{"collectionById":{"76d23fbd-a202-403b-8313-b3bc36d46679":{"id":"76d23fbd-a202-403b-8313-b3bc36d46679","name":"Research Post","fieldSchemas":[{"id":"1d8c181d-1ccd-446e-87b0-482ccb3ee240","name":"Title","type":"plain_text","role":"primary"},{"id":"571c37b5-f5fc-4742-a5de-b4da6f3415b3","name":"Slug","type":"slug","role":"slug"},{"id":"3fcb25bf-f8b6-47c4-84d6-226369594160","name":"Content (HTML)","type":"rich_text"}],"itemById":{"cbc9abca-deba-448f-8482-33d23533f329":{"id":"cbc9abca-deba-448f-8482-33d23533f329","index":"\"NNNNNNNNNNO","collectionId":"76d23fbd-a202-403b-8313-b3bc36d46679","fields":[{"id":"25aad452-f917-4c85-9396-ac2f27778d39","value":"{\"root\":{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"What Muse and Grok Bot could mean for server CPUs, memory and data-center capacity.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"A cloud computer for every personal-agent user turns adoption into demand for server CPUs, memory and power. We estimate that demand for Muse, then apply the same framework to Grok Bot’s larger cloud-computer configuration.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At 100 million continuously running cloud computers, we estimate 1.56 million server CPUs and 1.63 GW of facility power for Muse. The same number of cloud computers on Grok Bot’s configuration would require 6.25 million CPUs and 5.97 GW.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The calculations below use dedicated physical cores. They are resource estimates under stated assumptions, not observations of either company’s deployed fleet.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"linebreak\",\"version\":1},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Start with the Assumptions\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We model one continuously running cloud computer per user, with one physical CPU core per vCPU and no CPU overcommit. For memory sizing, we provision the configured capacity rather than the observed idle usage.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Muse configuration used in this analysis is 2 vCPUs and 8 GB of memory per user-level VM. For Grok Bot, we treat the supplied eight-core configuration as 8 vCPUs, with 15 GiB of configured memory and observed idle usage of about 1.6 GiB. The Grok Bot figures are inputs supplied for this research, not a verified official specification.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Several Grok Bot bots can share the same user-level cloud computer. The unit we count is that computer. Neither the number of bots nor the number of daily tasks can be substituted for the number of VMs.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"From Today’s Footprint To 100M Users\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"In our September 22 estimates, Muse’s iOS downloads were below 1 million. Including WhatsApp, web, Android and Mac, we put total accounts at 1–2.5 million. Most accounts would not run continuously, so we use close to 1 million live VMs as our starting point.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We consider three growth paths:\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If Muse remains mainly a US product, it could reach 30 million users in a year.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If WhatsApp distribution works, it could reach 25 million users in six months and 100 million in a year.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"If Meta pushes Muse through both WhatsApp and Instagram, it could reach 250 million users.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"listitem\",\"version\":1,\"value\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"list\",\"version\":1,\"listType\":\"bullet\",\"start\":1,\"tag\":\"ul\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We use 100 million as the middle scenario. To estimate the hardware footprint, we assume one continuously running cloud computer for each user at that scale, then apply the same user counts to Grok Bot for comparison.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"What The Configurations Require\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Our reference server has two 128-core CPUs, or 256 physical cores. It holds 128 Muse VMs at 2 vCPUs each, compared with 32 Grok Bot cloud computers at 8 vCPUs each.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At 100 million users, the Muse configuration requires about 781,250 servers and 1.56 million CPU packages. Grok Bot requires 3.125 million servers and 6.25 million CPU packages.