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.
Over the past 18 months, AI-related capital expenditure in the U.S. has increasingly relied on debt markets, with investment in equipment and software as a share of GDP set to surpass late-1990s peaks. Consensus estimates for the five largest hyperscalers’ capex have risen to nearly $690 billion for 2026 and $870 billion for 2027, up from about $480 billion at the start of 2026. Year-to-date index-eligible new debt issuance from hyperscalers reached approximately $136 billion, exceeding full-year 2025 totals, alongside an additional $58 billion tied to data centers and $822 billion in future lease commitments.
🔗 Source: Summary based on View Source from pimco.com | Found on May 30, 2026
Data infrastructure—including data centers, fiber networks, and towers—has become central to infrastructure investing due to strong demand from cloud computing, mobile usage, and AI. Despite 96% global mobile broadband coverage, a usage gap of over three billion people persists. Infrastructure development is constrained by access to power, land, and network connectivity; projects often take 20% longer than planned and can exceed budgets by up to 80%. Successful investments combine access to scarce inputs with business models offering revenue visibility and expansion capability. Returns are increasingly differentiated by the ability to deliver infrastructure in constrained environments.
🔗 Source: Summary based on View Source from brookfield.com | Found on May 30, 2026
Inflation has cooled from the 4%–6% levels of 2022, but recent oil price increases due to the Iran situation have created a near-term headwind. Government spending has been a major driver of growth and inflation, peaking in the first half of 2026 and expected to turn neutral or negative without new stimulus. Housing costs, which comprise about one-third of the Consumer Price Index, are softening; if rent and home price growth trends toward zero, inflation could drop by 80–100 basis points. Large tech companies may raise $500 billion to $1 trillion in bonds for AI infrastructure, mainly through longer-dated bonds.
🔗 Source: Summary based on View Source from wellington.com | Found on May 26, 2026
In May 2026, Anthropic reached $47 billion in run-rate revenue (RRR), up from $14 billion in February 2026 and $1 billion in January 2025. The company introduced Mythos Preview, a new model class with significant improvements in intelligence, coding, long-running agents, and security. Anthropic launched dozens of new product features in the past quarter, expanding Claude’s use cases from coding to financial analysis, design, slide generation, legal, security, and life sciences. Productivity per engineer within Anthropic has increased by hundreds of percentage points and is accelerating as the company scales.
🔗 Source: Summary based on View Source from coatue.com | Found on May 29, 2026
Data centres are emerging as a key theme in digital infrastructure, driven by sustained growth in cloud, streaming, e-commerce, and AI. According to Amundi on May 28, 2026, the investment focus is shifting from space and storage towards power, cooling, connectivity, and resource access as the next phase of development.
🔗 Source: Summary based on View Source from amundi.com | Found on May 28, 2026
The article highlights the complexities international businesses face due to differing AI regulations: the EU’s Artificial Intelligence Act uses a risk-based model with strict requirements for high-risk systems, the US employs a sector-specific approach, and China emphasizes state control and data sovereignty. In September 2025, Anthropic settled a class action over pirated training data for about $1.5 billion. By year-end 2025, over 70 generative AI copyright lawsuits had been filed, including The New York Times v. OpenAI and Microsoft. Key risks include bias, discrimination, intellectual property rights, privacy concerns, and training data issues.
🔗 Source: Summary based on View Source from alliancebernstein.com | Found on May 29, 2026
Recent Federal Reserve communications have become more hawkish due to concerns that persistent supply-driven price pressures could affect inflation expectations. U.S. Personal Consumption Expenditures (PCE) annualized inflation, both headline and core, has accelerated in recent months to over one percentage point above the Fed’s 2% target. Minutes from the April Fed meeting show most participants support further policy firming if inflation does not moderate, and some officials have stated that rate hikes cannot be ruled out. The Fed’s Taylor rule estimates suggest the policy rate is currently 75 to 100 basis points too accommodative.
🔗 Source: Summary based on View Source from pimco.com | Found on May 30, 2026
In 2026, AI spending reached $2.5 trillion, surpassing the GDP of 186 countries, while crypto assets fell 50% and traded at a 42% discount to their long-term trend. Blockchain has matured beyond cryptocurrencies into infrastructure for trust and ownership in decentralized systems, with convergence between AI and blockchain seen as a major innovation wave. World (formerly Worldcoin) uses iris scans to issue unique World IDs on its blockchain for proof of personhood. Pantera launched Fund V for growth-stage blockchain investments, offering co-investment rights to LPs with $25 million commitments and donating 1% of revenue to charity.
🔗 Source: Summary based on View Source from panteracapital.com | Found on May 29, 2026
This survey, authored by Ruizhe Zhou and six others and submitted on 11 May 2026, examines determinism issues in financial AI systems used for credit risk, fraud detection, and anti-money laundering. The paper highlights reproducibility failures in tabular models (post-hoc explanation variance), graph networks (stochastic sampling and temporal asynchrony), and LLM-based workflows (batch-dependent divergence and trajectory drift). First-party experiments on public financial datasets quantify explanation rank instability in credit scoring, prediction flip rates in GNN-based fraud detection, and tensor-parallel-induced output divergence in LLM entity extraction. A layered evaluation framework linking modality-specific metrics to audit readiness is proposed.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 26, 2026
The article, authored by Eleni Straitouri, Cheol Woo Kim, and Milind Tambe and accepted at the ICML 2026 LM4Plan Workshop, introduces a novel algorithm that generates robust portfolios of optimization models using large language models (LLMs). The proposed method utilizes a single LLM in two roles—as a stochastic generator and as a reasoning evaluator—within a unified framework. Theoretical guarantees are provided to ensure that if either the generator or evaluator aligns with human preferences, the portfolio will include high-quality candidates. Empirical validation demonstrates strong performance across various optimization modeling tasks.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 27, 2026
The article by Shuchen Meng and Xupeng Chen demonstrates that as AI-driven investment strategies become more widely adopted, excess returns are compressed through signal crowding, performative signal erosion, and Red Queen competition. The authors derive an alpha half-life formula showing that at current adoption levels ($\phi \approx 0.7$, $\rho \approx 0.6$), signal half-lives have dropped to 18 months from pre-AI levels of 5-7 years. Empirical analysis of SEC Form 13F filings (99.5 million holdings, 2013-2024) reveals a 42% increase in institutional portfolio convergence, with simulations highlighting declining return dispersion among AI funds and increased market fragility.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 26, 2026
The article introduces KTD-Fin (Knowing-To-Doing Financial Benchmark), developed by Taojie Zhu and colleagues, as an end-to-end stock-market trading benchmark for evaluating large language model (LLM) agents. KTD-Fin addresses two main evaluation failures: overlap with LLMs’ knowledge cutoffs and the inadequacy of raw returns as a measure of stock-selection ability. It employs a data-side masking protocol to anonymize identifiers and calendar information, separating market memory from investment decision-making, and uses a Barra-style attribution framework to decompose portfolio returns. Testing ten frontier LLM agents on the Chinese CSI300 (2024–2026), results show limited evidence of persistent stock-selection alpha.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 28, 2026
The article titled "MadEvolve: Evolutionary Optimization of Trading Systems with Large Language Models" by Yurii Kvasiuk, Tianyi Li, Owen Colegrove, and Moritz Münchmeyer presents MadEvolve, a general-purpose algorithm optimization framework inspired by DeepMind's Alpha-Evolve. The authors demonstrate MadEvolve's effectiveness in optimizing algorithmic trading strategies and alpha generation for Bitcoin trading through simulation and backtesting. The study reports significant improvements across tasks such as evolving feature sets for signal generation, optimizing individual strategy components, and jointly evolving the feature pipeline with execution strategy. Comparisons to agentic search methods like Claude Code are included, alongside careful evaluation of p-hacking probabilities.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 25, 2026
