Insights & Announcements
A weekly Newsletter on technology applications in investment management with an AI / LLM and automation angle. We combine 100% human curation/selection with LLM standardisation, summarisation, and more deterministic search/collection, classification and workflow - powered by Kubro(TM). Curated news, announcements, and posts, primarily directly from sources (Arxiv papers, major AI/Tech/Data companies, investment firms). It's an evolving project.
Disclaimers: content not fully human-verified, with AI summaries below. AI/LLMs may hallucinate and provide inaccurate summaries. Select items only, not intended as a comprehensive view. For information purposes only. Please DM with feedback and requests.
Capital Group’s Jared Franz outlines four AI futures based on two variables: the pace of adoption and the supportiveness of policy and financial conditions. The most bullish outcome, an AI supercycle, features broad adoption, supportive policy, sustained infrastructure investment and stronger growth. A balanced path sees real but uneven progress across sectors. In a bubble-bursts scenario, investment outpaces returns as financing tightens and regulation increases. A return to the pre-ChatGPT world would leave AI useful but economically marginal. For investors, Franz argues the key is tracking productivity gains, integration, capital spending and policy signals rather than betting on one outcome.
🔗 Source: Summary based on capitalgroup.com View Source | Found on Apr 02, 2026
The Iran war caused oil prices to surge 58% in a month, heavily impacting energy-importing economies and introducing new supply pressures for the AI industry. High bandwidth memory was already sold out for the year, and advanced chip packaging faced 1-2 year backlogs. Iranian strikes damaged Qatari infrastructure, reducing global helium output by 14%, with repairs potentially taking up to five years. Major Asian chipmakers hold three to six months of helium inventory, but most liquid helium lasts only 35-48 days in storage. South Korea’s industrial power prices rose 39-55% year-to-date, while Taiwan imports 97.7% of its energy.
🔗 Source: Summary based on am.jpmorgan.com View Source | Found on Apr 02, 2026
Rising demand for critical microelectronics, driven by AI capital expenditures and defense spending, is creating upside risks to global goods inflation. Developed Asia—specifically Japan, South Korea, and Taiwan—dominates advanced chip and memory chip production, benefiting from improved terms-of-trade that support corporate earnings and debt sustainability. The BlackRock Tactical Opportunities Fund maintains short positions in global government bonds due to inflation risks from upstream commodity prices. DRAM chip prices have risen 17-fold over the past year, marking a sharp reversal from decades of price declines. Slower price declines since 2010 further increase inflationary pressures on consumer goods.
🔗 Source: Summary based on blackrock.com View Source | Found on Apr 01, 2026
According to a T. Rowe Price survey published on March 16, 2026, 71% of Americans with an advisor prefer their advisor use AI for at least one professional purpose, and among 182 advisors surveyed, 83% use AI at least monthly while only 23% consider themselves advanced users. Only 5% of advisors are unsure about AI’s opportunity for their practice. The article outlines five steps for integrating AI: start with a specific business need, learn prompt techniques, leverage specialists or team members, connect with peers for insights, and ensure outputs are accurate and unbiased.
🔗 Source: Summary based on troweprice.com View Source | Found on Mar 31, 2026
The US 4Q 2025 earnings season ended with S&P 500 year-on-year earnings growth exceeding 14%, surpassing the projected 8.3% and marking five consecutive quarters of double-digit growth. Nearly three-quarters of companies beat expectations for both EPS and revenues, with the "Magnificent 7" achieving 27.2% earnings growth compared to 9.8% for the other S&P firms. Information technology, communication services, and healthcare led sector performance, while hyperscaler capex is now projected to surge by 70% in 2026 to $737 billion from $160 billion in 2023, driven by major investments from Amazon ($200 billion), META ($125 billion), and Microsoft (58% capex increase).
🔗 Source: Summary based on am.gs.com View Source | Found on Apr 01, 2026
From 2018 to 2022, software accounted for 30% to 40% of private equity activity, driven by high retention rates and growth with low perceived disruption risk. John Zito, Apollo Asset Management Co-President, noted that AI disruption is accelerating and investors are reassessing software valuations. Apollo has been underweight in the sector for about 18 months due to concerns over capital overallocation and emerging risks. The marginal cost of producing software is decreasing, which may impact margins and valuations. Apollo focuses on investment grade senior lending with hard collateral and prioritizes structure, seniority, and downside protection for credit investors.