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Put another way, 25 million Grok Bot cloud computers would require the same CPU capacity as 100 million Muse cloud computers under this model.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"08872b6c96e21d3158c9c890244e9e4a89e6a4fa\",\"src\":\"https://s3-alpha-sig.figma.com/img/0887/2b6c/96e21d3158c9c890244e9e4a89e6a4fa?Expires=1791158400\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=fnFyOxC~3VP9YnzB16se-Yl4xR45dm1zn02jZM8N1B87QCbJhKKa5CkgGBlOLAtmhREVDjdb2D47zH0cM-iMx9iLCvEy8GWeAbTthDXc8f7PmSExdpZHh~OS0VvcEO9T3AhSsW6EAu44OYsvA4ITAd3xuUqs6px9I6lQuWgFMgMz41j8jGgG7OS9eqTmc717e9I9YNQtde-DJ9NwD1vFrGMS6ypEcQmJE2z53MlqARYvk-ZLjy3aew4VV4cw8EEfZ~T6LN-Mxa9KwaapMY-X-XZYeGguUNLMh~pijbRnamKzEBHnPxISxDoGCiIEMJYonCHWj8lbj4oy~q9XfDM5Fw__\",\"altText\":\"\",\"originalImageWidth\":1280,\"originalImageHeight\":960,\"isFillWidth\":false}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Figure 1. At the same user scale, Grok Bot’s configuration requires four times as many server CPUs as Muse’s. At 100M continuously running cloud computers, the two configurations would require 15.6% and 62.5% of our estimated AMD annual shipment benchmark, respectively.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"From Servers To Gigawatts\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For Muse, we allow 1 kW for two 500 W EPYC 9755 CPUs and about 0.25 kW for roughly sixteen 64 GB DDR5 RDIMMs. Including local drives, the motherboard, networking and fans, the model budgets 1.60 kW of IT power per server. PUE 1.3 is applied separately to obtain facility power.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For Grok Bot, 32 cloud computers at 15 GiB each require 480 GiB of configured guest memory. We estimate roughly 512 GB of installed memory, or eight 64 GB DIMMs at around 15 W each. Replacing the Muse memory allowance with 0.12 kW, while holding CPU power at 1 kW and other components at 0.35 kW, gives 1.47 kW IT and about 1.91 kW facility power per server.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That allocates about 46 W IT and 59.7 W facility power to each Grok Bot cloud computer. The observed 1.6 GiB idle usage is not a peak measurement and is not used to reduce configured memory or DIMM counts.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"With PUE at 1.3, the modeled facility load is 1.63 GW for Muse and 5.97 GW for Grok Bot. Grok Bot needs four times as many servers, but slightly less power per server, so its total is about 3.7 times higher.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"These are provisioning estimates. They are not measured fleet power, and they exclude the separate compute required to serve the agents' AI models.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Implications For CPU And Memory Supply\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We estimate annual AMD server-chip shipments at around 10 million, against roughly 39 million across the industry. Against that AMD benchmark, the 100 million-user scenarios imply CPU requirements equivalent to 15.6% of annual shipments for Muse and 62.5% for Grok Bot. These are shipment estimates; the comparison shows scale rather than purchases that must occur within one year.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"There is already market talk that AMD server chips are sold out, although we have not verified that claim. Higher core counts in the next generation would let each CPU host more VMs and reduce the number of CPU packages required.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Memory demand is substantial too. The Muse scenarios require 200 PB at 25 million users, 800 PB at 100 million, 2 EB at 250 million and 8 EB at 1 billion. We estimate that 100 million Muse users would represent about 1% of annual DRAM bits, rising to 10% at 1 billion users.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Using Grok Bot’s supplied 15 GiB configuration, memory demand is approximately 357.6 PiB at 25 million cloud computers, 1.40 EiB at 100 million, 3.49 EiB at 250 million and 13.97 EiB at 1 billion. Muse and Grok Bot retain the different memory units used in the research.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"As a stress test, the dedicated-core model at 1 billion users reaches roughly 15.6 million CPU packages and 16.3 GW of facility power for Muse, versus 62.5 million CPUs and 59.7 GW for Grok Bot.