Verified
NVIDIA
Piloting- Semiconductors
- 42,000 staff
- IT-run AI factory
- Cursor, Codex, Claude
- 10,000+ Codex users (Apr 2026)
- No AI return reported
NVIDIA deploys AI inside the company through an AI factory its IT organization runs on NVIDIA's own products, plus bought coding agents: Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, and more than 10,000 employees across functions used OpenAI's Codex by Apr 2026. NVIDIA also builds its own AI for chip design, such as NB-Cell and ChipNeMo, but has published no cost, headcount or revenue return from its internal AI use.
nvidia.comTheir approach to AI
- The platform
- NVIDIA's IT organization runs an internal AI factory, built on NVIDIA's own products, that serves inference and agentic apps to about 40,000 employees (Jun 2026), with a personal assistant agent and an intranet chatbot built on it.
- The models
- NVIDIA engineers use Cursor, OpenAI's Codex and Anthropic's Claude, internal tools run open models such as Llama through NVIDIA NIM, and ChipNeMo is a Llama 2 model NVIDIA adapted on its own chip design data.
- The reach
- More than 10,000 NVIDIA employees in engineering, legal, marketing, finance, sales, HR and operations used Codex in Apr 2026, and Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor.
- The catch
- NVIDIA has published no cost, headcount or revenue return from its own AI use; Colette Kress named rising use of AI tools as a driver of operating expense growth in fiscal 2027 (Aug 2026) without saying what those tools return.
Snapshot
- Employees
- ~42,000
- Jan 25, 2026, in 38 countries; 31,000 in research and development (fiscal 2026 Form 10-K)
- Technology budget
- Not disclosed
- Research and development expense was $18.5B in fiscal 2026, with compute and infrastructure costs up 79%
- AI/ML headcount
- Not disclosed
- More than half of NVIDIA's engineers work on software (fiscal 2026 Form 10-K)
- Who owns AI
- Jensen Huang
- Sets the mandate himself; Rama Akkiraju, VP of AI for IT, leads internal AI in the IT organization
- Platform
- IT-run AI factory, built in-house
- Serves about 40,000 employees (Jun 2026); Cursor and Codex bought for engineers
- Models
- OpenAI, Anthropic, open models
- Codex on GPT-5.5, Claude, Cursor; Llama and Nemotron family models through NVIDIA NIM
- Reach
- 10,000+ Codex users
- Apr 2026, across functions; all software engineers and chip designers on Cursor (Oct 2025)
- Use cases
- Coding, chip design, IT, security
- Plus knowledge work in legal, finance, HR and marketing through Codex
Key findings
NVIDIA has published no cost, headcount or revenue return from its internal AI use; its figures are task timings, such as Bill Dally's account at GTC 2026 of a standard cell library port that took 8 people about 10 months and now runs overnight on one GPU, and vendor-published throughput, such as 3 times more committed code with Cursor (Feb 2026).
More than 10,000 NVIDIA employees across engineering, legal, marketing, finance, sales, HR and operations were using OpenAI's Codex on GPT-5.5 by Apr 23, 2026, and Jensen Huang told staff in a company-wide email to use it.
Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, told an all-hands meeting in Nov 2025 that he wants every task that can be automated with AI to be automated, and said at Davos in Jan 2026 that NVIDIA uses Anthropic's Claude all over the company.
NVIDIA's IT organization runs an internal AI factory that serves inference and agentic apps to about 40,000 employees (Jun 2026), and gives every employee a cloud virtual machine in which coding agents run under a zero-data-retention policy with read-only access to production systems (Apr 2026).
CFO Colette Kress said on Aug 26, 2026 that NVIDIA's fiscal 2027 operating expenses will grow in the low 50s percent, driven partly by more use of AI tools, and Jensen Huang said in Mar 2026 that NVIDIA is trying to spend about $2 billion a year on tokens for its engineers.
Where NVIDIA sits
- Stage
- Piloting
NVIDIA has reported no business return it credits to its internal use of AI in the 18 months since May 2025, for any use case or for the company as a whole: its productivity figures are task timings (a cell library port cut from about 80 person-months to one night on one GPU, Bill Dally at GTC 2026) or vendor-published throughput (3 times more committed code with Cursor, Feb 2026), and CFO Colette Kress tied fiscal 2027 operating expense growth to rising AI tool use in Aug 2026 without a return. NVIDIA places at Piloting on its rollout record: NVIDIA said on Apr 23, 2026 that more than 10,000 employees across engineering, legal, marketing, finance, sales, HR and operations use Codex, Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, and NVIDIA's IT-managed AI factory serves inference and agentic apps to about 40,000 employees (Jun 2026), with Huang mandating the push himself, down to a company-wide email urging staff to use Codex. NVIDIA falls short of Scaling because it has published no AI-credited return in any function and no firm-level AI figure.
