〔Day Digest〕Bernstein Puts OpenAI-Linked 2030 Compute at $614B, Below $750B — Product Moats and Open-Source Robotics Advance as Data-Center Approvals Become the Next Test

0. Daily Arc

Across the China International Import Expo, the clearest operating theme was a migration from model demonstrations to deployable systems: Qualcomm, Samsung, AMD, Intel and Synopsys all framed AI around edge devices, industrial workflows, cloud-edge-device inference and robotics ^[1]. The counter-theme was scarcity of permission and capital: Bernstein analyst Mark Moerdler put potential OpenAI-linked infrastructure outlays through 2030 at $614B, below the $750B figure often cited, while former White House AI director David Sacks argued that data-center permitting could produce a 2028 construction and capital-expenditure stall ^[2][3].

Product evidence was tangible but uneven: Hitch Open was formally open-sourced, a demonstration involving the platform reached 142 hits and a world record, and Cambridge reported complete survival for vaccinated mice and ferrets in its AI-designed avian-flu vaccine experiment ^[4][5]. The packet’s source quality is mixed: named institutional and company reports support the operational details, while the Anthropic filing figures, U.S. market-cap estimate and AI-policy headlines rely on secondary or headline-only items ^{[4][6][7][8][2][5][9][10][11]}.

1. Infrastructure, Capital Expenditure and Permits

• **[NEW] OpenAI estimate:** Bernstein analyst Mark Moerdler estimated that infrastructure spending tied to OpenAI’s disclosed compute-cooperation agreements could total $614B cumulatively by 2030, versus the $750B figure widely cited in the market ^[3]. His calculation assumes every identified contract executes as scheduled, while construction delays, project cancellations and changes in hardware procurement could reduce or defer spending ^[3].
• **[ESCALATED] U.S. permitting:** Former White House AI director and PayPal founding-team member David Sacks said approving data-center projects has become a bipartisan “political poison,” predicting that no ambitious politician would approve new projects over the next four years and that construction and capital expenditure could stall in 2028 once the approved pipeline is absorbed ^[2]. This is an attributed forecast in one item, not an observed policy outcome ^[2].
• **[NEW] Post-election market claim:** A Wall Street Journal report said newly elected governors paused new data-center permits before new rules, and that Nvidia, Tesla, AMD and Micron then recorded their largest one-day market-cap loss, near $4T ^[11]. The same report said All-In Podcast co-host Chamath Palihapitiya disclosed 90-day mutual exit clauses in SpaceX and Anthropic compute-supply contracts, allowing SpaceX in theory to terminate and reprice at 2-3x spot if supply tightens ^[11]. Because the $4T figure appears only in this secondary report within the packet, it should be quoted as a reported estimate rather than an audited market statistic ^[11].
• **[NEW] Brazil expansion:** ByteDance is discussing a phase-two expansion of its data-center layout in Ceará, northeastern Brazil, and is looking for a developer for the Pecém port campus, potentially at a scale above the first phase ^[10]. Omnia, owned by Patria Investments, is building that first phase with about $10B of investment, roughly 200 MW of IT capacity, 300 MW of total power demand, 20 data halls and planned operations from 2027; Casa dos Ventos has signed a 20-year wind-power purchase agreement ^[10].
• **[ONGOING] Neocloud asymmetry:** Serenity argued that tighter financing and potential permit limits could favor established neocloud providers with supplier relationships, contracted capacity, permits and well-funded partners ^[12]. Serenity cited IREN’s report that a five-year agreement’s prepayment covers 90% of GPU capital spending, and said older H100 GPUs were being re-contracted at twice the prior price; both claims remain secondary in the packet ^[12].