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The Investment And Who Pays For It\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"The 100 million-user Muse scenario also provides a way to examine the scale of the investment.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Reuters reported\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":null,\"title\":null,\"url\":\"https://www.investing.com/news/stock-market-news/meta-to-put-ai-chip-into-production-in-september-as-it-looks-to-double-computing-capacity-memo-shows-4787372?ampMode=1\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\", citing an internal memo, that Meta planned 7 GW of computing infrastructure in 2026 and 14 GW in 2027. If the 14 GW figure refers to IT power, our 100 million-user Muse estimate of 1.25 GW would equal 8.9% of it. If it refers to facility power, our 1.63 GW estimate would equal 11.6%. The report does not specify the boundary, so the comparison is conditional. Meta’s \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"earnings call\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":null,\"title\":null,\"url\":\"https://s21.q4cdn.com/399680738/files/doc_financials/2026/q2/META-Q2-2026-Earnings-Call-Transcript.pdf\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\" also makes clear that capacity serves model training, the core business and new products.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"We use a third-party hardware estimate of $30 billion to $50 billion for 100 million continuously running Muse VMs. This is an input to our analysis, not a disclosed Meta budget.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For scale, Meta reported $31.1 billion of capex in Q2 2026 and guided to $130 billion to $145 billion for the full year, including finance lease principal payments. Quarterly revenue was $60.8 billion, of which 97.6% came from advertising. \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Meta Q2 results\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":null,\"title\":null,\"url\":\"https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/default.aspx\"}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Ten million subscribers paying $20 a month would generate $2.4 billion a year before costs. To compare that annual revenue with an annual hardware expense, we use a 5.5-year straight-line depreciation scenario with no residual value. Meta \",\"type\":\"text\",\"version\":1},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"extended the estimated useful life of certain servers and network assets to 5.5 years from 2025\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"link\",\"version\":1,\"rel\":null,\"target\":null,\"title\":null,\"url\":\"https://www.sec.gov/Archives/edgar/data/1326801/000132680125000017/meta-20241231.htm\"},{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"; we apply that life to the full $30–50 billion hardware estimate as a simplifying assumption.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"That implies $5.5 billion to $9.1 billion in annual depreciation. The $2.4 billion subscription scenario covers only 26–44% of it. At $20 per month, about 23–38% of the 100 million users would need to pay just to cover hardware depreciation, before electricity, operations or model inference.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"\",\"src\":\"https://shixiangtech.feishu.cn/space/api/box/stream/download/asynccode/?code=NWMzYzNlN2U2MGRlOWM1Zjk2MzI1OGQ3OWMwNWQyNmRfRDVjR2hCdEc3WWhGQTlCeGRVU29CajhjYmVDdDFIdnNfVG9rZW46Tkh5UGJvUnVFb3ZibDZ4bWdsWmNyR1FUbnRjXzE3OTAyNjYwNTc6MTc5MDI2OTY1N19WNA\u0026add_watermark=true\u0026scene_type=CCM\",\"altText\":\"\",\"originalImageWidth\":0,\"originalImageHeight\":0,\"isFillWidth\":false}],\"direction\":null,\"format\":\"center\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"type\":\"image\",\"version\":1,\"hash\":\"1383a89473af8b80230009176e55a93bc7ee78ba\",\"src\":\"https://s3-alpha-sig.figma.com/img/1383/a894/73af8b80230009176e55a93bc7ee78ba?Expires=1791158400\u0026Key-Pair-Id=APKAQ4GOSFWCW27IBOMQ\u0026Signature=l33GvtG8zFqdj8bcUZgnS5if1l8kadAS1jtFUveWEnEdJQoKBJUVN4NeSM3taEla7WeccGpC7oCO~~Gdh~eYKWMegTTR5S~8ITuMNCW~5pDvg6RtiOXBiO-bdSxi7eWRSAStJEjZwSlovbboL34P7OXhJVBWKd-SB5WRpzRe~VZDSY3U5~AhMfiA-aL3Vw~O4GH4YJGZJjl5GmPN2tQ2tRf46~FqDQ3VI-yM5mjp~JcduwjlvzyrtJTLf~qNagelzBLLiwdmX37ubGQ~K6DI0WGdDf6dEiBtEAgNWJ9fX21WSrUakaupIiCcOvUgr6yfu14GiYHWsUHo~piwxQrndw__\",\"altText\":\"\",\"originalImageWidth\":1280,\"originalImageHeight\":960,\"isFillWidth\":false}],\"direction\":null,\"format\":\"center\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"center\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":2,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Figure 2. At 10% paid conversion, annual subscription revenue covers only 26–44% of modeled hardware depreciation. Covering depreciation alone would require about 23–38% of users to pay $20 per month.