Why Cheaper AI Models Could Slow Nvidia's Growth Engine
Flat AI token spending growth and a shift to cheaper models raise doubts about demand keeping pace with Nvidia's massive chip buildout.

Businesses are quietly switching to cheaper AI models, and that’s a problem for Nvidia because its entire buildout depends on spending growing fast enough to justify the chips, power, and data centers already ordered. Recent data from Ramp, which tracks business AI spending, shows nearly flat token spending growth across August and September, with a rise of less than 2% after that. That pace falls far short of the tenfold growth many AI labs are implicitly counting on to pay for committed capacity.
TL;DR
- Token spending growth has nearly stalled, with Ramp reporting close to zero growth in August and September and under 2% growth in the months after.
- Companies are defaulting to cheaper models once they confirm a less expensive option handles a task well enough, reducing the bill even as usage climbs.
- Google’s new Carbon and Hydrogen models, paired with an Orchestrator feature that auto-switches between them, show how labs themselves are building tools that push customers toward cheaper execution.
- Jensen Huang reportedly described incoming chip and memory commitments as an “avalanche” in an August meeting, reflecting internal concern at Nvidia about capacity outpacing paying demand.
- The most expensive frontier models still take roughly 70% of spend in Ramp’s data, meaning the premium tier isn’t disappearing, but its growth rate relative to total capacity is the open question.
- A new technical analysis found numerical fingerprints of restricted Nvidia Blackwell chips inside open-source models from Chinese labs, adding pressure behind a bipartisan push for tighter export controls.
- Nobody outside the data centers themselves knows the real margins, so claims that the buildout is already overbuilt or perfectly sized are both guesses.
Nothing gets lost.
What’s actually happening to AI token prices?
AI token prices have been falling as models get more efficient and competition between providers intensifies. That’s good news for anyone buying AI capability and bad news for anyone trying to build a revenue case around ever-expanding spending. The mechanism is simple: a business tries a cheaper model on a task it used to send to an expensive one, finds the cheaper model does the job, and switches by default. Ramp’s data shows businesses are increasingly setting internal defaults that steer employees toward less expensive models precisely because those models now clear the bar for most routine work.
This isn’t businesses rejecting AI. It’s businesses getting more deliberate about which model they pay for and when. Frontier intelligence scores only matter if they enable something a company couldn’t do before. For work a business already knows how to do, cheaper execution is the more attractive upgrade, not a smarter model.
Why does flat spending growth matter for Nvidia?
Nvidia’s buildout assumes that demand for AI compute will keep expanding at a pace that justifies the chips, wafers, and memory already committed years in advance. The Wall Street Journal reported that Jensen Huang told his finance team in August that incoming chip wafers and memory were arriving like an “avalanche,” and that those commitments needed paying customers to make sense. The report also described concern inside Nvidia about major cloud providers eyeing alternative chips as commitments piled up.
If customers keep shifting workloads to cheaper models, token usage can rise sharply while total spending barely moves. A business might use four times as many tokens but pay a quarter of the price per token, leaving the bill flat. That’s a perfectly rational outcome for the customer and a serious complication for anyone who financed massive compute capacity on the assumption that spending would scale alongside usage.
How are AI labs responding to the shift toward cheaper models?
Ironically, the labs themselves are building the tools that make switching easier. Google recently introduced two new Gemini models, Carbon for difficult reasoning and Hydrogen for faster execution, inside its Antigravity development software. A feature called Orchestrator automatically switches between the two during a task, using the expensive model only when needed and defaulting to the cheaper one otherwise.
In Google’s own testing, Orchestrator reportedly came close to Carbon’s success rate while using 24% less quota and finishing 19% faster. Open-source alternatives have shown even larger savings in some cases, with reports of 80-90% cost reductions at comparable speed and quality for certain tasks. The pattern shows up clearly in coding workflows: a capable model diagnoses a problem and plans the fix, then a cheaper, faster model executes the rest of the work.