🔗 Source: Summary based on apollo.com View Source | Found on Apr 02, 2026
On President Trump’s first full day of his new term, Sam Altman of OpenAI, Larry Ellison of Oracle, and Masayoshi Son of SoftBank announced a $500 billion investment in Stargate, an OpenAI-led US megaproject. SoftBank committed $100 billion to what Son described as “the largest infrastructure project in history,” aiming to build a public-private network of datacentres and strategic partnerships to strengthen US AI leadership. Son emphasized ambitions beyond artificial general intelligence (AGI), envisioning a future “golden age” of artificial super-intelligence (ASI) capable of solving unprecedented human challenges.
🔗 Source: Summary based on bailliegifford.com View Source | Found on Mar 31, 2026
In the six months to end February 2026, the software and services sector had its worst relative performance to the S&P 500 in three decades, despite a positive S&P 500 return. The market’s new focus is on ‘HALO’ stocks—heavy asset, low obsolescence companies—seen as resistant to AI disruption. Axon Enterprises was added to US Growth portfolios in February after a share price decline of about 40% since October 2025. Ensign grew EPS by over 40% in the past 18 months and owns roughly 350 facilities with about a 4% market share. Knife River operates across 14 states as a top ten aggregates producer.
🔗 Source: Summary based on bailliegifford.com View Source | Found on Apr 04, 2026
LSEG announced a multi-year collaboration with Dell Technologies to optimize its on-premises infrastructure and build a new private cloud platform, enhancing resilience and performance for LSEG’s Data & Analytics and Markets platforms outside its public cloud environments. Dell will design and build the secure, high-performance private cloud using Dell servers, storage, and automation software as part of LSEG’s multi-cloud strategy. The partnership aims to improve flexibility, resilience, and operational control while meeting security and regulatory requirements. Irfan Hussain is LSEG’s Chief Information Officer; Doug Schmitt is President of Dell Technologies Services.
🔗 Source: Summary based on lseg.com View Source | Found on Mar 31, 2026
Executives highlighted increased scrutiny in private credit, emphasizing the need for selectivity, structure, and disciplined exposure amid market sensitivity and concerns over leverage, liquidity, and transparency. Europe’s private credit remains a smaller share of lending compared to the U.S., supporting more disciplined origination. Regulatory complexity and rising capital requirements have imposed a “silent tax” on Europe’s economy, limiting investment capacity. Proposed reforms under the Savings and Investments Union (SIU) aim to deepen capital markets by channeling household savings into productive investments through securitization reform and reduced market fragmentation. AI deployment is driving operational efficiencies with potential cost reductions of 20% or more.
🔗 Source: Summary based on morganstanley.com View Source | Found on Apr 02, 2026
OpenAI said it has closed a $122 billion funding round at an $852 billion post-money valuation to expand AI infrastructure, products, and compute capacity. The company framed the raise as support for a flywheel linking consumer adoption, enterprise deployment, developer usage, and access to compute. It highlighted rapid growth, including 900 million weekly ChatGPT users, over 50 million subscribers, and revenue of $2 billion per month, with enterprise contributing more than 40%. Backers include Amazon, NVIDIA, SoftBank, Microsoft, and other major investors. OpenAI also expanded its credit facility and outlined a multi-cloud, multi-chip strategy plus plans for a unified AI superapp.
🔗 Source: Summary based on openai.com View Source | Found on Apr 01, 2026
Gemma 4, released on April 2, 2026 by Google Cloud, is an open model family built from Gemini 3 research and licensed under Apache 2.0. It supports context windows up to 256K, native vision and audio processing, over 140 languages, and excels at complex logic and code generation. Gemma 4 can be deployed on Vertex AI with models ranging from E2B for edge tasks to a dense 31B model for enterprise orchestration. It is available on Cloud Run with NVIDIA RTX PRO 6000 GPUs, GKE with vLLM serving engine, TPUs via GKE/GCE/Vertex AI, and all Sovereign Cloud offerings.
