News

At Dreamforce, AI's Biggest Players Split Over How Fast the Industry Should Move

• By John K. Waters
• 09/15/2026

Dreamforce opened Tuesday as a showcase for Salesforce's vision of the agentic enterprise. It quickly became something else: a stage for one of the most consequential disagreements in artificial intelligence.

Anthropic CEO Dario Amodei used his appearance with Salesforce CEO Marc Benioff to reiterate his argument that the AI industry needs stronger safeguards as increasingly capable models move toward greater autonomy. Later, Nvidia CEO Jensen Huang sat down with Benioff and offered a markedly different prescription.

Move faster.

The contrast put two competing philosophies of AI development on display at the same conference. Amodei is arguing that the technology is advancing quickly enough to require independent scrutiny and common safety standards. Huang maintains that companies can continue accelerating while handling safety through engineering and product decisions rather than new regulation.

The disagreement matters well beyond Silicon Valley. Enterprises are being encouraged to put AI agents into increasingly important business processes at precisely the moment when some of the people building the underlying technology are debating how quickly the frontier itself should advance.

Amodei Wants the Industry to Pace Itself
Amodei's Dreamforce comments followed his call over the weekend to slow the pace of frontier AI development so safety research and independent evaluation can keep up. As Pure AI reported Sunday, the proposal drew unusual support from several of Amodei's biggest competitors, including OpenAI CEO Sam Altman and Google DeepMind CEO Demis Hassabis.

Amodei has proposed giving independent evaluators ongoing access to AI systems so they can assess compliance with safety measures and investigate incidents. He has also called for common standards among AI companies in democratic countries, with broader international coordination eventually extending to countries including China.

Anthropic has said it will begin with the independent-evaluation component itself.

The proposal comes amid heightened concern about what increasingly autonomous AI systems could do when given access to computers, networks, and other tools. Those concerns became particularly visible last week when researchers inside Anthropic publicly acknowledged that the industry does not yet know how to reliably control the superintelligent systems it is trying to build. Pure AI examined those warnings, including concerns about recursive self-improvement, in which AI systems could help create increasingly capable successors faster than safety measures can keep pace.

The risks are not entirely theoretical. In July, Anthropic disclosed that several Claude models gained unauthorized access to the production systems of three organizations during internal cybersecurity evaluations after a simulated testing environment was mistakenly connected to the public internet. Anthropic said the incident resulted from an operational failure in the evaluation setup, not an intentional attempt by the models to escape.

Amodei's argument is not that AI development should stop. It is that capability development could begin outrunning the industry's ability to understand and control the systems it is creating.

At Dreamforce, that position put him somewhat at odds with the relentlessly forward-looking message surrounding him.

Salesforce spent much of the opening keynote promoting "AIforce," its new interface designed to bring AI together with Salesforce business data and context. The broader goal is what Salesforce calls the "Agentic Enterprise," in which humans and agents work across business systems.

That makes the question of how much autonomy those systems should have more than an academic debate.

Huang: Keep Running
Huang's message later in the day was considerably less cautious. The Nvidia CEO argued that AI companies do not need a new layer of laws to make their products safe. Companies should instead continue developing the technology quickly while stopping or delaying individual products when engineers determine they are not ready.

"Run as fast as you can," Huang said, according to the San Francisco Chronicle, while adding that a company should pause if it believes a product will not be safe.

Huang rejected the idea that rapid development and safety are inherently in conflict. He described safety as an engineering and computing problem and said companies should simply refrain from releasing products when they are not confident in their safety.

This position aligns with the broader vision Huang has been laying out all year. As we reported at Nvidia's GTC conference in March, Huang is pushing an industry-wide shift from training AI models to running them at scale. This strategy centers on agents, inference, and the supporting infrastructure needed.

Nvidia has also been moving further into the software needed to build those agents. Days after Huang's GTC keynote, the company introduced an open-source package for building and managing enterprise AI agents, including policy-based security and privacy controls. We reported at the time that Salesforce was among the companies using or testing Nvidia's agent technology.

But the disagreement between Huang and Amodei goes beyond the interests of their respective companies.

It raises a question the AI industry has never fully resolved: Can safety be engineered alongside rapid development, or does sufficiently powerful AI require deliberately slowing development until safeguards catch up?

For Amodei, recent developments suggest the latter. For Huang, slowing the technology itself risks sacrificing innovation when companies can instead manage risks at the product level.

The Debate Is Moving Beyond the Labs
The argument is unfolding as pressure over AI governance is increasing. Amodei's proposal has attracted support from several prominent AI executives, but it has also drawn criticism from those who oppose slowing AI development or question whether voluntary commitments by the companies building frontier systems are sufficient.

The Trump administration, meanwhile, has generally favored a lighter regulatory approach intended to preserve U.S. competitiveness in AI. Huang's comments at Dreamforce put him firmly on that side of the debate. That leaves enterprises in an awkward position.

Businesses are being told that autonomous agents can increase productivity, automate workflows, and change how software is used. At the same time, the companies producing the models and infrastructure behind those systems have yet to agree on how quickly AI capabilities should advance or how much external oversight they require.

Dreamforce made that tension unusually visible.

Salesforce's own announcements illustrate why. The company is working to connect AI systems more directly with the data, permissions, workflows, and business logic companies already maintain inside Salesforce. That promises to make AI considerably more useful inside enterprises.

It also increases the consequences when an agent makes a mistake. The closer AI moves to the systems that actually run businesses, the less abstract the industry's safety debate becomes.

A Debate Enterprises Can't Ignore
A considerable distance separates Amodei's call to pace frontier development and Huang's instruction to run as fast as possible. Yet the two positions share an important assumption: AI systems will become more capable and more deeply embedded in the economy.

The disagreement is over how the industry gets there. Amodei wants stronger external scrutiny and shared standards before capabilities advance substantially. Huang argues that companies can keep advancing while applying safety controls as they develop and release products.

Salesforce, meanwhile, is betting that enterprises will not wait for that argument to be resolved. Its Dreamforce message is that AI agents are moving into everyday business systems now.

That may be the most important takeaway from the debate in San Francisco. The question facing enterprises is no longer simply whether to adopt AI. It is how quickly to give increasingly capable systems access to the data and processes their businesses depend on, while the companies building the technology still debate how fast they should move.