Fast or Exponential? The Word That Decides NVIDIA's Multi-Trillion Bet

Let's start with the deals everyone points to. NVIDIA and OpenAI signed an MOU last year describing an intention to invest up to $100 billion progressively, tied to deployments. Worth being precise here, because the number gets repeated as fact: that plan stalled. It was never finalized, never a binding commitment, and NVIDIA's own CEO later said such a build is unlikely — not in the cards. There have been other announced financings across the sector, staged over time. Everyone reads all of it as proof AI demand is exploding. But look closer and part of that "demand" may be a chip supplier's own money coming back to it.

So the real question isn't whether AI is growing. It's whether it's growing fast, or growing exponentially. Those sound like the same thing. They aren't. And the difference is worth more than most of the market cap.

Let me start with a dumb-sounding question. Why would one adjective matter that much?

Because of what a trillion-dollar-plus market cap actually is. It's not a bet on selling chips. When you buy the stock at that price, you're betting that datacenter revenue holds at a very high annualized run-rate, and holds it for years. That run-rate is the bet. Everything else is a story you tell to justify it.

So back it out. What has to be true underneath for that run-rate to survive? You can't model it in dollars, because dollars hide the physics. Model the thing dollars are made of: revenue is roughly tokens processed, times price per token, times how much of that runs on NVIDIA silicon.

Now the fast-vs-exponential thing gets concrete. For revenue to keep compounding, token volume has to grow faster than price-per-token falls. If token demand compounds at 30 to 40 percent a year — genuinely fast — the terminal value doesn't reach the current multiple. You need something much higher, on the order of 70 percent-plus, to close the gap. One of those numbers justifies the price. The other leaves a hole hundreds of billions of dollars wide. That's the whole ballgame, sitting inside the choice between two words.

Sit with that, because it's the crux. At this price, "growing" isn't the win condition. The stock already has spectacular growth baked in. It needs growth to accelerate, not just continue. So every reassuring chart about rising usage is answering the wrong question. The question is the second derivative. Now watch how the second derivative gets faked.

The next part gets technical, but it matters, because the whole bull case rests on AI getting cheaper — and the cheapness and the revenue live on opposite ends of the same curve.

You've probably heard that AI is getting radically cheaper. One common industry estimate is that cost per token has fallen roughly 10x a year. Treat that as an approximation, not a measured rate, but the direction is real. It comes from three specific tricks: reusing the computed context across a conversation instead of recomputing it (KV-cache reuse), having a cheap model draft tokens and the expensive one just check them (speculative decoding), and only firing a fraction of the model's parameters per token (mixture-of-experts sparsity).

But watch the sleight of hand. That deflation is at fixed capability. Nobody buys fixed capability. They buy the frontier. And the frontier just moved to reasoning and long-context models — the ones that think in long chains and hold huge context windows. By rough estimate those cost on the order of 10 to 100 times more compute per query, and their attention cost grows roughly with the square of the input. Not a published figure — call it directional — but the shape holds even if the multiple is off.

So cost per token keeps falling, while cost per useful task stays flat or rises, because the tasks people actually pay for got more expensive. You cannot take the token-price deflation and paste it onto the products generating the revenue. They're on opposite ends of the curve.

That's why the labs keep raising money even as usage climbs. Exponential usage with inverted margins isn't a bull case. It's a way to burn capital faster. Call it the margin trap.

Now the part everyone skips. To the extent these financings are structured as prepaid compute plus equity rather than clean cash — and some reporting suggests parts are — they aren't independent demand. Prepaid compute is money contractually pointed at buying capacity, which means buying GPUs. So a slice of it could flow back and re-book as chip-supplier datacenter revenue, on a lag. That's a structural risk, not a confirmed line item; the exact structure isn't fully disclosed, which is itself the point.

