SiliconANGLE theCUBE Sep 18, 2026
Gartner projects worldwide AI spending to grow 49% in 2026, and the argument on this episode is that the money is no longer following the best model.
The familiar framing is a race between frontier labs. The reading here is that Wall Street has moved on to whoever can get capacity into production: data centers, chips, power and fiber. The enterprise software layer above it is being rebuilt at the same time.
"Moore's law was deflationary. Jensen's law is actually inflationary."
This is theCUBE's own analysts talking, two days out of Dreamforce 2026 and an AI infrastructure summit, with customer research from their data partner Qualitate and an on-the-ground read of the Nscale listing from the floor of the New York Stock Exchange.
The full episode is covered here so you can skip it.
Here are the 10 arguments that matter.
Key Takeaways
Gartner projects 49% growth in worldwide AI spending in 2026, and the thesis has shifted from the smartest model to the production stack
Salesforce customers expect spending to rise because of headless access, not fall, because the agent activity runs on Salesforce data
Moore's law was deflationary and Jensen's law is inflationary, with forward pricing locked a year ago now 30% to 90% below the market
Investors are financing the chips and the buildings together on non-investment-grade terms, which is what makes a neocloud bankable
Safety becomes a layer of the stack rather than a brake, on the argument that nothing can be secured before it is built
Salesforce guided revenue out of the 50 billions into the 60 billions and the stock fell
Agentforce customers report 30% to 40% improvements and are cutting out half of their poor inbound leads
Salesforce's remaining advantage is data, workflow, identity and business context, which is why Slack matters as the front door
Open source will not stop, and the lag behind the frontier compresses from six months toward two
The system of record may become commoditized once agents learn how a company works well enough to be trained on it
1. Dreamforce Stagecraft
The episode opened on the theater of the conference. Marc Benioff calls customers and employees up on stage, and every speaker after him tries to match his energy.
Why the keynote works
Benioff is the keynote master. He is very performative.
A host
Jensen Huang was there, as was Anthropic's chief executive Dario Amodei, whose company has invested alongside Salesforce. Huang told the room this was the best sales conference, which landed as a jab, and Benioff put an arm around him and said everybody was following him.
The pattern one of the hosts has seen at every show
Jensen's got to be the alpha. Whatever show he goes to, he's the alpha.
A host
The precedent given was CrowdStrike's conference a few weeks earlier, where Huang and George Kurtz traded lines about who had the bigger guns.
2. Headless Changes the Client
The substance behind the show, in this reading, is work George Gilbert started a couple of years ago: that Salesforce was not a disconnected pile of acquisitions such as Tableau and Slack, but was doing engineering to bring them together into a system of intelligence. Dreamforce is where that became visible. Claude now generates the user interface over headless Salesforce, and Slack can generate one too.
The research from Qualitate, which dropped its Agentforce study the same week, cuts against the obvious conclusion.
Customers expect the bill to go up, not down
What it showed is that customers actually expect spending to go up as a result of headless
A host
The mechanism is that a customer working in another client surface is still pulling Salesforce data and still generating Salesforce agent activity, so consumption rises. What is unsettled is how any of it is charged for — seats, consumption, users, outcomes or an enterprise license agreement.
And that is the variable everything else depends on
The big question is pricing.
A host
Software companies have a record of protecting their revenue through every one of those transitions, which is the reason the hosts do not read headless access as a threat to the top line. The larger claim is about the shape of the industry rather than one vendor.
Agents are not a feature cycle
They're forcing a complete reconstruction of the enterprise stack.
A host
That reconstruction runs from the applications and the interfaces down through security, infrastructure and the capital being spent on it, a point one host said he had been hearing from Boston Consulting Group people and investors the same week.
3. The Safety Argument
Both hosts rejected the idea that the industry is pausing. The week's safety noise arrived while agents were getting more autonomous, and in their account nobody slowed down. One called the go-slower narrative false.
The politics were dismissed in both directions: David Sacks saying the concern is nonsense, Bernie Sanders saying AI has to be stopped, with the truth somewhere between them. Nobody, one host said, has actually argued against governing AI at all.
Safety as a layer, not a brake
Safety just becomes another layer of the stack in my opinion. It just can't slow down. You can't make it safe until you build it.
A host
The design-it-in objection was answered on the same logic: you cannot design in protection for something that does not exist yet, because you do not know what you are protecting. The only force either host could see slowing it is external and political, Trump and Xi agreeing to it, which one host said he would not trust, and even then open weights are already loose.
On the open-source question, the view was that the community likely needs the frontier labs to lead, but the gap between a frontier release and an open equivalent compresses: six months to four, four to two, without reaching zero. The conclusion drawn was that open source is close enough to matter.
4. Betting on Agents
The framing question was whether Salesforce is betting the company.
The premise neither host disputed
Salesforce is essentially betting the company on agents.
A host
The answer given was that headless becomes the system of intelligence: the layer where data is harmonized and process knowledge is injected, which they called one of the first true enterprise harnesses. The distinction drawn is against what exists inside the models.
