When Dennis Hannusch needed an internal podcast recording app at NVIDIA, his first thought was not to build one. The privacy constraints meant procuring external software like Riverside would take weeks. Then he pointed Codex at the problem. A few hours later, the app was running, complete with video and audio recording features that Codex had tested autonomously through its desktop computer interaction capabilities.
"I didn't have to do anything," Hannusch says. "It was built and tested completely autonomously."
The story sounds like a productivity anecdote. It is not. What Hannusch discovered, and what NVIDIA is now experiencing at scale across 40,000 employees with Codex access, is something more fundamental: the threshold for what is worth building has shifted. Projects that previously failed the cost-benefit analysis because of procurement friction, setup overhead, or simply the number of hours required, now clear the bar. When code generation cost approaches zero, the question stops being "can we afford to build this?" and becomes "can we afford not to?"
This is not a story about AI making engineers faster. It is a story about AI expanding the radius of what engineers can act on, and what an organization like NVIDIA, which designs the hardware that runs these models, learns when it turns its own tools on itself.
The Coding Agents Team: Infrastructure, Not Adoption
NVIDIA did not simply distribute Codex licenses and hope for the best. The company maintains an internal coding agents team whose job is to help engineers across the company adopt and use AI tools effectively in real-world development workflows. This is a structural investment: instead of leaving AI tool adoption to individual preference and organic spread, NVIDIA created a dedicated function to ensure the tools are integrated into how the organization actually ships software.
This mirrors a broader pattern we have observed in enterprise AI deployment. Organizations that treat AI coding tools as a product to be deployed see adoption plateau at the enthusiast level. Organizations that build internal infrastructure around AI tool adoption, including dedicated teams, standardized workflows, and feedback loops, see adoption rates that enterprise software rarely achieves. Sea Limited reported 87% weekly active usage of Codex across its developer organization after making similar structural investments.
At NVIDIA, the coding agents team serves as the connective tissue between the tool and the engineering organization. They help engineers understand what Codex can do, they identify workflows where AI assistance creates the most leverage, and they feed usage patterns back into how the tools are configured and deployed. This is not a support function. It is a capability construction function.
From MVP to Production: The Amplifier in Practice
Hannusch, a senior software engineer on NVIDIA's agents team, has been using Codex with GPT-5.5 as his default tool for complex engineering work. His experience illustrates the core mechanism of AI amplification: the tool does not replace his engineering judgment, it extends the distance that judgment can travel.
"I've personally found Codex with GPT-5.5 to be way more autonomous, with much less handholding," Hannusch explains. "I'm able to go for long sessions with multiple compactions and find that it still performs with top accuracy and manages to keep the work in context. And it's great at tactically selecting the right tools as well as the right skills."
The key phrase is "multiple compactions." In a long coding session, Codex periodically compresses its context window to maintain coherence across extended tasks. Earlier models would lose the thread after one or two compactions, degrading in accuracy and requiring increasing human intervention. GPT-5.5's ability to maintain accuracy through multiple compaction cycles means the human can set direction, step back, and return to find substantive progress rather than a trail of partially completed work that needs to be restarted.
This is the shift from individual contributor to manager thinking applied to AI. Hannusch defines what needs to be built and what "done" looks like. Codex handles the execution across sessions that span hours rather than minutes. The leverage comes not from typing speed but from the clarity of the specification and the quality of the validation criteria.
His most significant achievement with the tool illustrates this dynamic. Hannusch used Codex to evolve an internal platform from a minimum viable product into a production-ready system, improving scalability and reliability. This was work that had proven difficult with earlier models, which could handle individual tasks but could not maintain the coherence needed to refactor an entire system architecture. The difference was not that GPT-5.5 writes better individual functions. It maintains context across the interconnected decisions that system-level refactoring requires.
The Threshold for Building Has Changed
The podcast recording app is worth examining in detail because it reveals something about how AI changes engineering economics.
NVIDIA has strict privacy constraints around internal tools. Procuring external software that handles audio and video recording requires security reviews, data processing agreements, and compliance checks. The timeline is measured in weeks. When Hannusch realized he could build an equivalent tool in hours using Codex, the calculus inverted. Building became cheaper than buying, even before accounting for the ongoing cost of vendor management.
But the more interesting insight is what happened during the build process. Using the Codex desktop app with computer interaction, the system was able to test the video and audio recording functionality as it was being built. It clicked through the interface, verified that recording started and stopped correctly, confirmed that audio levels were acceptable, and validated the end-to-end workflow. This is the 70-80% completion principle in action: the AI handled both construction and validation autonomously, with the human setting the boundary conditions rather than executing each test step.