2. Product Moat and Corporate Read-Through

• **[NEW] Product over model:** Tencent chief AI scientist Yao Shunyu said at the 2026 Future Science Prize dialogue that a large model’s lead can change in 3-6 months and that the harder-to-replace layer is the product, especially user context, proprietary data and trust ^[13]. Chinese Academy of Engineering foreign member Zhang Yaqin compared frontier models to operating systems and joked that Anthropic might not be worth $1T-$2T if models ultimately commoditize ^[13].
• **[NEW] AppLovin:** A Cailian Press company-research headline said AppLovin shares rose 70% over four days as nearly 100 agents entered real workflows and the platform had served more than 10,000 enterprises ^[14]. The supplied item is truncated, so the agent and customer figures should be treated as headline-level rather than fully verified ^[14].
• **[NEW] Samsung balance sheet:** Industry data cited by 36Kr and Sina Finance linked the AI-chip supercycle to Samsung Electronics’ Q3 cash and cash equivalents reaching KRW 214.8T, up KRW 24.8T from KRW 190T in three months and almost twice KRW 108.5T at the end of Q3 2025 ^[6]. The same report put total assets at KRW 791.8T, up KRW 32.4T quarter on quarter, and said observers estimated the cash could cover the upper end of Samsung’s 2026 shareholder-return commitment of KRW 90T-KRW 110T ^[6].
• **[NEW] Token economics:** China Mobile chairman Chen Zhongyue said the operator is advancing communications, computing and intelligent networks together, released the MoMA model service platform, aggregated more than 300 mainstream large models and materially reduced overall token-use costs ^[9]. Its 2027 Pan-Alliance strategy centers on 5G terminal upgrades, 6G terminal guidance, AI-era customized phones and a shift from byte-oriented to token-oriented operations ^[15].
• **[NEW] Consumer pricing:** TweakTown’s Google News headline said U.S. NVIDIA RTX 5070 Ti prices rose $160 in eight days to 79% above MSRP ^[16]. Because the packet contains only the headline, the pricing move is best treated as a single-source retail signal ^[16].

3. Robotics, Biotech and Financial Services

• **[NEW] Physical AI:** Global physical-intelligence platform Hitch Open was formally launched and open-sourced at the China International Import Expo, with material describing it as a data foundation for embodied and world-model training ^[17][5]. A team from Shanghai Jiao Tong University and Shanghai Innovation Institute led by Professor Gao Yue, working with Hitch Open, completed a 142-hit autonomous-decision robot table-tennis challenge that the report called a world record ^[5].
• **[NEW] Robotics output:** Coolwa Technology displayed the D0 heavy-load quadruped and X1 city-service robot, both using the same edge-side world model, and plans to release the D2 community-delivery robot in 2027 ^[18]. Qiyuan Robotics rolled off its 10,000th personal robot on November 8, doubling cumulative output from 5,000 units on October 25 in 14 days ^[19].
• **[NEW] Vaccine design:** Cambridge University said an AI-designed broad-spectrum avian-flu vaccine protected mice and ferrets against several genetically distinct variants: every vaccinated animal survived, compared with only about half of controls receiving an existing vaccine ^[4]. The research team plans human clinical trials, and the results were published in Cell Reports Medicine ^[4].
• **[NEW] Brain-computer interface:** Shanghai’s first NEO brain-computer-interface implant recipient marked two years since surgery at Huashan Hospital, with improvements including picking up a training ball by day nine, controlling a pneumatic glove after one month and writing “thank you” by the one-year mark ^[20]. The patient also said he had gone two years without a urinary catheter and that spontaneous bowel movements were increasing ^[20].
• **[NEW] Drug discovery:** MegaStream CEO Lei Ming argued that AI and biotechnology should be designed as one integrated system rather than used only as an outsourced discovery accelerator ^[21]. He said AI’s commercial reach often stops before preclinical candidate selection, and that even screening tens of thousands of targets can leave few candidates because of off-target effects ^[21].
• **[NEW] Cross-border finance:** 2000 Nobel economics laureate James J. Heckman said AI can improve cross-border transaction efficiency and lower costs, but it also opens room for forgery and fraud, requiring more transparent and synchronized regulation rather than blanket deployment ^[22].

4. Falsifiers and Source Control

• **[NEW] Anthropic filing figures:** Two Google News items attributed a $16.3B revenue figure to an Anthropic IPO report and a narrowing loss to $1.3B from $5.2B to an S-1 report ^[7][8]. The packet does not include the filing itself, so those figures should remain explicitly unverified rather than become base-case financial inputs ^[7][8].
• **[ONGOING] Infrastructure test:** The $614B compute estimate should be reduced if identified contracts are canceled, delayed or altered through hardware-procurement adjustments, exactly the risks Moerdler flags ^[3]. The 2028 U.S. construction-stall thesis weakens if states keep issuing data-center permits rather than freezing approvals ^[2][11].
• **[ONGOING] Product test:** Yao Shunyu’s product-layer thesis would weaken if superior models repeatedly retain users without proprietary context, data or trust ^[13].
• **[ONGOING] Clinical test:** Cambridge’s vaccine case remains an animal result until human trials begin ^[4].
• **[NEW] Next checks:** The observable tests are contract and permit decisions, Hitch Open adoption, Coolwa’s 2027 D2 plan, China Mobile’s token-cost performance and initiation of the vaccine’s human trial ^{[4][18][2][3][9]}.
• **[ONGOING] Bottom line:** The packet supports a two-speed AI cycle: deployment and product differentiation are advancing, while the buildout of compute faces a permit-and-capital bottleneck ^{[4][13][2][3][5]}.