Three of the most powerful people in artificial intelligence shared a stage in San Francisco this week, and they did not agree on much. At Dreamforce 2026, Salesforce’s annual conference at the Moscone Center, Nvidia CEO Jensen Huang, Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman each took turns telling a crowd of tens of thousands what they think the industry should do next. One argued for caution. One argued for speed. One landed somewhere in between.
The three-day event, running Tuesday through Thursday this week, drew a crowd that different outlets pegged at roughly 43,000 to as many as 50,000 attendees. Whatever the exact figure, Salesforce CEO Marc Benioff used the scale to make a point: the AI safety argument that usually plays out in essays, congressional hearings and X threads was now happening in front of a live enterprise-software audience, with three rival CEOs standing a few feet apart.
Three CEOs, One Stage: What Happened at Dreamforce 2026
According to NBC Bay Area, Dreamforce 2026 opened with AI safety as its organizing theme rather than a side session. Huang and Amodei both appeared as part of Benioff’s opening keynote, while Altman joined Benioff separately for an afternoon fireside chat, a structure confirmed by ITPro’s live coverage of the event. The staging mattered: Huang and Amodei shared the same platform and, by multiple accounts, pushed back on each other directly, while Altman’s remarks came later in a more controlled one-on-one format with Benioff.
The San Francisco Chronicle described the gathering as bringing “AI’s biggest names” to the city just as a public fight over the technology’s pace was escalating, and Adweek framed the Huang-Amodei exchange explicitly as a continuation of an ongoing dispute the two men have been having in public for months. Neither company disputed that characterization. That framing is also why Dreamforce, normally a showcase for Salesforce’s own product roadmap, ended up generating headlines about AI risk instead of CRM software.
Jensen Huang’s Case for Running Faster
Huang’s position at Dreamforce tracked with what he has said in other venues this year. He has consistently argued that capability and safety work move together rather than trading off against each other, a framing he has used to push back against calls to slow model development. In his commencement address at Carnegie Mellon University, Huang put it this way:
“The responsibility of our generation is not only to advance AI – but to advance it wisely.”
Jensen Huang, CEO, NVIDIA, NVIDIA blog
He followed that with a related point in the same address:
“Scientists and engineers have a profound responsibility to advance AI capabilities and AI safety together.”
Jensen Huang, CEO, NVIDIA, NVIDIA blog
At Dreamforce, that philosophy translated into a rejection of the “pause and regulate” instinct that has gained ground among some of his peers. Reports from the show describe Huang as unwilling to entertain the idea of slowing down, instead pointing to Nvidia’s own commercial momentum as evidence that rapid deployment and safety work are not mutually exclusive. It is a position that lines up with Nvidia’s business incentives as much as its stated philosophy: the company just posted a record quarter built almost entirely on AI infrastructure demand, and any industry-wide slowdown would hit its revenue directly.
Dario Amodei’s Warning: Slow Down and Verify
Amodei arrived at Dreamforce having already spent much of 2026 publicly arguing that the industry is moving faster than its safety infrastructure can support. That argument predates this week’s keynote. In his 2024 essay “Machines of Loving Grace,” later widely republished, Amodei wrote:
“I think that most people are underestimating just how radical the upside of AI could be, just as I think most people are underestimating how bad the risks could be.”
Dario Amodei, CEO, Anthropic, Times of India
He has also cautioned against treating model capability as inherently self-limiting. As Amodei put it in another widely cited remark:
“Intelligence may be very powerful, but it isn’t magic fairy dust.”
Dario Amodei, CEO, Anthropic, Times of India
That backdrop shaped how Amodei’s Dreamforce remarks were read. Axios described the conference as becoming “ground zero” for the AI slowdown debate, and tied the timing partly to an Anthropic employee’s public resignation earlier this year over fears that the industry was moving toward catastrophic risk without adequate guardrails. On stage, Amodei reportedly kept returning to a consistent structure: examine your own company’s safety record first, build industry-wide standards, then extend those standards through international cooperation, rather than using safety concerns as a competitive weapon against rivals. Diginomica’s coverage of the session summarized the split bluntly, describing Amodei as sticking to his position that the industry needs to slow down and regulate itself, in direct contrast to Huang’s response.
