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). See disclaimers at the bottom. Please DM with feedback and requests.
Investec is deploying Microsoft Copilot to all 8,000 employees globally, the first publicly announced full-workforce rollout by a South African organisation. Covering South Africa, the UK and other markets, the move advances Investec’s “higher tech, higher touch” AI strategy. The bank already uses over 800 AI agents, saving more than 350,000 staff hours annually by automating repetitive, analytical and administrative work. Investec says AI will amplify, not replace, human judgement, freeing employees to focus on clients, trust and complex problem-solving. The rollout includes staff training, governance, agentic AI development and Microsoft as its enterprise AI partner.
🔗 Source: Summary based on View Source from investec.com | Found on Jun 24, 2026
Amundi argues that AI disruption is creating software’s toughest downturn in over a decade, with valuations, dealmaking and exits under pressure. Yet it sees parallels with the shift from boxed software to SaaS: a shakeout will produce both losers and winners. Low-barrier B2C apps are most exposed, while B2B incumbents benefit from proprietary customer data, trust, embedded workflows, switching costs and the need for precision. For investors, old sector assumptions and passive exposure are insufficient. Amundi says manager selection, diversification across managers, vintages and geographies, operational value creation and access to data-rich co-investments are critical to navigating AI-driven market change.
🔗 Source: Summary based on View Source from amundi.com | Found on Jun 25, 2026
Nuveen’s midyear 2026 outlook argues that the AI boom, especially hyperscaler data-center spending, has become the dominant market force, creating concentration risk across equities, credit, real estate and infrastructure globally. It still sees AI opportunities, but stresses differentiation: winners include proprietary-data businesses, infrastructure, utilities, storage, transmission, asset-backed securities and sustainable data-center platforms, while overvalued or poorly monetized AI plays face pressure. Nuveen favors broad diversification without abandoning U.S. assets, selective alternative credit and private markets, municipals, and recovering private real estate. Best ideas include dividend growers, preferreds, senior loans, health-care municipals, senior housing, medical offices, data centers and clean-energy infrastructure
🔗 Source: Summary based on View Source from nuveen.com | Found on Jun 25, 2026
US consumption has typically been about 66–70% of GDP since 2000, yet personal consumption grew only 1.6% annualised in Q1 2026 while real wages rose just 0.5% in the year to end‑May 2026, leaving “no discernible increase in spending power for most Americans for half a decade.” Technology investment is surging: Morgan Stanley estimates USD 800 billion capex by hyperscalers this year, Q1 investment rose 43% in tech equipment, 23% in software and 22% in data centres, and the WSJ estimated the “AI economy” grew 31% in Q1 2026. The S&P 500 forward P/E was 21.5x on 11 June versus an 18.2x average.
🔗 Source: Summary based on View Source from gam.com | Found on Jun 19, 2026
FTSE Russell will reconstitute its indices on June 26; it ranks companies by market capitalization and categorizes them into growth and/or value buckets. This year’s reconstitution will be larger for value versus growth and for large versus small. Russell is the benchmark for U.S. value, growth, large-cap, and small-cap style investing; its rebalances have been annual and are moving to semi-annual this year for large/small (not for value/growth). An 18-month AI runup—from hyperscalers to data centers and semiconductors—has raised future earnings expectations, and the spread in forward EPS growth for Russell 3000 Value pre- vs. post-reconstitution is the largest this decade.
🔗 Source: Summary based on View Source from pgim.com | Found on Jun 25, 2026
Author Sébastien Page, CFA, published May 29, 2026: the Asset Allocation Committee entered 2026 overweight international, U.S. small- and mid-cap (U.S. SMID-cap), emerging market (EM) equities, and real asset equities, plus a short-duration fixed income stance. After U.S. large-cap growth leadership, the committee reduced developed international to neutral, added to U.S. large-caps and EM, and shifted EM toward growth (EM value vs. growth to neutral). The committee remains overweight U.S. SMID-cap, overweight real asset equities, and long a large-cap barbell (overweight U.S. large-cap growth and value, underweight core); it cites the largest oil supply disruption and the second-largest fertilizer disruption in history.