Model harnesses do not behave predictably
You see harnesses within the LLMs, but they're not deterministic.
A host
AI Force, announced at the conference, is what lets any client surface build the interface. The maturity assessment was blunt: Agentforce is early and mostly proofs of concept. The value was not dismissed.
What customers report so far
They're knocking out half of the inbounds that are crap inbound leads.
A host
Customers are also reporting improvements of 30% to 40%, and agents handling more calls. On budgets, there is no sign yet of Agentforce spending being taken out of existing Salesforce budgets, so no cannibalization, in early days.
What the company told investors, and what the market did
They did announce that they're moving out of the 50s, 50 billion into the 60s. They upped their projections. The stock went down quite a bit.
A host
5. Where the Moat Is
With agents able to sit anywhere, the question is what is left that is defensible.
The four things they say Salesforce still owns
Yeah, the whole moat is based on the data workflow identity and the context.
A host
Salesforce could be the agent interface, one host said, without owning every agent interface, because agents have standards and Salesforce has the data. That is also the reason Slack keeps coming up: unstructured conversation is the raw material graphs and agents work well on.
Which makes Slack the entry point to watch
That's why I think Slack becomes a potential agentic front door.
A host
The conditional that follows is the whole SaaS apocalypse argument in one sentence, and the customer research supports it. A company using Salesforce for its workflows, deeply embedded in its processes, will do more inside Salesforce. A company using it only as a system of record can reach that data through headless access and rebuild the functionality elsewhere, with Claude Code, Cursor, OpenAI or Databricks, wherever its developers already are. Some of that business then gets siphoned off. The open question is whether the incremental business Salesforce wins early in the cycle offsets the drift later.
6. Commoditizing Records
The thought exercise one host put up is what happens to that data advantage over time. Today roughly 1% of enterprise data is unlocked, because the frontier models have not ingested the proprietary part. If agents learn how enterprises actually work, that becomes a new training and inference data set.
His example was a bankruptcy. In the Spirit Airlines case there was a bid for the airline's historical transaction data, and the buyers wanted it for training.
Which puts a question over the systems themselves
So that kind of brings up the commoditization of the system of record.
A host
The consumer version is already running. An agent app called Instinct is completing transactions through messaging interfaces, including working a support line to change a seat assignment.
He has seen it done
I just talked to someone who had their middle seat changed by Instinct, the new AI agent app.
A host
Salesforce's answer to the fast-moving specialists is the Finn acquisition, a standalone agent aimed at Sierra, and one the hosts rate on ease of use.
7. Is There a Bubble
Every investor either host asked that week said no. The one they found most persuasive was a managing director at DigitalBridge who lived through the internet and telecom bubbles and says demand is too large, while acknowledging that the financing runs on non-investment-grade assets.
What makes those assets bankable, in his account, is that the chips and the physical buildout are financed together: the real estate and the plant alongside the silicon, analyzed as one credit. He was an early investor in two CoreWeave financing rounds, of two billion and seven billion dollars, on that basis. The host's summary is that the financial community is turning non-investment-grade deals into bankable ones, and that this is the neocloud model.
He then read out the textbook definition on air: asset prices surging far above intrinsic value on speculation and emotion rather than fundamentals, and the anatomy that follows, from displacement to euphoria to profit-taking. Profit-taking, he noted, has not started.
The pricing point that cuts against the bubble reading
Moore's law was deflationary. Jensen's law is actually inflationary.
A host
The evidence offered was Huang's own line to Benioff, that "People are bummed out because they locked in forward pricing a year ago and now they could get 30, 50, 90% higher prices." Monetization is real for anyone who owns compute, including SpaceX, the hyperscalers and Nvidia. The open question is whether mainstream enterprises are monetizing, where the answer given was that some are and most are not.
And the historical objection
So, like I've always said, bubbles don't always burst.
A host
His example was sports franchises, up for 30 years without a break, which he said is a market rather than a bubble.
8. Who Owns Production
The investment argument is that capital spending has moved past the foundation models, and that "the investment thesis is migrating from who has the smartest model to who has the production stack". The risk therefore sits with delivery: a neocloud with a hundred billion dollars of offtake orders has to build, and the question is whether it can put that capacity into service.
Two kinds of deals sit inside that, with different risks — building the data centers, and financing the chips. Huang's argument, relayed on the show, is that the chips hold their value better than people assume, which is what makes them creditworthy collateral.
The knock-on effects run through commodities nobody was excited about. Fiber optic cable is the most commoditized product there is, and land and power acquired for a data center is worthless without it, so fiber matters again. Copper prices are high because data centers consume it, which turns the copper telephone companies are pulling out of the ground into an asset.
Which is where one host says the enterprise now lives
The data center is the new enterprise
A host
On-premises is getting its own unlock, visible at Dell and HPE, with much of Nvidia's hardware going into private data centers and the neoclouds serving as an alternative to hyperscale. The one point of friction between the hosts was whether the model still matters.