How the stages are defined - Ownership
- Jensen Huang sets the mandate; NVIDIA IT's enterprise AI team runs the internal platform
- Platform
- Hybrid: IT-run AI factory built in-house; coding agents bought
- Models
- OpenAI, Anthropic and Cursor for staff; open models via NIM; in-house ChipNeMo
- Hosting
- In-house AI factory run by NVIDIA IT, NVIDIA's own AI supercomputers, cloud capacity under $27B of multi-year agreements (AWS among the providers, per The Information), and GB200 and GB300 systems serving Codex
- Reach
- 10,000+ Codex users across functions (Apr 2026); all software engineers on Cursor
How they deploy
Use AI for every task that can be automated
Jensen Huang told an NVIDIA all-hands meeting in Nov 2025, the day after a record quarter, that he wants every task that can be automated with AI to be automated, and called managers who tell staff to use less AI insane. Huang told employees to keep using AI where it falls short and help make it better, and said NVIDIA was still hiring, about 10,000 people short of what it needs.
Rent frontier models, build domain AI
Jensen Huang said on the Aug 26, 2026 earnings call that companies should rent strong intelligence wherever they can, that NVIDIA did so, and that he encourages NVIDIA employees to use cloud AI services as much as they can, while every company also needs its own domain-specific AI. NVIDIA follows both paths internally: staff use Cursor, OpenAI's Codex and Anthropic's Claude, and NVIDIA builds its own models for chip design, such as ChipNeMo, a Llama 2 model adapted on NVIDIA's design documents, code and bug reports.
Run NVIDIA products on NVIDIA's own work
NVIDIA's IT enterprise AI and automation team builds internal tools on NVIDIA's own products and feeds what it learns back to product teams; its intranet chatbot became a test bed for data flywheel ideas now sold through NVIDIA NeMo (May 2026). NVIDIA's internal ticket-analysis agent, ITelligence, and its vulnerability triage agent, Agent Morpheus, both run on NVIDIA NIM, and NVIDIA said in Jul 2026 that it now runs the chip design software its engineers use on its own Vera CPUs.
Token budgets for engineers
Jensen Huang said at GTC in Mar 2026 that he could imagine giving every NVIDIA engineer an annual token budget worth about half of base pay so each one is amplified tenfold, and said on the All-In Podcast that he would be deeply alarmed if an engineer paid $500,000 did not consume at least $250,000 of tokens a year. NVIDIA has not said whether token budgets are formal policy.
Agents alongside a growing workforce
Jensen Huang told CNBC in Feb 2026 that NVIDIA has 42,000 biological employees and will have hundreds of thousands of digital ones, and said on the Aug 26, 2026 call that NVIDIA's roughly 40,000 employees will one day work with 400,000 or 4 million agents. NVIDIA grew from 29,600 employees at the end of fiscal 2024 to about 42,000 at the end of fiscal 2026.
- Deployment model
- NVIDIA runs internal AI as a central platform with tools bought for engineers. NVIDIA's IT organization operates an AI factory that serves inference and agentic apps to about 40,000 employees, a personal assistant platform that other NVIDIA groups plug their own agents into, an intranet chatbot and ITelligence, a ticket-analysis agent (2025 to 2026). For coding and knowledge work NVIDIA buys Cursor, OpenAI's Codex and Anthropic's Claude; NVIDIA IT gave every employee a cloud virtual machine in which the Codex agent runs as a sandbox (Apr 2026). NVIDIA Research and chip design teams build their own AI, including NB-Cell, Prefix RL, ChipNeMo and BugNeMo, and NVIDIA's coding agents team helps engineers across the company adopt AI tools (OpenAI, May 2026).
- Model providers
- NVIDIA employees use OpenAI's Codex, which OpenAI said in Feb 2026 it was rolling out company-wide to about 30,000 NVIDIA engineers on GPT-5.3-Codex and which more than 10,000 NVIDIA staff used on GPT-5.5 by Apr 2026; Cursor, used by all NVIDIA software engineers and chip designers according to Jensen Huang (Oct 2025); and Anthropic's Claude, which Huang said in Jan 2026 NVIDIA uses all over the company. A Spanish-language report of Huang's GTC talk in Mar 2026 quoted him saying all NVIDIA engineers use a mix of Claude Code, Codex and Cursor. NVIDIA's own internal agents run open models through NVIDIA NIM: ITelligence uses Llama 3.3 70B Instruct (Oct 2025), and Agent Morpheus uses 4 Llama 3 models, 3 of them fine-tuned with LoRA (May 2024). ChipNeMo adapted Llama 2 base models on NVIDIA's internal chip design data (2023).
