Akshat Bubna

Akshat Bubna is a cofounder of Modal, alongside Erik Bernhardsson, and was its CTO at the time of his July 2026 Latent Space interview. His work at Modal centers on infrastructure for AI workloads: elastic model inference, sandboxes for executing agent-generated code, and training environments. He has emphasized keeping infrastructure configuration beside application code and making execution and debugging tools accessible to agents.[2][1]

By July 2026, Bubna said Modal had changed its SDK team’s focus to “agent experience.” He argued that agents benefit from the same compact configuration surface as humans: changing a decorator beside a function lets them provision resources and observe the result without navigating large collections of Kubernetes YAML. Modal Bench helped the team identify difficulties such as finding logs and making the right correction after a failure. The team also moved logs and metrics previously available only in the UI into the CLI.[1]

This approach also appears in Michael Waskom and Rebecka Storm’s November 2025 article, “Agents need good developer experience too”. They described tight feedback loops, programmatic access, actionable errors, and consistent examples as foundations for both human and agent development. Their concrete naming example was changing concurrency_limit to max_containers, distinguishing the number of containers from concurrency within each container.[7]

Bubna’s enthusiasm for agent-operated infrastructure comes with two qualifications. Humans still need observability to interpret behavior and make judgment calls; he suggested that this could become more important than reading the code itself. Sandbox permissions need hard boundaries to prevent exfiltration, supplemented by softer guardrails. He remained neutral about whether an agent harness runs inside a sandbox or connects from outside it, emphasizing production requirements such as persistent files, snapshot and restore, networking controls, and GPU access.[1]

Bubna identified inference for custom models as Modal’s first product-market fit, citing Suno’s audio models, Runway’s video models, robotics, and computational biology. He described the engineering problem as scaling different models across regions as daily traffic cycles and launches change demand. GPU snapshotting preserves GPU state, including a compiled model, to shorten subsequent cold starts. He also distinguished ordinary agent workloads from reinforcement-learning rollouts, which he said can sometimes require 100,000 sandboxes.[1]

As Modal expanded into LLM inference, Bubna emphasized speculative decoding: a smaller draft model proposes tokens, and the larger target model verifies them in batches. This can use compute more efficiently than repeatedly loading weights to generate one token at a time. The number of accepted draft tokens—acceptance length—is a central lever for speed. Modal’s written explanation qualifies the method’s losslessness: it preserves the target model’s sampling distribution, subject to numerical nondeterminism.[1][8]

Bubna described Modal’s work on the block-based speculator DFlash and its practice of open-sourcing and upstreaming inference improvements. The June 2026 article “Speculation Is All You Need”, by Charles Frye, David Wang, and Shankha Biswas, credits Z Lab with the DFlash architecture and describes collaboration with Z Lab on speculator training and SGLang on integration. Bubna presented Auto Endpoints as a way to deploy optimized inference through the UI or CLI while retaining runnable code that users can modify. Adapting draft models as customer data changes remained a prospective direction.[1][8]

Bubna defended concentrating on the software layer rather than owning data centers. In July 2026, he reported that Modal pooled capacity across 17 cloud providers and built a reliability layer over that capacity. He also explained how a networking primitive gained uses beyond its original purpose: Modal built a private IPv6/TCP overlay for distributed training, including the key exchange needed before establishing RDMA networking, and users subsequently adopted the TCP network for other purposes. He described Modal’s multi-node training focus as smaller post-training and elastic research runs.[1]

Within Modal, Bubna described an “auto inference” harness that launches configuration sweeps, profiles workloads, and changes GPU choices. He characterized this as hyperparameter search guided by model intuition rather than architecture search. His separately documented technical contributions include feedback on early drafts of Modal’s GPU Glossary, alongside Nathan Wang and Colin Weld. Charles Frye wrote most of the glossary, building on Matthew Nappo’s initial internal document.[1][10]

These agent and inference tools followed Modal’s earlier work on a general-purpose runtime. Bubna joined Bernhardsson in 2021, when the prototype was called Polyester; his account of that period explains the configuration and execution choices underlying the later AI platform.[3][1]

Bubna recalled meeting Bernhardsson through an investor, when Bernhardsson was already considering a new runtime. Their initial interest included bursty processing, job queues, and data pipelines, alongside inference. Bubna said they added GPUs a year before ChatGPT appeared. Placing resource configuration in code was an important early decision: decorators reduced YAML and allowed users to manipulate configuration as expressive, dynamic code while retaining ownership of their application logic.[1]

Bubna recalled Modal building sandboxes in May 2023 after customer requests, then adapting swyx’s smol developer into a loop that could iterate on its own output. He remembered models diverging after roughly ten iterations. Kenny Ning’s September 2024 article, “Open-source AI agents”, names Modal’s extension devlooper and describes it as iteratively generating and testing code in sandboxes; its illustrated example took 11 iterations to create a Voronoi-diagram library.[1][9]

In a January 6, 2025 post, Bubna said Modal initially hired only when a need became evident. He argued that premature hiring could distract an early startup from prioritization while it was still finding product-market fit. He reported that the team grew threefold during 2024, after customer traction justified more product directions, while continuing to seek people who enjoyed building, focused on customers, and worked with high agency.[3]

Before joining Modal, Bubna worked at Scale AI. By June 2021, he was an engineering manager leading a team building labeling software and systems for text, audio, and computer-vision ground truth. His talk “Humans-in-the-Loop Don't Perform Gradient Descent”, posted that month, framed human-in-the-loop labeling as a game-theoretic optimization problem and connected predicted problems with operational observations.[5][3]

Bubna studied computer science and mathematics at MIT and previously worked at D. E. Shaw and database and fintech startups. His International Olympiad in Informatics record lists a bronze medal representing India in 2013 and a gold medal in 2014. Scale’s historical biographies describe him as the first competitor from India to receive an IOI gold medal.[4][5][6]