The dissent on the smartest-model question
I actually think smartest model still matters.
A host
The reasoning was that compute access is what produces the best model, which is what Sam Altman has spent years locking up and what Amodei is now doing, and that the frontier is what drives demand for the compute underneath everything else.
9. Models Get Specialized
Where the two hosts agreed was on what the models become.
Past the chatbot era
So models are becoming specialized and embedded.
A host
The examples were Anthropic's biology work and Salesforce's own launch alongside it.
What replaces the chat window
These are workflows and domain expertise and operating environments.
A host
The stack that emerged from Dreamforce, in their telling, is models, agents, context and data, with orchestration, security and governance around it and infrastructure underneath to put it into production. Context is the piece they are watching most closely, after a Neo4j graph event in New York where the audience was engineers.
Where the graph fits
The context graphs are huge, context layers, that's a data thing.
A host
Their conclusion on infrastructure is that enterprises will not stand up gigawatt facilities themselves, so the neoclouds become the enterprise's data center, offering multi-tenant economics with a single-tenant experience. The interface progression Benioff described, from command line to graphical interface to web to mobile to an interface generated on the fly, is what drives token demand, and that demand lands back in the data center, including on private infrastructure.
10. Nscale's IPO
The news that broke during the recording was Nscale filing to list on the New York Stock Exchange. One host had been at the company's investor meeting and seen the numbers.
Where the company came from
Again, they came out of the whole cryptocurrency mining firm that they started.
A host
The comparison is to the playbook CoreWeave built: a fully integrated business rather than a reseller of capacity. Nscale bought Anyscale, which the host read as evidence the management understands disaggregated serving, and it adds a vertically specialized competitor to Nebius and CoreWeave. Its board now includes Fiji Simo, formerly of OpenAI, Instacart and Meta.
The timing is unknown, weeks rather than months if the filing is clean, and the hosts said they would cover it from the exchange floor. Qualitate already holds one customer interview mentioning the company, describing it as a burst option for when jobs spike beyond provisioned capacity, which is the kind of single data point the hosts said they can put in context rather than treat as a finding.
Bonus Insights
CoreWeave was worth less than a billion dollars when theCUBE first covered it. One host said they knew the company before it was clear it had been funded at all, which is the span the current valuation has to be judged against.
Their own research process has changed. Both hosts have been onboarded to Qualitate and are running surveys by speaking to the platform, one on ServiceNow control-tower customers and one on CoreWeave ahead of an event, with results back within days.
Dell is up about 250% for the year. One host is keynoting a Dell symposium and expects Michael Dell at the New York Stock Exchange, and credited Dell's calls on the market to the company listening rather than to luck.
Matt Baker of Dell has written a book. "Unconventional Wisdom, an unlikely career in tech" was described as unusually candid on personal history and internal detail, from a career that started in technology at Intel and ran to head of strategic planning at Dell.
The episode's bottom line is that the money in AI has moved from model quality to production capacity, financed against chips and buildings together, while the enterprise software layer above it is being rebuilt so that the interface can live anywhere — which for Salesforce means more consumption from customers who work outside its screen, and an unresolved question about what happens to a system of record once agents have learned how the business runs.
Products, Companies & Tools Mentioned
Salesforce (Headless access, AI Force and Agentforce; guided revenue out of the 50 billions into the 60 billions and the stock fell; the Finn acquisition aimed at Sierra)
Qualitate (theCUBE's data partner, whose Agentforce study found customers expect headless access to raise their Salesforce spending)
Slack (Can now generate the interface over Salesforce data, and the reason one host calls it a potential agentic front door)
Anthropic (Dario Amodei was at the conference, the company has invested alongside Salesforce, and its biology work is the example of a specialized model)
Nvidia (Jensen Huang's pricing argument, that chips hold value and forward prices locked a year ago are 30% to 90% below today, is the core of the no-bubble case)
CoreWeave (The neocloud template, financed across chips and buildings on non-investment-grade terms, worth under a billion dollars when theCUBE first covered it)
Nscale (Filed to list on the New York Stock Exchange during the recording; came out of cryptocurrency mining, bought Anyscale, board includes Fiji Simo)
Nebius (The other horizontal neocloud Nscale now competes with)
Dell and HPE (Where the hosts see the on-premises unlock showing up; Dell is up about 250% for the year)
Databricks and Snowflake (Composable elements inside the Salesforce ecosystem, and the environments a customer could rebuild Salesforce functionality in)
Neo4j (The graph event that convinced one host context graphs are the layer to watch)
CrowdStrike (Where Huang and George Kurtz traded lines a few weeks earlier, the precedent for the Dreamforce stage dynamic)
Gartner (Source of the 49% worldwide AI spending growth projection for 2026)
DigitalBridge (The managing director who financed two CoreWeave rounds and argues demand is too large for a bubble)
Books & Resources Mentioned
Unconventional Wisdom, an unlikely career in tech – Matt Baker (A Dell executive's account of his own career, which one host was reading on a Kindle and recommended for the internal detail)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: