〔Day Digest〕SpaceX AI Capacity Heads to 2.7GW as a More Than £20bn UK Data-Center Project Is Announced and Anthropic’s $1.72 Compute Yield Turns Down — Buildout Meets Unit-Economics Stress

0. Daily Arc

• **[NEW] Capacity-versus-monetization split:** SpaceX is expected to reach roughly 2.7GW of AI compute by year-end, and the Drax-Brookfield-Radiant project carries a more than £20bn investment plan, yet Anthropic’s revenue return per $1 of compute fell 17.3% q/q to $1.72 from $2.08 ^[1][2][3]
• **[ESCALATED] Service reliability:** DeepSeek repaired a partial web/API interruption, but its latest update still showed degraded performance in the Chat Service, DeepSeek V4 Pro API and DeepSeek V4.1 Flash API ^[4][5]
• **Net read:** Infrastructure commitments accelerated while monetization, compute efficiency and uptime became harder tests ^[1][4][2][3]

1. Infrastructure and Capital Buildout

• **[NEW] SpaceX capacity:** Deutsche Bank expects SpaceX’s AI compute capacity to reach about 2.7GW by year-end, inside the prior 2.5-3.0GW target, as new Nvidia GB300 deployments support cloud-compute leasing and Grok inference; the bank sees room for more large contracts or expanded orders and maintains a Buy rating with a $250 target ^[1]
• **[NEW] Cluster scale:** Each cluster of 220,000 GB300 chips corresponds to about 440MW of compute ^[1]
• **[NEW] UK project:** Drax, Brookfield and Radiant agreed to build a phased, gigawatt-scale data center at Drax’s power station in North Yorkshire, with financing from the Brookfield-Qai AI infrastructure joint venture ^[3]
• **[NEW] Funding and timetable:** Total investment is expected to exceed £20bn, including about £2.5bn for the first phase; that phase targets up to 100MW in Q1 2028, and capacity is planned to rise to 500MW during 2028-2031 ^[3]

2. Unit Economics and Service Reliability

• **[ESCALATED] Growth and profit:** Anthropic’s Q3 revenue grew 47% q/q, versus 144% in Q2, while operating performance swung from a $974m profit in Q2 to a $567m loss in Q3 ^[2][6]
• **[NEW] Compute yield:** Revenue return per $1 of compute fell to $1.72 in Q3 from $2.08 in Q2, a 17.3% decline and the first downward turn since the metric began being tracked ^[2]
• **[NEW] Cost and mix explanation:** The cost of emergency spot compute was 110%-160% above long-term locked capacity; clearing high-compute, low-value distilled traffic, which represented 18% of compute calls but only 6% of revenue, reduced utilization; and average API prices for leading enterprise clients fell 17% q/q ^[2]
• **[ONGOING] DeepSeek:** A partial web/API interruption began at 16:58, affecting the Chat Service, DeepSeek V4 Pro API and DeepSeek V4.1 Flash API; DeepSeek said access had been repaired, but degraded performance remained under monitoring ^[4]
• **[NEW] Security release:** CrowdStrike launched on AWS a Falcon sensor for Amazon ECS managed instances as part of a native AI security release ^[7][8]
• **Falsification:** The test for the economics warning is a subsequent Anthropic disclosure showing $1.72 rebound toward $2.08, while the report cautions that one quarter cannot establish a durable inflection ^[2]
• **Operational test:** DeepSeek’s operational test is normalized performance across all three affected services, rather than only restored web/API access ^[4]

3. China Deployment and Products

• **[NEW] Investment scale:** Industry estimates cited by Yicai put 2026 embodied-AI investment in China’s power sector above RMB 10bn ^[9]
• **[NEW] Grid robotics:** China Southern Power Grid’s competition, described as the power industry’s first embodied-intelligence application contest, tested substation inspection, emergency outage operation and verification of de-energized status before electricity-meter replacement, with six teams advancing to the final ^[9]
• **[NEW] Xiaomi deployment:** Xiaomi Group received an Outstanding Contribution Award, its embodied-intelligence foundation-model technology received a Leading Technology Award, and the company said its humanoid robot had completed preliminary validation of production-line work capability ^[10]
• **[NEW] NetEase security:** NetEase Zhichi released an AI-native security suite based on the Yidun bidirectional architecture, covering AI content and data security, AI interaction and identity security, and AI model and agent security; the three releases were a CMA review agent, DeepGuard face anti-spoofing and AI application and agent security ^[11]
• **[NEW] Government and enterprise agents:** Sugon Cloud launched OneBox, presented as the first trusted desktop-level AI agent for government and enterprise users, integrating model services, internal knowledge and business tools with private deployment, permission management and trusted execution ^[12]

4. Policy and Corporate Signals

• **[NEW] Global institution:** CCTV News, relayed by Yicai, said China is advancing the operationalization of the World Artificial Intelligence Cooperation Organization ^[13]
• **[NEW] OPC pilot:** Zhejiang’s market regulator and four other departments issued a notice supporting pilot development of AI OPC in Hangzhou, Ningbo, Wenzhou and Jiaxing, with support spanning space, compute, data, talent, finance and services and applications in manufacturing, healthcare and digital creativity ^[14]
• **[NEW] Governance framing:** Foreign Ministry spokesperson Mao Ning said internet technology should enhance people’s capabilities, expand their choices and increase their welfare, rather than turn people into appendages of technology or widen development disparities ^[15]
• **[NEW] OpenAI policy:** Semafor reported that OpenAI is trying to reshape the data-center concept and set a policy tone after the U.S. midterms, but some circles including lawmakers responded cold; reported tactics include renaming data centers as compute campuses, consulting communities on development plans, comparing AI regulation to vehicle speed limits and promoting global AI standards ^[16]
• **[NEW] Cybersecurity hire:** A Wall Street Journal headline reported that OpenAI hired former Army Secretary Dan Driscoll for a cybersecurity role, but the packet contains no further detail ^[17]
• **[NEW] Talent signal:** AI scientist Yao Shunyu denied rumors that he was preparing a recursive self-improvement company and said he remains a Google DeepMind employee ^[18]
• **[NEW] IPO denial:** Baidu denied reports that the Kunlunxin IPO was delayed, said the chip company is in the confidential-filing stage, and said regulatory requirements prevent disclosure of process progress ^[19][20]

5. Supply Chain and Falsification

• **[NEW] Headline-only Nvidia signals:** A Techzine Global headline said Nvidia imposes strict requirements on cloud providers, while a Financial Times headline carried by Yahoo Finance said Nvidia tells data-center partners to limit China sales; the supplied material provides no details to test the scope, timing or enforcement of either claim ^[21][22]
• **[NEW] Single-report HBM mechanism:** The Information reported that U.S. restrictions introduced at the end of 2024 left a packaging loophole: standalone advanced HBM was restricted, while HBM packaged with a processor could be judged by processor compute, allowing Samsung or SK Hynix memory to travel with low-performance processors and be removed in mainland workshops after transit through Hong Kong ^[23]
• **[NEW] Reported HBM economics:** An HBM3E stack rose from hundreds of dollars to about $4,000 in one year; at 6-8 stacks per AI chip, memory cost reaches $24,000-$32,000, and some memory ultimately reached Huawei customers ^[23]
• **[NEW] Huawei supply pressure:** The same report said Huawei’s next-generation Ascend 960DT was planned for Q1 2027 amid insufficient domestic HBM supply ^[23]
• **[NEW] Domestic chip capital:** Moore Threads’ wholly owned Beijing Mobi Generating Technology, founded in December 2023, raised registered capital from RMB265m to RMB345m, a roughly 30% increase ^[24]
• **Decision points:** The clearest tests are whether SpaceX reaches its year-end 2.7GW estimate with added cloud contracts, whether Drax hits its Q1 2028 100MW target, whether Anthropic’s next disclosed compute yield rebounds from $1.72 and whether DeepSeek restores normal performance across all affected services ^[1][4][2][3]