Quick answer
Cheaper AI models are a problem for NVIDIA because businesses are using far more AI without paying much more for it. In his latest analysis, Nate B Jones points to Ramp data showing business AI spending barely grew in August and September, while companies move routine work to cheaper models. NVIDIA and the AI labs have committed to huge amounts of chips and data-centre capacity on the assumption that spending will keep multiplying. For buyers like you, falling prices are good news. For the companies financing the build-out, they raise a hard question about who pays for it all.
Key takeaways
- Usage is up, bills are flat: Ramp’s data showed almost no growth in business token spending over August and September.
- Trading down is now normal: companies test cheaper models, find they do the job and set them as the default.
- Tools are automating it: Google says its Orchestrator in Antigravity got close to Gemini 4 Carbon’s success rate while using 24% less quota and finishing 19% faster.
- Supply is piling up: the Wall Street Journal reports Jensen Huang described NVIDIA’s committed chip and memory purchases as an “avalanche”.
- Reliability matters as much as price: Nate uses the Microsoft 365 outage to argue that pushing AI use without protecting engineering know-how is risky.
In this guide
- Why are cheaper AI models a problem for NVIDIA?
- What does the AI spending data show?
- How are businesses cutting their AI bills?
- What is NVIDIA’s “avalanche” of supply?
- Why can usage grow while spending stays flat?
- What are chip fingerprints and why do export controls matter?
- What does the Microsoft 365 outage have to do with AI?
- What does this mean for your business?
- Watch: the source video
- Frequently asked questions
- Related posts
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Prepare Your Small Business for a Microsoft 365 Outage — a simple plan for email, calls and files when Outlook or Teams goes down.
- How to Track What Your Business Spends on AI Each Month — a one-sheet tracker so you know your AI costs and can spot savings.
This is our plain-English summary of Nate B Jones’s video on AI getting cheaper, with practical takeaways for small businesses. It builds on our earlier article, Do You Need the Most Expensive AI Model? Why Cheaper Is Often Enough, which covered why most everyday work doesn’t need the priciest model.
Why are cheaper AI models a problem for NVIDIA?
Demand for AI is real and enormous. Nate’s question is whether spending is growing fast enough, something like tenfold, to pay for the chips, power and data centres being built.
Here’s the catch. Each time a business discovers a cheaper model does most of the job, it stops paying for the expensive one. The lab usually keeps the customer, and token usage goes up, but the bill doesn’t rise nearly as fast. Multiply that across thousands of companies and you get lots of AI activity without the revenue growth the build-out was planned around.
What does the AI spending data show?
Ramp, a US company that tracks business card and bill spending, publishes an AI index. According to Nate, its data showed virtually no growth in token spending across August and September, followed by a rise of less than 2%.
- Nate is careful to say Ramp’s numbers bounce around: some months show 30–40% growth, others are flat.
- The most expensive frontier models still make up roughly 70% of spend in Ramp’s data, and actually gained share in October.
- Customer success and support grew from under 1% of identifiable AI spend in June to more than 13% in October. That kind of work (“where’s my parcel?”) usually needs a fast, correct answer, not deep reasoning.
📌 Note: Ramp’s data mostly covers US businesses that use Ramp. Treat it as a useful signal, not the whole market.
How are businesses cutting their AI bills?
Nate says the pattern he hears over and over is simple: companies test whether a cheaper model does the job, and as models improve, it often does. Many now set defaults that steer staff towards cheaper models.
Google has just made this easier. Gemini 4 Carbon handles harder reasoning and Gemini 4 Hydrogen is built for fast execution. Inside Google’s Antigravity coding tool, a feature called Orchestrator switches between them during a task. In Google’s own testing, it got very close to Carbon’s success rate while using 24% less quota and finishing 19% faster.
Nate adds that some teams using open-source models report savings of 80–90% or more. That’s his reported observation rather than a published benchmark, so treat it as a direction of travel.
💡 Tip: You don’t need special software to do this yourself. Use a strong model to plan, then a cheaper one for the routine steps. Our guide How to Split a Project Between a Fast, Cheap AI Model and a Premium One shows how.
What is NVIDIA’s “avalanche” of supply?
The Wall Street Journal reported on an August meeting where Jensen Huang reportedly described the chip wafers and memory NVIDIA had committed to buying as coming at him like an “avalanche”. The concern was about commitments piling up, and big cloud providers looking at alternative chips.
NVIDIA has also been deeply involved in financing the AI build-out for its customers. Nate is clear that this isn’t evidence the industry has overbuilt; unlike the dot-com era, token demand is still soaring. But it’s a reason to check whether the capacity being ordered matches what businesses are willing to pay.
📌 Note: NVIDIA has since had a leadership change. We cover what that could mean for your costs in Jensen Huang Resignation: What It Could Mean for Your AI Bill.
Why can usage grow while spending stays flat?
