OpenAI scraps its Dune-inspired AI model
OpenAI quietly killed an internal AI model code-named Arrakis earlier this year after it failed to meet efficiency targets, according to The Information. Built alongside GPT-4 to drastically lower the eye-watering compute costs of running apps like ChatGPT, the setback left some Microsoft executives disappointed. It shows that slashing inference expenses remains a brutal hurdle, even for AI's market leader.
Even OpenAI isn't immune to hitting a computational wall.
According to a report from The Information, the company recently killed off an ambitious internal AI model code-named Arrakis, named after the harsh desert planet in "Dune." The project, kicked off shortly after ChatGPT debuted last November, was designed to run alongside GPT-4 while dramatically cutting the astronomical costs of powering AI apps.
Turns out, Arrakis couldn't deliver the efficiency gains OpenAI needed, and the company pulled the plug earlier this year.
The $700,000-a-day headache
Operating frontier-grade AI is outrageously expensive. Independent estimates suggest keeping ChatGPT running burns over $700,000 every single day, requiring oceans of expensive cloud compute.
While OpenAI secured billions in Azure compute through its expanded partnership with Microsoft in January, resources are hardly infinite. The Information reported that certain Microsoft executives were left genuinely disappointed by the shortcomings and eventual scrapping of Arrakis.
It marks a rare public speed bump for OpenAI during an otherwise meteoric year. CEO Sam Altman reportedly projects annualized revenue to hit $1.3 billion this year, up from just $28 million in 2022. But mounting compute costs, Google's impending release of Gemini, and heightened regulatory scrutiny ahead of next month's AI safety summit mean the pressure is mounting on all fronts.
Why it matters
Arrakis' demise shows that making models smarter is only half the battle; making them cheap enough to run at scale is where the real war is being fought.
For developers and enterprise customers, this explains why top-tier frontier model inference hasn't dropped in price as quickly as many had hoped. Everyone wants GPT-4 capabilities at a fraction of the compute, but compressing massive architectures without gutting performance is proving notoriously difficult.
We think killing an underperforming model early was the right call from an engineering perspective, but it exposes real friction in OpenAI's relationship with Microsoft. Cloud sponsors want high-margin efficiency, not endless hardware subsidization. With Google breathing down its neck with Gemini, OpenAI will need to crack the cost puzzle soon before enterprise margins start looking like an actual desert wasteland.
Sources (1)
- businessinsider.com(opens in a new tab)Originalbusinessinsider.com