Microsoft's $3.2B Anthropic Gain Against a $600M OpenAI Markdown in the Same Quarter
In the fourth quarter Microsoft’s Anthropic investment produced a $3.2 billion gain while its OpenAI position was marked down by roughly $600 million, against a $5 billion gain on that same OpenAI position for the full year. Two frontier lab stakes, held by one company, moving $3.8 billion in opposite directions inside ninety days. There is no operating business inside Microsoft that swings that hard that fast, which makes the other income line the most volatile input in the model and the least analyzed.
The mechanics are worth stating plainly because they are counterintuitive. These are fair-value equity positions in private companies. The marks are set by primary rounds and secondary transactions Microsoft may have no part in, they flow through earnings rather than through equity, and the positions cannot be sold to realize or hedge them. Microsoft’s reported profit is therefore partly a function of what other investors paid other parties for shares Microsoft happens to also own. A quarter in which the mark moves against you is not a quarter in which anything about your business changed.
What makes the OpenAI markdown interesting is that it coincides with an unambiguously improving revenue picture at the marked company. In an internal meeting, finance chief Sarah Friar told employees that annualized recurring revenue in July exceeded the level for the whole of the second quarter. Revenue accelerating, mark falling. That combination points at the instrument or the round structure rather than at the operating trajectory, which is exactly the sort of distinction that disappears when a $600 million number is reported as a loss on an AI bet.
There is also a circularity here that the SpaceX and Alphabet situation has already made familiar. Microsoft books cloud revenue from OpenAI. Microsoft also marks its equity in OpenAI. Money flows out of one company as compute spend, is recognized as revenue by the vendor, contributes to the valuation the vendor then marks its own stake against, and the vendor discloses both lines in the same filing without the reader being able to net them. Nobody is doing anything improper. But a valuation partly underwritten by your own revenue is not an independent price, and the disclosure format invites treating it as one.
The comparison that puts a number on how loose private AI pricing has become is OpenRouter. Reporting put it at roughly $140 million in annualized revenue, about $12 million a month and nearly three times its April level, against Stripe’s $10 billion offer. That is somewhere around seventy times annualized revenue for a routing layer, which is a business with real switching-cost characteristics and essentially no proprietary model asset. Growth of 3x in three months justifies a large multiple. It does not obviously justify that one, and the offer is the more informative data point than the revenue.
Microsoft’s competitive position in this is genuinely singular and it is not about either lab individually. It is the only public company carrying a mark on two frontier labs at once, which means it holds an option on the outcome rather than a bet on a winner. If the leading position rotates, the aggregate stake does not care. That optionality is real and it is unpriced, because the sell side models the operating segments and treats other income as noise to be normalized away. The counterargument is equally real: an asset you cannot sell, cannot hedge, and must remark every quarter is a source of earnings volatility that a company with Microsoft’s investor base gets no credit for carrying.
The stock rose more than 7% after hours on the composite print, driven by Azure and Copilot rather than by any of this. Which is the point. A $3.8 billion swing across two positions passed through the results without moving the narrative, and the same line will be treated as an anomaly the next time it goes the other way.
The disclosure to watch is whether the OpenAI instrument itself changed. A markdown against rising revenue is either a structural repricing of the security or a change in the valuation methodology, and only one of those two is worth adjusting a model for.