# OpenAI Broadens Its Stack as Agent Infrastructure and AI Biology Advance *By AI News Digest • March 15, 2026* Sam Altman outlined a broader OpenAI strategy around enterprise coding, chips, supply chains, and a less-exclusive Microsoft partnership. Elsewhere, new agent infrastructure and open computer-use data arrived, AI biology drew unusual attention, and Nando de Freitas called for limits on autonomous weapons. ## Platform strategy ### OpenAI leans further into coding, chips, and a broader partner model Sam Altman said ChatGPT is growing strongly and that Codex has shown especially strong momentum, with most enterprise demand still centered on coding and broader knowledge-work adoption expected over the coming year [^1]. He also said OpenAI now expects to rely on a richer semiconductor portfolio than it first thought—partnering with Nvidia and Cerebras while building its own inference chip—and warned that the AI stack is tight enough that one broken layer could cause knock-on effects [^1]. > "The partnership between Microsoft and OpenAI remains of paramount importance." [^1] Altman added that the Microsoft relationship is still crucial but less exclusive on both sides than it was a few years ago, with OpenAI working with other infrastructure partners and Microsoft using other model families too [^1]. *Why it matters:* OpenAI is talking less like a single-model lab and more like a company managing enterprise demand, chip supply, and a diversified infrastructure ecosystem [^1]. ### Perplexity gets a new distribution lever Perplexity said its Android app has passed 100 million cumulative downloads, and that figure does not yet include the broader rollout of Samsung native integration that Aravind Srinivas said is still ahead [^2]. That gives the company both a large installed base and an additional handset-driven distribution channel [^2]. *Why it matters:* Consumer AI competition is increasingly about distribution as well as models, and Samsung integration could materially extend Perplexity's reach [^2]. ## Agent infrastructure ### Pydantic launches Monty for safer, lower-latency agent code execution Pydantic launched Monty, a Rust-based Python interpreter for AI agents, positioned between simple tool calling and full sandboxes [^3]. Samuel Colvin said the focus is safe, self-hostable execution with tight control over what code can do: the system uses registered host functions and type checking, while in-process execution can run in under a microsecond in hot loops versus roughly one second to create a Daytona sandbox in his comparison [^3]. Early traction is notable, with 6,000 GitHub stars, 27,000 downloads last week, and serializable agents defined in TOML coming to Pydantic AI [^3]. *Why it matters:* Monty is built around practical production constraints—latency, self-hosting, and controllable execution—rather than just agent demos [^3]. ### Markov AI opens a large computer-use dataset Markov AI said it is releasing what it calls the world's largest open-source dataset of computer-use recordings: more than 10,000 hours across tools including Salesforce, Blender, and Photoshop, aimed at automating more white-collar work [^4]. Thomas Wolf's brief "wow!" response showed the launch quickly drew notice [^5]. *Why it matters:* The release packages large-scale recordings from real software workflows into open data explicitly aimed at computer-use automation [^4]. ## High-stakes applications and safety ### A canine cancer-vaccine story becomes a rallying point for AI biology A case amplified by Greg Brockman, Demis Hassabis, and Aravind Srinivas described an Australian with no biology background who paid $3,000 to sequence his rescue dog's tumor DNA, used ChatGPT and AlphaFold to identify mutated proteins and design a custom mRNA cancer vaccine, and then received ethics approval to administer it [^6][^7][^8][^9]. According to the shared account, the first injection halved the tumor and improved the dog's condition; Hassabis called it a "cool use case of AlphaFold" and "just the beginning of digital biology" [^9][^7]. > "Cool use case of AlphaFold, this is just the beginning of digital biology!" [^7] *Why it matters:* Whatever one makes of the broader rhetoric around the story, the level of attention from Greg Brockman, Demis Hassabis, and Aravind Srinivas made AI-enabled biology one of the day's clearest discussion points [^6][^7][^8]. ### Nando de Freitas calls for a moratorium on autonomous weapons Nando de Freitas called for a moratorium on AI autonomous weapons, arguing that cheap drones have already shown destructive effectiveness and that turning them into more capable agentic weapons is now technically feasible [^10]. > "It’s time to have a moratorium on AI autonomous weapons." [^10] *Why it matters:* As the ecosystem pushes agent capabilities into software and biology, leading researchers are also arguing that the same technical progress has immediate military implications [^10]. --- ### Sources [^1]: [OpenAI's Sam Altman on supply chains: "one layer of stack failing could cause real knock on effects"](https://www.youtube.com/watch?v=XqYYCfXE8W4) [^2]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2032902150477775132) [^3]: [⚡️Monty: the ultrafast Python interpreter by Agents for Agents — Samuel Colvin, Pydantic](https://www.youtube.com/watch?v=nxnQl4AcqFg) [^4]: [𝕏 post by @DevvMandal](https://x.com/DevvMandal/status/2032104265121140924) [^5]: [𝕏 post by @Thom_Wolf](https://x.com/Thom_Wolf/status/2032865647374356911) [^6]: [𝕏 post by @gdb](https://x.com/gdb/status/2032983723579523248) [^7]: [𝕏 post by @demishassabis](https://x.com/demishassabis/status/2033010713413754987) [^8]: [𝕏 post by @AravSrinivas](https://x.com/AravSrinivas/status/2032885558020812887) [^9]: [𝕏 post by @IterIntellectus](https://x.com/IterIntellectus/status/2032858964858228817) [^10]: [𝕏 post by @NandoDF](https://x.com/NandoDF/status/2032946531670884623)