Salesforce's annual Dreamforce landed in San Francisco this week, pulling in about 50,000 people to what's long been called the Super Bowl of software. The marquee chats paired CEO Marc Benioff with leaders from Anthropic, OpenAI and Nvidia. On stage with Benioff, Nvidia chief Jensen Huang told top AI labs to "run as fast as you can."
The safety conversation arrived just after an Anthropic researcher quit and said leading labs were "gambling with our lives," prompting Anthropic's Dario Amodei and OpenAI's Sam Altman to float safety plans and argue for easing the tempo of model releases.
Away from the spotlight, the vibe was more nuts and bolts. Many guests said the real challenge is putting current AI to work. "It's already hard enough to keep up," said Alec Bronston, senior Salesforce director at Chicago-based retail data firm Spins. A slower cadence would create "a lot of opportunity to just even catch up and get our feet wet," he said.
Consulting firm SummitX vice president Jaya Rohit Vuyyuru offered a blunt assessment, saying, "The frontier models are way ahead already," while many customers are still figuring out what AI agents can do.
The expo floor had its own spectacle. Anthropic's booth on Tuesday drew heavy traffic, staffed by youthful reps in oversized white sweaters giving live demos. Around the Campground, Salesforce's cartoon mascots sparked viral memes. Photos showed arrivals for Dreamforce on Sept. 15, 2026 in Santa Clara, California.
Salesforce used its August earnings call to debut Claudeforce, a way for sales teams to pull critical data within Anthropic's Claude. Some attendees said their companies are using Salesforce's Agentforce to field customer support and sales requests, and Salesforce notes on a support page that Agentforce bots aren't built on Anthropic's Claude Fable 5.1 or OpenAI's GPT-6 Astra.
Plenty of teams said they don't always need the newest, flashiest models. Kevin Lee, CTO at cloud contact center vendor Nice, said most tasks don't require high-end options like Fable. "In large part, with the models that are out there already today, and even one generation behind, they are highly performant and effective at doing the things that our customers need," he said. "It's almost like everything beyond this point is icing on the cake."
DocuSign, an exhibitor, is mixing approaches. CEO Allan Thygesen said the company taps "all the big frontier models as well as some of the open-weight models." DocuSign's site says it saves bigger, more expensive systems for judgment-heavy work such as complex clause analysis, multi-document reasoning and summarization, where the extra reasoning is worth the higher per-call bill. Frontier systems are typically closed and can't be tuned or inspected, while open-weight models can be pulled down and run on whatever infrastructure a company prefers.
When new technologies complicate choices, patient planning helps protect and grow your savings. Join Briefs Finance CEO Jaspreet Singh on September 29th for a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, where he shows how we're spotting investment opportunities as the dollar falls. Save your spot.
DocuSign set up shop across from Moscone Center at the Canopy by Hilton, taking the entire hotel to host quiet meetings. Thygesen said DocuSign uses model routing to steer individual prompts to the lowest-cost AI capable of doing the job, a tactic Nice and many others now use to rein in spending.
Databricks, which builds tools for AI agents, made GPT-6 Astra available to its full group of 3,500 software developers this week. "Astra unambiguously out performs our previous highest-end models (Opus 5, Sol 5.6) on highly complex tasks," wrote Databricks engineering VP Patrick Wendell on X, with Opus 5 and Sol 5.6 pointing to Anthropic and OpenAI offerings, respectively.
Not everyone races to the edge. At systems integrator Nagarro, Ram Reddy - the technology lead for consumer industries - said the buzziest topic at their booth was Claudeforce, not the latest frontier release. Engineers there typically wait around three months before wiring in the newest models. "We are not one of those first early adopters jumping at it," he said.
The economics are shifting fast. Sanders said cloud software vendors must adjust to a token-based world where customers pay for usage and outputs rather than fixed subscriptions. One token is roughly three quarters of a word.
"SaaS up to now has had no variable cost to deliver services," he said. "The profit margin structure changed." He added that moving from a traditional SaaS setup to agentic services could tug gross margins down from north of 85% to somewhere near 45%.
That lands as Salesforce keeps arguing AI is a tailwind, not a threat. Benioff has pushed back on the "SaaSpocalypse" storyline and framed AI as fuel for growth. Still, Salesforce is down 8% year to date after a big jump in August, while peers like Adobe and Autodesk have fallen by even larger percentages.
The takeaway from Dreamforce is less apocalypse, more arithmetic: which models to use, what they cost and how to route work so you don't light money on fire. Many execs called AI transformative, but the practical playbook looked very human - pick the right model for the task, mix frontier and open-weight tools, and watch the meter.
If you own cloud software names or companies building AI agents, keep an eye on two levers: when firms choose pricier frontier systems for judgment-heavy jobs, and how quickly they shift customers to tokens that introduce real variable costs. Those choices showed up in product launches, booth buzz and even who booked entire hotels, and they will shape pricing and profitability far beyond this week's show.
Staying steady with a thoughtful strategy can preserve capital and create long term gains. Our CEO Jaspreet Singh is hosting a FREE live investor workshop, How to Profit From A Dollar That's Losing its Value, on September 29th. Sign up free to join him live.
