

Developers may need to modify software, test workloads, resolve compatibility problems and optimise applications for a different architecture.

AMD’s answer is ROCm, its software platform for developing and running applications on AMD GPUs.

The Information reported that improvements to ROCm have reduced the technical support burden for some AMD-focused cloud providers.

TensorWave, a cloud provider specialising in AMD infrastructure, said the amount of time its employees spent helping customers resolve ROCm issues had declined by 80% to 90% compared with the previous year.

This is important because software usability directly affects the commercial attractiveness of a chip.

If customers can deploy applications more easily, they may become more willing to experiment with alternative hardware.

However, ROCm still competes with a mature CUDA ecosystem that benefits from years of developer adoption, software optimisation and industry support.

AMD-Focused Cloud Providers Are Seeing More Interest

The growing interest in AMD hardware is also creating opportunities for specialised cloud providers.

These companies offer computing infrastructure to businesses that need access to AI accelerators without purchasing and operating their own large data centres.

TensorWave is one example.

The company expects to reach an annualised recurring revenue run rate of approximately $210 million by the end of 2026, compared with $125 million at the end of 2025, according to The Information.

Annualised recurring revenue, or ARR, is a measure that converts recurring revenue into a yearly figure based on the current business run rate. It is not the same as audited annual revenue.

Another AMD-focused cloud provider, Sciforium, has also reported rapid expansion.

Its chief executive said the company had tripled its customer base since signing its first commercial contract in June. It had also increased computing capacity five-fold over the preceding year and nearly tripled its team.

Sciforium is raising additional funding to expand its infrastructure.

These examples suggest that demand for AMD-based AI computing is developing beyond product demonstrations and technical testing.

However, growth at individual cloud providers does not establish that AMD has displaced Nvidia across the broader AI accelerator market.

It demonstrates that a commercial ecosystem around alternative hardware is expanding.

Google Is Turning Its TPU Chips Into a Larger Business

Google is taking a different approach to challenging Nvidia.

Instead of primarily selling general-purpose GPUs, the company has developed custom processors called Tensor Processing Units.

TPUs are designed to accelerate machine-learning and AI workloads.

Google originally developed these chips to support its own services, including its AI models and cloud infrastructure.

It is now expanding their availability to external customers.

The Information reported that Google has been investing heavily in expanding the market for its TPUs, including supporting the development of data centres outside its own Google Cloud infrastructure.

The company has also been improving software compatibility to make it easier for external customers to use its chips.

One important development is the growing interest in Google’s TPU 8 systems.

According to The Information, Cirrascale, one of the early providers of Google TPUs outside Google Cloud, said customers had expressed interest in spending hundreds of millions of dollars on computing using the TPU 8 chip.

This indicates that Google is beginning to attract demand from businesses that previously relied more heavily on Nvidia hardware.

Google is targeting specialised AI workloads

Google’s strategy includes developing processors around specific computing requirements.

In November 2026, the company introduced two specialised TPU designs.

One is aimed at confidential computing, which is relevant for businesses processing sensitive information.

The other is designed for workloads that require handling larger volumes of data, including applications involving AI agents.

AI agents are systems that can perform multi-step tasks using models, software tools and external information.

These applications can require substantial memory, data movement and computing resources.

By developing hardware around particular workloads, Google is attempting to make its chips attractive to customers with specific infrastructure requirements.

This approach could help it compete without needing to replace Nvidia across every category of AI computing.

Why Lower-Cost AI Chips Matter to Businesses

The growing interest in AMD and Google hardware reflects a larger change in the economics of AI.

During the initial AI boom, many companies focused on obtaining enough computing capacity to train and deploy advanced models.

Availability was often more important than cost.

As AI applications become part of regular business operations, the financial equation changes.

Companies must consider how much it costs to generate each response, process each request or run an AI application continuously.

This makes inference economics increasingly important.

Consider a business operating an AI assistant that processes millions of requests every month.

