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.
The article argues that AI capital expenditure is likely pushing up real yields by raising desired investment, but this is not the same as conventional crowding out. It says roughly $1 trillion in hyperscaler capex is expected this year, regardless of whether it is financed by debt, retained earnings, or equity, because financing does not change the real resources consumed. The paper also says the evidence for portfolio-rebalancing crowding out is weak: over the past 12 months it identified six surprise AI debt deals, and the two-day change in 10-year Treasury yields around those events was not statistically significant. Amazon’s 10 March deal was the main outlier.
🔗 Source: Summary based on View Source from pimco.com | Found on Sep 21, 2026
In a September 10, 2026 newsletter from Northwestern University’s Kellogg School of Management, AQR co-founder Mr. Kabiller and Kellogg Dean Ms. Cornelli argue that AI is making analytical and technical capabilities easier to automate, so creativity and emotional intelligence are becoming durable differentiators. They cite Kellogg’s 2020 MBAi program and AQR’s 2015 Quanta Academy as efforts to bridge technical skills with leadership and human development. They also reference Kellogg Professor William Brady, the 2025 Kabiller Science of Empathy Prize recipient, whose research suggests AI’s effect on empathy depends on design choices, incentives, and culture.
🔗 Source: Summary based on View Source from aqr.com | Found on Sep 16, 2026
In an Investment Insights episode published on September 21, 2026, leaders from TPG Transition Infrastructure discussed themes shaping infrastructure investing across power, transport and environmental services. Scott Lebovitz and Steven Mandel said the strategy combines middle-market infrastructure investing with TPG’s broader platform, including operational expertise, capital markets resources and relationships from the TPG Rise Climate ecosystem. They said rising power demand, including needs linked to technological advancements and AI, is creating opportunities in power generation and energy infrastructure, alongside renewable energy and battery storage. JD Vargas highlighted opportunities in waste and water infrastructure and the investment needed to modernize these systems.
🔗 Source: Summary based on View Source from tpg.com | Found on Sep 21, 2026
The AI investment story is extending beyond US hyperscalers and model developers to Asia’s semiconductor and AI ecosystem. US hyperscalers are increasing capital expenditure, and there have been no reported capacity cuts or order delays. Asia accounts for more than 90% of global foundry capacity, 85% of assembly and packaging, and most silicon wafer and semiconductor equipment production. The region’s advantage rests on decades of engineering expertise across Taiwan, South Korea, Japan, and parts of China. AI is described as a generational increase in compute demand, with opportunities across the full AI value chain and China’s semiconductor localization efforts.
🔗 Source: Summary based on View Source from wellington.com | Found on Sep 18, 2026
Consumer AI agents are shifting from conversational tools to action-oriented programs that may access calendars, passwords, and credit cards to complete tasks such as buying event tickets, booking hotels in Portugal, and managing appointments. It says the mass market will likely be monetized through advertising and subscriptions, with premium tiers for advanced models, and that AI can improve ad creation, placement, and measurement. It also says the capex cycle is expected to stay elevated through end-2027, with US hyperscalers projected to deploy $1.4 trillion in capital in 2027, though supply-chain constraints, memory chips, power, and land could limit deployment.
🔗 Source: Summary based on View Source from goldmansachs.com | Found on Sep 18, 2026
The AI infrastructure buildout is being treated as one large credit trade, with spread dispersion across the financing chain still limited despite sharp differences in underlying risk. It notes that, quarter-to-date, AI-related debt has underperformed broader indices in both investment grade and high yield credit, according to Bloomberg index data. Debt investors receive mainly contractual returns through coupon, principal, and possible spread compression, while facing leverage, execution, utilization, technological obsolescence, and refinancing risks. It also says AI capex will absorb large amounts of capital, the funding gap is likely to persist, and debt supply should keep growing.
