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.
Janus Henderson’s 2026 Investor Survey shows investors view AI as a major long-term market opportunity but remain wary of near-term risks. While 61% expect AI to benefit markets over time, 90% have concerns, led by execution risk, misuse and overvaluation. Two-thirds worry about an AI bubble or correction within 12 months, though younger investors are more optimistic about outsized returns. In advisory settings, investors are relatively comfortable with AI-created educational content but skeptical of AI-driven recommendations and automated client communications. Transparency is critical: 79% would object to undisclosed AI use, and 85% hold advisors responsible for AI output.
🔗 Source: Summary based on View Source from ir.janushenderson.com | Found on May 20, 2026
The AI-native enterprise services firm led by Anthropic, Blackstone, Hellman & Friedman, and others announced the acquisition of Fractional AI, an applied AI services company founded in 2024 by Chris Taylor, Eddie Siegel, and Travis May in San Francisco. Fractional AI’s team will serve as the operational centerpiece of the new company and collaborate with Anthropic’s Applied AI organization from day one. The new firm is backed by Goldman Sachs, General Atlantic, Leonard Green & Partners, Apollo Global Management, GIC, and Sequoia Capital. Terms of the acquisition were not disclosed.
🔗 Source: Summary based on View Source from blackstone.com | Found on May 21, 2026
Year-over-year data center capacity additions are projected to reach 13.6 GW in 2026 and 36.3 GW in 2027, compared to realized additions of 6.4 GW in 2024 and 8.5 GW in 2025, according to Goldman Sachs Research using Aterio’s facility-level data. Historically, only about 72% of scheduled activations occur on time, with current forecasts adjusted for risk predicting approximately 60% of next year’s scheduled capacity will materialize on time, dropping to roughly 50% over two years. Regional power reliability risks are highest in the Mid-Atlantic, Mid-Continent, and Northwest markets due to limited generation capacity relative to demand.
🔗 Source: Summary based on View Source from goldmansachs.com | Found on May 20, 2026
The article discusses the concept of the "token tax," which refers to the increased costs associated with embedding AI into software, including expenses for compute, tokens, memory, inference, storage, and networking. This shift challenges traditional software models by raising operating expenses and potentially lowering incremental margins. Equity investors have seen valuation multiples compress due to these pressures, while credit investors face rising loan-to-value ratios and refinancing difficulties as earnings profiles change. The article emphasizes that semiconductors and AI infrastructure are structurally advantaged due to their proximity to system constraints. Active management is recommended to distinguish between companies that benefit from AI and those challenged by it.
🔗 Source: Summary based on View Source from troweprice.com | Found on May 22, 2026
Modern mercantilism is evolving with the weaponization of strategic choke points, while artificial intelligence is driving an inflection point in software disruption, according to Co-CIO Greg in a Q1 CIO call excerpt. The science bet on AI is paying off, leading to accelerating resource grabs as compute demand now outpaces supply and may require rationing. These developments have significant implications for investors navigating today’s environment.
🔗 Source: Summary based on View Source from bridgewater.com | Found on May 18, 2026
Franklin Templeton argues enterprise software is entering a major AI-driven reset, with value shifting away from traditional seat-based SaaS models. AI has sharply lowered software development costs and is automating knowledge work, challenging vendors whose revenue depends on expanding employee licenses. The sector is splitting into three groups: AI workload beneficiaries, pressured seat-model incumbents and AI operating platforms embedded in workflows, orchestration and systems of record. Investors should prioritize pricing power, AI-driven revenue acceleration, stable gross margins, internal AI leverage and consumption- or outcome-based monetization. The conclusion: software demand should grow, but value distribution will materially change for active investors over time.
🔗 Source: Summary based on View Source from franklintempleton.com | Found on May 21, 2026
Earnings expectations for U.S. companies benefiting from AI infrastructure investment surged to over 50% as of late April 2026, following the public release of ChatGPT and accelerated AI adoption. Despite this growth, price-to-earnings multiples remained contained, near 27 times earnings—only modestly higher than the S&P 500. Demand for computing power and data center infrastructure has led to supply shortages and significant price increases; DRAM chip prices rose by more than 1,700% since early 2025. Hyperscaler capex budgets have sharply increased, with analyst estimates suggesting rolling 12-month free cash flow could approach zero by early 2027.