🔗 Source: Summary based on cloud.google.com View Source | Found on Apr 03, 2026
IBM researchers have focused on developing AI systems that simulate physical reality, a concept now gaining traction among major industry figures. Yann LeCun left Meta in late 2023 to found AMI Labs in Paris, aiming to build “world models” that learn the structure and dynamics of reality rather than just text patterns. AMI Labs raised USD 1.03 billion at a USD 3.5 billion pre-money valuation, backed by investors including Bezos Expeditions and NVIDIA. World Labs, founded by Fei-Fei Li, raised USD 1 billion from AMD and others. Google DeepMind is also investing in world-model research programs.
🔗 Source: Summary based on ibm.com View Source | Found on Apr 02, 2026
From 2025 to 2029, Microsoft will invest $5.5 billion in cloud and AI infrastructure and ongoing operations in Singapore, as announced by Vice Chair and President Brad Smith on April 1, 2026. Every tertiary student in Singapore—over 200,000 individuals—will receive free access to Microsoft 365 Premium with Copilot for 12 months. Microsoft Elevate programs are being expanded to provide free AI training for educators and upskilling opportunities for nonprofit leaders. These initiatives align with Singapore’s National AI Strategy 2.0 and respond to a more than 70% year-on-year increase in demand for AI literacy skills.
🔗 Source: Summary based on news.microsoft.com View Source | Found on Apr 03, 2026
The article describes a research project by Thomas Jiralerspong and Trenton Bricken of the Anthropic Fellows Program, which developed a Dedicated Feature Crosscoder (DFC) tool for cross-architecture model diffing to identify unique behavioral features in AI models. The DFC successfully isolated features such as “Chinese Communist Party Alignment” in Qwen3-8B and DeepSeek-R1-0528-Qwen3-8B, “American Exceptionalism” in Meta’s Llama-3.1-8B-Instruct, and a “Copyright Refusal Mechanism” in OpenAI’s GPT-OSS-20B. These features were validated through steering experiments, with the CCP alignment feature rediscovered five out of five times and American Exceptionalism four out of five.
🔗 Source: Summary based on anthropic.com View Source | Found on Apr 04, 2026
Researcher—Microsoft 365 Copilot's deep research agent—has introduced two new multi-model capabilities, Critique and Council, to enhance research accuracy and depth. Critique separates generation from evaluation by using models from Frontier labs, including Anthropic and OpenAI; one model generates drafts while another reviews for source reliability, report completeness, and strict evidence grounding. Evaluated on the DRACO benchmark of 100 complex tasks across 10 domains using GPT-5.2 as judge, Critique showed significant improvements over single-model approaches: Breadth and Depth of Analysis (+3.33), Presentation Quality (+3.04), Factual Accuracy (+2.58), with p < 0.0001 in most domains.
🔗 Source: Summary based on techcommunity.microsoft.com View Source | Found on Mar 31, 2026
Anthropic’s Claude Code allows developers to determine its level of autonomy, ranging from manual approval of every action to automatic classification of safe versus risky actions. By default, Claude Code requests permission before modifying files or executing commands. Anthropic documents its agent safety approach, including trust design, access boundaries, and human control, in its research.
🔗 Source: Summary based on anthropic.com View Source | Found on Apr 01, 2026
This research by Jay Chen and Royce Lu, published April 3, 2026, examines Amazon Bedrock Agents’ multi-agent collaboration from a red-team perspective. The study demonstrates how attackers can exploit prompt injection vulnerabilities in multi-agent applications by progressing through operating mode detection, collaborator discovery, payload delivery, and exploitation. Attacks included extracting agent instructions and tool schemas and invoking tools with attacker-supplied inputs; for example, creating fraudulent tickets in the Energy-Efficiency Management System demo. No vulnerabilities were found in Bedrock itself; enabling Bedrock’s Guardrail feature blocked all attacks. The authors recommend layered defenses including guardrails, input validation, capability scoping, vulnerability scanning, and least privilege principles.
🔗 Source: Summary based on unit42.paloaltonetworks.com View Source | Found on Apr 04, 2026
Planning is identified as the most crucial step in an enterprise’s AI readiness journey, followed by prototyping, integration, and scaling. McKinsey’s 2025 Global AI Survey reports that 88% of organizations use AI in at least one business function, but only about 7% have fully scaled it across the enterprise. Gartner predicts over 40% of agentic AI projects will be canceled by 2027 due to failure to achieve intended business value. Companies like Mac Haik Restaurant Group, Zoom Communications, and Itaú Unibanco achieved success with Intel processor technology by aligning AI initiatives with defined business objectives and focusing on robust planning and integration.