🔗 Source: Summary based on View Source from pimco.com | Found on Sep 15, 2026
Major cloud providers are expected to spend more than $1.4 trillion on AI infrastructure next year, while compute capacity could quadruple from 2025 to 2028 to about 120 gigawatts. He said Morgan Stanley sees roughly 25% to 50% ROIC across three AI business models: renting compute power, where a large next-generation data center could earn about 30% ROIC; AI labs owning both model and infrastructure, with about 75% incremental operating margin and 40%+ ROIC; and renting infrastructure, with about 30% incremental operating margin and 25% post-tax return potential.
🔗 Source: Summary based on View Source from morganstanley.com | Found on Sep 17, 2026
Artificial intelligence can improve efficiency and productivity while lowering costs, but it also increases cybercrime risk as AI agents become more capable. Unlike traditional software, AI agents can run continuously, act faster than humans, and connect information and actions across tools, systems, and data sources. Reports of agents bypassing safeguards to compromise networks have intensified security efforts. Identity security is expected to be especially affected, with the enterprise market projected to more than double from about $24 billion today to $60 billion over the next two years. Demand is also expected for tools to discover, register, govern, and monitor AI agents.
🔗 Source: Summary based on View Source from morganstanley.com | Found on Sep 17, 2026
The U.S. borrowing is again extremely large, with Treasury issuance excluding bills running near $5 trillion over the past 12 months and corporate debt issuance around $2.6 trillion. The 10-year Treasury yield is near 5%, unlike 2020, when the federal deficit reached $3.1 trillion and long-term yields stayed near historic lows. In 2020, the Federal Reserve increased its Treasury holdings by roughly $1.8 trillion in five months. J.P. Morgan estimates hyperscalers will spend about $700 billion on capital expenditures in 2026, and their bond issuance rose from $17 billion in 2024 to $109 billion in 2025 and $194 billion in the first half of this year.
🔗 Source: Summary based on View Source from doubleline.com | Found on Sep 17, 2026
AI infrastructure spending is expected to reshape fixed income markets, with total data center capital expenditures potentially reaching USD 5.5 trillion from 2026 through 2030. By late June 2026, total AI-related debt issuance had already exceeded full-year 2025 levels, according to J.P. Morgan. Funding is expected to come through five markets: investment-grade corporate bonds, securitized credit, private credit and hybrid debt, leveraged finance, and equity-linked debt. J.P. Morgan estimates leveraged finance could absorb about USD 150 billion over five years. A recent example is Hut 8’s Beacon Point, which issued about USD 4.3 billion of secured bonds to finance a Texas data center leased solely to NVIDIA.
🔗 Source: Summary based on View Source from troweprice.com | Found on Sep 18, 2026
US securitization is providing more capital for infrastructure than European markets. Since 2018, US data-center securitizations totalled $81.4 billion, including $18 billion in the first half of 2026, versus $1.7 billion in the EU and $2.3 billion in the UK. The disparity applies to solar financing: $29.3 billion in US securitizations versus $1.1 billion in the EU. The author attributes Europe’s weakness to restrictive securitization and insurance rules, which have deterred life insurers: EU insurers allocate only 0.33% of assets to securitizations, compared with about 17% in the US. Modest EU reforms may leave Europe unable to finance AI infrastructure.
🔗 Source: Summary based on View Source from apollo.com | Found on Sep 17, 2026
Apollo Partners’ Brian Weinstein and John Cortese said AI-related capital expenditures are driving a record issuance boom across public and private debt markets. They said four companies have issued $187 billion in the last year, AI-related exposure in the Bloomberg Agg has risen from 0% to 4.5%, and total issuance could reach $3 trillion to $5 trillion by 2030, with about $800 billion of room in indexes. They said spreads are widening, dispersion is rising, and investors must consider duration, convexity, factor risk, and correlations as AI financing spreads across high-grade, high yield, and loans.
