In this post 8 sections
At the World Economic Forum in Davos on 21 January 2026, Jensen Huang told BlackRock CEO Larry Fink that most of AI's economic benefit will come from the application layer, where AI is used in finance, healthcare and manufacturing. This was a spoken conversation, summarized on NVIDIA's blog, not an article he wrote. Funding data supports part of it and complicates the rest.
Update, 13 November 2026: NVIDIA announced on 12 November that Huang resigned as CEO, effective immediately. This post covers what he said at Davos while CEO.
Key Takeaways
- Format: this is a mainstage interview, reported by NVIDIA's own blog. The wording is the blog's summary, so I paraphrase and do not treat anything as a verbatim quote.
- Huang described AI as a five-layer stack and said the application layer is where most economic benefit lands. He later wrote the same stack up in a March 2026 essay, which I treat as his authored version.
- He said 2025 was among the biggest venture years on record. Crunchbase confirms about $425 billion in global venture funding in 2025, with AI at roughly half.
- The money is concentrated. Five companies raised about $84 billion, 20% of 2025 venture funding, per Crunchbase. "Applications win" does not describe where most dollars went.
- He argued AI automates tasks, not the purpose of a job, and told people to learn to direct, manage and evaluate AI.
- For founders and marketers, the useful takeaway is where to build, and it comes with a warning: the application layer is also the most crowded.
What Huang said, as reported
NVIDIA's summary of the conversation attributes these points to him. I label them as claims.
| Topic | His claim, paraphrased |
|---|---|
| Stack | AI is five layers: energy, chips and compute infrastructure, cloud data centers, models, applications |
| Value | The application layer, such as finance, healthcare and manufacturing, is where most economic benefit will come from |
| Models | Models are good enough to build on in several industries |
| Work | AI automates tasks, not the purpose of a job; radiology and nursing examples show hiring rising |
| Funding | 2025 was one of the biggest venture years on record, mostly AI-native companies, with a global total above $100 billion |
| Skills | AI is the easiest software to use ever, with nearly a billion users in two to three years; people should learn to direct and evaluate it |
| Countries | Nations should treat AI like electricity or roads and build capability grounded in local language and culture |
| Bubble | Fink said evidence did not suggest a bubble; Huang agreed large investment is needed |
Two cautions. First, the "nearly a billion users" figure and the nursing shortage of roughly 5 million are as relayed by NVIDIA's blog and I did not verify them. Second, Fink runs an asset manager and argued that savers should share in AI-driven growth, so the "no bubble" exchange was between two parties with a stake in the answer.
Checking the funding claim
Crunchbase's year-end report, published 7 January 2026 with data as of 4 January:
- Global venture funding in 2025 was $425 billion across more than 24,000 private companies, up about 30% from $328 billion in 2024.
- AI took about $211 billion, roughly 50% of the total, up 85% from $114 billion in 2024.
- OpenAI, Scale AI, Anthropic, Project Prometheus and xAI together raised $84 billion, about 20% of the total.
So "above $100 billion" is true and understated: the AI total was about double that. Crunchbase's first-quarter 2026 report goes further. Global venture investment reached about $300 billion in the quarter, an all-time quarterly high, with AI at $242 billion, about 80%. OpenAI ($122 billion), Anthropic ($30 billion), xAI ($20 billion) and Waymo ($16 billion) raised a combined $188 billion, about 65% of the quarter. Early-stage deal counts were down 30% year over year.
That is the complication. By the companies named, much of the money labeled "AI" went to model builders, which is layer four in Huang's stack, not to applications (Waymo is the clearest application among them). Crunchbase itself warns that totals are driven by a few mega-rounds, so headline figures overstate what a typical startup can raise.
Where Huang's framing helps and where it does not
Helps: the layer model is a decent checklist for any AI product. Ask which layer you depend on, who controls it and what happens if its price changes. In my post on compute costs I show why that matters for budgets.
Does not help: it says applications capture value but not which applications, or whether the application builder or the layer owner keeps the margin. The company giving the talk sells a lower layer. If applications win, demand for chips rises, so the frame suits NVIDIA's business. That does not make it wrong. It means you should test it with your own numbers.
What this means for marketers, founders and SEOs
- Build on your own data and distribution. If models are good enough to build on, as he says, the difference between products is workflow, data, brand and reach. For an SEO team that means original research, first-party data and clear expertise. See original research as a content moat on a solo budget.
- Stay model-agnostic. Keep prompts, evaluation sets and data portable so you can switch layers when prices or quality change. My model routing strategy and evaluation protocol help.
- Write your role around purpose, not tasks. His "tasks, not purpose" point is the most practical idea in the talk. List the ten tasks you do each week, mark which an AI could draft, and write down the outcome each serves.
- Do not copy AI-funded startup playbooks. Funding is concentrated in a few giants. A seed-stage founder should plan for efficiency, not for raising a mega-round. See SEO for startups on a budget.
- Learn to evaluate AI output. His call for AI literacy fits the verification skill that employers need. See my AI content workflow.
- Leave alone any strategy that treats a keynote as a forecast. Davos conversations are public relations for the speakers' businesses.
What we do not know yet
- Whether application-layer companies keep their margins as model and compute prices move.
- How many of the 2025 and 2026 funding dollars went to applications rather than infrastructure and models. Crunchbase's totals do not break this out in the pages I read.
- Whether the user and nursing-shortage figures are correct. I did not find a source for either.
- Whether the Davos remarks changed after the March essay. The essay covers similar ground.
FAQ
Did Jensen Huang write an article at Davos?
No. At Davos on 21 January 2026 he was interviewed on stage by BlackRock CEO Larry Fink, and NVIDIA's blog reported on it. His authored 2026 piece is the March essay "AI Is a 5-Layer Cake".
What did Huang say about the application layer?
He said the application layer, including AI in finance, healthcare and manufacturing, is where most economic benefit will come from. He also said models are good enough to build on in several industries.
How much venture funding went to AI in 2025?
About $211 billion of roughly $425 billion, around half, according to Crunchbase's January 2026 report. About 20% of the total went to five companies.
Is there an AI bubble?
Huang and Fink said the evidence did not show one at Davos, but both have a financial interest in AI. I would treat the question as open, since several large companies reported negative free cash flow in October 2026.
What does "AI automates tasks, not purpose" mean?
He means AI can do individual tasks inside a job but does not replace the reason the job exists. It is a reasonable way to review your own role, but it is a framing, not a forecast.
Should startups build at the application layer?
Maybe, but the layer is crowded and depends on model and compute prices set elsewhere. Build around data, distribution or workflow that a model provider cannot easily copy.
Want help with this?
I'm Younus Fardeen, an organic growth, SEO and AEO strategist with 4+ years of marketing experience across edtech and startup teams. If you want help turning AI tooling into a defensible organic growth plan built on your own data and expertise, see my work and reach me through the contact form at younusfardeen.com.
Found this useful? Pass it on.