The company needs to pay for computing resources every time users interact with the application.

If an alternative processor can deliver the required performance at a lower total cost, the savings can accumulate over time.

However, purchase price is only one part of the calculation.

Businesses also need to consider:

  • Performance for the specific AI model being deployed.
  • Power consumption and electricity costs.
  • Software compatibility and developer support.
  • Availability of processors and delivery schedules.
  • Networking and memory requirements.
  • Reliability and technical support.
  • The cost of migrating existing applications.

A cheaper chip is not automatically the better business choice if it requires expensive software changes or additional engineering resources.

The competitive opportunity for AMD and Google is to demonstrate attractive total cost of ownership, not simply a lower hardware price.

Nvidia Still Has a Major Competitive Advantage

Despite the progress made by AMD and Google, Nvidia remains the leading supplier in the AI accelerator market.

The Information reported that customers continue increasing their Nvidia chip orders even as they experiment with alternative processors.

Several factors explain why Nvidia remains difficult to displace.

CUDA’s software ecosystem

Nvidia’s CUDA platform has a large developer base and extensive software support.

Many AI frameworks, research tools and production applications have been optimised for Nvidia hardware.

Switching to another platform can involve engineering work, compatibility testing and performance optimisation.

Manufacturing scale

Nvidia’s large production volumes also provide an advantage when advanced chip manufacturing capacity is limited.

The Information reported that Nvidia’s relationship with Taiwan Semiconductor Manufacturing Company gives it important access to manufacturing capacity during periods of supply constraints.

This can influence how quickly customers receive the processors they need.

Strong customer support and relationships

Nvidia has also developed a broad commercial ecosystem around its products.

Industry executives cited by The Information described the company’s sales and customer-support organisation as particularly proactive.

Its support can extend beyond supplying processors to helping customers with technical integration, infrastructure planning and connections to other service providers.

These relationships make it harder for competitors to win customers solely through lower prices.

Availability can matter more than price

The Information also reported examples in which customers considered AMD hardware but ultimately chose Nvidia because the AMD systems would take longer to become available.

One cloud provider said a customer would have needed to wait approximately six months for AMD hardware, compared with 12 to 14 weeks for Nvidia systems.

In another example, a startup abandoned discussions about AMD hardware after learning that access could be delayed until the end of the year.

These examples highlight an important reality: customers need computing capacity when their projects require it.

A lower-priced processor is not useful if the business cannot obtain it on time.

AI Chip Market Is Moving Towards Multiple Suppliers

The developments involving AMD and Google point towards a more diversified AI infrastructure market.

Instead of choosing one supplier for every workload, businesses can increasingly evaluate different processors based on the task they need to perform.

For example:

CompanyAI hardwareCompetitive approach
NvidiaGPUs and integrated AI systemsBroad software ecosystem, performance and scale
AMDInstinct acceleratorsAlternative GPU platform with a focus on price-performance
GoogleTensor Processing UnitsCustom chips designed around specialised AI workloads

This approach can help customers reduce supplier concentration and negotiate better commercial terms.

It can also provide flexibility if one supplier faces production delays or capacity constraints.

However, managing multiple hardware platforms creates additional complexity.

Companies may need different software environments, development tools and infrastructure configurations for different processors.

The success of a multi-vendor strategy will depend on whether the benefits of lower costs and greater flexibility outweigh the additional operational burden.

AMD’s Big AI Customer Deals Strengthen Its Position

AMD’s progress is also supported by major commercial agreements with AI companies.

Reuters reported that AMD had agreements with OpenAI and Meta involving substantial future chip purchases.

Meta agreed to purchase up to $60 billion worth of AMD AI chips over five years, subject to the terms of its agreement.

The deal includes warrants that could allow Meta to acquire up to approximately 10% of AMD if specified share-price, shipment and commercial milestones are met.

AMD also has a similar arrangement with OpenAI.

