NVIDIA is rebuilding its growth story around AI agents, and the material, almost all of it dated mid-to-late October 2026, supports the direction of that story better than its size. The argument has three links. Agents consume far more tokens than chatbots. That moves the contest from peak chip performance to cost per token and tokens per watt. And NVIDIA is reaching up the stack into agent runtime, security and possibly licensed software. Rivals and idle-capacity data cut against the story, but in this evidence they do not overturn it. Most of the supporting numbers are vendor claims from single sources. Treat them as a compass heading, not a surveyed distance.
The demand thesis: agents multiply the tonnage
The master resource in this business is the token, and the case for agents is that they burn it quickly. Jensen Huang says agents "think, reason, iterate, use tools, and as a result consume an enormous amount of tokens" 36. He also says token consumption has gone "through the roof" 36. NVIDIA's own figure is that agentic systems use up to 15 times more tokens than traditional AI applications. That is among the better-corroborated claims in the material, with two sources reporting from September into mid-October 1,29.
The outside estimates agree on direction but disagree on scale:
- Anthropic reports that its agents use roughly four times the tokens of a chat 24. Its multi-agent research system uses roughly 15 times 24.
- Gene Munster estimates 10–15x 36.
- Shay Boloor says agents already consume about five times more tokens than human users on OpenRouter 36.
- An NVIDIA-side speaker estimated in August that an agent probably needs 15 to 100 times the compute of a human doing the same task 32.
My reading is that 15x is the upper end of what the evidence shows. It describes multi-agent or heavy workloads, not the typical single agent, and the independent single-agent data points (4x and 5x) sit well below it. A planner who builds capacity on the headline multiple is building for the heaviest user. The direction is still the same everywhere, and one asset manager has already adopted the unit. Generali Asset Management treats tokens as the main measure of AI demand, in place of seats or subscriptions 28.
From peak performance to cost per token
Once volume is the driver, the buyer's question changes. NVIDIA now stresses time to first token and tokens per watt as the measures of an AI infrastructure investment 31. An industry article notes that AI service providers are focused on cost per token 10. This is the economics of a mill: what matters is the cost of each ton out of the furnace, within a fixed power supply.
NVIDIA's proof points are built on that logic:
- Nscale's GB300 NVL72 evaluation showed 49% higher inference throughput inside the same 264-kW provisioned power budget 37, and 37% more managed GPU capacity 37.
- A Lambda proof of concept showed 24% more cluster token throughput 37 and 19% more GPU capacity 37 within a fixed power budget.
- On AA-AgentPerf, at 100 tokens/s per user for DeepSeek V4.1 Flash, GB300 ranks first in throughput per megawatt. Two sources report this 21.
NVIDIA also argues that the installed base keeps improving through software, claiming a 5x Pareto shift on DeepSeek R1 on GB200 over 12 months on identical hardware 25. All of this is NVIDIA-sourced, so I would call it marketing evidence. It is consistent, but it has not been tested by an independent party.
The opportunity is wider than the GPU
Huang says the agent architecture needs dedicated compute for the LLM, for managing knowledge data and for executing tools 36. That matters because tool execution is CPU work. In Modal's own tests, NVIDIA's Vera CPU handled up to twice as many tasks as an Intel server 7. Intel itself reported that agentic AI was supporting server CPU demand 26. Agentic demand is therefore not exclusive to NVIDIA's GPUs, and the same CPU market is open to its rivals. Intel's D-Series GPU is one example of a competitor in high-memory inference parts 9.
Competitors: credible on cost, unproven on displacement
The sharpest challenge in the material is on cost. In a Qwen3.5 397B-A17B AgentX FP8 comparison at 201 output tokens/sec/user, the modeled cost per million total tokens is $0.018 on NVIDIA's GB300 NVL72 23 and $0.008 on Google's TPU7x 23. That is a single source, but it is the only like-for-like cross-vendor cost comparison in the material, and it runs against NVIDIA. Other vendors make efficiency claims of their own:
- Microsoft says Maia 300 delivers 40% more tokens per megawatt than GPU-based systems 8,22.
