Perplexity Will Commit At Least 10,000 GPUs to Building an Automated Researcher
Perplexity has launched a Frontier Research division and pledged at least 10,000 GPUs to it. The goal is an automated researcher: a system in which several AI agents split tasks among themselves and keep learning as they work, rather than freezing after training. Aravind Srinivas said the company will publish evaluations, papers, and blog posts, and release models and tools as open source. It's a signal that the AI search startup is aiming at the research labs.
AI-processed from X: @AravSrinivas; edited by Hamidun News
Perplexity, which most people know as an AI search engine that answers questions with links to its sources, has announced a new division. Its CEO, Aravind Srinivas, wrote on X that Perplexity Frontier Research will receive at least 10,000 GPUs. The team will work on two problems: orchestrating multiple agents and continual learning systems. The ultimate goal is to build an automated researcher. The post drew about 173 thousand views, so the topic clearly caught the industry's attention.
For now, this is only an announcement. It includes no timelines, no budget, no model names, and no information about the team. But the claim stands out: 10,000 GPUs is the scale at which large labs typically operate, not product startups.
What an "automated researcher" means
Imagine a graduate student who reads papers on their own, comes up with hypotheses, runs experiments, analyzes the results, and moves on to the next step. An automated researcher is an attempt to build such a "graduate student" out of an AI system. A single model with a long prompt usually won't do the job, because the task is too long and branches in too many directions.
So Perplexity is betting on two pillars.
The first is orchestrating multiple agents. One agent breaks a large task into parts, and others take the pieces: they search for sources, write code, check figures. The results are then assembled into a whole. The difficulty is that the agents must not get in each other's way, must not repeat work, and must notice their neighbors' mistakes in time. Anyone who has built multi-agent chains knows how quickly they fall apart over a long run.
The second is continual learning. A normal model is "frozen" after training: it only learns what fits into the context of a given request. A continual-learning system has to accumulate experience: what worked, what didn't, which techniques produced results. For a researcher, this is critical. If you start every day from a blank slate, you won't get far. At the same time, this is one of the hardest unsolved problems in AI: a model that learns new things tends to forget old ones.
What they promise to open up
Srinivas writes that Perplexity will publish evaluations (that is, the tests used to measure quality), papers, and blog posts, and will release its models and tools as open source. For practitioners, this is the most interesting point. If the promise holds, developers will get not only numbers but also working code for building multi-agent systems.
The company has released open models before, for example a modified version of DeepSeek R1 called R1 1776. But promising to open its own models and agent tools at this scale of compute is a step of a different magnitude. What exactly will be released, and under which license, is not yet known.
What to keep in mind
First, "at least 10,000 GPUs" is a commitment, not a cluster that is already running. It is unclear whether Perplexity is renting capacity or building its own, or when that capacity will be available.
Second, an automated researcher is a goal, not a product. Other labs have made similar claims, and so far none has shown a system that reliably conducts independent research from idea to result.
Third, open-source promises have a mixed track record in the industry. Publishing evaluations and code will be the real test of the words.
For product managers and founders, the takeaway is simple. Companies that started as wrappers around other people's models are increasingly building their own research layer. If you build agents on top of APIs, keep an eye on what Perplexity releases publicly. Its orchestration tools could become a ready-made foundation, and its evaluations could serve as a useful benchmark for testing your own agent systems.
- Nov 21, 2026Perplexity returned risky shops in 15% of shopping answers, ChatGPT in 8%
- Nov 21, 2026Perplexity Releases Decider v1.3: 70 Points on the Decision Index and $68 per Million Tasks
- Nov 21, 2026OpenAI Names 4 Reasons GPT-6 Astra Hides Its Reasoning from Monitors
- Nov 21, 2026Perplexity Launches Frontier Research, Plans to Train Its Research Agent on 10,000+ GPUs
Key Questions Answered
What an "automated researcher" means?
Imagine a graduate student who reads papers on their own, comes up with hypotheses, runs experiments, analyzes the results, and moves on to the next step. An automated researcher is an attempt to build such a "graduate student" out of an AI system. A single model with a long prompt usually won't do the job, because the task is too long and branches in too many directions. So Perplexity is betting on two pillars. The first is orchestrating multiple agents.
What they promise to open up?
Srinivas writes that Perplexity will publish evaluations (that is, the tests used to measure quality), papers, and blog posts, and will release its models and tools as open source. For practitioners, this is the most interesting point. If the promise holds, developers will get not only numbers but also working code for building multi-agent systems. The company has released open models before, for example a modified version of DeepSeek R1 called R1 1776.
What to keep in mind?
First, "at least 10,000 GPUs" is a commitment, not a cluster that is already running. It is unclear whether Perplexity is renting capacity or building its own, or when that capacity will be available. Second, an automated researcher is a goal, not a product. Other labs have made similar claims, and so far none has shown a system that reliably conducts independent research from idea to result. Third, open-source promises have a mixed track record in the industry.
Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews
Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.