Tag
swyx argues the market hasn't priced in that every GPU sold gained ~30% in lifetime value, citing Latent.Space's breakdown of DeepSeek's production numbers: 155T tokens/day with high cache hit rates and ~80% GPU cost margins.
The author predicts that frontier AI providers could experience slowing revenue growth even as overall AI usage continues to rise, due to task economics pushing developers to route easy tasks to cheaper models while reserving premium models for complex work. The post frames this as a hypothesis with counterarguments, such as new high-value workloads potentially sustaining premium pricing.
Kyle Corbitt argues that AGI won't truly arrive until H100 rental prices rise about 10x to roughly $30/hour, using a supply-and-demand calculator to show that AGI-capable agents would demand ~28x more compute than currently exists at today's $3.52/hour rate.
Tidelines.ai 的 Token 价格指数显示，2026 年 OpenRouter 上路由的平均 token 价格几乎减半，主要因 DeepSeek V4.1 Flash 等廉价模型占据市场；同时日均 token 用量从约 1 万亿增至 23 万亿，估算日支出达到约 4700 万美元的历史新高。
Rohan Paul highlights SemiAnalysis estimates that Anthropic's inference business may enjoy ~88% post-compute margins, with subscription users receiving over 4X more compute per dollar than API buyers, since the 'heavily subsidized' subscriptions still run at an estimated 50% compute margin within that blended figure.
开发者分享其 LLM 驱动的人生模拟项目 Slow Vale 的工程实践：中文服已有 800+ 持久化 AI 居民共处一个持续运行的城市，文章详述并发决策、动态动作空间、上下文缓存与推理成本。当前运行时使用托管的 DeepSeek Flash 而非本地推理。
Omar Moore argues that personal AI assistants' high COGS ($3k-$7k per user/year) stem from early adopters being tech workers using agents for coding, and predicts model costs will fall as true consumer use cases are cheaper.
The article presents research on the rapid decline in AI performance costs, averaging a 47% quarterly drop across key benchmarks, highlighting a transformative economic trend for AI technology.
Router by Ramp is a tool that reduces AI inference costs by up to 40% by routing requests to the lowest-cost model that meets performance needs, offering a single endpoint for multiple models.
Canva cut its expected revenue growth rate by a third to 20% due to unexpectedly high costs of delivering AI features, highlighting how AI inference costs are undermining traditional SaaS economics. Figma similarly saw margin compression, signaling a broader industry challenge.
A blog post explaining that cache read costs dominate LLM inference spending for agentic workloads, with cumulative costs growing quadratically as context is re-read each turn, and advice to reduce tool call count to cut costs.
Analyzes how AI inference costs are eroding software's traditional high-margin economics, forcing founders to choose between product quality and unit profitability.
The article analyzes factors that could drive AI compute costs up 10x in coming years, including rising lab revenue, margins, and spot prices, while suggesting that inference spending may signal stalled progress.
Token prices have fallen dramatically but enterprise AI bills are rising due to increased token consumption from agents and background inference, illustrating Jevons paradox.
This article analyzes the implications of Chinese open weights AI models like Kimi K3, arguing that while R&D is free for users, inference costs (COGS) remain significant, challenging the notion of free AI and highlighting a return to marginal costs in the era of zero marginal cost software.
Levie argues that lowering the cost of AI intelligence (tokens) is key to large-scale agentic adoption, and predicts most information work will involve agents in the future, citing innovation in both open and closed models.
A discussion on how rapidly falling inference costs (280X for GPT-3.5-level in under 2 years) will lead to Jevons paradox, driving far more demand for AI and making multi-agent systems ubiquitous with cheaper open-source models.
Benedict Evans analyzes the current instability in AI token pricing, noting a supply crunch and uncertain future as infrastructure investment surges and use cases like software development drive demand. He argues that foundation models may become low-margin commodity providers.
GLM 5.2 from Z.ai emerges as a strong open-weights competitor to frontier models like Opus and GPT, but the real story is the impending margin collapse in AI inference as costs decrease and competition intensifies.
OpenAI reportedly cut inference costs for guest ChatGPT users by more than half, reducing GPU requirements to a few hundred, though it's unclear if these optimizations apply to the full product.