Read that again. A chip vendor's revenue going up and the labs' ARR going up are being cited as two independent proofs of demand. If they partly share a source, they aren't independent. It's the same shape that showed up in telecom in 1999 to 2001 — a supplier lends the buyer money, the buyer buys the supplier's gear, the supplier books it as organic demand. Companies like Lucent and Nortel leaned on vendor financing on the way up. Vendor financing wasn't the main thing that broke telecom — the collapse had many causes — but it was documented, and while it lasted it made the boom look bigger than the underlying demand.

This doesn't mean the demand is fake. Most of a big lab's revenue is consumer subscriptions and outside API calls — real people paying real money. The circularity doesn't poison the base. It would poison the margin — the newest, incremental demand, which is exactly the part the exponential is being extrapolated from.

And then there's the demand side, which is where the doubt should actually sit. Reported ARR is often padded with committed-spend minimums — contractual floors a company agreed to, whether or not it uses them. Committed is not consumed. As a rough heuristic — not an industry-standard number, just a way to sanity-check it — that gap in year one can run 40 percent or more. Check it, don't take it from me. The number in the headline is what someone promised to spend, not what got processed.

The tell isn't logos or top-line growth. It's usage-based net revenue retention — do the same customers consume more next year on their own. As an illustrative threshold — again, a line in the sand, not a published benchmark — if that's under about 130 percent, I'd get nervous about the exponential thesis no matter what the press prints. And what actually happens on the ground tends to be an S-curve, not a hockey stick: pilots deploy fast, then flatten. Agents fail on reliability, seat expansion stalls, the rollout saturates.

Here's the cruel part that ties it together. The fix for unreliable agents is more inference — verification passes, running the model several times and voting, longer reasoning chains. So the thing that's supposed to unstick demand pushes cost per successful task right back up. The demand ceiling and the margin trap are the same wall.

So the exponential you're being sold may be manufactured on three floors at once. Possibly vendor-financed demand at the top. Committed-but-not-consumed ARR in the middle. Margin-negative token growth at the bottom. Any one of those unwinding turns "exponential" into "fast, then flat" overnight.

Here's what to notice: none of this requires AI to fail. It can be genuinely, impressively fast and still not be exponential. And at these prices, fast isn't good enough.

So what can you actually check, tonight, instead of taking anyone's word?

First, when you see a big AI financing announced, go find how it's structured — and whether it even closed. The structure lives in the primary sources: the company's own press release, the funder's 8-K and quarterly filings on the SEC's EDGAR site, and the earnings-call transcript. An MOU or a "letter of intent" is not a signed deal; search for whether it was finalized at all. Then search those documents for "prepaid compute," "capacity commitment," or "equity plus." If the money is pointed back at buying the funder's own hardware, treat the demand it "proves" as partly a mirror.

Second, stop reading ARR as consumption. When a lab or vendor reports revenue, look in the same filings and transcripts for whether the figure is committed or consumed, and hunt for usage-based net revenue retention. No NRR number, or one hidden behind "logos" and "bookings," is itself the answer.

Third, separate cost-per-token from cost-per-task in your head. When someone says AI got 10x cheaper, ask: at fixed capability, or on the frontier reasoning models people actually pay for? If it's the former, they've proven nothing about the products driving the growth.

So, three things to do tonight:

① On the next AI mega-deal, open the funder's latest 8-K and the deal press release on EDGAR, confirm whether it's a binding agreement or just an MOU, then search them for "prepaid compute," "capacity commitment," and "equity," and flag any prepaid-compute recycling before you believe the demand.

② For any ARR headline, pull the earnings-call transcript, search it for "net revenue retention," "consumed," and "committed," and ignore the top-line number until you find the consumption figure.

③ Every time you hear "AI is getting cheaper," write down one question — cheaper at fixed capability, or cheaper on the frontier reasoning models people actually pay for — and don't accept the claim until it's answered.

One thing genuinely can't be called from the outside: what fraction of the last two years' datacenter beat traces back to funder-linked entities. If it's a meaningful slice of the incremental, the exponential read is contaminated at the source. If you've seen a real number on that — related-party revenue, a recycle ratio, anything — where did it come from? That's the one figure I'd trade all the narrative for.