OpenAI trains GPT-6 Astra on 100,000 GPUs, plans larger 400,000-GPU run
Միայն փաստեր

OpenAI trains GPT-6 Astra on 100,000 GPUs, plans larger 400,000-GPU run

Ամփոփում

OpenAI used a 100,000-GPU cluster to train its GPT-6 Astra model, a project that may have cost up to $1 billion, and NVIDIA CEO Jensen Huang announced a forthcoming 400,000-GPU training effort.

OpenAI completed a training run for its GPT-6 Astra model on a cluster of 100,000 GPUs, a process that analysts estimate could have cost between $540 million and $1 billion depending on hourly GPU pricing. The training likely took place at OpenAI’s Stargate data center in Abilene, Texas.

NVIDIA chief executive Jensen Huang congratulated OpenAI on the achievement and disclosed plans for a subsequent training run that would employ 400,000 GPUs. The projected effort would require roughly 1.2 gigawatts of power, based on the specifications of NVIDIA’s GB 200 NVL72 server racks.

According to OpenAI, GPT-6 Astra features a multi-agent architecture that enables the model to execute tasks across browsers, spreadsheets, websites and desktop applications, generating documents and presentations without relying on separate APIs for each function. The system is described as capable of delegating sub-tasks to specialized agents, allowing it to debug code and adjust strategies autonomously.

Some experts remain skeptical about whether Astra represents artificial general intelligence. Independent benchmarks show that Anthropic’s Fable 5.1 model outperforms Astra on several coding and reasoning tests, including SWE-bench Pro, the Artificial Analysis Intelligence Index, the Artificial Analysis Coding Agent Index and ProofBench v1.1.

Աղբյուր

Wccftech

Շարունակեք կարդալ — անվճար

Ստեղծեք անվճար հաշիվ՝ նորություններին հետևելու համար։ Քարտ չի պահանջվում։