These agreements provide a significant commercial opportunity for AMD and indicate that major AI customers are interested in diversifying their computing supply.

However, the arrangements also have financial implications.

AMD recorded a $1.07 billion non-cash charge in its third quarter of 2026 related to warrants granted under its OpenAI and Meta agreements.

The agreements could also dilute existing shareholders if the warrants vest and are exercised.

This demonstrates that winning large AI infrastructure customers can involve complex financial incentives in addition to chip performance and pricing.

What Happens Next in the Nvidia Competition?

The next stage of competition will depend on whether AMD and Google can turn customer interest into sustained commercial adoption.

For AMD, the most important indicators will include Instinct chip shipments, ROCm software adoption, performance improvements and the expansion of its AI cloud ecosystem.

For Google, the focus will be on external TPU customers, the availability of its next-generation processors and the economics of supplying complete TPU systems.

Nvidia will continue defending its position through new chip generations, software development, manufacturing relationships and customer support.

The competitive landscape will also depend on the types of AI models businesses choose to deploy.

Smaller open-weight models and inference-heavy applications may create opportunities for alternative processors, while large-scale model training and demanding workloads may continue relying heavily on Nvidia’s integrated systems.

The Bigger Picture

The AI chip market is gradually moving from a period in which Nvidia dominated customer choices towards one where businesses are more actively comparing hardware suppliers.

AMD is benefiting from improvements in its software platform and growing interest in lower-cost AI computing.

Google is using its custom TPU technology to build a larger external business beyond its own data centres.

The shift is being driven by a practical business concern: companies need to control the cost of operating AI applications as usage increases.

Nvidia still has major advantages in software, manufacturing scale, availability and customer relationships. But the growing adoption of alternative chips shows that customers increasingly want flexibility and more negotiating power.

What Happens Next?

The coming quarters will reveal whether AMD and Google can convert early customer interest into larger deployments and recurring revenue.

AMD’s ability to improve ROCm, deliver Instinct accelerators on schedule and fulfil major customer agreements will be important. Google will need to expand TPU availability, demonstrate competitive economics and make its software environment easier for external developers to use.

For Nvidia, the challenge will be to protect its market position while competing against customers that increasingly want multiple suppliers.

The emerging contest is not simply about which company makes the fastest AI chip. It is about performance, price, software, availability and the total cost of running AI at scale.

Frequently Asked Questions

1. Why are companies looking beyond Nvidia for AI chips?

Businesses want to reduce AI computing costs, improve supply flexibility and avoid depending on a single chip supplier. AMD and Google’s processors are attracting interest for workloads where they offer suitable performance and economics.

2. What are AMD Instinct chips?

AMD Instinct is a family of data-centre accelerators designed for demanding computing workloads, including AI model training and inference.

3. What is Google’s TPU?

A Tensor Processing Unit is a custom AI processor developed by Google to accelerate machine-learning workloads. Google is expanding its availability to customers outside its own infrastructure.

4. What is the difference between Nvidia CUDA and AMD ROCm?

CUDA is Nvidia’s software platform for programming and running applications on its GPUs. ROCm is AMD’s software platform for developing and deploying workloads on AMD accelerators.

5. Is AMD replacing Nvidia in the AI chip market?

No. AMD is gaining customer interest and commercial opportunities, but Nvidia remains dominant. Many customers are adopting alternative processors for selected workloads rather than replacing Nvidia completely.

6. Why are AI inference chips becoming important?

Inference is the process of running a trained AI model to generate responses or perform tasks. As AI applications gain users, the recurring cost of inference becomes an important part of operating expenses.

7. Which companies are buying AMD and Google AI chips?

AMD has major agreements involving OpenAI and Meta. Google has expanded TPU access to external customers, with companies including Anthropic and OpenAI linked to its growing TPU ecosystem. Cirrascale has also reported customer interest in Google TPU 8 computing.

Get the day’s top stories in your inbox

One concise email. No spam, unsubscribe anytime.