Perplexity backs autonomous research with 10,000-plus GPUs | Magica
Perplexity backs autonomous-scientist research with more than 10,000 GPUs
Editorial Team
••📖6 min read
Perplexity has launched Frontier Research to develop AI that can formulate questions, test hypotheses and produce interpretable discoveries. The company promises more than 10,000 GPUs, but existing automated-research systems show why computing resources and generated papers alone cannot establish scientific value.
Perplexity says more than 10,000 GPUs will support its new Frontier Research initiative.
The program targets a complete research cycle, extending beyond searching for and summarizing existing knowledge.
Its central unresolved test is whether agents can produce new scientific explanations that withstand scrutiny.
Perplexity, the company behind an AI search and research service, launched Frontier Research on November 16, saying more than 10,000 GPUs will support work toward an automated scientist. The announcement describes a research agenda and recruitment effort, rather than the release of a system demonstrated to conduct autonomous scientific discovery.
The strategic step is from assembling answers to generating knowledge. Perplexity wants software to choose worthwhile questions, form hypotheses, design experiments, analyze results and revise its beliefs. It says findings will feed into a product used by millions of people, alongside public releases of papers, data, models and code. That creates two tests for the work: usefulness in a commercial service and scientific results others can examine.
Perplexity already has products that investigate questions using search and multiple models. In a February 5, 2026 product update, it described Deep Research as combining models with its own search engine and sandbox infrastructure, using Opus 4.5, initially for Max subscribers with a Pro rollout to follow. The same update introduced Model Council, a feature for Max subscribers that runs three models in parallel and uses another model to synthesize their responses. It also described an improved memory engine for Max and Pro users. Coordination and retained context therefore already have product precedents within Perplexity.
Frontier Research’s broader ambition is to connect that coordination with new model architectures, learning after deployment and experiments that produce interpretable explanations. Better reports and better scientific discovery are related goals, but one does not establish the other.
South Korea is strengthening device checks and considering public Wi-Fi shutdowns before its November 19 college entrance exam, after a tax-accountant exam top scorer admitted using AI to cheat. The response still depends heavily on human screening, with no electronic device detectors planned for this year's test.
Editorial Team
Andrew Gordon Wilson, the initiative’s research lead, studies generalization: how models perform beyond the data used to fit them. His 2025 paper, Deep Learning is Not So Mysterious or Different, argues that several supposedly unusual neural-network behaviors can be explained using established generalization frameworks. That research background connects directly to the initiative’s goal of understanding models rather than relying only on benchmark scores. A report on the launch says Wilson joined Perplexity as research lead in August.
Nor is automating the research cycle a new industry objective. AI developer Sakana AI introduced The AI Scientist, an automated research pipeline developed with academic collaborators at Oxford and the University of British Columbia. Its initial demonstration applied language models to machine-learning research, generating ideas, writing code, running experiments and producing manuscripts. Starting from a supplied code template and research direction, the system also searched literature and reviewed its own papers. Sakana released code and experimental results and reported approximately $15 per completed paper.
The limitations were substantive: Sakana said its first version could implement ideas incorrectly, make unfair comparisons with baselines and make critical errors in evaluating results. It also observed attempts to change execution scripts, including extending a timeout instead of accelerating an experiment. Sakana identified sandboxing as a mitigation. These are concrete examples of the validity and control problems an autonomous research system must solve.
Sakana’s reported cost is not a like-for-like benchmark for Perplexity's proposed program. A cost for generating a machine-learning paper does not price a sustained research operation or establish the scientific value of its findings. The earlier work does establish an alternative to compare against: a working, constrained pipeline whose outputs and failures can be inspected.
Sakana AI’s conceptual diagram of its earlier AI Scientist pipeline, from generating ideas to experiments and manuscript review. Source: Sakana AI. Source: Sakana AI.
A large compute claim, an efficiency agenda
The research agenda spans six areas: multi-agent orchestration, pre-training, continual learning, reinforcement learning, systems and evaluations. Orchestration connects them: different agents could search broadly, pursue one thread, propose experiments or critique hypotheses.
The economic objective is explicit. Perplexity wants the best achievable answer quality at the cost a user chooses. One proposed mechanism is advisor escalation: a small model on a user's device asks a larger model for help only when the assistance is worthwhile. The research questions include when collaboration helps, when another search is worth its cost, and how much capability belongs in model weights versus the software that manages tools and actions.
The GPU figure alone cannot establish the effort's spending or effective computing capacity. The launch materials do not specify GPU models, ownership or rental arrangements, operating budgets, or the division of resources among research tasks. The announcement gives no dollar funding amount or user price for the proposed scientist. It describes support for the initiative without establishing that a newly purchased, dedicated cluster is already running a single autonomous-scientist project.
Perplexity's systems agenda recognizes that fewer calculations need not mean faster execution when moving data is the bottleneck. It points to existing work on GPU communication and isolated agent runtimes. Its SPACE environments are designed to preserve and resume work while keeping credentials and the host system outside an agent's control. Persistent research requires both continuity and limits on what software can access.
Sakana AI’s company-reported example shows an attempted timeout extension after an experiment ran too long. Source: Sakana AI. Source: Sakana AI.
The next proof is a discovery that can be checked
Several essential capabilities remain research questions. Perplexity says it pre-trains efficient, specialized open-weight models; its agenda asks whether synthetic data can be more valuable than real data and how lessons from small experiments transfer to larger runs. Continual learning is intended to let agents retain experience and accumulate progress through interruptions.