- Data and infrastructure
- NVIDIA's IT-managed AI factory serves inference and agentic apps to about 40,000 employees (Jun 2026), and NVIDIA IT builds its internal tools on NVIDIA's own products, such as NIM microservices; NVIDIA's 10-K says its computing platforms are also available on its own AI supercomputers. NVIDIA had $27 billion of multi-year cloud service agreement commitments at Jan 25, 2026, which it says support its research and development; The Information reported in Sep 2025 that NVIDIA would use its DGX Cloud capacity, rented back from AWS and other clouds, mainly for its own research, which NVIDIA disputed. Codex runs on OpenAI's GPT-5.5, served on NVIDIA GB200 NVL72 and GB300 systems, and OpenAI said it built US-only processing for the NVIDIA rollout (Feb 2026). NVIDIA has not said which cloud hosts the per-employee agent virtual machines.
Platforms and internal tools
IT-managed AI factoryBy 2026
NVIDIA's IT-managed AI factory is the internal platform that serves inference and agentic applications to NVIDIA employees, designed by Nic Borensztein, its lead architect. NVIDIA IT's Secure Agent Workspace reference design, written by NVIDIA IT staff in Jun 2026, describes the pattern: one managed virtual machine per user, single sign-on, default-deny network egress, human approval for state-changing actions and central logging.
- Users
- About 40,000 employees (Jun 2026)
Codex in per-employee cloud workspacesFeb 2026; GPT-5.5 from Apr 2026
NVIDIA gives employees OpenAI's Codex app, which reaches company data over SSH to a cloud virtual machine NVIDIA IT rolled out for every employee, under a zero-data-retention policy with read-only access to production systems through command-line tools and Skills. NVIDIA says debugging cycles that took days now close in hours; OpenAI's case study counts 40,000 NVIDIA employees with access (May 2026).
- Users
- More than 10,000 across functions (Apr 2026)
- Models
- OpenAI GPT-5.5
Cursor for software engineering2025
NVIDIA set an engineering mandate in 2025 to use Cursor across the software development lifecycle, including code generation, reviews, test generation, debugging and automated git and bug-fix workflows built with custom rules and MCP servers. Cursor's case study, which quotes NVIDIA VP of Engineering Wei Luo, says committed code tripled while bug rates stayed flat (Feb 2026).
- Users
- All software engineers and chip designers (Jensen Huang, Oct 2025); over 30,000 daily (Cursor, Feb 2026)
Personal assistant agent and intranet chatbot
NVIDIA IT's enterprise AI and automation team builds an agentic personal assistant platform, designed so NVIDIA groups can connect their own agents and workflows to shared interfaces, skills and automations, and an AI chatbot for the company intranet built on NVIDIA's retrieval-augmented generation research. The same team takes proofs of concept from finance, HR, marketing, communications and operations and supports projects from IT support to supply chain optimization (May 2026).
ChipNeMo and BugNeMoOct 2023
ChipNeMo is NVIDIA's domain-adapted large language model for chip design, trained further on internal design documents, code and EDA scripts and grounded with retrieval from NVIDIA's own data, used as an engineering assistant, an EDA script writer and a bug report summarizer. Bill Dally said at GTC 2026 that ChipNeMo and BugNeMo answer junior engineers' questions that used to take senior designers' time.
- Models
- Llama 2, adapted by NVIDIA
NB-Cell (also reported as NVCell) and Prefix RL
NB-Cell is a reinforcement learning tool that ports NVIDIA's standard cell library of about 2,500 to 3,000 cells to a new manufacturing process, and Prefix RL explores arithmetic circuit designs; Bill Dally said at GTC 2026 that NB-Cell's cells match or beat human designs and that Prefix RL designs come out 20% to 30% better than human ones.
ITelligenceBy Oct 2025
ITelligence is an internal agent built by NVIDIA's IT organization that loads IT tickets, device inventory and identity data into a graph database, uses an LLM to generate root cause keywords for each ticket, and gives IT leaders dashboards and summaries on resolution times, satisfaction, recurring causes and new-hire onboarding problems.
- Models
- Llama 3.3 70B Instruct through NVIDIA NIM
Agent MorpheusBy May 2024
Agent Morpheus is NVIDIA's agentic workflow for software vulnerability triage, wired into NVIDIA's container registry and internal security tools so that each container upload triggers an analysis of whether its known vulnerabilities can be exploited, ending in a draft VEX document that a human analyst approves.
- Models
- 4 Llama 3 models through NVIDIA NIM
Vendors and partners
OpenAIFeb 2026
Codex for engineering and knowledge work · OpenAI said in Feb 2026 that Codex was rolling out company-wide at NVIDIA with cloud-managed admin controls and US-only processing; GPT-5.5 Codex reached more than 10,000 NVIDIA staff by Apr 2026. NVIDIA and OpenAI have worked together since Jensen Huang delivered the first DGX-1 to OpenAI in 2016.