This is not a new position for Amodei. shattered.io covered his earlier call for the industry to pace the frontier with a three-step plan, and the market reaction that followed when investors briefly rotated out of Nvidia and into security stocks like CrowdStrike. Dreamforce did not introduce a new Amodei doctrine so much as put the existing one in front of a bigger, more commercially minded audience.
Sam Altman Splits the Difference
Altman’s fireside chat with Benioff drew less direct confrontation than the Huang-Amodei exchange, partly because of the format and partly because his public position sits closer to Amodei’s than to Huang’s. Coverage of the session describes Altman as largely sympathetic to the case for regulation and industry-wide safety checks, while stopping short of endorsing a broad slowdown in model development. That is consistent with statements Altman has made elsewhere this year. In one widely cited comment, he argued:
“Every frontier AI lab has a responsibility to demonstrate that it can develop increasingly capable systems safely.”
Sam Altman, CEO, OpenAI, Business Standard
That framing lets Altman occupy a middle position: acknowledge the responsibility Amodei is pointing at without committing OpenAI to the kind of public slowdown call Anthropic has made. It also reflects OpenAI’s own competitive position. The company has been running enormous training workloads this year, and a public commitment to slow down would carry real costs while Nvidia and Anthropic each stake out clearer, more differentiated brand positions on the safety question.
Marc Benioff’s AIforce Bet and Why He Needed All Three CEOs
The safety debate did not happen in a vacuum. Benioff used the same keynote slot to introduce AIforce, a new interface layer that Salesforce is positioning as a way for people and AI agents to reach Salesforce data, workflows and business logic from wherever they already work, rather than through a traditional CRM dashboard. Salesforce’s framing, echoed by CX Today’s coverage of the keynote, treats AIforce as an “interface revolution”: the value Salesforce sells was never the screen, the argument goes, it was the customer data, permissions and workflow logic sitting behind it.
That is why Benioff needed both Huang and Amodei on stage rather than one or the other. Nvidia supplies the compute and the open-weight model family underneath Salesforce’s new reasoning model. Anthropic supplies the AI interface, Claude, that Salesforce wants customers using to reach that data. Putting the two men on stage together was as much a product-marketing decision as an editorial one, even though the safety disagreement between them ended up dominating coverage of the show.
Koa: Salesforce’s New CRM Reasoning Model Built on Nvidia Nemotron
The technical centerpiece of Dreamforce 2026 was Koa, described in Salesforce’s own announcement as the company’s first CRM reasoning model. Koa was built by post-training Nvidia’s open-weight Nemotron-3-Super-120B foundation model using GRPO reinforcement learning, according to Unite.AI’s technical breakdown of the release. Salesforce says the model was trained on a proprietary synthetic dataset modeled on close to three decades of CRM deployment patterns, and that no customer data was used in training.
Salesforce is positioning Koa specifically for the kind of multistep, tool-using workflows its Agentforce platform is built around, rather than as a general-purpose chat model. The company’s own CRM benchmark claims Koa matches or exceeds leading model performance on CRM-specific actions while producing three times fewer errors, though that figure comes from Salesforce’s internal testing rather than an independent third-party benchmark. Koa is available now to a limited set of Agentforce pilot customers, with general availability planned for U.S. regions in winter 2026, per ITPro’s reporting on the launch.
The choice to build on Nemotron rather than a closed frontier model is itself a signal. Salesforce previously leaned on Nvidia’s open-weight releases for other internal tooling, and shattered.io covered the underlying model family when Nvidia shipped Nemotron 3.5 Lightning as a free, faster open-weight release earlier this year. Using an open-weight base lets Salesforce fine-tune and audit the model without depending on a single closed API provider, a decision that fits with Benioff’s broader pitch that Salesforce, not any one AI lab, should own the customer relationship.