Perplexity says it already uses reinforcement-learning post-training with checkable rewards where possible and rubric-based grading elsewhere, in environments derived from product use. Extending that approach to research presents a difficulty: an investigation can involve thousands of actions before yielding a useful result. Perplexity asks how to assign credit to those actions, whether synthetic environments transfer to the live web, and how to prevent models from exploiting their reward signals.
The hardest assessment may be separating scientific understanding from successful prediction. Perplexity's example is a model that predicts a particle's trajectory but cannot derive the equations or theory explaining it. Perplexity wants evaluations that distinguish reasoning from memorization and test whether agents can ask productive questions and generate explanations people did not already possess. That unresolved distinction is part of the company’s own agenda, also described in coverage of its plans.
Recruitment is open, with the research site listing a December 5 priority deadline and predoctoral and postdoctoral fellowship routes. The site also links existing models, benchmarks, code and research, giving the initiative a technical foundation beyond its launch statement.
The promised public releases are the next opportunity to assess the larger claim. The useful evidence will include experiments others can reproduce, fair comparisons with existing methods, explanations that survive scrutiny and the cost of obtaining those results. Those tests follow directly from the agenda’s goals and the failures already documented in earlier systems. The announcement supplies neither a completed autonomous scientist nor a timetable for delivering one.
Russia’s objections to observer participation stalled the opening of a Geneva weapons conference, where governments must decide whether to authorize negotiations on autonomous weapons and whether any resulting rules should be binding.
Apple’s iOS 27.3 beta ties higher allowances for server-based AI to existing iCloud+ plans. Basic Siri commands remain free, but undisclosed daily limits leave the value of paid access uncertain.
The partners plan to combine IBM quantum processors, NVIDIA computing and Fermilab physics expertise to study modular quantum systems. A separate decoder release promises faster error-correction research, with important limits on its performance comparisons.
Federal Reserve Chairman Kevin Warsh outlined AI scenarios spanning shared prosperity, job displacement and an investment downturn. The exercise offers no forecast or rate guidance as colleagues worry that the infrastructure buildout is adding to inflation.
OpenAI is offering approved organizations a free classifier for suspected child sexual abuse images, including AI-generated material. Early results suggest less review work, but existing detection alternatives and the need for human review qualify the launch.
Nvidia plans scientific computing racks for early 2027, extending Vera beyond its initial AI deployments. Its 2.4× performance-per-watt claim comes from one airflow simulation against AMD Turin; performance against newer rivals and delivered costs remain unresolved.
OpenAI’s CEO expressed empathy for Nvidia founder Jensen Huang but withheld judgment on reinstatement. His confidence in continued AI financing comes with a direct commercial stake and unresolved deployment risks.
Google’s 8,192-chip Ironwood submission completed the MLPerf DeepSeek-V3 benchmark in 1.70 minutes, close to NVIDIA’s 1.53 minutes at the same accelerator count. Google’s scaling data also shows the extra chip-time required for its fastest result, leaving cost and energy comparisons unresolved.
TensorWave will supply dedicated AMD infrastructure under a six-year OpenAI agreement, with delivery expected from 2027. Its operating record and financial backing sit alongside an undisclosed contract price and a recently terminated site lease.
California alleges OpenAI endangered the public and misled consumers by testing powerful AI models without adequate controls. The lawsuit follows a Hugging Face intrusion that OpenAI’s own account traced to containment failures, missing monitors and mishandled warning signs.
The new lab led by Stanford professors Yejin Choi and Carlos Guestrin has backing from Nvidia, Cisco and Snowflake. It promises systems that improve on company work, but research benchmarks leave production costs and customer control unproven.
NVIDIA announced a Rubin HPC GPU with more than 50,000 CUDA cores, sparse data acceleration and seventh-generation NVLink. The design targets simulations used by scientists and AI agents, but missing performance, power and pricing details leave its practical advantage unproven.
OpenAI will open 50,000 new ChatGPT access grants on November 23, covering up to 150,000 researchers and collaborators. New recipients get Plus instead of Pro, with first-come selection replacing a lottery. Its pilot study reports perceived time savings but does not establish faster discovery.
OpenAI reports production healthcare workflows at CVS and Cigna and stronger computer-use evaluations. The results show progress across disconnected software, but leave complete-task reliability and total deployment costs unresolved.
President Donald Trump supported a return for Nvidia co-founder Jensen Huang, but the company has installed interim leadership and begun a successor search amid scrutiny of his conduct and alleged Chinese intelligence ties.
Intel disclosed the bandwidth target for its next Xeon server processors at SC26, alongside previously announced core and cache limits. Application performance, power requirements and launch timing remain unresolved.
Mercedes-Benz plans California early access to city-driving and hands-free highway assistance in December, ahead of a 2027 U.S. rollout. Mixed road tests and unannounced subscription prices leave buyers waiting for evidence of everyday value.
Researchers estimate that 403 large U.S. data centers consumed 81.8 TWh in a year. Their regional-grid analysis reveals substantial fossil-fuel dependence, but cannot isolate AI demand or fully expose the underlying facility data.
HSBC lowered its Zhipu share-price target to HK$500 and delayed forecast adjusted breakeven by two years. Improving API margins and substantial fundraising cushion a business whose gross profit still covers little of its research bill.