Meta halves internal use of Anthropic's Claude Code to about 30,000 staff

At the pace of that 28-day bill ^[6], Meta's spending on Claude Code ran to about $1.37 billion a year ^[17]. That is the same order as the more than $1 billion Microsoft once projected for all its internal Anthropic spending ^[7]. Crypto Briefing's account of The Information's October 5 report ^[3] does not say where in the year the period fell. If it came early, when Meta had about 60,000 users, it works out to about $1,750 a head per 28 days. At 30,000 users it is about $3,500 ^[19][1].

Microsoft's cut is easier to size, or rather its floor is: more than a third off a projection above $1 billion is more than $333 million a year ^[14]. Its monthly budgets fell by 90% ^[13]. Both figures can be true if the smaller budgets cover only part of the spending, or if full-price months earlier in the year hold the annual total up. The cuts gathered pace in spring and summer ^[12].

The prospectus figure, about 25% of revenue from two unnamed customers ^[10], means different things depending on who they are. If they are the cloud channels, the cuts sit outside the concentration, and demand through those channels is growing ^[4][8]. Should one be Meta buying seats directly, the halving at Meta ^[1] lands on a quarter of revenue. Microsoft booked as a single customer, internal and Azure together, would net the cut against the growth inside one line (a steady 25% share could hide a changing mix).

I think the exposed revenue is the direct seat sold to a customer that also builds models. As Anthropic's products overlap with what Meta and Microsoft sell or build, those buyers have reasons of cost and data control to keep code on their own systems ^[9]. Meta has already moved. Its two replacement tools count more than 36,000 users between them ^[15], though one engineer can appear in both counts.

A second Crypto Briefing write-up reads the pullback as a possible decline in Claude's competitive standing among AI models ^[18]. That reading and mine split at the cloud channels. A model losing on quality would lose resale demand as well, and customer spending through Azure and Bedrock is rising ^[4]. The view is wrong if that resale growth stalls while seat cuts spread to buyers that do not build their own models.