🔗 Source: Summary based on View Source from troweprice.com | Found on Jun 22, 2026
The paper proposes a model-grounded, RAG-based AI economist that combines knowledge graphs and LLM-based agents to plan analyses, retrieve evidence, select models, and generate reports; it avoids producing quantitative claims solely from the language model by grounding narratives in explicit model-based computations linked to retrieved evidence. The framework was evaluated on U.S. inflation persistence and Federal Reserve policy reports and on bank stress-test narratives for U.S. commercial real estate refinancing stress, showing improved economic coherence and traceability.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 19, 2026
Johnston et al. (submitted 25 Jun 2026) analyze OpenAI Codex usage across external personal accounts, external organizational accounts, and OpenAI workers, finding active users grew more than fivefold in the first half of 2026, with fastest growth outside software developers; within OpenAI Codex use is nearly universal and has largely replaced business ChatGPT usage; over 10% of users manage three or more concurrent Codex agents weekly and 26.6% use shared “skills”; the share of users submitting at least one request estimated to require >8 hours rose nearly tenfold since the start of 2026; median OpenAI legal staff produced 13× and median researchers >50× more monthly output tokens in June 2026 versus November 2025.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 26, 2026
Submitted 22 Jun 2026 by Mostapha Benhenda, the paper introduces IPO Finance Agent to extend Finance Agent v2 across task domain and retrieval architecture for IPO due diligence, noting Finance Agent v2’s narrow focus on SEC 10-K/10-Qs and that its original harness failed to produce outputs on the SpaceX S-1 due to document length, prompting contextual retrieval. The authors built a 1,000-question IPO-diligence dataset, publicly released 70 SpaceX S-1 questions, introduced an evaluator-optimizer pipeline for automated rubric generation, and report Alibaba Qwen 3.7 Max at 79.4% accuracy ($0.30/query) and Xiaomi MiMo-2.5 Pro at 76.8% ($0.05/query); both outperform Google Gemini 3.5 Flash (57.9%, $2.51) and cost-undercut MiniMax M3 (48.3%, $0.32). Code and data are on GitHub.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 23, 2026
Submitted 21 Jun 2026, Moran Koren’s "Theorist Toolbox" proposes a verification-first protocol for LLM-assisted economic theory and instantiates three reusable methods differing by checking: a single disciplined pass; an adversarial prover–verifier pair (Claude Opus 4.8 proposing, OpenAI Codex refuting, author triaging); and a structured multi-agent project with a reviewer gate. The protocol was evaluated on one worked example—designing a Groves/Pigouvian mechanism for the Gans–Kominers eigengrade model—and none of the three runs produced a strict direct-revelation VCG/Clarke mechanism. Three recurring phenomena were reported: convergent discovery, adversarial verification caught three false claims and a rejected sub-goal, and the most polished output was least verified; the takeaway is that external verification, not model capability, is the design variable.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 23, 2026
Submitted 10 Jun 2026, "AI Tokenomics: The Economics of Tokens, Computation, and Pricing in Foundation Models" by Quanyan Zhu develops a framework treating tokens as the accounting unit linking information processing, computation, memory use, energy expenditure, pricing, and economic value. It studies how tokens are generated, consumed, priced, allocated, and optimized across AI systems, connects token-level technical costs to workflow production functions, enterprise resource allocation, measurement/instrumentation, and market-design, and states token expenditure differs from economic value, which depends on marginal productivity, workflow position, hidden reasoning, risk, and downstream effects; it identifies open research directions.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 24, 2026
Submitted 22 June 2026, the paper by Vamshi Krishna Bonagiri, Juan Nicolas Sepulveda-Arias, Abdoul Jalil Djiberou Mahamadou, and Monojit Choudhury defines cognitive digital twins (CDTs) as dynamic computational representations of a specific person's cognition updated from behavioral, contextual, or physiological data to model, predict, simulate, or act as that person's proxy. It makes four contributions: CDT definition and distinction, a 5A governance framework (authority, autonomy, access and control, accountability, availability), identification of CDT-specific risks (misrepresentation, epistemic authority shifts, shadow twins, simulated participation, proxy action, proxy-power asymmetries), and governance requirements (strengthened consent, purpose limitation, validity, traceability, contestation, independent review, model retirement).
