Two enterprise AI vendors beat their quarters this week and both got sold

On September 2, C3.ai reported a quarter that its founder called evidence a turnaround is working. Free cash flow turned positive at $2.1 million, the non-GAAP operating loss narrowed by a third sequentially, and bookings jumped 73% over the prior quarter. The stock fell 5.8%. The next afternoon Zscaler reported revenue of $898.2 million and annual recurring revenue up 25%, beat on every line, and dropped 4% the following morning. Both companies did what they said they would do. Both got sold. The reason showed up in a third interview the same week, when Capgemini's chief executive explained why his company quietly stopped publishing its AI bookings figure.

Siebel is selling arithmetic, and the market wants a number he cannot give yet

Thomas Siebel came back as chief executive of C3.ai in May 2026, after a year in which health problems kept him out of the job and the company lost $470 million. His predecessor had already cut headcount from 1,075 to about 700, work Constellation Research tallied at roughly $135 million of annualized savings. Siebel spent his prepared remarks on the September 2 call listing management practices rather than products. Clear ownership. Hard deadlines. Weekly reviews.

We're not going to ask the market to underwrite a story. Our plan is to deliver results quarter over quarter and let those results speak for themselves.

Thomas Siebel, September 2, 2026, C3.ai fiscal Q1 2027 earnings call

I like the line. The trouble is what sits under it. Revenue of $52.4 million is down 25% from $70.3 million a year ago. The 73% bookings figure is sequential, not annual. Free cash flow of $2.1 million follows a quarter of negative $54.8 million. Every number carrying the turnaround is either a cost line or a quarter-over-quarter comparison against a very bad quarter, and the one forward number investors actually price came in light: a Q2 midpoint of $53 million against a $56.6 million consensus. Siebel knows exactly what he is asking for.

I think it's highly likely this company will not be trading at a revenue multiple of like 3.5. They'll be trading at a revenue multiple of 10, 15, 20, or 25.

Thomas Siebel, September 2, 2026, C3.ai fiscal Q1 2027 earnings call

That is a bet that consistency gets re-rated. Nothing this week suggests the market is ready to pay for it in advance.

Capgemini stopped publishing an AI number because it could not define one

In February, Capgemini told investors that generative and agentic AI made up more than 10% of group bookings, up from around 5% earlier in the year. Speaking in India on September 4, chief executive Aiman Ezzat explained why that number is no longer reported.

So if you don't have a definition, you can throw any number behind it. And I think this is a bit where we stopped putting a number out, and we're thinking carefully about what should be measured, because without a definition, it's meaningless.

Aiman Ezzat, September 4, 2026, source

This is the most useful sentence any enterprise vendor said all week, and it indicts the whole category, Capgemini included. If nearly every delivery now has AI inside it, an AI revenue line measures the labelling convention, not the business. Ezzat says the firm will publish a defined figure eventually. Until someone does, an investor comparing two AI disclosures is comparing two marketing decisions.

He was more concrete about the thing that is measurable, which is how the work gets priced. Time and materials was about half of Capgemini's sales three years ago, fell to 35% last year, and he expects it to settle near 20% to 25%. Fixed price puts the productivity gain on the vendor's side of the ledger, which is where he wants it. On the size of the prize he was blunt in a way most services chief executives are not.

So I'd say without the AI opportunity, we'd have pretty much a flat industry.

Aiman Ezzat, September 4, 2026, source

Zscaler's record quarter was priced as a deceleration

Jay Chaudhry went on CNBC the morning after his numbers with the stock already red and led with superlatives.

First of all, a record quarter from any measurement. Our revenue, our ARR growing at 25% at scale. ARR exceeded $3.8 billion, our non-GAAP operating margin 24%. These are spectacular numbers.

Jay Chaudhry, September 4, 2026, CNBC interview

All true. What moved the stock was fiscal 2027 guidance of 16.6% to 17.5% growth against the 25% just delivered, alongside a restructuring that cuts about 3% of the workforce. Shares closed September 4 at $170.25, down 21% for the year. Chaudhry pushed past the guidance question to the demand he sees coming, and his argument is a good one: the Hugging Face breach has boards asking two questions at once, whether frontier models can find holes in software they never adopted, and whether the agents they do adopt can be hijacked.

Tomorrow, agents will be the weakest link, or if I may say, the most dangerous link.

Jay Chaudhry, September 4, 2026, CNBC interview

He may well be right. But that sentence is a thesis, and this was the week the market stopped accepting theses in place of a growth rate.

Perplexity's answer is to take the inference off the cloud entirely

Aravind Srinivas spent his September 4 appearance describing a product that runs Perplexity's agent on hardware the customer already owns. It shipped on August 25 with Nvidia, and local work consumes no billing credits.

The first version is a fully local implementation where you own the entire agent stack, literally a mini data center living on your desk.

Aravind Srinivas, September 4, 2026, CNBC interview

His buyers, he says, are banks that want an air-gapped box. Asked whether the industry is overbuilding, he did not defend the build-out so much as route around it, pointing at the memory and power already sitting idle in offices and homes.

If a billion people need to run 24/7 agents, they're going to need a terawatt of power and a lot of memory. And so you're not going to be able to do this just with data centers.

Aravind Srinivas, September 4, 2026, CNBC interview

Strip the framing and this is the same move the other three made. Perplexity is shipping a feature whose selling point is a smaller bill.

Cathie Wood named the casualty in passing

Cathie Wood spent most of her September 4 monthly update arguing that AI is nothing like the railway mania, because the money is already coming in.

The railroads were built on a hope and a prayer that a new world would develop and that revenues ultimately would come. In AI, the revenues are screaming.

Cathie Wood, September 4, 2026, source

Then, three minutes later, while making the case that credit markets look calm, she dropped this.

The credit markets have not been disturbed much by the debacle in SaaS for the private credit funds, and so there's nothing systemic happening here, and that's a good thing to take away.

Cathie Wood, September 4, 2026, source

She offers it as reassurance. I read it the other way. The most committed AI bull on the tape now treats a software debacle as a settled fact, worth mentioning only to note that the lenders have not been hurt yet. That is the trade underneath everything above: capital is flowing to the compute and away from the applications that were supposed to justify it.

The discount rate is doing some of the work

Mohamed El-Erian, talking about the Treasury market on September 4, offered the cleanest explanation of why unprofitable software is expensive to own right now, and it has nothing to do with software.

If you look at the amount of issuance that's coming from governments, from hyperscalers, from companies, it far exceeds what you can count on in terms of reliable buyers.

Mohamed El-Erian, September 4, 2026, source

He is describing supply and demand for bonds, not sentiment about AI. But the effect lands on Siebel and Chaudhry anyway. The AI build is issuing debt into a market with a shrinking pool of reliable buyers, that pushes long rates up, and a higher long rate is exactly what discounts a promise of profitability three years out. The capex boom is raising the cost of capital for the companies whose revenue is supposed to justify it.

What would settle this

A definition, then a number. Ezzat set the test and his own firm has not met it yet, which is more honest than the alternative. Siebel set a checkable one for himself: consistent revenue growth starting in fiscal Q3, positive free cash flow, non-GAAP profitability. His Q2 guidance midpoint of $53 million against $52.4 million delivered is the first place to look, and it is a very small step. Chaudhry's is the simplest of the three, because fiscal 2027 either lands above his 17% guide or it does not.

None of these are hard tests. That is the point. After three years of AI narrative, the enterprise software trade has come down to whether three executives hit numbers they have already published, and to whether anyone can say what an AI dollar is before counting it.