Claudeforce and the Anthropic-Salesforce Partnership
Alongside Koa, Salesforce expanded its partnership with Anthropic through a product it is calling Claudeforce, which connects Claude Enterprise directly to Salesforce data, skills and security controls. In practice, that means an employee can ask Claude about an account, a deal or a customer’s service history and take an authorized action without opening Salesforce’s own interface at all. Salesforce has described the pairing as bringing together what it calls the top AI assistant and the top CRM platform under one workflow, delivered through AIforce alongside earlier integrations into Slack and Tableau.
For Anthropic, Claudeforce is a distribution play as much as a product one. Enterprise software buyers who already run Salesforce get a reason to route more of their day-to-day work through Claude rather than a rival assistant, without Anthropic needing to build CRM functionality of its own. It also reinforces the split on display during the keynote: Anthropic is happy to embed Claude everywhere it can, even while its CEO argues in public that the industry underneath those integrations needs tighter safety standards.
How the AI Safety Debate Reached Dreamforce
The Huang-Amodei clash did not start this week. Amodei has spent much of 2026 publicly pressing for the industry to slow down, culminating earlier this year in what shattered.io reported as a three-step “pace the frontier” plan aimed at getting labs to coordinate on safety testing before racing ahead on capability. That call rattled markets briefly: shattered.io also covered how the slowdown message briefly pressured Nvidia’s stock while lifting security vendor CrowdStrike, as investors weighed whether a coordinated pause was actually plausible.
Axios’ framing of Dreamforce as “ground zero” for that debate reflects how much the argument has moved from AI-industry trade press into mainstream business coverage. A year ago, a disagreement between Huang and Amodei over safety pacing would likely have stayed inside specialist outlets. In September 2026, it is the lead story out of one of enterprise software’s largest annual conferences, standing alongside product news that would normally dominate the week on its own.
Market and Industry Reaction
None of the three companies involved is a stranger to having its public statements move markets. Nvidia’s results and commentary have repeatedly swung AI-sector sentiment this year, and the company’s own comments about pacing model releases, which shattered.io covered when Nvidia said it would slow its release cadence to four-to-six-week cycles, show the company is not entirely deaf to the “slow down” argument even as Huang rejects it rhetorically. OpenAI, for its part, has been the subject of reporting on its own massive compute buildout, including the 100,000-GPU training run behind GPT-6 Astra, a scale of investment that makes a genuine industry-wide slowdown commercially difficult for OpenAI to embrace regardless of Altman’s rhetoric.
Salesforce shares saw modest movement around the announcement window, consistent with how investors have treated prior AI product launches from the company: enthusiasm for the growth story tempered by skepticism about how quickly AIforce and Koa will translate into measurable revenue. The bigger market question raised by Dreamforce is not about Salesforce’s stock but about whether Nvidia’s core customers, increasingly vocal about safety risk, will keep buying at the same pace while also asking Nvidia to build in more guardrails.
Nvidia vs Anthropic vs OpenAI: Comparing Three Public Safety Postures
Stripped of the keynote theater, the three companies occupy genuinely different points on the same spectrum. The table below summarizes how each CEO’s Dreamforce 2026 position compares, based on reporting from NBC Bay Area, Diginomica, Axios and ITPro.
| Company | CEO | Dreamforce 2026 session | Reported public stance | Commercial incentive at stake |
|---|---|---|---|---|
| Nvidia | Jensen Huang | Opening keynote with Benioff | Advance capability and safety together, opposes a broad industry pause | AI infrastructure demand tied directly to fast model iteration |
| Anthropic | Dario Amodei | Opening keynote with Benioff | Slow down, verify safety records, build shared industry standards | Brand built on safety leadership, Claude distribution via Claudeforce |
| OpenAI | Sam Altman | Afternoon fireside chat with Benioff | Broadly agrees labs need safety accountability, resists a hard pause | Massive active training compute makes a slowdown costly |
| Salesforce | Marc Benioff | Host and moderator across both sessions | Frames AI responsibility as core to enterprise adoption of AIforce | Needs enterprise buyers to trust agentic AI enough to adopt Koa and Claudeforce |
Timeline: The 2026 AI Slowdown Debate
The Dreamforce exchange is the latest entry in a running argument that has played out across the industry all year. Here is how the public debate escalated ahead of this week’s keynote.
| Period | Event | Source |
|---|---|---|
| Earlier 2026 | Dario Amodei publishes his “pace the frontier” three-step plan urging coordinated safety checks across labs | shattered.io |
| Earlier 2026 | Amodei’s slowdown call briefly pressures Nvidia shares while lifting security stocks like CrowdStrike | shattered.io |
| 2026 | Nvidia signals it will slow its own model release cadence to four-to-six-week cycles | shattered.io |
| 2026 | An Anthropic employee resigns publicly, citing fears the industry is moving too fast toward catastrophic risk | Axios |
| Sept 15, 2026 | Huang, Amodei and Altman appear at Dreamforce, where Huang and Amodei clash on stage over pacing and regulation | NBC Bay Area, Diginomica, Adweek |
| Sept 15, 2026 | Salesforce launches Koa (built on Nvidia Nemotron) and Claudeforce (built on Anthropic Claude) at the same event | Salesforce, ITPro |
Competitive Comparison: AIforce vs Microsoft Copilot and Google’s Enterprise AI
Salesforce is not the only enterprise vendor trying to make its platform reachable from an outside AI assistant instead of forcing users into its own interface. Microsoft has spent the past year pushing Copilot into the operating system layer itself, an approach shattered.io examined when Google’s Gemini app landed on Windows to compete directly with Copilot’s existing install base. Google, meanwhile, has been packaging Gemini for enterprise buyers through its own agent and workspace tooling, an area shattered.io covered in the context of Gemini Enterprise’s push into regulated industries.
AIforce’s pitch is narrower but arguably more defensible: rather than competing to be the default AI assistant on every device, Salesforce is betting that whichever assistant a customer already prefers, whether that is Claude, Copilot or Gemini, still needs to reach Salesforce’s underlying data and workflow logic to be useful in a sales or service context. That is a bet on Salesforce’s data moat rather than on any single AI model winning the assistant war, which is also why Benioff was comfortable putting Nvidia and Anthropic leadership on stage together rather than picking one AI partner exclusively.
Historical Context: From AI Hype Cycle to Safety Standoff
Enterprise tech conferences have hosted AI keynotes for years without turning into safety standoffs. What changed in 2026 is that the companies supplying the underlying models, not just the companies deploying them, now have public, divergent positions on how fast the technology should move. Two years ago, an event like Dreamforce would have treated AI safety as a compliance checkbox tucked into a session on responsible AI governance. This year, it was the headline act, ahead of product news that in a normal year would have carried the show on its own.
That shift also reflects how much bigger the stakes have gotten. Nvidia is now one of the most valuable companies in the world on the back of AI infrastructure spending, Anthropic and OpenAI are both running training operations at a scale that would have been unthinkable five years ago, and enterprise buyers like Salesforce’s customer base are the ones actually deciding whether agentic AI gets trusted with real business workflows. A disagreement among the three CEOs is no longer an academic argument about model alignment. It is a live input into how quickly companies like Salesforce can sell AI-driven automation to risk-averse enterprise buyers.
What Comes Next: Five Predictions
- Koa gets extended beyond CRM actions. If Salesforce’s internal benchmark claims hold up under independent scrutiny once Koa reaches general availability this winter, expect Salesforce to expand the reasoning model into adjacent Agentforce workflows beyond sales and service.
- Amodei keeps making the safety case in commercial venues, not just policy ones. Dreamforce suggests Anthropic sees enterprise conferences, not just congressional testimony, as a venue for its safety argument, since that is where buyers actually decide whether to trust agentic AI.
