NVIDIA's earnings were out of this world: revenue up 106% on the same quarter last year. The company is also guiding for 70% growth in fiscal 2028, which runs mostly through calendar 2027… and even that number is capped by what NVIDIA can physically build (not by what customers want to buy). Vera Rubin, its newest GPU, is shipping on schedule.

But it isn't all rainbows and butterflies for the company now carrying the entire AI trade on its back.

Let’s start with the elephant in the room. A large share of NVIDIA's revenue comes from customers who can't pay for these chips out of their own cash flow. So NVIDIA sells the GPUs, guarantees the loans that fund them, and takes equity in the buyers to keep the orders coming. OpenAI and Anthropic, two of its biggest end customers, are burning cash and expected to go public soon so they can raise more and spend more.

Then there's memory. NVIDIA's commitments to buy components jumped from $119B to $279B in a single quarter, almost entirely memory chips. In plain English: as those chips come off the factory line, NVIDIA has agreed to buy them, whether or not a customer is waiting on the other side. The logic is sound, since memory is the scarcest thing in the industry right now and locking it up starves the competition. But it means NVIDIA now owns the downside if demand ever softens.

The cherry on the cake is that NVIDIA is giving its largest, most creditworthy customers longer to pay: 60 days instead of 45. As a result, free cash flow fell the previous quarter, even as revenue and paper profits climbed, because a growing share of those profits sits in invoices rather than in the bank.

None of this makes NVIDIA a Ponzi scheme, and it doesn't prove we're in a bubble. It may well be the most efficient way to get AI infrastructure built at speed. But it does make NVIDIA fragile in a way it wasn't two years ago: it is now on the hook for its customers' mistakes as well as its own.

So if you own NVIDIA, or you're thinking about it, four numbers will tell you which way this is going.

Salesforce and Veeva Systems both reported last week.

Salesforce's growth was unremarkable, with organic revenue only up 7% (11% including acquisitions). What stood out was bookings, up 14%. When signed orders grow faster than the revenue being recognised, growth is usually about to pick up: the sales are already in hand, they just haven't been billed yet. The CFO said new orders had their strongest quarter in four years.

Veeva's quarter was even cleaner. Revenue grew 18%, and management called it the best CRM quarter in the company's history, with Vault CRM still winning the largest drugmakers.

Put the two together and AI has flipped from a reason to avoid software to a reason to own it. The evidence isn't AI revenue, which every company defines differently and redefines whenever it suits them. It's orders and competitive wins. With each quarter it looks less like the big AI models will replace the incumbents, and more like the incumbents will rent them, paying for a model to sit inside a platform customers already use. Of those two positions, we'd argue the model holds the weaker hand.

Software has rallied hard on this shift in sentiment. Three names still screen as undervalued on our database, with a Value Rating above 80. Click here to see them.

Upgrade to unlock full access to the RatedA Portfolio (with weekly updates on every buy/sell and rationale), and the complete database of companies + ratings.