🔗 Source: Summary based on View Source from apollo.com | Found on Sep 18, 2026
TypeSafe AI announced its first System One Model, Jev, after two years in stealth. The company said the model is built for fast, structured decisions that software can use directly, using a new architecture, a parallel sampler, and Reinforcement Learning for Calibrated Decisions (RLCD). Jev is available in early access and is described as achieving similar intelligence on System One tasks as existing LLMs while being 40x-200x faster, with end-to-end response times of 70ms-500ms. It is optimized for structured outputs, cannot hallucinate, and always returns calibrated probabilities and confidence scores.
🔗 Source: Summary based on View Source from typesafe.ai | Found on Sep 16, 2026
Anthropic says AI systems are increasingly automating their own development, and it proposes three measurements to track this: an automation index for AI R&D, oversight metrics for AI agents, and compute allocation. Using its automation scale, Claude was not fully autonomous for any measured AI R&D subset, and more than 90% of work was at or above “AI collaborates.” As of August 2026, about 30,000 agents were doing research and engineering work on Anthropic’s most-used internal platform. In a July 13–20 compute snapshot, about 6% of AI R&D compute and 12% of AI-driven AI R&D compute went to safety.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 18, 2026
AI labs are using their systems to help build better AI, while researchers warn that more autonomous development could amplify mistakes and make failures harder to control. OpenAI says fully autonomous recursive self-improvement is not happening today, and Anthropic says its systems cannot yet autonomously build their successors. Researchers including Michael Littman, Gabe Goodhart and Nathalie Baracaldo say current tools can write software, use tools and assist with coding and experiments, but have not shown an autonomous cycle that reliably creates more capable successors. Weco AI reported in July that its AIDE² system modified another research agent’s software framework and classified the result as “Level 1” recursive self-improvement.
🔗 Source: Summary based on View Source from ibm.com | Found on Sep 17, 2026
The article presents an 8-step generative AI customization spectrum on AWS, advising users to start with the simplest approach and escalate only when needed. Steps 1–2 use existing foundation models on Amazon Bedrock and improve prompts with system instructions, few-shot examples, and chain-of-thought reasoning. Steps 3–5 add external grounding, prompt caching, or model distillation. Steps 6–8 change model weights through fine-tuning, continued pre-training, or Amazon Nova Forge. Most workloads do not need to go beyond Step 3, and the right choice depends on accuracy, latency, data, and domain requirements.
🔗 Source: Summary based on View Source from aws.amazon.com | Found on Sep 15, 2026
Silicon Valley Power sent more than 200 demand signals to NVIDIA partner Emerald AI’s Conductor platform, which automatically shifted workloads at NVIDIA’s Eos AI factory, reducing power from 4 megawatts to 3 megawatts while high-priority inference kept running. In a separate validation by cloud provider Lambda, a five-rack, 19-node cluster on NVIDIA HGX B200 GPU Servers achieved 24% more token throughput—about 4 million to 5 million tokens per second—and 23% better performance per watt within the power budget of 16 full-power nodes. NVIDIA said the first dedicated DSX Flex commercial deployment will be a 96-megawatt Vera Rubin AI factory in Manassas, Virginia.
🔗 Source: Summary based on View Source from blogs.nvidia.com | Found on Sep 16, 2026
ATLAS data shows that in OECD countries, computer and mathematical and business and financial operations occupations lead AI use, while in non-OECD countries office and administrative support, arts, design, entertainment, sports, and media, and educational instruction and library occupations rank highest. AI adoption generally rises with income, though Brazil and the UAE exceed expectations. For manual tasks such as real-time equipment diagnostics and troubleshooting, AI use is 7% in Brazil and Germany and 4% in Japan. A Google, Google DeepMind, and MIT FutureTech study of 2,600 specialized AI models and over 600 U.S. and U.K. scientists found nearly half use AI daily and save just under seven hours a week.