🔗 Source: Summary based on community.intel.com View Source | Found on Apr 03, 2026
Lakebase, introduced as a third-generation database architecture by Databricks, is designed to support the rapid, agent-driven transformation of the software development lifecycle enabled by LLMs. In Lakebase production environments, each database project averages about 10 branches, with some reaching over 500 evolutionary iterations. For approximately half of agentic applications, database compute lifetimes are less than 10 seconds. AI agents now create roughly four times more databases than human users in Lakebase. Lakebase achieves near-zero cost branching using an O(1) metadata copy-on-write mechanism and stores data in open Postgres page formats directly on cloud object storage.
🔗 Source: Summary based on databricks.com View Source | Found on Mar 31, 2026
Uber’s Base design system supports thousands of engineers and hundreds of components across seven implementation stacks, including UIKit™, SwiftUI™, Android® XML, Android® Compose, Web React®, Go, and SDUI. To address the challenge of keeping component specifications accurate and current, Uber’s design systems team uses uSpec—an agentic system that connects an AI agent in Cursor® to Figma® via the open-source Figma Console MCP. This setup enables automated generation of up-to-date spec pages directly in Figma within minutes while ensuring all data remains local for security. The process eliminates manual spec writing delays and inconsistencies.
🔗 Source: Summary based on uber.com View Source | Found on Apr 01, 2026
Agentic AI introduces a fundamental shift in enterprise security, as AI agents interpret language, reason at runtime, and act autonomously across systems, expanding the attack surface beyond traditional cloud security controls. In SAP environments, this shift requires layered security measures including data masking, prompt templating, input/output filtering, orchestration policy enforcement, runtime isolation, human checkpoints for high-risk actions, and audit logging. SAP addresses these needs through capabilities such as SAP AI Core, Generative AI Hub with Harmonized API access to foundation models, orchestration controls for interaction governance, CAP connectivity patterns, and principal propagation to maintain user context and authorization boundaries during agent-driven workflows.
🔗 Source: Summary based on community.sap.com View Source | Found on Apr 04, 2026
The article "LLM-as-a-Judge for Time Series Explanations" by Preetham Sivalingam, Murari Mandal, Saurabh Deshpande, and Dhruv Kumar addresses the challenge of evaluating factual correctness in large language model (LLM) generated explanations for time series data. The authors introduce a reference-free evaluation method where LLMs assign ternary correctness labels based on pattern identification, numeric accuracy, and answer faithfulness. They construct a synthetic benchmark of 350 time series cases across seven query types with varying explanation correctness. Results show generation accuracies from 0.00–0.12 for some queries to 0.94–0.96 for others, while evaluation remains stable and reliable across tasks.
🔗 Source: Summary based on arxiv.org View Source | Found on Apr 03, 2026
The article "The Self Driving Portfolio: Agentic Architecture for Institutional Asset Management" by Andrew Ang, Nazym Azimbayev, and Andrey Kim describes an agentic AI system that transitions the investor’s role from analytical execution to oversight. The proposed pipeline involves approximately 50 specialized agents generating capital market assumptions, constructing portfolios using over 20 competing methods, and critiquing and voting on each other's outputs. A researcher agent introduces new portfolio construction methods, while a meta-agent evaluates past forecasts against realized returns and updates agent code and prompts. The process is governed by the Investment Policy Statement used by human portfolio managers.
🔗 Source: Summary based on arxiv.org View Source | Found on Apr 03, 2026
The article, "Reinforcement Learning for Speculative Trading under Exploratory Framework" by Yun Zhao, Alex S.L. Tse, and Harry Zheng, submitted on April 2, 2026, investigates a speculative trading problem using the exploratory reinforcement learning framework of Wang et al. [2020]. The problem is formulated as a sequential optimal stopping problem with general utility functions and price processes. A relaxed version models stopping times via Cox processes with bounded intensity controls. The agent's randomized control is defined by probability measures over jump intensities and regularized by Shannon's differential entropy, resulting in closed-form exploratory HJB equations and Gibbs distributions as optimal policy. Error estimates and convergence are established, and an RL algorithm is implemented in a pairs-trading application.
🔗 Source: Summary based on arxiv.org View Source | Found on Apr 03, 2026