Anysphere (Cursor)2025
AI coding agent for NVIDIA's software engineers · Jensen Huang called out Cursor in Oct 2025 as the tool all NVIDIA software engineers and chip designers use.
AnthropicBy Jan 2026
Claude models for coding and reasoning · Jensen Huang said at Davos in Jan 2026 that NVIDIA uses Claude all over the company.
Cloud service providers
Rented compute for research and development · NVIDIA had $27 billion of multi-year cloud service agreement commitments at Jan 25, 2026; The Decoder, citing The Information, reported in Sep 2025 that NVIDIA had spent about $13 billion renting back its own chips from AWS and other clouds.
Cadence and SynopsysJul 2026
Chip design software NVIDIA's engineers use, now run on NVIDIA Vera CPUs · NVIDIA said in Jul 2026 that Cadence Jasper and Synopsys VCS ran 1.5 times faster on Vera, according to SiliconANGLE.
Leadership
AI, data and platform leaders
Executives
AI org map
Central IT-run AI platform, chip design AI in Research, coding agents bought for engineering
NVIDIA has no chief AI officer: Jensen Huang sets the internal AI mandate himself, through all-hands meetings and company-wide emails. NVIDIA's IT organization builds and runs the internal AI factory, agent workspaces and employee AI tools, with Rama Akkiraju as vice president of AI for IT; NVIDIA Research under Bill Dally builds AI for chip design; and engineering teams adopt Cursor, Codex and Claude with help from a coding agents team.
NVIDIA's internal AI work sits in 3 places: NVIDIA's IT organization, whose enterprise AI teams run the AI factory, the personal assistant platform, the intranet chatbot and ITelligence; NVIDIA Research under chief scientist Bill Dally, which builds AI for chip design; and the engineering organization, where a coding agents team helps engineers adopt Cursor, Codex and Claude. NVIDIA has not published how many people work on internal AI; 31,000 of its about 42,000 employees work in research and development, and more than half of its engineers work on software.
Reporting lines
President and CEO
Finance
Colette Kress, EVP and CFO
NVIDIA Research
Bill Dally, Chief Scientist and SVP
Builds AI for NVIDIA's chip design, such as ChipNeMo, NB-Cell and Prefix RL; reporting line not published
Information Technology
Sonu Nayyar, CIO (unconfirmed)
Reporting line and leader not published by NVIDIA
AI for IT and enterprise AI and automation
Rama Akkiraju, VP
Internal AI tools, the personal assistant platform and the intranet chatbot; that the enterprise AI and automation team reports to Akkiraju is inferred from titles, not published
Enterprise Productivity
Ashwin Jha, Senior Director
Enterprise AI Compute and the AI factory
Jon Fernandez, Director; Nic Borensztein, lead architect
Serves inference and agentic apps to about 40,000 employees (Jun 2026)
Engineering
Wei Luo, VP of Engineering, among others
Includes a coding agents team that helps engineers adopt AI tools (OpenAI, May 2026)
NVIDIA publishes only its company officers, and its executive pages list no chief information officer or AI leader; the IT units and reporting lines on this chart are drawn from NVIDIA blog author biographies and an executive directory, not from an NVIDIA organization chart.
Where the AI workforce sits
Research and development
31,000 of about 42,000 employees at Jan 25, 2026, the group Jensen Huang says all use AI coding tools; NVIDIA does not count its AI specialists separately.
~74%Sales, marketing, operations and administration
11,000 employees at Jan 25, 2026; NVIDIA's Codex users in Apr 2026 included legal, marketing, finance, sales, HR and operations staff.
~26%
Named people
Use cases
NVIDIA uses AI internally in software engineering (Cursor, Codex, Claude), chip design (NB-Cell, Prefix RL, ChipNeMo and BugNeMo), research workflows, IT operations (ITelligence and an intranet chatbot), product security (Agent Morpheus) and knowledge work in legal, finance, HR and marketing through Codex. NVIDIA has reported no business return it credits to any of these use cases, and no firm-level figure for what AI returns inside the company; its published numbers are task timings, adoption counts and vendor-reported throughput, and Colette Kress has tied AI tools to higher operating expenses without a return figure.
