LLMs in Business the Huang Interview - CEO & Founder of NVIDIA

 

What the Huang Interview Shows, How Ready Organizations Are, and What to Expect Next Year

In a long conversation on The Ezra Klein Show on 23 September 2026, Nvidia CEO Jensen Huang made the case that AI has moved from promise to everyday use. This post sums up what he said about how large language models (LLMs) work in business, checks how ready organisations actually are, and predicts how fast things will change over the next year.

1. How LLMs are used and what they actually do in business

Where the value sits. Huang describes AI as a "five-layer cake": energy, chips, data centres ("AI factories"), models, and applications. He argues the application layer matters most. That is where LLMs become legal, healthcare, manufacturing, finance and customer-service tools, and where the economic return actually shows up. For most businesses, building models matters less than putting them inside existing workflows.

What LLMs change in practice. Huang places AI in a long line of general-purpose technologies: "200 years ago, we were able to power anything and everything, electricity. And then... with the internet, we were able to find anything. Today, or soon, we'll be able to know everything and do anything" (Klein & Huang, 2026, ~3:52). In practice, he says, this means that instead of searching and clicking through links, "you just ask it a question and it comes back with an answer. You give it a project, comes back with a solution. You give it a task, it comes back and gets it done." He also says this is a recent change: "In the last six months, A.I. has become, if you will, useful — the inflection point of A.I. Previous to that, we spent 15 years trying to make it work" (Klein & Huang, 2026, ~5:55).

Because LLMs work in plain language ("you just have to speak human"), employees no longer need programming skills to use computing power. This explains why workers have taken them up so fast. MIT's NANDA research found employees at over 90% of surveyed firms used personal AI tools for work, even where official projects had stalled (MIT NANDA, 2025, via Spacedaily, 2026). The "six months" claim is contested, though. Maynard's fact-check rates it "mostly accurate," pointing to a sharp commercial jump in 2026. But he notes that earlier AI products were already useful and that "inflection point" is a recurring Nvidia talking point (Maynard, 2026).

Tasks versus purpose. Huang's main point about jobs is that LLMs automate tasks, not the purpose of a job. In radiology, AI reads the scans, radiologists handle more cases, and hospitals need more of them. Software engineers keep their purpose (solving problems and building products) even if agents write much of the code. The exception is jobs where the task is the job, such as phone-based customer service. Klein pushed back: LLMs are general-purpose and move without the "friction" that slowed earlier waves of automation. Companies should expect both effects. McKinsey's 2026 survey found 80% of respondents say AI improved their personal productivity. But 39% now expect AI-related falls in their organisation's headcount, up from 32% the year before (McKinsey, 2026).

Agents and tools. Huang strips agents of any human qualities. To him, an agent is just software given an objective that plans and optimises toward it. He says the biggest recent gain in usefulness came from LLMs using tools, meaning existing enterprise software, not replacing it. In his view, this is what finally made LLMs useful software for businesses (Klein & Huang, 2026, ~1:00:00). He expects more use of products like Salesforce and Adobe, not a "SaaS apocalypse." He names two other technical drivers:

• Inference-time scaling: models produce better answers by iterating, searching and reasoning longer before answering.
• Skills and memory: agents record what worked so they do repeated tasks better.

McKinsey's data partly supports and partly challenges his view on software. About one in five organisations are scaling coding agents. And 32% have decided not to buy at least one software product because they could build it with agentic coding tools (McKinsey, 2026).

Open versus closed models. Huang says businesses need both:

• Closed frontier models (OpenAI, Anthropic, Gemini, Grok) offer the most advanced capability as a service.
• Open-weight models let a company fine-tune on its own data, build a "data flywheel," keep control of its infrastructure, and avoid depending on someone else's service.

He claims token usage flipped during the year from about 70/30 closed-to-open to about 70/30 the other way. Nvidia's roughly $12.9 billion purchase of Hugging Face is a bet on that trend (TechCrunch, 2026). Many U.S. startups build on Chinese open models, adapted inside their own agent harnesses and sandboxes.

Operational lessons. Huang draws several rules for deploying LLMs at scale:

• Release discipline: "No enterprise is able to operate in an environment where the underlying software is literally changing all the time." Companies must test each new model before putting it into operations.
• Verification as core work: Nvidia spends about 80% of its engineering effort on verification and 20% on design. He expects AI labs, and by implication business users, to move toward similar ratios.
• Containment: the OpenAI agent incident shows agents need sandboxing, external monitors ("watchdogs") and limited access to outside systems until they are proven safe.
• Cost as a factory input: generating answers takes far more computation per user than retrieving files did, so it costs more. Huang judges AI by productivity rather than price. Even so, about 20% of McKinsey respondents say token and operating costs already limit their AI use (McKinsey, 2026).