PortBench is introduced as a benchmark for LLM-driven portfolio management, addressing gaps in existing benchmarks by incorporating cross-asset correlation structures and evaluating the full portfolio management decision pipeline. It spans six heterogeneous asset classes over ten years and includes a static QA dataset of 6,269 correlation-based questions across seven task templates, as well as a dynamic five-stage allocation pipeline. Two dedicated metrics are proposed: a dual-layer correlation score and CEPS, which measures compounding reasoning errors. Evaluation of ten frontier LLMs shows that 90% of model-profile combinations fail to outperform equal-weight allocation and suffer catastrophic drawdowns under stress.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 28, 2026
KairosAgent is a novel agentic framework for multimodal time series forecasting, proposed by Kun Feng and eight co-authors and submitted on 28 May 2026. The framework integrates an LLM-based reasoner with a TSFM-based forecaster, dynamically invoking analytical tools to enhance numerical understanding and semantic reasoning in large language models. Reasoning results are fused into the TSFM pipeline for improved prediction accuracy. The authors curated a large-scale corpus of high-quality trajectories and introduced reinforcement learning from forecasting with multi-turn refinement and turn-level credit assignment. Experiments show KairosAgent achieves superior zero-shot forecasting performance.
🔗 Source: Summary based on View Source from arxiv.org | Found on May 29, 2026
Claude Opus 4.8, released as an upgrade to Opus 4.7 at the same price ($5 per million input tokens and $25 per million output tokens; fast mode at $10 and $50 respectively), offers improved benchmarks, reliability, and judgment in agentic tasks. New features include user effort control, dynamic workflows enabling codebase-scale migrations, and a Messages API update for mid-task instruction changes. Opus 4.8 is three times cheaper in fast mode than previous models, scored 84% on Online-Mind2Web, broke 10% on the Legal Agent Benchmark all-pass standard, and delivers more efficient tool use with lower rates of misaligned behavior.
🔗 Source: Summary based on View Source from anthropic.com | Found on May 28, 2026
NVIDIA has launched the high-performance, energy-efficient Vera CPU, designed for agentic AI, reinforcement learning, and data processing across industries. Vera powers standalone servers, NVIDIA Vera Rubin systems, and BlueField-4 STX AI storage platforms. It delivers 1.8x faster task completion than x86 CPUs and features 88 Olympus cores with up to 1.2TB/s LPDDR5X memory bandwidth. Global AI labs such as Anthropic, OpenAI, SpaceXAI, and hyperscalers including ByteDance and Oracle Cloud Infrastructure plan to adopt Vera. Manufacturers like Dell Technologies, HPE, Lenovo, Supermicro and others will offer Vera-based systems starting this fall.
🔗 Source: Summary based on View Source from nvidianews.nvidia.com | Found on Jun 01, 2026
Anthropic has raised $65 billion in Series H funding led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital, resulting in a post-money valuation of $965 billion. The round was co-led by Capital Group, Coatue, D1 Capital Partners, GIC, ICONIQ, and XN, with significant investments from firms including AMP PBC and Baillie Gifford. The funding includes $15 billion of previously committed investments from hyperscalers such as Amazon’s $5 billion. Anthropic’s run-rate revenue crossed $47 billion earlier this month. Claude is now available on AWS, Google Cloud, and Microsoft Azure; AWS remains the primary cloud provider and training partner.
🔗 Source: Summary based on View Source from anthropic.com | Found on May 28, 2026
NVIDIA has released a major open source collection of physical AI agent skills and tools spanning Omniverse, Cosmos, Alpamayo, and Metropolis for robotics, autonomous vehicles, vision AI, and industrial digital twins. Industry leaders such as Agile Robots, Cadence, Dassault Systèmes, Delta Electronics, Foxconn, Pegatron, PTC, Siemens, Synopsys and TSMC are using these tools to accelerate development. Pegatron reduced model training and deployment time by 67% with synthetic data; Delta Electronics improved defect detection rate by 17%; Inventec reduced defect data collection effort by 30%; Foxconn boosted first pass yield by about 3%.
