Stability AI faces leadership fragmentation and a compute-to-revenue cliff

Leadership fragmentation and technical loss

I find the continuous loss of technical talent to Black Forest Labs and the massive compute-to-revenue gap to be the primary threats to Stability AI’s survival. Mostaque resigned from his position as CEO to pursue decentralized AI, a decision that highlights a deep rift between the company’s original vision and the demands of its investors. This departure follows months of tension with senior staff and investors who previously called for his removal. Stability AI appointed COO Shan Shan Wong and CTO Christian Laforte as interim co-CEOs. The loss of three major researchers to Black Forest Labs in early 2024 remains a significant technical setback. Mostaque also faced a lawsuit from Cyrus Hodes alleging he tricked the founder into selling a stake before a large fundraising round. Mostaque also faced scrutiny for making exaggerated statements regarding his past in the UK government. An investment from Intel in November 2023 did not stop the instability, as disagreements over the use of copyrighted material for training continued. Will the interim leadership reconcile the decentralized vision with investor demands?

The compute-to-revenue trap

The financial structure of Stability AI creates a high risk of insolvency because of the massive gap between compute costs and revenue. In 2023, an internal forecast reported that the company planned to spend $99M on compute against $11M in projected revenue. This 9:1 ratio illustrates a massive disparity between infrastructure costs and monetization. The company also faces $300M in future contractual obligations to cloud providers. These take-or-pay capacity agreements mean the company cannot reduce spending if revenue fails to reach its targets. Stability AI reported $11.4M in underlying revenue for 2024, though trade reports cited figures as high as $50M. The company’s cumulative deficit reached $239M, and net liabilities stood at $52M. In late 2023, the reported monthly burn reached $8M while monthly revenue stayed between $1.2M and $3M. You should observe that the 2022 funding round only added $76.3M in cash, despite a public announcement of $101M. This discrepancy exists because $11.3M of the $101M total consisted of converted debt and interest rather than new cash. The company also manages a heavy burden from the loss of $300M in future obligations through a 2024 recapitalization.

Market erosion from Black Forest Labs

Black Forest Labs challenges the dominance of Stable Diffusion with its FLUX.1 model family. FLUX.1 dev wins on prompt adherence and text rendering when compared to SD 3.5 Large. FLUX.1 schnell provides a fully commercial option via an Apache 2.0 license. SD 3.5 Large remains a practical pick for users who need a permissive license for revenue under $1M. FLUX.1 dev outperformed Google Gemini 2.0 Flash and OpenAI GPT Image 1 in benchmarks for character consistency and image editing. The FLUX.1 Kontext models further expand this capability by focusing on character consistency and local or global alterations.

The hardware requirements for these models also vary based on the specific GPU used for inference.

An RTX 5080 16 GB can run SD 3.5 Large in FP16, but FLUX.1 dev requires FP8. An RTX 5090 32 GB handles FLUX.1 dev using FP8 plus LoRAs, while SD 3.5 Large fits in FP16 plus an ensemble. While the SD 3.5 Large ecosystem remains mature with robust ControlNet support, FLUX.1 dev offers native capabilities for new deployments.