Marc Benioff has spent years watching artificial intelligence reshape the technology industry from the inside.

At Salesforce's Dreamforce conference this week, however, the company's longtime CEO offered an unusually enthusiastic assessment of the company powering so much of the AI boom.

Nvidia, Benioff said, is “exquisite.”

The comment was not simply corporate flattery.

Benioff's argument goes directly to one of the biggest questions facing AI investors: why has Nvidia been able to capture such an extraordinary share of the value created by the generative-AI infrastructure boom?

His answer is essentially scarcity.

Nvidia has the specialized products that the biggest AI data-center operators need at exactly the moment demand for advanced computing has exploded. Benioff told Yahoo Finance that Jensen Huang's company has a relatively limited number of customers, but those customers are building enormous data centers where Nvidia's products are “material” to the business.

That is what makes Nvidia unusual.

It does not need millions of customers.

It needs the world's largest AI companies to keep building.

At first glance, having a limited number of major customers sounds like a weakness.

Normally, investors prefer diversified customer bases.

But the AI-chip market is unusual.

The world's largest technology companies are spending tens of billions of dollars on AI infrastructure.

Microsoft.

Amazon.

Google.

Meta.

OpenAI and its cloud partners.

Those companies have enormous computing requirements.

A small number of customers can therefore represent enormous demand.

Benioff's point is that Nvidia sells into a highly concentrated part of the economy where each customer is making historically large infrastructure investments.

A single large data-center project can require thousands of advanced accelerators, high-speed networking components, memory and associated systems.

That is why Nvidia's customer concentration has not prevented the company from becoming the defining hardware supplier of the AI boom.

The relationship between the two companies has become deeper.

At Dreamforce, Salesforce and Nvidia unveiled Koa, a new CRM reasoning model designed specifically for Salesforce's Agentforce platform.

Koa was developed using Nvidia's Nemotron 3 Super model and post-trained by Salesforce for enterprise workflows. It is trained using synthetic scenarios modeled on nearly three decades of Salesforce CRM experience.

That partnership illustrates where enterprise AI is heading.

Companies increasingly want AI models that understand specific domains.

A general-purpose model can reason about a physics problem, write a poem or summarize a document.

But enterprise software requires something different.

It must understand customer records.

Sales pipelines.

Service cases.

Business processes.

Permissions.

Compliance.

And the precise sequence of actions that must happen to complete a workflow.

Koa is designed around that environment.

The first wave of generative AI was dominated by general-purpose models.

The next wave may be dominated by specialized models.

Financial-services models can understand financial workflows.

Healthcare models can reason about medical operations.

Manufacturing models can understand supply chains.

CRM models can understand sales and service processes.

That creates an opportunity for software companies such as Salesforce.

They already possess the data, workflows and customer relationships.

The AI model becomes another layer on top.

Salesforce's strategy is to turn 27 years of CRM knowledge into specialized intelligence rather than asking customers to rely entirely on a general-purpose chatbot.

The Salesforce CEO's praise for Nvidia is also revealing because Salesforce operates on a very different part of the AI stack.

Salesforce sells enterprise software.

Nvidia sells the computing infrastructure required to run the models behind that software.

Benioff therefore has a front-row seat to both sides.

He can see customers demanding more AI functionality.

He can also see what it takes to deliver that functionality.

The computing requirements are enormous.

Agentic AI systems need to process more information, reason across multiple steps and interact with enterprise applications.

That increases demand for computational resources.

The more capable the agents become, the more infrastructure customers may require.

This is one of the mechanisms connecting enterprise software growth to Nvidia's hardware demand.

Salesforce says Koa runs inside its own trust boundary and was built without using customer data to train the model. Instead, the company constructed a proprietary synthetic dataset reflecting CRM reasoning and workflows.

That is an important point for enterprise customers.

Businesses are understandably cautious about sending sensitive customer information into external AI systems.

A model designed specifically for enterprise environments can provide greater control over data, security and deployment.

That may become a competitive advantage.

And it reinforces Benioff's broader argument about the future of AI.

The value is not simply in having the largest general-purpose model.

It may be in combining powerful models with specialized knowledge.

Nvidia has increasingly moved beyond simply selling chips.

The company has been developing and promoting its own open models and software frameworks, including the Nemotron family.

That expands Nvidia's role in the AI ecosystem.

Instead of being merely the manufacturer of the hardware, Nvidia can also provide model technology that customers use to build specialized systems.

Koa is an example.

Salesforce takes Nvidia's Nemotron model, adds its own enterprise knowledge and post-training, and turns it into a CRM-specific reasoning system.

That creates a form of vertical integration.

