Anthropic 宣布 Mythos 只会有限度发布，原因是安全风险，这引发了一场争论：这项限制反映的是安全顾虑，还是单纯的容量限制。批评者指出 Anthropic 的可靠性问题，并表示这个模型看起来并不像许多人预期中那种令人畏惧的突破；而 David Sacks 称它「不是末日装置」，Sophos CEO Joe Levy 则说这并不是一场戏剧性的网路安全转变。更大的背景是，Anthropic 在购买算力方面历来比竞争对手更保守，尽管随著用户数增长，推理需求已变得更重要。

文中引用的证据显示，Anthropic 的基础设施压力如今已在产品上显现。Jefferies 分析师 Brent Thill 表示，Claude Code 的正常运行时间为 99.18%，而 OpenAI 的 Codex 为 99.98%，他把原因归咎于 token 限制、算力限制与技术债。Zapier 的 Wade Foster 说，Anthropic 取得的算力远少于 OpenAI，而且整个产业都受算力约束；Anthropic 也承认，Claude 的需求以前所未有的速度上升，公司的基础设施承受了压力，尤其是在高峰时段。之后，公司透过提高 Claude Code 的速率限制，并与 Elon Musk 的 SpaceX 达成一项新的容量协议，来缓解瓶颈。

如今这场容量竞赛规模惊人：Anthropic 表示，透过扩大的 Amazon 协议将新增 5 gigawatts 的算力，并自 2027 年起从 Google 和 Broadcom 获得 3.5 gigawatts，而 The Information 报导称，与 Google 相关的支出在 5 年内将达到 $200 billion。OpenAI 的 Greg Brockman 表示，公司今年单是算力就会花费 $50 billion，而 4 家最大的 hyperscalers 正在投入超过 $700 billion 的资本支出，其中很大一部分与 AI 有关。文章的结论是，对企业买家而言，模型品质可能不如可靠性、价格和晶片取得那么重要；这意味著，除非 Anthropic 迅速缩小算力差距，否则其领先地位可能会被侵蚀。

Anthropic’s announcement that Mythos would get only a limited release because of security risks set off a debate over whether the restriction reflected safety concerns or simple capacity limits. Critics pointed to Anthropic’s reliability problems and said the model did not look like the feared breakthrough many expected, while David Sacks called it “not a doomsday device” and Sophos CEO Joe Levy said it was not a dramatic cybersecurity shift. The broader context is that Anthropic has historically been more conservative than rivals in buying compute, even as inference demand has become more important as user counts grow.

Evidence cited in the article suggests that Anthropic’s infrastructure strain is now visible in products. Jefferies analyst Brent Thill said Claude Code uptime was 99.18% versus 99.98% for OpenAI’s Codex, blaming token constraints, compute limits, and technical debt. Wade Foster of Zapier said Anthropic has procured a lot less compute than OpenAI and that the whole industry is compute constrained, while Anthropic acknowledged that demand for Claude had risen at an unprecedented rate and that its infrastructure had been stretched, especially at peak hours. The company then moved to ease the bottleneck by increasing Claude Code rate limits and striking a new capacity deal with Elon Musk’s SpaceX.

The capacity race is now massive: Anthropic said it was adding 5 gigawatts of compute through an expanded Amazon deal and 3.5 gigawatts from Google and Broadcom starting in 2027, while The Information reported $200 billion of Google-related spending over 5 years. OpenAI’s Greg Brockman said the company will spend $50 billion on compute this year alone, and the 4 largest hyperscalers are investing more than $700 billion in capital expenditures, much of it tied to AI. The article’s bottom line is that model quality may matter less than reliability, price, and chip access for enterprise buyers, meaning Anthropic’s lead could erode unless it closes its compute gap quickly.