🔗 Source: Summary based on View Source from troweprice.com | Found on May 22, 2026
Semiconductor providers are reducing the unit cost per computing token for AI inference by 60%-70% annually, driven by improved chip and data center efficiencies. Despite rapid demand growth, especially for agentic AI, chip production cannot keep pace immediately, leading to expected shortages over the next 12 months but potential equilibrium in two years. Hyperscalers currently see most free cash flow consumed by capital expenditures, but falling costs are projected to improve gross margins within 3-12 months. Daily LLM queries are forecasted to grow at a 40% CAGR to reach 11 billion by 2030. Enterprise adoption of agentic AI is slower and uneven.
🔗 Source: Summary based on View Source from goldmansachs.com | Found on May 21, 2026
The article "Agentic Trading: When LLM Agents Meet Financial Markets" by Yihan Xia and six co-authors presents an audit-oriented evidence map of 77 studies screened through March 9, 2026, focusing on the integration of Large Language Models (LLMs) in trading systems. Of these, a primary empirical subset of 19 studies meets the minimum criteria for Action Output plus Closed-Loop Evaluation. Key findings include protocol incomparability: only 2 out of 19 studies report extractable time-consistent split protocols, one reports an explicit transaction-cost model, one documents universe or survivorship handling, and no study achieves R3 reproducibility.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 20, 2026
The article, authored by Mona H. Albaqawi, Eman M. Albalkhi, Joud A. Albaiti, and Enrico Lopedoto, presents an Arabic NLP framework for large-scale financial sentiment analysis tailored to the Saudi market. The framework integrates official financial news and social media to capture both institutional and public investor sentiment, constructing a large Arabic financial corpus of 84,000 samples through a multi-stage pipeline including data collection, cleaning, deduplication, entity linking with transformer-based NER and a curated company lexicon, and five-class sentiment annotation. Experimental results demonstrate reliable and scalable company-level sentiment analysis relative to stock market behavior on the Saudi Exchange.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 20, 2026
The article introduces BacktestBench, the first large-scale benchmark for automated quantitative backtesting, developed by Zhensheng Wang and colleagues and accepted by KDD 2026. BacktestBench is constructed from over 6 million real market records and contains 18,246 annotated question-answering pairs across four categories: metrics calculation, ticker selection, strategy selection, and parameter confirmation. The authors also present AutoBacktest, a multi-agent baseline system that translates natural language strategies into reproducible backtests using a Summarizer, Retriever for SQL generation, and Coder for Python implementation. Evaluation on 23 mainstream LLMs identifies key factors affecting performance.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 19, 2026
The article "WorkstreamBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance" by Thomson Yen and 11 co-authors, submitted on 21 May 2026, presents one of the first evaluations of large language model (LLM) agents on end-to-end spreadsheet tasks relevant to finance, such as modeling and scenario analysis. The authors introduce an evaluation taxonomy with three dimensions—Accuracy, Formula, and Format—to assess solution quality according to professional standards. Results show that while the Claude family leads the benchmark with the most professional-looking outputs, all agents struggle with complex workflows and do not consistently meet professional finance standards.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 22, 2026
In the paper "Memory-Induced Supra-Competitive Outcomes Between Deep Reinforcement Learning Agents in Optimal Trade Execution" by Christos Spyridon Koulouris and Carlo Campajola, submitted on 19 May 2026, the authors investigate deep reinforcement-learning agents in a shared optimal-execution environment. They study a two-agent Almgren-Chriss liquidation game and find that supra-competitive outcomes—lower implementation shortfalls than the competitive benchmark—are more frequent and persistent when agents have access to intra-episode history, including recent prices and their own past actions. The results show that feedback, memory, and state-contingent interaction drive these outcomes.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 21, 2026
On average, after one round trip of LLM edits (two interactions), 18% of a document’s content was corrupted, increasing to over 50% after 20 interactions. The study evaluated 19 large language models using the DELEGATE-52 dataset, which spans 52 professional domains and documents of about 3–5,000 tokens each. Repetitive, numerical, and structurally dense documents were less affected; Python code was almost perfectly preserved. Even the top three models—Gemini 3.1 Pro, Claude 4.6 Opus, and GPT-5.4—degraded documents by 25% on average. Agentic tool use increased token consumption by two to five times and worsened document corruption slightly.
