DeepSeek’s $1B Revenue: What It Means for AI’s Margin War
DeepSeek’s $1B revenue run rate exposes the brutal economics of AI. Plus: GPT-4’s null-handling quirk, Google’s private AI compute, and new dev tools.
The story of the day
DeepSeek’s annualized revenue hit $1B, doubling in months, as it finalizes a $7.5B raise at a $500B valuation. The numbers, shared by CEO Liang Wenfeng with investors, reveal a model where price hikes and scale—not just performance—drive margins. For builders: expect more aggressive monetization from frontier models, and watch for DeepSeek’s IPO to pressure public peers on cost discipline. The question isn’t whether AI can scale—it’s who can do so without bleeding cash.
New tools & launches
Litellm — Open-source router to abstract LLM providers from your code. [GitHub]
Phyll — UX review tool for AI apps that preserves design intent. [GitHub]
CastTake — Turns single product photos into AI-generated UGC ad videos. [Website]
Models & research
GPT-4 vs. GPT-3.5 null handling — A paper shows GPT-4 returns empty responses for "null" concepts (e.g., "Be the null") 30/30 times, while GPT-3.5 never does. The difference isn’t prompt adherence; it’s a shift in how the model interprets existential constraints. [Paper]
Google Private AI Compute — New server-side memory tech for on-device AI keeps data isolated during inference. Proprietary, aimed at enterprises needing compliance without cloud latency. [Blog]
Industry pulse
DeepSeek’s revenue surge underscores a market where scale and pricing power matter more than SOTA claims. The $7.5B raise—targeting Shanghai’s exchange—signals China’s willingness to fund AI at US-style valuations, despite export controls. Builders should note: the capital is flowing, but the margins are razor-thin.
Stat of the day
30/30 — GPT-4’s success rate in returning empty responses for "null" prompts, vs. 0/30 for GPT-3.5. A rare, measurable behavioral shift between generations. [Paper]
Get the full analysis — subscribe to Gradient Pro.