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 23, 2026
The paper "Can LLM Coding Agents Reason About Time Series?" (submitted 15 Jun 2026) by Filip Rechtorík, OndÅ™ej Dušek, and ZdenÄ›k Kasner compares three approaches—providing raw numerical data, using the LLM as a coding agent, and a combination—where the coding agent iteratively queries data using Python. Using two time series understanding benchmarks, coding agents outperform models processing raw data by up to 10%, yet the best agent still answers about 22–34% of questions incorrectly; analysis with a strong LLM judge found coding agents choose appropriate statistical tests but often miss important nuances, while raw-data models sometimes succeed via back-of-the-envelope calculations.
🔗 Source: Summary based on View Source from arxiv.org | Found on Jun 16, 2026
OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom Intelligence Processor, designed specifically for large-language-model inference and the first chip in a multi-generation compute platform. Co-developed with Broadcom and Celestica, the accelerator aims to improve cost, speed, reliability and energy efficiency across ChatGPT, Codex, API and future agentic products. Early lab tests reportedly show performance per watt well above current state of the art, with samples running GPT-5.3-Codex-Spark. OpenAI says full-stack control over chips, kernels, networking and deployment will lower compute costs, while Broadcom highlights gigawatt-scale data-center deployment beginning in 2026 with Microsoft and partners.
🔗 Source: Summary based on View Source from openai.com | Found on Jun 25, 2026
On June 22, 2026, IBM announced it joined the OpenAI Daybreak Cyber Partner Program and launched a new application security service that uses OpenAI cyber capabilities to identify and validate software vulnerabilities, assess application code, and prioritize high-risk areas. The service, delivered as a managed offering and available today, operates within client environments with read-only access to code repositories and bounded execution, connects via IBM Consulting Advantage, supports focused evaluations and continuous monitoring, and builds on Project Lightwell, which is supported by a $5 billion commitment from IBM and Red Hat.
🔗 Source: Summary based on View Source from newsroom.ibm.com | Found on Jun 23, 2026
Claude Tag is a new Slack-based way for teams to work with Claude, available today in beta for Claude Enterprise and Team customers and built on Opus 4.8. Administrators specify channel-scoped access to tools, data, and memories, can set token-spend limits and view logs, and can opt in to migrate from the existing Slack app within 30 days. Claude Tag is multiplayer, learns over time, can take initiative with “ambient” behavior, works asynchronously and can schedule tasks, supports private direct messages, and—per Anthropic—creates 65% of the product team’s code; eligible organizations receive an introductory launch credit.
🔗 Source: Summary based on View Source from anthropic.com | Found on Jun 24, 2026
NVIDIA and AWS introduced EC2 G7 instances with NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, offering up to eight GPUs, 256 GB total GPU memory, 700 Gbps EFA-enabled networking, and up to 7.6 TB local NVMe SSD storage; G7 delivers up to 4.6x AI inference and up to 2.1x graphics performance versus G6 and faster GPU analytics via NVIDIA cuDF on Amazon EMR. AWS OpenSearch Serverless now uses NVIDIA cuVS for GPU-accelerated vector indexing by default, claiming up to 10x faster indexing at one-quarter the cost versus CPU-only, and AWS achieved NVIDIA Exemplar Cloud status on NVIDIA GB300 for training workloads.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Jun 24, 2026
On June 26, 2026, Deloitte, IBM, and Red Hat announced a collaboration in which Deloitte joins the Lightwell initiative as an integration collaborator to protect the software supply chain from automated, AI-driven threats; Lightwell, supported by IBM and Red Hat, decouples open source remediation from upgrade cycles by coordinating upstream disclosures, developing, testing, backporting, and delivering validated patches to pinned in‑use software versions; Deloitte will maintain a bench of Forward Deployed Engineers and combine orchestration with IBM/Red Hat automated patch validation to provide continuous discovery, contextual prioritization, machine‑speed remediation, and ecosystem trust and compliance.
