NVIDIA Broke Its Own Rule: What It Means When the Supplier Has to Vouch for Demand

Yesterday afternoon NVIDIA did something it had never done. Not the quarter, although the quarter was enormous. What it said afterward.

The numbers

The earnings release NVIDIA filed with the SEC reports USD 96.2 billion in revenue for the second quarter of fiscal 2027, up 106% year over year. Data center alone contributed USD 89.0 billion, up 117%. GAAP net income was USD 59.7 billion and GAAP diluted earnings per share came in at USD 2.46. For the current quarter the company guides to USD 108.0 billion, plus or minus 2%.

That document contains no fiscal 2028 outlook at all. The outlook was delivered out loud, on the analyst call: CFO Colette Kress said fiscal 2028 revenue will grow roughly 70%. On a base of approximately USD 405 billion, that implies USD 690 to 700 billion in a single year, against an analyst consensus of about USD 570 billion.

Jensen Huang was explicit about how unusual the gesture was: «we have never forecasted, never guided to a year in advance». Axios and Fortune agree on both the figure and the circumstances.

One more line from that call drew little attention: Kress said the company remains supply constrained. The 70% does not describe how much the market wants to buy. It describes how much NVIDIA can ship.

Why it matters

A company does not break a twenty-year communications norm because it feels good. It breaks it because staying quiet has started to cost something. And what it is costing has a name: circular financing.

The bear case holds that much of the growth comes not from end demand but from a closed loop: NVIDIA invests in AI companies, those companies buy its GPUs, and the revenue returns to its origin wearing the costume of a third-party sale. It is an accounting accusation, not a technical one, which is exactly why no good quarter can refute it: a good quarter is precisely what the thesis predicts.

Against that, the only available play is to commit to a number that has not happened yet. Twelve-month guidance does not prove demand is final, but it puts the company's credibility behind that claim for four consecutive quarters. It is a defensive move with the aesthetics of a victory lap. The market bought it: the stock rose roughly 6% in premarket trading on August 27 according to CNBC; CoinDesk reported 8% over the same window.

What it means if you run a technology company

This is where the headline stops being a market datapoint and becomes an operating constraint.

First: the 2027 constraint is not budget, it is queue position. If the CFO says the company is supply constrained, the growth she announced is not a demand ceiling: it is a capacity floor. For anyone sizing a project that depends on dedicated compute, the risk has shifted from getting spend approved to having the hardware on the date the plan assumes. That means contracting earlier, signing commitments longer than feels comfortable, and designing for the real possibility that capacity arrives late.

Second: compute will not get cheaper through competition any time soon. A supply-constrained vendor does not compete on price. Any product model that treats falling compute costs as a 2027 margin lever should revisit that assumption now, not when it breaks.

Third, the uncomfortable one: the same day produced a reason to diversify. The Information reported that NVIDIA agreed to acquire Hugging Face for USD 12.9 billion. Precisely stated: neither company confirmed it, and TechCrunch, which asked both and got no response, notes that Business Insider maintains there is no signed agreement. It is not a fact; it is a report serious outlets disagree on. But even as a hypothesis it forces a question: if the vendor selling you silicon also owned the registry your pipeline pulls weights from every night, how many links in your deployment chain would sit under a single owner?

My read

I do not read the 70% guidance as proof that the bubble is not real. I read it as proof that the argument got far enough to force the strongest actor in the cycle to do something it did not want to do.

For those of us who build software on this infrastructure, that is good news and a warning at once. Good, because a vendor willing to commit twelve months out is a vendor you can plan around. A warning, because that commitment rests on the capex of a handful of customers who are funding it with debt: SoftBank is weighing a USD 10 to 20 billion bond sale to repay the USD 40 billion bridge loan it took against its OpenAI position, per Bloomberg, and its credit default swaps widened the same day the report landed.

When a supplier's stock rises and its customer's bond falls on the same day, that is not consensus. That is two markets reading the same story backwards from each other.

My working rule did not change yesterday: you plan against contracted capacity, not projected capacity. The 70% belongs to NVIDIA. The delivery date belongs to you.