🔗 Source: Summary based on View Source from ibm.com | Found on May 23, 2026
Enterprise AI is advancing rapidly, as highlighted at Dell Technologies World on May 18, 2026. Dell and NVIDIA introduced the Vera Rubin NVL72, offering agentic AI inference at one-tenth the cost per token and 50% faster sandbox performance than traditional CPUs. The PowerEdge XE9812 delivers up to 10x lower cost-per-token than NVIDIA Blackwell, while Vera CPUs provide 1.2 TB/s memory bandwidth and complete workloads 50% faster than x86 processors. Over 5,000 enterprises—including Lilly, Samsung, and Honeywell—are running AI workloads on Dell AI Factories with NVIDIA. Worldwide AI infrastructure spending may reach $3-4 trillion by 2030.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on May 19, 2026
Driven by robust AI demand, Alibaba’s Cloud Intelligence Group’s external revenue growth accelerated to 40% in the final quarter of fiscal 2026, with AI-related products accounting for 30% of this revenue. The company has achieved production at scale for its proprietary T-Head AI chips and released three updates to the Qwen model family within three months, including Qwen3.7-Max engineered for agents. In November 2025, Alibaba launched the Qwen app as an all-in-one personal AI assistant and introduced Wukong, an enterprise agentic platform. Alibaba is scaling up investments in full-stack AI capabilities across infrastructure, models, and applications.
🔗 Source: Summary based on View Source from alibabagroup.com | Found on May 20, 2026
The article discusses the shift in board-level conversations about AI from ambition to accountability, highlighting that most CIOs are unprepared due to the rapid proliferation of AI agents without adequate governance. It notes that teams have independently deployed agents across multiple platforms, resulting in ungoverned assets, unclear ownership, and untracked costs. The article identifies three main challenges: compliance with regulations like the EU AI Act requiring inventories and risk documentation; demonstrating ROI by linking agents to business process improvements; and establishing control through ownership, authorization, and continuity plans. The SAP AI Agent Hub is presented as a solution for comprehensive governance.
🔗 Source: Summary based on View Source from community.sap.com | Found on May 21, 2026
The article by Edward Li, published on May 19, 2026, explains that evaluating an AI model focuses on static benchmarks like MMLU, GSM8K, and HumanEval to measure cognitive and linguistic potential, while agent evaluation assesses end-to-end behavior in dynamic environments using benchmarks such as GAIA, SWE-bench, and WebArena. Agent evaluation tracks Task Success Rate (TSR), Tool Call Accuracy, and Trajectory Efficiency to measure intent resolution and efficiency. The article provides five practical tips for agent evaluation and highlights the NVIDIA NeMo Agent Toolkit as a solution for integrating evaluation metrics into existing frameworks without a full rebuild.
🔗 Source: Summary based on View Source from developer.nvidia.com | Found on May 20, 2026
The article by Peihan Huo, published May 21, 2026, details the use of NVIDIA’s NeMo Agent Toolkit and Nemotron open models to automate quantitative signal discovery in finance. The system employs three specialized agents—signal generation, code execution, and evaluation—in a self-improving loop managed via YAML configuration. Two momentum-based stock-selection signals were generated: ExpVolume-Adjusted Momentum and Rank-Adjusted Return Momentum. The latter was backtested against S&P 500 forward returns, achieving a mean IC of -0.0134 over 3,504 trading days with a t-statistic of -5.37 and positive-IC ratio of 46.4%.
🔗 Source: Summary based on View Source from developer.nvidia.com | Found on May 22, 2026
KPMG, operating in 138 countries with over 276,000 employees, has formed a global alliance with Anthropic to integrate Claude AI into its core business functions. Claude will be embedded in KPMG’s Digital Gateway platform—built on Microsoft Azure—initially for tax and legal clients, and rolled out to all employees worldwide. Anthropic named KPMG a preferred partner for private equity, collaborating on new AI-powered products for portfolio companies. The alliance includes using Claude to enhance cybersecurity and modernize IT systems through offerings like KPMG Blaze. Joint research with UT Austin highlights the importance of human judgment alongside AI deployment.
🔗 Source: Summary based on View Source from anthropic.com | Found on May 19, 2026
In the first month of Project Glasswing, Anthropic and about 50 partners used Claude Mythos Preview to identify over 10,000 high- or critical-severity vulnerabilities in key software. Cloudflare found 2,000 bugs (400 high/critical), Mozilla fixed 271 Firefox vulnerabilities, and open-source scans revealed 6,202 high/critical issues out of 23,019 total. Of 1,752 assessed open-source vulnerabilities, 90.6% were valid; about 530 high/critical bugs have been disclosed to maintainers and 75 patched so far. The bottleneck is now patching speed rather than discovery. Anthropic has released tools like Claude Security and collaborates with organizations to improve vulnerability management.