🔗 Source: Summary based on View Source from newsroom.ibm.com | Found on Jun 27, 2026
Amazon announced an additional $13 billion to expand AWS data center capacity in Mumbai and Hyderabad and a $35 billion investment across its India businesses; it said it is investing over $48 billion in the coming five years and that cumulative investments from 2010–2030 now exceed $88 billion. Amazon plans more than 20 new fulfillment centers and over 100 new last‑mile delivery stations this year, announced a $300 million operations/associate‑wellbeing investment tied to the "Sammaan" program, and reported digitizing 12 million small businesses, $20 billion in e‑commerce exports, 2.8 million jobs supported, and training over 10 million Indians on cloud skills; through 2030 it committed to supporting 3.8 million jobs, $80 billion in cumulative exports, enabling AI for 15 million small businesses, and AI education for 4 million government school students.
🔗 Source: Summary based on View Source from aboutamazon.com | Found on Jun 26, 2026
Confluent’s 2026 Data Streaming Report, from a survey of 4,625 IT leaders in 14 countries, finds investment in data streaming has overtaken AI/ML as a top strategic priority: 88% ranked data streaming a high investment priority while only 32% said they have agentic AI in production. The report cites 72% naming insufficient real-time data infrastructure as an AI barrier (up from 61% in 2025), nearly half reporting indefinite delays or abandonment of agentic AI projects, and 71% citing a skills gap (up from 66%). VP Andrew Sellers urges fixing siloed, poorly governed data, using shift-left integration and annotating critical data with schema and ownership.
🔗 Source: Summary based on View Source from ibm.com | Found on Jun 27, 2026
Author Justin Boitano published on June 23, 2026, a Nemotron Labs blog post describing NVIDIA Agent Toolkit as an open, modular foundation comprising models, tools and skills, blueprints and a runtime to build specialized agents. The toolkit interoperates with third‑party harnesses including Hermes Agents and OpenClaw. The post cites NVIDIA BioNeMo Toolkit, saying it enables work that previously took months to be completed in days, and mentions CrowdStrike’s specialized security agents triaging alerts with 98.5% accuracy. Cadence, Synopsys, Palantir, SAP, ServiceNow, Siemens and Dassault Systèmes are named as adopters.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Jun 23, 2026
In an Amazon Staff article published June 24, 2026, Garman said he told a room of CIOs “a couple of months ago” that 90% raised their hands when asked if they were seeing materially positive ROI today or had a path to high ROI soon, unlike a year earlier. He urged maximizing the “C” in ROIC, citing durable investments in land and power and server and chip commitments made only months out when customer visibility is high. He advised using the right (not the most expensive) model, noted AWS tools including Amazon’s agentic development environment route tasks to lighter models for code generation and higher-reasoning models for complex requests, cited Kiro as an example that picks models and helps customers budget to get results faster and less expensively, and recommended scaling successful use cases and quickly shutting down those that do not.
🔗 Source: Summary based on View Source from aboutamazon.com | Found on Jun 24, 2026
Authors Melanie Nakagawa and Juan Lavista Ferres published on June 15, 2026 a Joule study finding that a typical AI query to large LLMs uses 0.16–0.60 watt‑hours of electricity and, under conservative assumptions, 0.0–0.067 mL of water (median about one‑hundredth of a teaspoon). The study reports these energy figures are 4–20× lower than prior estimates, estimates serving 1 billion conversational queries uses ~0.7 GWh at baseline and ~0.3 GWh after efficiency improvements, and projects combined near‑term per‑query energy reductions of 8–20× via optimized models (e.g., Fara‑7B, Phi), smarter serving (Model Router), and better hardware (1.5–2.5× or Maia 200).