🔗 Source: Summary based on View Source from nvidianews.nvidia.com | Found on Jun 01, 2026
Alibaba Group announced a multi-year partnership with UEFA and UC3, becoming the official and exclusive AI, Cloud Computing Services, and E-commerce partner for the UEFA Champions League, Europa League, Conference League (2027/2028–2032/2033), and UEFA EURO 2028TM. The collaboration will deploy Alibaba Cloud infrastructure and Qwen Large Language Model to enhance fan engagement, media management, and personalized digital experiences. Fans worldwide will gain seamless access to official merchandise via Alibaba’s global e-commerce platform. The partnership was facilitated by Relevent and will be managed by CAA11 for UEFA EURO 2028TM.
🔗 Source: Summary based on View Source from alibabagroup.com | Found on Jun 01, 2026
TSMC is utilizing NVIDIA CUDA-X libraries, AI models, and GPUs to accelerate workloads in lithography, transistor and process simulation, advanced process control, and fab operations optimization. TSMC reports a 20-50% improvement in cost effectiveness or cycle time for computational lithography compared to CPU-based methods and achieves 50x faster chemistry simulations for semiconductor material design. The company uses NVIDIA Metropolis and TAO Toolkit with vision AI to improve nanometer-scale defect detection while reducing repeated labeling and retraining. TSMC is also exploring NVIDIA Omniverse libraries to build FabTwin, a virtual fab environment for planning efficiency.
🔗 Source: Summary based on View Source from nvidianews.nvidia.com | Found on Jun 01, 2026
On May 28, 2026, CrowdStrike announced the next phase of Project QuiltWorks, expanding its framework to address both technical and financial risks associated with frontier AI. In collaboration with Coalition, Liberty Mutual Insurance, Lockton, Resilience, and Marsh, the initiative integrates actuarial intelligence and underwriting expertise to provide organizations with a comprehensive model for identifying, prioritizing, remediating, and financially mitigating frontier AI risk. Powered by models from OpenAI and Anthropic and supported by leading insurers and brokers, QuiltWorks offers financial risk modeling, exposure prioritization using adversary intelligence and vulnerability telemetry, continuous visibility for underwriting confidence, and coordinated risk mitigation at scale.
🔗 Source: Summary based on View Source from crowdstrike.com | Found on May 29, 2026
Workday and Google Cloud announced an expanded strategic partnership on May 28, 2026, integrating the Sana Self-Service Agent from Workday directly into Gemini Enterprise. Gemini is now the default AI model for Sana for Workday, enhancing HR and finance workflows with advanced reasoning, multilingual support, and multi-modal capabilities. The collaboration enables employees to access HR and finance services within Gemini Enterprise, supports zero-copy data sharing between Workday Data Cloud and Google Cloud Lakehouse, and involves global system integrators like Accenture, Deloitte, and KPMG to accelerate customer outcomes. The Sana Self-Service Agent is available in early access for eligible customers.
🔗 Source: Summary based on View Source from newsroom.workday.com | Found on May 29, 2026
A recent study in Nature Neuroscience, led by Nai Ding of Zhejiang University, found that the human brain processes language more efficiently than large language models (LLMs) by reducing next-word prediction at sentence and phrase boundaries and compressing completed linguistic units into higher-order conceptual representations. Unlike LLMs, which use vast computational resources to analyze all possible word relationships, the brain prioritizes key information and adapts rapidly to context. Researchers such as Stanislaw Wozniak of IBM Research suggest that adopting brain-inspired hierarchical representations could significantly reduce model size and improve AI efficiency.