"Codex has completely changed the threshold for what's worth building," Hannusch says. This statement is more precise than it appears. The threshold he is describing is not a productivity metric. It is a decision threshold: the point at which the expected value of building something exceeds the cost of building it. When that cost drops by an order of magnitude, an entire category of projects that were previously uneconomical becomes viable. These are typically not the glamorous projects. They are internal tools, process automation, one-time data transformations, and quick prototypes that solve real problems but never justified a sprint allocation.
This is the practical manifestation of a deeper shift. When code is cheap, cognition becomes the scarce resource. The engineer's value shifts from "how fast can I write this?" to "is this the right thing to build, and what does 'right' mean?" The first question is about execution. The second is about judgment. AI amplification rewards the second.
The Research Loop, Automated
NVIDIA's research teams are experiencing a parallel transformation, but the mechanics are different. For engineering teams, Codex extends the distance that engineering judgment can travel. For researchers, it compresses the cycle between hypothesis and evidence.
Shaunak Joshi, an AI researcher at NVIDIA, describes a workflow that would have been laborious to orchestrate manually. He points Codex at a large corpus of research papers in areas like reinforcement learning. GPT-5.5 processes the corpus, traces evidence chains across papers, and generates a knowledge graph that visualizes how concepts connect. From this graph, Joshi identifies hypotheses worth testing. Codex then writes the training scripts, connects to remote machine learning infrastructure via SSH, and runs the experiments.
"It's been a 10x speed improvement just in terms of running experiments," Joshi says, "because it's able to handle the whole end-to-end machine learning research workflow."
The 10x figure deserves context. Speed improvements in research workflows rarely come from a single bottleneck being removed. They come from the elimination of handoff friction: the time spent switching between reading papers, writing code, configuring remote machines, monitoring training runs, and analyzing results. When a single agent can handle this entire chain, the researcher's role shifts from orchestrating logistics to making creative and strategic decisions about where to direct attention.
"GPT-5.5 seems to be much more creative compared to competitors," Joshi says. "It helped me trace snippets of evidence throughout the entire chain, and suggested a knowledge graph of the ideas that really helped me visualize how concepts tied together."
The creativity claim is worth scrutinizing. What Joshi is describing is not creativity in the human sense of generating novel ideas from nothing. It is the ability to synthesize patterns across large volumes of text and surface non-obvious connections. This is a specific form of amplification: the model extends the researcher's reading capacity and pattern recognition across a corpus that would take days to process manually. The creative leap, deciding which connections are meaningful and which hypotheses to pursue, remains with the human.
This division of labor maps directly onto the axiom of cognition as asset. The researcher's durable value is not in the scripts that run the experiments or in the logistics of SSH connections to remote machines. It is in the judgment about which research directions are promising, which evidence chains are worth following, and which anomalies in the data deserve deeper investigation. Codex automates the logistics. The researcher invests in the judgment.
Python to Rust, 20x Faster
One of the most concrete efficiency gains Joshi describes is code migration across languages.
"If you have an old codebase that isn't that performant, Codex is really good at machine translation," he says. "So a lot of folks are taking their Python repository, sending it to GPT-5.5, and it's rewriting it into Rust and making it like 20x more efficient."
This is not a new capability in the abstract. Automated code translation has existed for years. What has changed is the reliability and scope. Earlier translation tools could handle syntactic conversion but struggled with idiomatic patterns, performance characteristics, and the subtle semantic differences between languages. GPT-5.5's ability to produce Rust code that is not just syntactically correct but actually 20x faster suggests it understands the performance semantics of both languages, not just their grammar.
For organizations like NVIDIA, where performance is the product, this capability has strategic implications. Legacy Python codebases that were too expensive to rewrite become candidates for migration. The bottleneck shifts from "can we afford to rewrite this?" to "can we verify that the rewritten version is correct?" This is the verification problem, not the production problem, and it is exactly where human engineering judgment adds the most value.
40,000 Engineers and Counting
NVIDIA has extended Codex access to 40,000 employees, spanning engineering, product, legal, marketing, finance, sales, HR, operations, and developer programs. The deployment runs on NVIDIA's own GB200 NVL72 and GB300 infrastructure, with a zero-data retention policy governing the arrangement. Agents access production systems with read-only permissions through command-line interfaces and Skills, the same agentic toolkit NVIDIA uses to run automation workflows across the company.
