NVIDIA reported US$96.2 billion in quarterly revenue in late August — up 106 percent year on year — and guided the next quarter to $108 billion. At GTC in March, Jensen Huang told the room they're looking at around a trillion dollars in orders for their next two chip generations through 2027.
I don't own NVIDIA stock and this isn't investment commentary. I care about one downstream question: what does this do to the price of the AI calls inside the systems I build?
The Quarter in Four Numbers
Revenue $96.2 billion. Data centre segment $89 billion of it. Gross margin 75 percent. Next-quarter guidance $108 billion. The margin is the number to sit with — three-quarters of every dollar spent on this hardware is profit for the vendor, and every layer above it has to earn that cost back from somewhere. That somewhere eventually includes your API bill.
Sold Out Means Something
On the earnings call, management described cloud GPU capacity as sold out, with even previous-generation chips fully utilised. The next platform, Rubin, ramps in the second half of 2026 — and the demand pipe is already longer than the supply pipe. Translation for the rest of us: serious compute stays scarce and priced for hyperscalers for the foreseeable future. Cheap, abundant inference is not around the corner.
The Desktop AI Box Went the Wrong Way
The counter-argument used to be: fine, run models locally. NVIDIA sells a desktop AI machine for exactly that — and it's moving in the wrong direction. The DGX Spark launched at $3,999; a new 64GB model now costs $4,999 with half the memory of the original, and the 128GB unit lists at $6,950. Blame the memory squeeze I wrote about separately. For a Malaysian SME, that's RM 20,000-plus of depreciating hardware versus pay-per-use APIs. It's not a contest.
Design API-Light
So the architecture answer isn't "avoid AI" — it's spend AI calls like they cost money, because they do and they'll keep doing so. In practice, in the systems I ship: cache aggressively so the same question never costs twice. Route routine steps to small, cheap models and reserve the frontier model for the few calls where its judgement earns the premium. Precompute what you can. Let deterministic code do everything deterministic code can do — it runs for free.
You Are the Margin
Seventy-five percent gross margin means every unnecessary AI call in your workflow is a small donation to the most profitable hardware company on earth. A well-designed system is just the discipline of not donating. That discipline is most of what you're paying for when someone builds you one.