🔗 Source: Summary based on View Source from ibm.com | Found on May 30, 2026
SilverTorch is a unified model-based recommendation retrieval system developed at Meta, replacing traditional microservice architectures with a single PyTorch neural network. In production tests on an 80-million-item workload, SilverTorch achieved up to 23.7× higher throughput and 20.9× greater compute cost efficiency compared to state-of-the-art CPU-based solutions, while maintaining sub-100-millisecond latency and improving recommendation accuracy. Key innovations include the integration of ANN search, eligibility filtering, and multi-task scoring as nn.Modules within one model, use of Int8 quantized embeddings for memory efficiency, Bloom index filtering for rapid eligibility checks, and real-time streaming updates for index freshness.
🔗 Source: Summary based on View Source from engineering.fb.com | Found on May 28, 2026
More than 30 percent of the US working-age population is using AI, marking a three percentage point increase since the end of 2025. Microsoft’s new report, released on May 28, 2026, provides the first state- and county-level review of AI adoption in the United States. Nationally, while the US leads in AI innovation, it ranks 21st globally in adoption rates. Metropolitan counties have an average AI usage rate of 32.9 percent compared to 16.2 percent in rural areas. Counties with higher shares of residents aged 18 to 24 show a usage rate of 28.6 percent versus 20.3 percent elsewhere.
🔗 Source: Summary based on View Source from blogs.microsoft.com | Found on May 29, 2026
Many employees are using unauthorized GenAI tools to handle sensitive data, while developers add AI features before security review, and SaaS platforms process business data by default, creating an expanding attack surface. Organizations lack visibility into which AI services are used and what data is exposed, leading to risks such as data leakage, compliance failures, and reputational damage. Traditional security tools cannot address threats like prompt injection or agent-to-agent communications. CrowdStrike Falcon® AI Detection and Response (AIDR) provides unified visibility and control across endpoint, identity, cloud, and AI to detect threats like prompt injection and credential abuse.
🔗 Source: Summary based on View Source from crowdstrike.com | Found on May 31, 2026
Google SRE is advancing agentic AI across the entire software development lifecycle, notably enhancing investigation and mitigation processes such as root cause analysis. AI agents monitor and improve playbooks, generate new documentation from incidents, and augment alerting with anomaly detection using models like TimesFM. These agents consolidate alerts, enrich context, and autonomously handle issues to reduce review workload. Incident management is supported by orchestration layers that summarize communication, facilitate handoffs, and draft postmortems. AI Insights leverages Gemini embedding models and vector databases to extract risk information from historical incidents. All SRE AI systems adhere to strict reliability, security, transparency, and compliance principles.
🔗 Source: Summary based on View Source from cloud.google.com | Found on May 29, 2026
A Microsoft study evaluated 19 large language models (LLMs) using the DELEGATE-52 dataset, which spans 52 professional domains and documents averaging 3–5,000 tokens. The researchers found that after just two LLM interactions, 18% of document content was corrupted; after six interactions, a third was corrupted; and after 20 interactions, over 50% was degraded. Even top models like Gemini 3.1 Pro, Claude 4.6 Opus, and GPT-5.4 degraded documents by an average of 25%. Introducing basic agentic tool use increased input token consumption by two to five times and slightly worsened document degradation due to context length limitations.
🔗 Source: Summary based on View Source from ibm.com | Found on May 28, 2026
Amazon's AGI Lab, led by Bryan Silverthorn, is focused on advancing AI agents from being useful under human supervision to operating independently in complex real-world scenarios. According to research engineer Gaurav Mishra, Amazon employs reinforcement learning in thousands of realistic simulated environments, enabling skills learned in one context to transfer to others and making agents production-ready. Since 2015, Amazon has designed its own chips for AI applications and estimates these custom chips reduce AI training costs by up to 50% compared to alternatives. Leading companies like Anthropic use Amazon's infrastructure for training advanced models.
🔗 Source: Summary based on View Source from aboutamazon.com | Found on May 29, 2026