🔗 Source: Summary based on View Source from microsoft.com | Found on Jun 16, 2026
Author Benjamin Wu (June 16, 2026) describes an NVIDIA developer example that trains transformer-based transaction foundation models and reproduces a near-50% lift in Average Precision (AP) over a strong XGBoost baseline on the IBM TabFormer fraud dataset. The notebook loads ~24.4M synthetic transactions (~0.12% fraud) split 80/10/10 by time, trains an XGBoost baseline on a 1M-row balanced sample, evaluates on a 100k stratified holdout (~0.1% fraud), uses a domain tokenizer (≈12 tokens/transaction, 6,251 vocab) vs BPE (≈39 tokens, 50,257 vocab), pretrains a decoder checkpoint (3,000 steps), extracts 512-dim embeddings, and reports a combined model AP uplift of 41.76% and ROC-AUC +0.41%.
🔗 Source: Summary based on View Source from developer.nvidia.com | Found on Jun 18, 2026
Published on June 17, 2026, 16:37, the article states that AI has accelerated identity-focused cyberattacks by automating reconnaissance, personalizing social engineering, analyzing leaked credentials, identifying privileged users, probing systems, and adapting tactics in real time. At RSA earlier this year Microsoft unveiled a unified identity risk score and an updated Microsoft Entra ID Protection experience that correlates signals across accounts, sessions, workloads, and applications. Microsoft plans an identity-focused RBAC role in public preview, extended the Security Alert Triage Agent to identity scenarios, and added Conditional Access Optimization recommendations including a “Block risky user agent” policy.
🔗 Source: Summary based on View Source from techcommunity.microsoft.com | Found on Jun 18, 2026
Author Naveen Panakkal published on June 23, 2026, argues that "agentic" systems loop perceive→reason→act→observe and introduce designed stochasticity; ten design patterns are described, with core patterns (Prompt Chaining, Tool-Use Routing, Human-in-the-Loop), common patterns (Plan-then-Execute, Agentic RAG, ReAct), and specialized patterns (Memory, Reflection, Multi-Agent). He warns data readiness blocks production, notes production systems in 2025–2026 moved to Adaptive RAG, and gives reliability examples: five-step chains at 95% per step yield ~77% end-to-end, at 90% per step yield ~59%. This post is Part 1 of a four-part series.
🔗 Source: Summary based on View Source from community.sap.com | Found on Jun 24, 2026
In "Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs,” the authors show that enabling reasoning traces in R-LLMs (Gemini-2.5 Flash and Pro, Qwen3-32B) improves recall on simple single-hop questions from SimpleQA Verified and EntityQuestions via two mechanisms: a computational buffer effect (dummy traces like repeated "Let me think" tokens increase pass@k recall versus reasoning off) and factual priming (generated related facts prime retrieval); extracting concrete facts from traces and conditioning on them recovers most of reasoning’s gains and helps even when reasoning is disabled.
🔗 Source: Summary based on View Source from research.google | Found on Jun 25, 2026
Author Anand Oswal wrote on June 23, 2026 that the June 22, 2026 White House Executive Order mandates U.S. federal civilian agencies migrate to NIST-approved post-quantum cryptography, sets milestones of 2030 for key establishment and 2031 for digital signatures, and directs support for critical infrastructure owners/operators, federal contractors, and cryptographic bill of materials guidance. The order targets “harvest now, decrypt later” risk and signals broader procurement momentum; organizations doing business with the government or in energy, financial services, and healthcare should expect accelerated post-quantum readiness expectations. The article lists five actions for readiness: visibility, prioritization, modernizing trust, automation, and governance.
🔗 Source: Summary based on View Source from paloaltonetworks.com | Found on Jun 24, 2026