🔗 Source: Summary based on View Source from blog.google | Found on Sep 16, 2026
Successful AI scaling depends less on technology than on organization, leadership and operating practices. It advises firms to prioritize high-value, cross-functional use cases rather than visible but low-value ones, and to redesign workflows before deploying AI so existing dysfunction is not automated. It also recommends running data readiness, governance and trust, capability building, and leadership alignment in parallel. Governance should be designed in from day one, with written answers on accountability, auditability and human review. Boards now want clear return, so each AI initiative should be tied to a business metric with baseline, current reading and target.
🔗 Source: Summary based on View Source from ibm.com | Found on Sep 18, 2026
Pharmaceutical organizations are moving AI from experimentation into core workflows across research, manufacturing, supply chain, and commercial functions. Novo Nordisk used a governed AI reasoning agent on Microsoft Azure to cut time-to-insight from weeks to minutes and raise quarterly idea evaluation from about 5-10 to more than 50 opportunities. Amgen, Almirall, and UCB deployed AI tools for searchable institutional knowledge and compliant assistant deployment. Heathrow Scientific cut order processing time by 20%, while Körber and PharmaGuardrails improved recipe management and task precision. Rohto Pharmaceutical reduced manual data-entry time by 50%, Astellas migrated 250 servers and 500 terabytes of data in six months, and Pierre Fabre’s platform was adopted by more than 50% of employees.
🔗 Source: Summary based on View Source from microsoft.com | Found on Sep 15, 2026
Modern search and recommendation systems are expected to return coherent, complementary result sets rather than a single best match. The article describes query fan-out, which decomposes a broad prompt into related sub-queries to improve diversity, coverage, complementarity, and coherence while staying grounded in a fixed database. In an ICML 2026 paper, “Efficient, Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion,” the authors propose a reward-to-data compilation framework called Retrieve-for-Train. It uses offline reinforcement learning to discover reward-aligned fan-outs, compiles them into supervision, and distills them into a lightweight diffusion retriever for efficient, single-pass query fan-out at inference time.
🔗 Source: Summary based on View Source from research.google | Found on Sep 16, 2026
The paper, submitted on 14 Sep 2026, reports a preliminary study of whether large language models can simulate financial users. In a controlled paper-trading environment with 120 volunteers, participants used non-redeemable virtual funds under real-time market conditions, with no real brokerage accounts, real-money positions, or real transaction records accessed. Using only information available before a prediction cutoff, the simulator predicts each participant’s next-trading-day action, traded security, and transaction quantity. Market context improves action and ticker prediction in the controlled ablation, but transaction sizing remains difficult. The models also overproduce hold actions, underpredict sell decisions, and simplify multi-security transactions.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 15, 2026
The paper introduces UNRESTSENT200K, a Bangla crisis sentiment dataset of approximately 200,000 Facebook and YouTube comments from the July-August 2024 Bangladesh uprising. It spans five event-aligned phases, from early escalation and internet blackout to regime transition and a later flood crisis, and links each comment to its parent post for context-aware evaluation. The comments were fully human-annotated by 14 native Bangla-speaking annotators with senior validation, reaching kappa 0.73, alpha 0.71, and 94.2% blind-audit agreement. Benchmarks of fine-tuned encoders, prompted LLMs, and LoRA-tuned LLMs showed that parent-post context improves performance, while temporal shift causes large drops.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 16, 2026
The paper studies univariate time series anomaly detection using a simple statistical pipeline with a small pool of sliding-window statistics, a transductive robust MAD model, and feature selection on a held-out tuning split. On TSB-AD-U, it achieves 0.529 per-series VUS-PR, exceeding the best neural entry at 0.45 and the best statistical entry at 0.44, and coming within 0.06 of the strongest pretrained foundation model. Ablations show the candidate pool matters most. The authors generate domain-specific pools with a multimodal LLM; selecting over the union of generated and hand-crafted pools lifts performance to 0.588, matching the best leaderboard entry.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 21, 2026