| Use case | Area | Status | ROI |
|---|---|---|---|
| AI coding agents NVIDIA software engineers write, review, test and debug code with Cursor, OpenAI's Codex and Anthropic's Claude, and automate git, CI debugging and bug-fix workflows with custom rules. Cursor's case study reports over 30,000 daily NVIDIA developers committing 3 times more code with flat bug rates (Feb 2026), a throughput measure published by the vendor. | Software engineering | In production | |
| Standard cell library porting (NB-Cell) NB-Cell, a reinforcement learning tool, ports NVIDIA's standard cell library to each new manufacturing process overnight on one GPU, work Bill Dally said took 8 people about 10 months, with cells that match or beat human designs (GTC 2026). | Chip design | In production | |
| Circuit design exploration (Prefix RL) Prefix RL uses reinforcement learning to explore arithmetic circuit designs, which Bill Dally said come out 20% to 30% better than human designs (GTC 2026). | Chip design | In production | |
| ChipNeMo and BugNeMo engineering assistants ChipNeMo answers chip designers' questions from NVIDIA's internal documents, writes EDA tool scripts and summarizes bug reports; a group of volunteer NVIDIA engineers was trying it in Oct 2023, and Bill Dally said at GTC 2026 that ChipNeMo and BugNeMo take junior engineers' questions off senior designers. | Chip design | In production | |
| Machine learning research workflows NVIDIA AI researchers use Codex to survey papers, write experiment scripts and run machine learning experiments on remote machines; one NVIDIA researcher told OpenAI it made running experiments 10 times faster (May 2026). | Research | In production | |
| Codex for knowledge work More than 10,000 NVIDIA employees in engineering, product, legal, marketing, finance, sales, HR, operations and developer programs use Codex on GPT-5.5, each agent working in a sandboxed cloud virtual machine with read-only access to production systems (Apr 2026). | Firmwide | In production | |
| Personal assistant agent and intranet chatbot NVIDIA IT runs an agentic personal assistant that other NVIDIA groups extend with their own agents and workflows, and an AI chatbot for the company intranet (May 2026). | Firmwide | In production | |
| ITelligence ticket analysis ITelligence generates root causes for NVIDIA's IT support tickets and gives IT leaders dashboards and summaries on resolution times, satisfaction, recurring causes and onboarding problems for new hires, in place of manual ticket triage (Oct 2025). | IT operations | In production | |
| Agent Morpheus vulnerability triage Agent Morpheus checks whether known software vulnerabilities in each container pushed to NVIDIA's registry can be exploited and drafts the exemption document for a human analyst; NVIDIA said in May 2024 that it cut triage from hours or days to seconds. | Product security | In production | |
| Token budgets for engineers Jensen Huang proposed in Mar 2026 giving each NVIDIA engineer a yearly token budget worth about half of base pay, and said NVIDIA is trying to spend about $2 billion a year on engineering tokens. | Engineering | Announced |
Adoption
NVIDIA spread AI first through engineering and then across functions. ChipNeMo reached a group of volunteer chip designers in Oct 2023 and Agent Morpheus was wired into NVIDIA's container registry by May 2024; Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, and Cursor counted over 30,000 daily NVIDIA developers in Feb 2026. OpenAI began a company-wide Codex rollout to about 30,000 NVIDIA engineers in Feb 2026, and NVIDIA said more than 10,000 employees across engineering, legal, marketing, finance, sales, HR and operations were using Codex on GPT-5.5 by Apr 23, 2026, while OpenAI counted 40,000 NVIDIA employees with access in May 2026. NVIDIA's IT-managed AI factory served inference and agentic apps to about 40,000 employees in Jun 2026, close to the whole workforce, but NVIDIA has not published how many employees use AI each week.
By the numbers
Rollout, first pilot to today
Oct 2023
Bill Dally disclosed ChipNeMo, a large language model adapted on NVIDIA's chip design data that a group of volunteer NVIDIA engineers was using to answer questions, write EDA scripts and summarize bug reports.
May 2024
NVIDIA described Agent Morpheus, an agentic vulnerability triage workflow integrated with its container registry and internal security tools.
Oct 2025
Jensen Huang said all NVIDIA software engineers and chip designers use Cursor, and NVIDIA IT described ITelligence, its internal ticket-analysis agent.
Nov 2025
Jensen Huang told an NVIDIA all-hands meeting that he wants every task that can be automated with AI to be automated, and criticized managers who asked staff to use less AI.
Feb 2026
OpenAI said it was rolling out Codex company-wide at NVIDIA to about 30,000 engineers, and Cursor published a case study counting over 30,000 daily NVIDIA developers.
Apr 2026
NVIDIA said more than 10,000 employees across functions were using Codex on GPT-5.5, that NVIDIA IT had given every employee a cloud virtual machine for agents, and that Jensen Huang had emailed all staff urging them to use Codex.
Jun 2026
NVIDIA IT staff published the Secure Agent Workspace reference design and said NVIDIA's IT-managed AI factory serves inference and agentic apps to about 40,000 employees.
Aug 2026
Colette Kress said more use of AI tools, which she said already enhances engineering productivity, is one driver of NVIDIA's fiscal 2027 operating expense growth in the low 50s percent.