- Huang’s rhetoric will keep tracking Nvidia’s earnings calendar. Expect Nvidia’s public safety commentary to stay upbeat and speed-focused as long as AI infrastructure demand keeps beating expectations, with any softening most likely to appear only if that demand cools.
- More enterprise vendors will host multi-CEO AI panels. Benioff’s format, putting rival AI lab leaders on the same stage rather than picking one exclusive partner, is likely to get copied by other large enterprise software vendors trying to hedge against picking the wrong AI partner.
- Regulators will cite Dreamforce as evidence of industry self-awareness. Policymakers weighing AI legislation are likely to point to public disagreements like this one as proof that the industry itself is unsettled on pacing, which could accelerate rather than slow momentum toward binding safety rules.
Why This Matters Beyond Dreamforce
It is easy to read the Huang-Amodei exchange as another round of AI-industry theater, two well-resourced CEOs restating positions they have held for months in front of a new audience. But the venue matters. Dreamforce is not an AI safety conference, it is a sales and CRM event attended largely by enterprise buyers deciding how much of their operations to hand over to AI agents. Watching the two men most responsible for the compute and the model behind those agents disagree publicly about how much to trust the technology is a different kind of signal than reading the same disagreement in a policy paper.
For enterprise IT buyers evaluating Koa, Claudeforce or any competing product, the practical takeaway is less about who is right on safety philosophy and more about vendor accountability: Salesforce is now standing behind a reasoning model built on infrastructure from a company that argues against slowing down, wrapped in an interface layer built with a partner that argues for exactly that. Buyers will have to make their own call on how much that tension matters to their own deployment plans.
Frequently Asked Questions
What is Dreamforce 2026?
Dreamforce 2026 is Salesforce’s annual enterprise technology conference, held over three days this week at the Moscone Center in San Francisco. It typically combines product announcements with keynotes from major technology figures, and this year’s edition centered heavily on AI safety and the launch of Salesforce’s Koa reasoning model.
Why did Jensen Huang and Dario Amodei disagree at Dreamforce?
Huang has argued throughout 2026 that AI capability and safety work can advance together, and that slowing model development is not necessary or productive. Amodei has spent much of the year calling for the industry to slow down, verify individual companies’ safety records, and build shared safety standards before racing further ahead on capability. Their Dreamforce appearance put those two positions on the same stage for the first time in front of a large enterprise audience.
What is Koa, and who built it?
Koa is Salesforce’s first CRM-specific reasoning model, built by post-training Nvidia’s open-weight Nemotron-3-Super-120B model. Salesforce says it used a synthetic dataset modeled on nearly three decades of CRM deployment patterns and no customer data. It is currently available to select Agentforce pilot customers, with general availability planned for U.S. regions in winter 2026.
What is Claudeforce?
Claudeforce is a Salesforce-Anthropic integration that connects Claude Enterprise directly to Salesforce data, workflows and security controls, letting employees query and act on CRM information from within Claude rather than the traditional Salesforce interface.
What is AIforce?
AIforce is the new interface layer Salesforce unveiled at Dreamforce 2026 that exposes its data, workflows, permissions and business logic to outside AI tools and agents, including Claude via Claudeforce and Slack, instead of requiring users to work inside a conventional Salesforce dashboard.
How many people attended Dreamforce 2026?
Reported attendance figures vary by outlet, ranging from roughly 43,000 to as many as 50,000 people. Salesforce has not published a single official confirmed figure as of this report.
Did Sam Altman side with Huang or Amodei?
Reporting on Altman’s Dreamforce fireside chat with Marc Benioff describes his position as closer to Amodei’s, broadly supportive of safety accountability and regulatory frameworks, while stopping short of endorsing a full industry-wide pause on model development.
Is this the first time Huang and Amodei have publicly disagreed on AI safety?
No. The Dreamforce exchange followed months of public back-and-forth in 2026, including Amodei’s earlier call for the industry to pace its frontier development and a period where that message briefly weighed on Nvidia’s stock while lifting cybersecurity vendors seen as beneficiaries of a more cautious AI rollout.