The infrastructure choice is not incidental. The GB200 NVL72 delivers 35x lower cost per million tokens and 50x higher token output per second per megawatt compared with prior-generation systems, according to NVIDIA's published benchmarks. These economics make frontier-model inference viable at enterprise scale, which is precisely the problem that has prevented most large organizations from deploying agentic AI tools company-wide.
The zero-data retention policy addresses the other major barrier: trust. NVIDIA's engineers are using Codex on production codebases and proprietary research. The assurance that OpenAI does not retain this data is not a nice-to-have. It is a prerequisite for the kind of deep, autonomous sessions that create the most value. If engineers had to sanitize every prompt to avoid leaking intellectual property, the tool's utility would be severely limited.
The deployment also reflects a decade of collaboration between the two companies. Jensen Huang hand-delivered the first NVIDIA DGX-1 AI supercomputer to OpenAI's San Francisco headquarters in 2016. The current Codex integration runs on hardware that NVIDIA designed for exactly this kind of workload, creating a feedback loop where NVIDIA's hardware improvements enable more capable AI agents, which in turn drive demand for more capable hardware.
What This Means for Engineering Teams
NVIDIA's experience offers three lessons that transfer beyond the specifics of one company or one tool.
The threshold matters more than the speed. A 10x speed improvement in research workflows is impressive, but the deeper change is that projects previously deemed uneconomical now get built. When you measure the impact of AI coding tools, look at the new capabilities that emerge, not just the time saved on existing tasks.
Autonomy requires trust infrastructure. Long, autonomous sessions where Codex handles compaction, tool selection, and self-testing only work when the surrounding infrastructure supports them. Zero-data retention policies, read-only production access, SSH integration, and computer interaction capabilities are not features. They are the preconditions for the kind of deep autonomy that creates real leverage.
Judgment compounds, execution does not. The researchers and engineers getting the most value from Codex are not the ones who write the best prompts. They are the ones with the deepest understanding of their domain, the clearest criteria for what constitutes success, and the discipline to define boundaries rather than micromanage execution. The tool amplifies these qualities. It also amplifies their absence.
Hannusch captures the trajectory: "We're just scratching the surface of what it can do. I'm really excited to keep building real systems and see how far it can go."
The surface, in this case, is 40,000 engineers at one of the most technically demanding companies on earth. What lies beneath is a reconfiguration of how engineering judgment gets converted into working systems, and a preview of what that looks like when the cost of that conversion drops toward zero.
FAQ
What is Codex and how does NVIDIA use it?
Codex is OpenAI's agentic coding application powered by GPT-5.5. NVIDIA uses it as the default tool for complex engineering work and end-to-end machine learning research workflows. Over 40,000 NVIDIA employees have access to Codex, which runs on NVIDIA's own GB200 and GB300 infrastructure with zero-data retention.
How has GPT-5.5 improved Codex for NVIDIA's engineers?
GPT-5.5 enables longer autonomous sessions with multiple context compactions while maintaining accuracy. NVIDIA engineers report it surfaces bugs that previous models missed, selects appropriate tools autonomously, and handles end-to-end workflows from hypothesis generation to experiment execution. Research teams report a 10x speed improvement in running ML experiments.
What production systems has NVIDIA built with Codex?
NVIDIA's coding agents team has used Codex to evolve an internal platform from MVP to production-ready status, improving scalability and reliability. The team also built an internal podcast recording app comparable to Riverside in just hours, with Codex autonomously testing video and audio recording features through computer interaction.
How does NVIDIA ensure data security with Codex?
NVIDIA's Codex deployment is governed by a zero-data retention policy. Agents access production systems with read-only permissions through command-line interfaces. The system runs on NVIDIA's own hardware infrastructure, providing both performance and control.
What is the business impact of NVIDIA's Codex deployment?
NVIDIA reports a 10x speed improvement in end-to-end ML research workflows. Engineers describe it as "completely changing the threshold for what's worth building," enabling projects that previously failed cost-benefit analysis. Code migration from Python to Rust via GPT-5.5 has yielded 20x performance improvements for legacy codebases.
References
- OpenAI, "How NVIDIA engineers and researchers build with Codex," May 12, 2026. https://openai.com/index/nvidia/
- NVIDIA Blog, "OpenAI's New GPT-5.5 Powers Codex on NVIDIA Infrastructure," April 23, 2026. https://blogs.nvidia.com/blog/openai-codex-gpt-5-5-ai-agents/
- StartupHub.ai, "NVIDIA Touts Codex GPT-5.5 Gains," May 2026. https://www.startuphub.ai/ai-news/artificial-intelligence/2026/nvidia-touts-codex-gpt-5-5-gains