2. How ready are organisations today?

Overall verdict: widely adopted, unevenly ready. Most organisations are ready for LLMs as assistants. Some are ready for narrow, well-defined agents. Few are ready for autonomous agents that work across departments.

Adoption is broad. 88% of organisations use AI in at least one function, and about 79% use generative AI (McKinsey, 2025; Stanford HAI, 2026, via Tensorway, 2026). Chatbots are the most widely scaled tool: 47% of respondents say their organisation is scaling them company-wide (McKinsey, 2026).

Financial results are thin. Only 37% attribute any EBIT impact to AI, unchanged from the previous year. "High performers," who get at least 5% of EBIT from AI, have stayed at about 6% (McKinsey, 2026). KPMG reports only 7% of leaders have established ROI (KPMG, 2026, via Articsledge, 2026). MIT NANDA's widely cited (and disputed) finding was that about 95% of integrated generative-AI pilots showed no measurable profit-and-loss effect. It also found that budgets went mostly to sales and marketing, while the clearest returns came from back-office automation (MIT NANDA, 2025).

Agent readiness is the weakest area.

• Size gap: 40% of large firms (over $1 billion in revenue) are scaling agents, up from 27%. Smaller firms are flat at 22% (McKinsey, 2026).
• Deployment stage: in Salesforce's survey, only 30% have agents in production, 47% are piloting and 23% are evaluating (Salesforce, 2026).
• Infrastructure: Cisco classes only about 13–14% of organisations as fully AI-ready. Only a third say their infrastructure can support the autonomous agents most plan to deploy (Cisco, 2025, via Digital Applied, 2026).
• Projected failures: Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027 because of cost, unclear value or weak risk controls. It also warns of "agent washing," estimating that only about 130 of the thousands of vendors calling themselves "agentic" are genuine (Gartner, 2025).

What separates successful firms is preparation, not speed. Salesforce found that clean, well-governed data and clearly scoped agents predicted ROI better than being an early adopter. Companies deploying agents reported reaching ROI in about eight months (Salesforce, 2026). The most common obstacle is connecting AI to existing systems, cited by 46% (Anthropic, 2026, via Enterprise DNA, 2026).

How this compares with Huang's view. The data supports his engineering framing. Readiness depends on integration, data, testing and governance, not on how smart the model is. But organisations are not where he implies. Most lack the "80% verification" discipline he describes. Gartner notes governance and risk controls "aren't really designed precisely for agentic systems" yet (CIO.com, 2026). And the OpenAI agent breakout shows that even the frontier labs can build more than they can control.

A caution about the evidence. Most of these figures are self-reported. Several come from vendors with a stake in the results (Salesforce, Anthropic, Google), and these tend to be more optimistic. Huang himself has a stake too: Nvidia gains from every extra unit of compute used.

3. How fast will LLMs improve in the coming year?

Capability will keep improving very quickly. METR measures how long a task, in human working time, an AI agent can finish on its own. On software tasks, that "time horizon" has been doubling about every 4.3 months since 2023 (METR, 2026, via Wikipedia, 2026). An early Claude Mythos model reached a 50%-success horizon of at least 16 hours, which is close to the most METR's task set can measure (Metavert, 2026). Three more doublings over 12 months would mean frontier agents could do roughly two to four working weeks of software work on their own, succeeding about half the time.

My forecast, with caveats:

1. Reliability will lag raw capability. The 80%-success horizon is much shorter and, by some measures, improves more slowly. Sources disagree on that rate: J.P. Morgan cites both about 7 months and about 3 months (J.P. Morgan AM, 2026). Errors and hallucinations remain the main barrier for high-stakes work. Expect agents to be trusted with multi-day coding and research tasks under supervision, but rarely with unsupervised work that affects customers or finances.
2. Outside software, progress will be slower. Time horizons drop sharply in fields without clear success criteria (J.P. Morgan AM, 2026). Outside IT, back-office functions such as finance, HR and procurement will gain the most.
3. Adoption will move more slowly than capability. I expect the share of large firms scaling agents to rise from about 40% to about 50–55%. I expect the share reporting EBIT impact to move only modestly, from 37% to perhaps 40–45%, with high performers still in single digits. Integration and data work, not model quality, will set the pace.
4. Spending on evaluation and governance will rise sharply. After the agent incident, both labs and enterprises will put more compute and staff into evaluation, sandboxing and monitoring. This matches Huang's view that the labs are "in transition" toward production engineering. Some projects will be cancelled, roughly in line with Gartner's forecast.
5. Costs will split. Price per token keeps falling; Google cites a 98% drop since 2024. But each reasoning-heavy agent workflow will cost more, and Gartner expects more than a fivefold rise through 2028 (Morningtick, 2026). Controlling costs will become a main concern when buying AI.