🔗 Source: Summary based on View Source from anthropic.com | Found on May 22, 2026
Google launched Gemini 3.5 Flash, a new AI model available via Google Antigravity, the Gemini API in Google AI Studio, and Android Studio, outperforming Gemini 3.1 Pro on benchmarks such as Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo), and MCP Atlas (83.6%). Gemini Omni was introduced for multimodal content generation starting with video, featuring SynthID watermarking and global rollout through the Gemini app and YouTube Shorts Remix at no cost for users aged 18+. Major upgrades include an AI-powered Search box supporting text, images, files, videos, Chrome tabs; Universal Cart for intelligent shopping; Daily Brief agent; Google Pics image tool; Docs Live voice editing; and new science tools like Hypothesis Generation and Computational Discovery.
🔗 Source: Summary based on View Source from blog.google | Found on May 21, 2026
Gemini Omni, introduced by Nano Banana and announced by Koray Kavukcuoglu on May 19, 2026, is a new multimodal model that enables users to create high-quality videos from any combination of images, audio, video, and text inputs. The first model in this family, Gemini Omni Flash, is being launched for the Gemini app, Google Flow, and YouTube Shorts. Users can edit videos using natural language instructions that build upon each other while maintaining character consistency and scene continuity. Future updates will expand output modalities to include image and audio generation.
🔗 Source: Summary based on View Source from blog.google | Found on May 20, 2026
AMD announced over $10 billion in investments across the Taiwan ecosystem to expand strategic partnerships and scale advanced packaging manufacturing for next-generation AI infrastructure, as stated on May 21, 2026. The company is collaborating with partners such as ASE, SPIL, and PTI to develop EFB-based 2.5D packaging technology supporting the 6th Gen AMD EPYC CPUs codenamed “Venice.” AMD has qualified the industry’s first 2.5D panel-based EFB interconnect with PTI. These innovations will support multi-gigawatt deployments of the AMD Helios rack-scale platform featuring “Venice” CPUs and Instinct MI450X GPUs beginning in the second half of 2026.
🔗 Source: Summary based on View Source from amd.com | Found on May 21, 2026
At the 2026 Google I/O conference, NVIDIA and Google Cloud announced enhancements to their joint developer community, which now supports over 100,000 developers with curated learning paths, hands-on labs, and events for building AI applications using the NVIDIA AI platform on Google Cloud. New offerings include a JAX library learning path for NVIDIA GPUs, a NVIDIA Dynamo codelab focused on inference optimizations, and monthly livestreams. The collaboration features tools such as NVIDIA cuDF in Google Colab Enterprise or Dataproc and supports deployment of multi-agent applications using Gemma 4 models and Nemotron open models on G4 VMs with RTX PRO 6000 Blackwell GPUs.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on May 19, 2026
Anthropic is acquiring Stainless, a company founded in 2022 that specializes in SDKs and MCP server tooling, to enhance the reach of its AI agents. Stainless has generated every official Anthropic SDK since the launch of Anthropic’s API, supporting hundreds of companies with SDKs, CLIs, and MCP servers for languages such as TypeScript, Python, Go, and Java. Katelyn Lesse, Head of Platform Engineering at Anthropic, emphasized Stainless’s role in shaping the Claude API experience. Alex Rattray, Founder and CEO of Stainless, highlighted their longstanding collaboration with Anthropic and commitment to developer experience.
🔗 Source: Summary based on View Source from anthropic.com | Found on May 18, 2026
NVIDIA and Ineffable have initiated a new engineering-level collaboration focused on reinforcement-learning agents, which are AI systems that learn by trial and error and can convert computation into new knowledge.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on May 19, 2026
IBM and the U.S. Department of Commerce announced a Letter of Intent on May 21, 2026, to establish Anderon, America’s first pure-play quantum chip foundry headquartered in Albany, New York. The initiative will receive $1 billion in CHIPS incentives from the Department of Commerce and $1 billion in cash from IBM, with additional intellectual property, assets, and workforce contributions by IBM. Anderon will operate as a 300-millimeter quantum wafer foundry supporting superconducting qubit wafers and aims to expand into other modalities. The project is expected to create thousands of high-paying jobs and strengthen U.S. quantum leadership.
🔗 Source: Summary based on View Source from newsroom.ibm.com | Found on May 23, 2026