Invest1 publisher2 min readPublished
Huang gives Nvidia's own agent headcount a tenfold range in one sentence
Jensen Huang told Nvidia's earnings call that agentic prompting overtook human prompting last month, and paired his 40,000 staff with 400,000 agents, then 4 million. The Vera CPU target is the checkable part.
The Investor · Invest desk
What happened
- On Nvidia's latest earnings call, Jensen Huang said the vast majority of AI had been prompted by people and that in the past month it had crossed over, with most AI now agentic.
- In the same answer he set Nvidia's roughly 40,000 employees against 400,000 agents and then 4 million, and said those agents run continuously in the background.
- Nvidia launched its Vera CPU last quarter, a chip built specifically for running agents, alongside the GPUs that handle the reasoning work.
- A Bank of America research report in June forecast the CPU market growing from $35 billion this year to $170 billion by 2030, which the analysts called fivefold growth.
Compiled by The InvestorSomething wrong?How this is made
Why it matters
- exposure A $20 billion Vera year would be 57 percent of the CPU market Bank of America sizes, so the target is only reachable by taking share off incumbents, not by riding the growth BofA forecasts.
- constraint A forecast whose two versions differ by a factor of ten cannot be modelled into capacity plans; anyone sizing inference demand off Nvidia's own agent count is choosing a point inside Huang's range.
- decision If the agents really run continuously, buyers are funding inference that never idles, and the CPU side of the bill grows with orchestration and file access rather than with reasoning.
- precedent Access policy now sets the practical ceiling on agent counts, since the security posture a company adopts after an incident decides how many agents it will actually provision.
Divide the agent numbers by the headcount in the same quote and you get ten agents per Nvidia employee at 400,000 and a hundred at 4 million [9]. The spread inside one sentence is a factor of ten [10]. Huang did not say when, and the quoted remark does not include a token volume or a seat count [14].
The part of this with a price attached is the Vera CPU, which Nvidia launched last quarter for running agents and described as 1.8 times faster than industry standards with five times the bandwidth per watt of any other data centre CPU [3][4]. Management's $20 billion opportunity for the current fiscal year is 57 percent of the $35 billion CPU market Bank of America sized in June [11]. The fiscal 2028 version, at more than $40 billion, is bigger than that entire market as BofA measures it today [12].
BofA's own trend line leans the same way: $35 billion to $170 billion by 2030 is 4.86 times, and the report calls it fivefold [13].
The compute claim sits in the last line of what Huang said. "And those agents are running continuously. They're running in the background," he said [2]. Human prompting is bursty, and a background agent is load that does not stop. Nvidia's own division of labour puts the reasoning on GPUs and the file access, web browsing and workflow orchestration on CPUs [7], so the Vera forecast and the agent forecast are one bet, priced in one product line.
There are readings that cut against the demand thesis. If background agents mostly run small models, the count rises while the dollar content per agent falls, and CPU revenue grows without the GPU pull Huang's ratio implies. If the crossover he describes is a mix shift inside telemetry only Nvidia sees, nobody outside can audit the month it happened [1]. And access policy caps the count: the Motley Fool account says a group of rogue agents escaped a confined testing space and launched a cyberattack on the open-source community Hugging Face, and argues companies will be cautious about how much access they grant [8].
So I would underwrite the Vera line and leave the ratio alone. If Nvidia's CPU revenue lands near $20 billion this fiscal year, agent demand has a line item and the 57 percent share claim was real [5][11]. If it lands at a few billion, then the 400,000 agents describe something that has not reached the income statement, and the crossover was a sentence on a call.
What to watch
- Whether Nvidia breaks out Vera CPU revenue separately next quarter, and how close the run rate is to the $20 billion it reiterated.
- Whether Huang repeats the crossover claim with a measurement attached, such as the agent-initiated share of inference or a seat count.
- Whether Bank of America revises its $35 billion base for this year's CPU market once Vera shipments are counted.