Nvidia supplies the computing foundation.

Salesforce supplies the business context.

The customer gets an AI system designed around its actual workflows.

The broader Dreamforce announcements reinforce this trend.

Salesforce has been expanding Agentforce, its platform for enterprise AI agents, with systems designed to pursue goals, use tools and work across business functions.

The company said Agentforce and Slack had already processed billions of “Agentic Work Units,” reflecting growing deployment of AI-driven workflows.

This represents an important shift.

Traditional software waits for a human to click through screens.

Agentic software can potentially perform the work itself.

That requires significantly more reasoning and computing.

Which brings the story back to Nvidia.

AI models are hungry for compute.

More complex reasoning means more processing.

More users mean more inference.

More autonomous agents mean more simultaneous workloads.

And enterprise customers expect those systems to work reliably at scale.

That creates an enormous market for accelerated computing.

Nvidia's advantage is not simply that it makes fast chips.

It has built an ecosystem around those chips.

Hardware.

Networking.

Software.

Libraries.

Developer tools.

AI frameworks.

The combination makes it harder for customers to switch suppliers quickly.

That ecosystem is one reason Nvidia has remained so influential.

The word may sound like a colorful CEO compliment.

But the underlying argument is financial.

Nvidia has products that match an extraordinary moment in technology.

Massive data centers are being built.

A small number of companies are spending enormous amounts of money.

And those companies need precisely the kind of accelerated computing Nvidia specializes in.

That combination gives Nvidia unusual pricing power and strategic importance.

The question for investors is how long that advantage lasts.

Competitors such as AMD and other semiconductor companies are developing alternative accelerators.

Major cloud companies are designing their own chips.

Chinese technology companies are developing domestic alternatives.

So Nvidia's lead is not guaranteed forever.

But Benioff's point is that the current market structure strongly favors Nvidia.

There is another important part of the relationship.

Some investors have argued that AI could undermine traditional software companies.

If AI agents can perform tasks automatically, why should businesses continue paying for large software platforms?

Benioff rejects that idea.

At Dreamforce, he dismissed the so-called “SaaSpocalypse” narrative and argued that software is evolving rather than disappearing.

His strategy is to place Salesforce's data, workflows, permissions and business logic underneath AI agents.

In that model, AI does not eliminate Salesforce.

It makes Salesforce more important.

The software becomes the trusted system of record while AI becomes the layer that acts on top of it.

The concept is straightforward.

Salesforce has customer data.

That data makes AI more useful.

More useful AI makes Agentforce more valuable.

More AI usage creates more demand for computing.

Nvidia provides the computing.

The relationship therefore forms a flywheel connecting enterprise software and semiconductor infrastructure.

Salesforce wants more businesses using its AI agents.

Nvidia wants more businesses requiring accelerated computing.

Their interests increasingly overlap.

The model is currently being tested by select customers, with general availability expected in the U.S. during winter 2026.

Companies participating in pilots include organizations such as Formula 1, Xero, Baxter Credit Union and UChicago Medicine, according to industry reporting.

That provides an early glimpse into how enterprise AI could develop.

Rather than buying one giant model and hoping it understands every industry, businesses may increasingly deploy specialized models optimized for particular workflows.

That could create a large market for smaller, domain-specific reasoning models.

The AI-chip market is competitive.

But Benioff's assessment captures why Nvidia remains at the center.

Its products are being purchased by the companies making the largest AI investments.

Its software ecosystem is deeply embedded in AI development.

And its technology is increasingly being integrated into specialized enterprise systems like Salesforce's Koa.

For now, the demand curve remains powerful.

The bigger question is whether that demand ultimately turns into sustainable economic returns for the customers buying all those chips.

That debate will become increasingly important as AI infrastructure spending continues to expand.

Benioff's comments provide a useful window into how the technology industry is changing.

Enterprise software companies increasingly need AI.

AI companies need enormous computing infrastructure.

Semiconductor companies are moving up the stack into software and models.

And the boundaries between these businesses are becoming less distinct.

Nvidia is not merely selling chips.

Salesforce is not merely selling CRM.

Together, they are building pieces of an enterprise AI system in which specialized models, business data and accelerated computing work together.

That may explain why Benioff calls Nvidia “exquisite.”

The company is sitting at the intersection of three extraordinary trends:

the growth of AI agents, the explosion of data-center investment and the transformation of enterprise software.

Whether Nvidia can maintain that position indefinitely remains an open question.

But for the AI economy of 2026, one thing is clear.

The chip king's greatest advantage may not simply be having the fastest hardware.

It may be having an ecosystem that keeps finding new reasons for the world's biggest companies to need it.