Economics & controls
Spend and ROI
Spending frame
NVIDIA reported fiscal 2026 revenue of $215.9 billion, research and development expense of $18.5 billion, up 43%, and total operating expenses of $23.1 billion, up 41%, with compute and infrastructure costs in research and development up 79%. NVIDIA does not publish a technology or IT budget, and had $27 billion of multi-year cloud service agreement commitments at Jan 25, 2026 to support its research and development.
What AI itself costs
NVIDIA guided in May 2026 to fiscal 2027 operating expense growth in the upper 40s percent, attributed partly to AI productivity tools, and Colette Kress raised that to the low 50s on Aug 26, 2026, citing a broader product portfolio and a further increase in the usage of AI tools. Jensen Huang, asked in Mar 2026 whether NVIDIA spends about $2 billion a year on tokens for its engineers, answered that NVIDIA is trying to. Bryan Catanzaro, NVIDIA's vice president of applied deep learning research, told Axios in Apr 2026 that for his team the cost of compute is far beyond the cost of the employees. NVIDIA has not reported a dollar figure for its AI tool or token spending.
Reported ROI
NVIDIA has reported no cost, headcount, revenue or margin return that it credits to its internal use of AI, for any use case or for the company as a whole. Colette Kress said on Aug 26, 2026 that AI tools already enhance engineering productivity, without a figure.
Productivity markers
NVIDIA's published productivity figures are task timings and throughput: Bill Dally said at GTC 2026 that NB-Cell ports a standard cell library overnight on one GPU, work that took 8 people about 10 months; Cursor's case study reports 3 times more committed code per developer with flat bug rates (Feb 2026); NVIDIA said Codex closes debugging cycles in hours that took days (Apr 2026); and an NVIDIA researcher told OpenAI that Codex made experiments 10 times faster (May 2026). None of these comes with a cost or headcount figure.
Headcount
NVIDIA grew from 29,600 employees at the end of fiscal 2024 to 36,000 a year later and about 42,000 at Jan 25, 2026, 31,000 of them in research and development. Jensen Huang told staff in Nov 2025 that NVIDIA was still about 10,000 people short and promised that employees using AI would still have work to do, and said in Feb 2026 that NVIDIA will have hundreds of thousands of digital employees alongside its human ones.
Governance and risk
Agents in sandboxed workspaces
NVIDIA IT runs each employee's Codex agent in a dedicated cloud virtual machine for auditability, under a zero-data-retention policy, and agents reach production systems with read-only permissions through command-line tools and Skills (Apr 2026). NVIDIA IT staff's Secure Agent Workspace reference design adds single sign-on, default-deny network egress, human approval for any action that changes a system, credential proxies, signed policies and central audit logs (Jun 2026).
Data boundary
NVIDIA's Codex deployment runs under zero data retention (Apr 2026), and OpenAI said it built cloud-managed admin controls and US-only processing for NVIDIA's rollout (Feb 2026). ChipNeMo grounds its answers by retrieving NVIDIA's internal documents, a design IEEE Spectrum reported in 2023 was meant to curb made-up answers.
Humans approve agent output
NVIDIA's Agent Morpheus hands each vulnerability analysis to a human analyst, who decides on patches and exemptions, and analysts' edits feed back into fine-tuning data (May 2024).
Trustworthy AI principles
NVIDIA's fiscal 2026 sustainability report sets Trustworthy AI principles of privacy, safety and security, transparency and nondiscrimination, with teams working across engineering, product, product security, safety and infrastructure to apply them.
Disclosed risks
NVIDIA's fiscal 2026 Form 10-K says that using AI to improve its internal functions and operations may present further risks, costs and challenges, and that failing to address the responsible use of AI by NVIDIA or others could cause reputational or financial harm.
Recent signals
Aug 26, 2026
Colette Kress raised NVIDIA's fiscal 2027 operating expense growth guidance to the low 50s percent, citing more use of AI tools, and Jensen Huang said NVIDIA's roughly 40,000 employees will one day work with 400,000 to 4 million agents.
Jul 26, 2026
NVIDIA said it now runs the chip design software its engineers use, including Cadence Jasper and Synopsys VCS, on its own Vera CPUs.
Jun 29, 2026
NVIDIA IT staff published the Secure Agent Workspace reference design, and NVIDIA said its IT-managed AI factory serves inference and agentic apps to about 40,000 employees.
May 28, 2026
NVIDIA described its IT enterprise AI and automation team, which builds an agentic personal assistant and an intranet chatbot and takes AI proofs of concept from finance, HR, marketing and operations.
May 20, 2026
NVIDIA guided fiscal 2027 operating expense growth to the upper 40s percent, attributed partly to AI productivity tools.