- AMD says Helios delivers up to 30% more tokens per dollar than the competition 12.
The software layer is also contested. One article argues that competition intensifies as frameworks become hardware-agnostic and that cloud-cost and performance trade-offs will face more scrutiny 17. The same article cautions that decades of GPU software, infrastructure and developer habits will not reverse overnight 17.
The buying evidence fits that caution. Epoch AI estimated NVIDIA used 60.3% of all CoWoS capacity in 2025, against 13.5% for Google 19. The NVIDIA–Anthropic arrangement is up to 1 GW 13. Reflection AI trained Beam on 10.5K GB300 GPUs 5, and doubleAI plans to buy 1,000 B300 GPUs 14.
This is a tension, not a contradiction. Rivals can plausibly claim lower modeled cost per token, while NVIDIA keeps its supply share and its anchor customers. Both can be true in the near term. The evidence does not say which prevails over a longer horizon. I would watch whether modeled cost gaps like the one between GB300 and TPU7x start showing up in where the large buyers place their orders.
A caution from the utilization data
A pure demand story would assume that a buyer's hunger tracks its usage, and the survey data complicates that. Across three VB Pulse surveys from July to September 2026, 172 of 218 respondents running AI in production (79%) said their GPUs are busy half the time or less 18. Even so, 120 of those 172 (70%) plan to adopt a new, additional or replacement AI compute platform within 12 months 18. The comparable figure for companies with busier GPUs is 34 of 46 (74%) 18, and the difference is statistically insignificant 18. Idle capacity does not appear to be holding back purchase intent.
This is a single-source survey, and the busier group is small. The result is still awkward for the efficiency narrative. Many workloads need only a fraction of a GPU 20, and NVIDIA Run:ai provides fractional GPU allocation 16. NVIDIA sells the capacity, and it also sells the tooling that makes less capacity necessary.
Extending the stack: security first, software fees possibly after
The most deliberate move up the stack is governance. NVIDIA acknowledged in September that recent incidents involved agents bypassing the controls of the applications running them 11. Its answer has two parts. OpenShell governs how agents execute 35. Sentry monitors agent behavior 34. Together they are meant to control capabilities and watch behavior to prevent incidents 33. NVIDIA says Sentry quarantines an agent that tries to leave its software boundary within milliseconds 15.
Partnerships give this layer some substance:
- Palo Alto Networks, for agent identity governance on Sentry 27.
- CrowdStrike, for extending agent security to silicon 6.
- ServiceNow, for multi-layer real-time checks 2.
- Microsoft, for permission control via Execution Containers 30.
Monetization is the open question. A reported contemplated fee would charge $2,000 a year per GPU for NVIDIA AI agent software 3. A Bluesky post attributes the plan to Huang 4. Each element comes from a single source and is described only as contemplated. If it happened, software revenue would scale with the installed base of GPUs. That is my interpretation, not a reported outcome. A moat built from chips, runtime and security together would be harder to dislodge than a chip lead. For now the security partnerships are the better-documented part of the strategy, and the fee is a rumor.
What this implies
- Token intensity is the demand engine, but its size is uncertain. NVIDIA's 15x claim is matched only by Anthropic's multi-agent figure 24, while independent single-agent estimates are 4–5x 24,36. For typical use, plan on the lower figures.
- The contest has moved to cost per token and per watt. NVIDIA's efficiency evidence is vendor-sourced. The one cross-vendor cost comparison models Google's TPU7x as cheaper than GB300 NVL72 on the Qwen3.5 case 23. Share and anchor-customer evidence still favors NVIDIA for now 19.
- Idle GPUs have not dampened buying intent. Firms with underused GPUs plan new platform purchases at rates statistically indistinguishable from busier firms 18. The evidence is one small survey.
- Security and software licensing are NVIDIA's route to higher-margin attachment. The reported $2,000 per-GPU fee is unconfirmed 3.
In short, NVIDIA is moving from selling the mill's machinery toward running the rail line that the agent traffic travels on. The direction is clear. The durable advantage will be settled by independently verified cost per token and by whether the software fee ever materializes.