Nate’s example is easy to follow:
| Before | After | |
|---|---|---|
| How much AI you use | 1x | 2x–4x |
| Price per unit | Full price | ½ to ¼ of the price |
| Your bill | £X | About £X |
The provider serves far more requests but collects no extra money. Provider costs are falling too, thanks to better hardware and more efficient models, so a cheaper model can still be a good business. But whether any single data centre pays for itself depends on its prices, how much work it wins and what it cost to build. Those figures are private, so nobody outside can say for sure, and Nate says anyone claiming to know is guessing.
Nate still believes businesses are under-using AI overall. His view is that spending will only climb once companies see clear returns, “the cash register ringing”, rather than just more activity.
What are chip fingerprints and why do export controls matter?
Nate also covers a bipartisan push in the US to tighten chip exports to China, now backed by both Anthropic and OpenAI (Bloomberg reports OpenAI told lawmakers it supports the package). One proposal, the Chip Security Act, would require location checks on advanced chips covered by export controls.
He highlights research by Will DePue, who examined six open models from four Chinese labs, including DeepSeek and Moonshot, and found numerical patterns consistent with NVIDIA Blackwell GB200 and GB300 chips, which are export-controlled. The idea is that the hardware used to set a model’s starting values leaves traces that survive training.
🔎 Unconfirmed: These “fingerprints” suggest access to Blackwell chips, but as Nate points out, they can’t show where the chips were, how access was obtained or whether the whole training run used them. Treat this as an open investigation, not a proven finding.
For most small businesses, the practical effect is indirect: tighter controls could shift chip supply and prices. See Nvidia China Chips Standoff: What It Means for Your Access to AI for more.
What does the Microsoft 365 outage have to do with AI?
On 12 November, a Microsoft 365 outage disrupted Outlook, Exchange Online, Teams and other services for nearly a day. Microsoft’s preliminary report blamed a bug introduced in a July networking update, triggered by extra traffic and reduced server capacity, and said Outlook took longest to recover. Microsoft has said it wasn’t related to cyber activity or AI agents.
Nate accepts the technical explanation. His wider concern is engineering culture: when firms reward staff for AI use and long-serving experts leave, they risk losing the people who know how complex systems really work. He also mentions Amazon’s March incident, where an engineer followed incorrect AI advice drawn from an outdated internal wiki.
🔎 Opinion, not a finding: There is no official evidence that staffing or AI incentives contributed to the Microsoft outage. Nate raises it as a question worth asking, not a proven cause.
🛠️ Practical guide: If your business runs on Outlook and Teams, read How to Prepare Your Small Business for a Microsoft 365 Outage.
Paying for AI tools? Prices are falling, so check what you’re actually spending. Our step-by-step guide How to Track What Your Business Spends on AI Each Month shows you how in under an hour.
What does this mean for your business?
This is mostly a story about big companies and big money, but there are useful lessons for a small business:
- Shop around. Nate calls 2027 a likely “buyer’s paradise” as token prices keep falling. Avoid locking into long contracts without checking; see Annual or Monthly AI Subscription: How to Choose When Prices May Fall.
- Test before you pay more. Use the cheapest model that does your task well. How to Test Whether a Cheaper AI Model Is Good Enough for Your Work walks you through it.
- Track your spend. Usage can grow while bills stay flat, or the other way round. Know your monthly figure.
- Plan for outages. Big providers fail occasionally. Have a simple backup for email, calls and key files.
- Keep your know-how. Write down how your key processes work, so AI tools support your experience rather than replace it.
Your next step: pull up last month’s AI bills and list every tool and what it costs. It’s the quickest way to see whether you’re overpaying.
Watch: the source video
What this video covers: Nate B Jones explains why cheaper AI models, flat business spending and NVIDIA’s committed supply raise questions about the AI build-out, plus chip export controls and the Microsoft 365 outage.
Video: Nate B Jones (AI News & Strategy Daily), “AI Just Got Cheaper. Here’s Why That’s a Problem for NVIDIA.”, published 16 November 2026. No sponsor. This summary and the practical advice are our own; the analysis and predictions are Nate’s.
Frequently asked questions
Are AI prices really falling?
Yes, the price per unit of AI (tokens) has been falling as hardware and models improve, and providers are releasing cheaper models that handle routine work well. What you pay depends on which models and plans you use.
Does this mean the AI bubble is about to burst?
Not necessarily. Nate says there’s no clear evidence of overbuilding so far, and demand is still growing fast. The open question is whether spending grows quickly enough to pay for the capacity being built.
Should I switch to a cheaper AI model?
Only after testing it on your own tasks. Many businesses find a cheaper model is good enough for routine work and keep a premium model for harder jobs.
What is Ramp’s AI index?
Ramp is a US finance platform that publishes data on how businesses using its cards and bill pay spend on AI. It’s a helpful signal but mostly reflects US companies.
Was the Microsoft 365 outage caused by AI?
Microsoft says no. Its preliminary report points to a software bug in a networking update. Nate’s point is a wider one about protecting engineering knowledge as companies push AI use.
Related posts
- Do You Need the Most Expensive AI Model? Why Cheaper Is Often Enough
- Jensen Huang Resignation: What It Could Mean for Your AI Bill
- How to Split a Project Between a Fast, Cheap AI Model and a Premium One
- Nvidia China Chips Standoff: What It Means for Your Access to AI