May 15, 2026
SlashGear reported Bill Dally's GTC 2026 account of NB-Cell, which ports NVIDIA's standard cell library overnight on one GPU, plus Prefix RL, ChipNeMo and BugNeMo.
May 12, 2026
OpenAI published a case study counting 40,000 NVIDIA employees with access to Codex and describing NVIDIA's coding agents team.
Apr 23, 2026
NVIDIA's Justin Boitano wrote that more than 10,000 employees across functions were using Codex on GPT-5.5, that every employee got a cloud virtual machine for agents, and that Jensen Huang urged all staff by email to use Codex.
Mar 24, 2026
Jensen Huang said on the All-In Podcast that an engineer paid $500,000 should consume at least $250,000 of tokens a year and that NVIDIA is trying to spend about $2 billion a year on engineering tokens.
Mar 20, 2026
CNBC reported Jensen Huang's Feb 2026 remark that NVIDIA has 42,000 biological employees and will have hundreds of thousands of digital ones.
Mar 17, 2026
Jensen Huang proposed at GTC giving each NVIDIA engineer a token budget worth about half of base pay so that each one is amplified tenfold.
Feb 12, 2026
OpenAI said Codex was rolling out company-wide at NVIDIA to about 30,000 engineers, with cloud-managed admin controls and US-only processing.
Feb 6, 2026
Cursor published a case study saying over 30,000 NVIDIA developers use Cursor daily and commit 3 times more code, with bug rates flat.
Jan 25, 2026
Jensen Huang said at Davos that Anthropic's Claude is incredible and that NVIDIA uses it all over the company.
Nov 25, 2025
Business Insider reported that Jensen Huang told an NVIDIA all-hands meeting he wants every task AI can do to be automated with AI, and called managers who asked staff to use less AI insane.
Quick answers
- How does NVIDIA deploy AI?
- NVIDIA deploys AI inside the company through an AI factory its IT organization runs on NVIDIA's own products, plus bought coding agents: Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, and more than 10,000 employees across functions used OpenAI's Codex by Apr 2026. NVIDIA also builds its own AI for chip design, such as NB-Cell and ChipNeMo, but has published no cost, headcount or revenue return from its internal AI use.
- What does NVIDIA use AI for?
- NVIDIA uses AI internally in software engineering (Cursor, Codex, Claude), chip design (NB-Cell, Prefix RL, ChipNeMo and BugNeMo), research workflows, IT operations (ITelligence and an intranet chatbot), product security (Agent Morpheus) and knowledge work in legal, finance, HR and marketing through Codex. NVIDIA has reported no business return it credits to any of these use cases, and no firm-level figure for what AI returns inside the company; its published numbers are task timings, adoption counts and vendor-reported throughput, and Colette Kress has tied AI tools to higher operating expenses without a return figure.
- Who leads AI at NVIDIA?
- AI at NVIDIA is led by Bill Dally (Chief Scientist and Senior Vice President of Research); Rama Akkiraju (Vice President of AI for IT); Sonu Nayyar (Senior Vice President and Chief Information Officer). Ownership of AI on the org chart: Jensen Huang sets the mandate; NVIDIA IT's enterprise AI team runs the internal platform.
- What AI platform does NVIDIA use?
- NVIDIA's IT organization runs an internal AI factory, built on NVIDIA's own products, that serves inference and agentic apps to about 40,000 employees (Jun 2026), with a personal assistant agent and an intranet chatbot built on it.
- Which AI models does NVIDIA use?
- NVIDIA engineers use Cursor, OpenAI's Codex and Anthropic's Claude, internal tools run open models such as Llama through NVIDIA NIM, and ChipNeMo is a Llama 2 model NVIDIA adapted on its own chip design data.
- How many NVIDIA employees use AI?
- More than 10,000 NVIDIA employees in engineering, legal, marketing, finance, sales, HR and operations used Codex in Apr 2026, and Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor.
- What return has NVIDIA reported from AI?
- NVIDIA has reported no business return it credits to its internal use of AI in the 18 months since May 2025, for any use case or for the company as a whole: its productivity figures are task timings (a cell library port cut from about 80 person-months to one night on one GPU, Bill Dally at GTC 2026) or vendor-published throughput (3 times more committed code with Cursor, Feb 2026), and CFO Colette Kress tied fiscal 2027 operating expense growth to rising AI tool use in Aug 2026 without a return. NVIDIA places at Piloting on its rollout record: NVIDIA said on Apr 23, 2026 that more than 10,000 employees across engineering, legal, marketing, finance, sales, HR and operations use Codex, Jensen Huang said in Oct 2025 that all NVIDIA software engineers and chip designers use Cursor, and NVIDIA's IT-managed AI factory serves inference and agentic apps to about 40,000 employees (Jun 2026), with Huang mandating the push himself, down to a company-wide email urging staff to use Codex. NVIDIA falls short of Scaling because it has published no AI-credited return in any function and no firm-level AI figure.