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":2,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"At 10 million subscribers paying $20 a month, subscription revenue would not cover the modeled hardware depreciation of this always-on deployment, before operating and inference costs.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Meta’s advertising revenue and capital-spending capacity remain central to the funding question. Spreading the hardware cost over its useful life also leaves the upfront investment to be financed.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"DAU Is A Better Starting Point Than Registered Users\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"heading\",\"version\":1,\"tag\":\"h2\"},{\"children\":[],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"Cost gives platforms an incentive to suspend inactive VMs and reclaim their host CPU resources. Estimating continuously running VMs from registered accounts can therefore overstate actual CPU demand.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"},{\"children\":[{\"detail\":0,\"format\":0,\"mode\":\"normal\",\"style\":\"\",\"text\":\"For both Muse and Grok Bot, we think DAU is a better proxy for active demand than registered users. But DAU is not the number of VMs running simultaneously. Physical CPU demand still depends on concurrency, active duration and instance-reclamation policies.\",\"type\":\"text\",\"version\":1}],\"direction\":null,\"format\":\"left\",\"indent\":0,\"type\":\"paragraph\",\"version\":1,\"textFormat\":0,\"textStyle\":\"\"}],\"direction\":null,\"format\":\"\",\"indent\":0,\"type\":\"root\",\"version\":1,\"textFormat\":2}}","itemId":"cbc9abca-deba-448f-8482-33d23533f329","fieldSchemaId":"3fcb25bf-f8b6-47c4-84d6-226369594160"},{"id":"5f37684a-71ea-4e5c-9ad0-7ff62e64dc4e","value":"What It Takes to Give Every Personal Agent a Computer","itemId":"cbc9abca-deba-448f-8482-33d23533f329","fieldSchemaId":"1d8c181d-1ccd-446e-87b0-482ccb3ee240"}]}}}},"slugByItemId":{"61387d11-a584-4b82-b223-7fd449542135":"the-ai-bubble-reckoning-1999-all","f650e52a-191a-4241-a8b4-12705cb62385":"ai-for-life-science-landscape","3db0c5d9-c44c-4950-a175-45c033132ebe":"ai-coding-landscape-how-agents-disrupt","c4aeef83-34d4-411e-b40c-b432dd225928":"continual-learning-next-paradigm","e632f659-8b25-4ec8-8c28-2032c89b4fab":"our-investment-in-phylo-building-the-intelligence-layer-for-biology","809bd036-e216-4bdb-ae70-dbd1d65b4804":"fable-5-expensive-intelligence-needs-expensive-work","a54b6460-7b13-40d1-b52f-59d326989a10":"openclaw-is-the-signal-our-thesis-on-long-horizon-agents","d495e0e3-e929-4163-9d3e-0fec19ef9c2c":"beyond-deepseek-what-chinas-model","5ea3ff6e-b7d9-440c-aa76-6f6ce74c432a":"our-investment-in-convex-the-backend-where-agents-build","deb68f7b-2d3f-40b3-bbe5-ca239d245f27":"our-investment-in-generalist-a-brain-for-many-bodies","7b886fb4-c171-49aa-877e-475cf8b97f18":"how-openai-could-turn-the-tables","598cfcc0-365f-40c6-9f63-6631c6f5a6ff":"will-chinese-ai-leap-ahead-or-follow","91e6e6f8-7396-4490-b421-e7e2ff664063":"rl-scaling-from-research-trick-to","d285731a-bce6-45c3-9dac-26edd7829535":"pulse-how-openai-starts-outrunning","b5bb1a96-7690-427f-ae92-b2b434675642":"claude-tag-is-underrated","1ac8ae85-9538-48b0-bf23-537de16e760c":"data-is-a-great-place-to-start-an-ai-company-and-a-dangerous-place-to-stop","35fee5fc-be53-402a-a10d-4144286ecc7b":"our-investment-in-agentrys-a-self-improving-layer-for-chip-design","afae6882-f966-42cb-b91c-354a1633a2ec":"building-robot-bodies-not-a-supply-chain-a-learning-loop","c23fe0ec-cf33-40e2-b090-b73d06099e9a":"generalist-and-the-270000-hour-advantage","80ff4031-030f-4caf-b7a9-1825fd8cb70f":"agi-2026-are-we-the-final-white-collar","940cd5e4-2e54-4b7a-941f-c0ea4e37df50":"beyond-the-cloud-are-llms-the-new","0c9ed8bc-cb92-4008-9f50-15867c8349d3":"our-thoughts-on-llm-part-one","fb158125-2004-4c70-972d-88d6846176ab":"decode-the-buzzword-why-harness-engineering-matters-now","cd412711-7095-43ca-8389-8011dab2f900":"inventing-anthropic-two-ingredients-behind-the-ai-winner","cbc9abca-deba-448f-8482-33d23533f329":"what-it-takes-to-give-every-personal-agent-a-computer"}}