This behavior isn’t confined to startups. It shows up across traditional companies too, and it’s spreading beyond engineering. Ramp’s November report found that customer success and support functions grew from under 1% of identifiable AI spend in June to more than 13% by October. Most of that support work, like answering a shipment status question, doesn’t need frontier-level reasoning. It needs a fast, correct answer, which is exactly what cheaper models are good at.
Is the AI capacity buildout actually overbuilt?
There’s no clear evidence of that yet, and the honest answer is that nobody outside the data centers knows for sure. Token demand is still growing quickly in absolute terms, unlike the dot-com era’s eventual collapse in demand. The tension here is different: usage can skyrocket while revenue per unit falls just as fast, leaving total spending close to flat. A data center can be extremely busy serving requests and still be a poor investment if the price it charges doesn’t cover what it cost to build.
Whether a given facility is profitable depends on proprietary numbers: construction and power costs, chip depreciation, the price charged per token, and how much business it actually wins. None of that is public, and it varies site by site. Ramp’s data shows the most expensive frontier models still account for roughly 70% of tracked spend, so there’s clearly still a substantial market for top-tier intelligence. The real question isn’t whether frontier models matter. It’s whether total spending across the market grows fast enough, and reliably enough, to match the scale of commitments already locked in by labs and chipmakers.
What role do chip export controls play in this story?
Export controls have become tangled up with the demand question because of a separate but related development. Researcher Will Depue examined six open-source models released by four Chinese AI labs, including DeepSeek and Moonshot, and found numerical fingerprints in the model weights consistent with Nvidia’s restricted Blackwell GB200 and GB300 chips. Those chips are subject to export controls and shouldn’t have been accessible to Chinese labs at all.
The method works because a model’s weights are initialized in patterns partly shaped by the hardware used during training, and some trace of that initial state can survive the full training process. Researchers can work backward from a released model to infer what hardware likely produced it. The fingerprints can’t prove where the chips were physically located, how access was obtained, or whether an entire training run happened on them, but they give investigators a technical signal independent of what any company claims about its own hardware.
That finding is landing at a politically sensitive moment. Bloomberg reported that OpenAI has told lawmakers it supports tighter chip export controls, joining Anthropic, which had already backed similar measures. One proposal, the Chip Security Act, would require location verification for advanced export-controlled chips, aiming to prevent shipments from disappearing from view and resurfacing somewhere they shouldn’t be. With both major labs and both political parties now aligned on tighter controls, the routes through which restricted compute has reportedly reached Chinese developers are likely to narrow.
Frequently Asked Questions
Why would cheaper AI models hurt Nvidia’s business?
- Self-hosted
- SOC 2
- HIPAA
- Audit log
- Role-based access
Nvidia’s chip and data center buildout assumes spending on AI will keep growing fast enough to justify the capacity already ordered. If businesses shift workloads to cheaper models, token usage can rise while total spending stays flat, which weakens the revenue case for that capacity even as AI usage itself keeps climbing.
What does Ramp’s data show about AI spending trends?
Ramp tracks AI token spending across businesses and reported close to zero growth in August and September, followed by growth of under 2% in the following months. Ramp’s monthly figures do fluctuate significantly, but the recent trend has been notably slower than the scale of growth needed to match current capacity commitments.
Are expensive frontier AI models losing market share?
Not necessarily. Ramp’s data shows the most expensive models still represent roughly 70% of tracked spend and even gained share in October. The issue isn’t that premium models are disappearing, it’s whether the overall market is growing fast enough relative to the capacity being built to serve it.
What is the significance of chip fingerprints found in Chinese AI models?
Researchers found numerical patterns in open-source Chinese models consistent with Nvidia’s export-controlled Blackwell chips. This gives investigators a technical way to infer what hardware trained a model without relying on company disclosures, and it’s adding momentum to bipartisan efforts to tighten chip export enforcement.
Is the AI industry in a bubble similar to the dot-com era?
The transcript evidence doesn’t support that comparison directly. Token demand is still rising sharply, unlike the demand collapse seen in the dot-com bust. The open question is narrower: whether revenue growth can keep pace with falling per-token prices and the scale of infrastructure commitments already locked in.