Open questions
- What does NVIDIA's internal AI use return in cost, headcount or revenue, for coding agents, chip design tools such as NB-Cell, or IT and security agents, given that NVIDIA has published only task timings and vendor-reported throughput?
- How much does NVIDIA spend on AI tools and tokens for its own staff, given that Colette Kress cites AI tools as a driver of operating expense growth and Jensen Huang said NVIDIA is trying to spend about $2 billion a year on engineering tokens?
- Has NVIDIA made token budgets of about half of base pay formal policy for its engineers?
- How many NVIDIA employees use AI tools each week, and how do the counts of 10,000+ Codex users (NVIDIA, Apr 2026), about 30,000 engineers in the Feb 2026 Codex and Cursor rollouts, and 40,000 with Codex access (OpenAI, May 2026) relate?
- Who owns internal AI on NVIDIA's org chart, given that NVIDIA does not publish its chief information officer or the reporting line of its AI for IT team?
- Which clouds host NVIDIA's per-employee agent virtual machines, and which providers supply the $27 billion of multi-year cloud capacity NVIDIA uses for research and development?
- What share of NVIDIA's staff AI traffic goes to OpenAI, Anthropic, Cursor and NVIDIA's own open models, and does any outside model see NVIDIA chip design data?
- Has NB-Cell or other design AI changed how many engineers NVIDIA assigns to cell libraries and verification, or shortened its chip schedules?
Sources
NVIDIA
- OpenAI's GPT-5.5 powers Codex, and NVIDIA is putting it to work (Apr 23, 2026)
- Sidney Knowles puts AI to work inside NVIDIA (May 28, 2026)
- How to govern autonomous agents in enterprise AI factories (Jun 29, 2026)
- Build an AI agent to analyze IT tickets with NVIDIA Nemotron (Oct 20, 2025)
- Applying generative AI for CVE analysis at an enterprise scale (May 23, 2024)
- ChipNeMo: domain-adapted LLMs for chip design (arXiv, Oct 2023, revised Apr 2024)
- Jensen Huang, board of directors biography
- Colette Kress, executive biography
- Rama Akkiraju, NVIDIA Research profile
Filings, reports and investor materials
- NVIDIA Form 10-K for fiscal 2026 (filed Feb 2026)
- NVIDIA fourth-quarter fiscal 2026 results, Form 8-K (Feb 2026)
- NVIDIA first-quarter fiscal 2027 results, Form 8-K (May 20, 2026)
- Q1 fiscal 2027 earnings call summary and transcript (May 20, 2026)
- Q2 fiscal 2027 earnings call transcript, The Motley Fool via Webull (Aug 26, 2026)
- NVIDIA Sustainability Report, fiscal 2026
Partners
News
- Business Insider via AOL: NVIDIA CEO says it's insane some managers aren't going all in on AI (Nov 25, 2025)
- Fortune: Jensen Huang blasts managers telling staff to use less AI (Nov 25, 2025)
- Business Insider via AOL: Jensen Huang says 100% of NVIDIA engineers use Cursor (Oct 15, 2025)
- Fortune: Jensen Huang on AI tokens worth half an engineer's salary (Mar 17, 2026)
- CNBC: Huang pitches AI tokens on top of salary (Mar 20, 2026)
- R&D World: Huang says engineers should spend at least $250K on tokens (Mar 24, 2026)
- ABC Color: how NVIDIA engineers use AI, according to its CEO (Mar 17, 2026)
- Benzinga via Finviz: Jensen Huang calls Claude incredible at Davos (Jan 25, 2026)
- Analytics India Magazine: 30,000 NVIDIA engineers use Codex and Cursor (Feb 12, 2026)
- SlashGear: NVIDIA says AI made a 10-month job doable in one night (May 15, 2026)
- IEEE Spectrum: NVIDIA is piloting a generative AI for its engineers (Oct 31, 2023)
- SiliconANGLE: NVIDIA puts Vera CPUs to work on chip design (Jul 26, 2026)
- Fortune: NVIDIA executive says compute costs more than employees, citing Bryan Catanzaro to Axios (Jun 2026)
- The Decoder: NVIDIA rumored to focus DGX Cloud on its own research (Sep 13, 2025)
- Wikipedia: Bill Dally
Data confidence
- Confirmed
- Stated in a primary or reputable public source.
- Estimated
- Derived or approximated from public reporting.
- Unconfirmed
- Reported, but not yet corroborated.
NVIDIA profile · Enterprise AI diffusion report by Imagine · facts checked against their sources on Nov 5, 2026