Bottom line: over the next 12 months, models will improve faster than most organisations can absorb. The firms that gain most will treat LLMs the way Huang treats chips: as engineering that needs heavy verification.

---

References

Anthropic. (2026). State of AI agents report. Cited via Enterprise DNA (2026). [link needed]

Cisco. (2025). AI readiness index. Cited via Digital Applied (2026). [link needed]

CIO.com. (2026, February 18). Why most agentic AI projects stall before they scale. https://www.cio.com/article/4132031/why-most-agentic-ai-projects-stall-before-they-scale.html

Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

J.P. Morgan Asset Management. (2026). Artificial intelligence market themes. [link needed]

Klein, E., & Huang, J. (2026, September). Jensen Huang thinks A.I. alarmism has gone too far [Audio podcast episode]. In The Ezra Klein Show. New York Times. https://podcasts.apple.com/us/podcast/jensen-huang-thinks-a-i-alarmism-has-gone-too-far/id1548604447?i=1000791251478

KPMG. (2026, Q2). AI pulse survey. Cited via Articsledge (2026), Enterprise AI statistics 2026. [link needed]

Maynard, A. (2026, September 26). Fact-check: Ezra Klein Show, "Jensen Huang thinks A.I. alarmism has gone too far." https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/huang/factcheck/factcheck.html

McKinsey & Company. (2025, November). The state of AI in 2025. [link needed] (McKinsey's state-of-AI page now shows the 2026 edition.)

McKinsey & Company. (2026, August 24). The state of AI: Global survey 2026. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
Full report (PDF), The state of AI in 2026: On the road to ROI: https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi_final.pdf

METR. (2026). Time horizon 1.1. Cited via Wikipedia and Metavert (2026). [link needed]

MIT NANDA. (2025, July). The GenAI divide: State of AI in business 2025. Cited via Fortune and Spacedaily (2026). [link needed]

Morningtick. (2026, October 9). Google gives Gemini agents their own company email. [link needed]

New York Times. (2026, September 23). Jensen Huang thinks A.I. alarmism has gone too far [Edited transcript]. [link needed]

Podcast Rex. (2026). Transcript: Jensen Huang thinks A.I. alarmism has gone too far [Verbatim transcript]. https://podcastrex.com/shows/the-ezra-klein-show/jensen-huang-vs-the-ai-doomers/transcript

Salesforce. (2026, August 27). New study of 2,025 agentic AI leaders: First to launch isn't fastest to ROI (State of agentic AI in the enterprise). https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/

Stanford HAI. (2026). AI index report 2026. Cited via Tensorway (2026). [link needed]

Mehta, I. (2026, September 3). Nvidia confirms it will buy Hugging Face for $12.9 billion. TechCrunch. https://techcrunch.com/2026/09/03/nvidia-confirms-it-will-buy-hugging-face-for-12-9-billion/

Read More

Augment Code. (n.d.). Scaling AI agents †. https://www.augmentcode.com/guides/scaling-ai-agents

CIO.com. (n.d.). Breaking the 5% ROI ceiling: Why enterprise AI stalls at the pilot stage †. CIO. https://www.cio.com/article/4144243/breaking-the-5-roi-ceiling-why-enterprise-ai-stalls-at-the-pilot-stage.html

CIO.com. (2026, February 18). Why most agentic AI projects stall before they scale. CIO. https://www.cio.com/article/4132031/why-most-agentic-ai-projects-stall-before-they-scale.html

Digital Applied. (2026). Agentic AI project cancellations: Gartner's 40 percent forecast †. https://www.digitalapplied.com/blog/agentic-ai-project-cancellations-gartner-40-percent-2026

Gartner. (2025, June 25). Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 [Press release]. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027

Klein, E. (Host). (2026, September). Jensen Huang thinks A.I. alarmism has gone too far [Audio podcast episode]. In The Ezra Klein Show. The New York Times. https://podcasts.apple.com/us/podcast/jensen-huang-thinks-a-i-alarmism-has-gone-too-far/id1548604447?i=1000791251478
(UK Apple Podcasts lists the episode as "Jensen Huang vs. the A.I. doomers": https://podcasts.apple.com/gb/podcast/jensen-huang-vs-the-a-i-doomers/id1548604447?i=1000791251478)

Maynard, A. (2026). We have been here before † [Article]. Late lessons: AI. https://andrewmaynard.net/late-lessons-ai-sept-2026/analysis/04-article-we-have-been-here-before.html

Maynard, A. (2026). Proof thresholds and liability (Red team A) †. Late lessons: AI. https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/synthesis/redteam/D03-proof-thresholds-liability-A.html

Maynard, A. (2026). Proof thresholds and liability (Red team B) †. Late lessons: AI. https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/synthesis/redteam/D03-proof-thresholds-liability-B.html

Maynard, A. (2026). Reading the analyses and article †. Late lessons: AI. https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/maynard-lens/M8-reading-the-analyses-and-article.html

Maynard, A. (2026). Warnings and thresholds †. Late lessons: AI. https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/synthesis/lens/LA2-warnings-thresholds.html

Maynard, A. (2026, September 26). Fact-check: Ezra Klein Show, "Jensen Huang thinks A.I. alarmism has gone too far." https://andrewmaynard.net/late-lessons-ai-sept-2026/supporting/huang/factcheck/factcheck.html

McKinsey & Company. (2026). The state of AI. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

McKinsey & Company. (2026). The state of AI in 2026: On the road to ROI [Report]. https://www.mckinsey.com/~/media/mckinsey/business%20functions/quantumblack/our%20insights/the%20state%20of%20ai/the-state-of-ai-in-2026-on-the-road-to-roi_final.pdf

Mehta, I. (2026, September 3). Nvidia confirms it will buy Hugging Face for $12.9 billion. TechCrunch. https://techcrunch.com/2026/09/03/nvidia-confirms-it-will-buy-hugging-face-for-12-9-billion/

Mowshowitz, Z. (2026). On Ezra Klein's podcast with Jensen Huang †. Don't Worry About the Vase. https://thezvi.substack.com/p/on-ezra-kleins-podcast-with-jensen

NVIDIA. (n.d.). Jensen Huang at Dreamforce †. NVIDIA Blog. https://blogs.nvidia.com/blog/jensen-huang-dreamforce/

O'Reilly Media. (n.d.). Jensen Huang gets it wrong †. O'Reilly Radar. https://www.oreilly.com/radar/jensen-huang-gets-it-wrong/

Podcast Rex. (2026). The Ezra Klein Show: Jensen Huang thinks AI alarmism has gone too far [Hard Fork feed episode page]. https://podcastrex.com/shows/hard-fork/the-ezra-klein-show-jensen-huang-thinks-ai-alarmism-has-gone-too-far

Podcast Rex. (2026). Jensen Huang vs. the AI doomers [Episode page]. https://podcastrex.com/shows/the-ezra-klein-show/jensen-huang-vs-the-ai-doomers

Podcast Rex. (2026). Transcript: Jensen Huang thinks A.I. alarmism has gone too far [Verbatim transcript]. https://podcastrex.com/shows/the-ezra-klein-show/jensen-huang-vs-the-ai-doomers/transcript

Prefactor. (n.d.). AI agent adoption statistics †. https://prefactor.tech/learn/ai-agent-adoption-statistics

Salesforce. (n.d.). Agentic AI ROI report †. Salesforce Blog. https://www.salesforce.com/blog/agentic-ai-roi-report/

Salesforce. (2026). Agentic enterprise index: Insights 2026 †. https://www.salesforce.com/news/stories/agentic-enterprise-index-insights-2026/

Salesforce. (2026, August 27). New study of 2,025 agentic AI leaders: First to launch isn't fastest to ROI. https://www.salesforce.com/news/stories/agentic-ai-leaders-survey-on-roi/

Semicon Alpha. (2026). Ezra Klein's extended interview with Jensen Huang †. Substack. https://semiconalpha.substack.com/p/ezra-kleins-extended-interview-with

TechRadar. (n.d.). "I've got big ambitions now": Nvidia CEO Jensen Huang says we're finally in the era of useful AI and things will only get faster †. TechRadar Pro. https://www.techradar.com/pro/ive-got-big-ambitions-now-nvidia-ceo-jensen-huang-says-were-finally-in-the-era-of-useful-ai-and-things-will-only-get-faster

The New Stack. (n.d.). Huang on AI agents and engineers †. https://thenewstack.io/huang-ai-agents-engineers/

Wikipedia. (n.d.). Jensen Huang. In Wikipedia. Retrieved [add date accessed], from https://en.wikipedia.org/wiki/Jensen_Huang