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.
AI tokens are an immaterial commodity whose input is cognition, with prices collapsing and no durable floor. It describes tokens as fungible, transparently priced feedstock used across industries, with prices falling from about US$30 per million tokens in early 2023 to below US$0.30 by late 2024. It says China’s daily token consumption rose from roughly 100 billion at the start of 2024 to over 140 trillion in early 2026. The article argues value is shifting to infrastructure, orchestration tools, and proprietary context, while professional services, finance, healthcare, education, and media face deflationary pressure as cognitive work becomes cheaper.
🔗 Source: Summary based on View Source from man.com | Found on Sep 25, 2026
Capital Group says AI’s capital-spending boom is driving record public and private debt issuance, led by hyperscalers. Alphabet, Amazon, Meta, Microsoft and Oracle issued $240.7 billion year-to-date through August; including SpaceX and Nvidia, total issuance reached $290.7 billion. Strong demand—from higher yields, annuity investors and foreign inflows—has kept broader US investment-grade spreads unchanged, but longer-dated hyperscaler and AI-related spreads have widened as supply grows. The authors warn that projected AI investment will exceed operating cash flow in 2027–28 and monetisation remains uncertain. They favour diversified issuers, fungible infrastructure, data-center and power suppliers, avoiding complex structures and chip debt.
🔗 Source: Summary based on View Source from capitalgroup.com | Found on Sep 24, 2026
Leaders of major AI companies have called for a slowdown in the race to build AI superintelligence, while the investment case has shifted from being long semiconductors and short software to stock selection within each sector. The AI race is moving from the smartest model to the cheapest to run, with cheaper open-source models, mainly Chinese, narrowing the gap to frontier models to six months. The AI supply-chain demand exceeds supply with order visibility to 2028. Cybersecurity is highlighted as a winner, while consumer AI adoption is being watched closely.
🔗 Source: Summary based on View Source from man.com | Found on Sep 22, 2026
AI-native services companies sell completed outcomes rather than software seats, using software, models, agents, and human operators to deliver jobs such as insurance, bookkeeping, benefits access, home maintenance, compliance, and procurement. The article argues SMBs are attractive customers because they often want the responsibility removed entirely and already spend heavily on outsourced services. Examples include Corgi, which provides startup insurance, Mindset Care, which handles SSI and SSDI applications at no upfront cost, and Casa, a home-services platform built around a digital home record, concierge layer, and managed vendor network. These businesses may initially resemble services revenue, but their marginal delivery costs can decline as software takes on more work.
🔗 Source: Summary based on View Source from wellington.com | Found on Sep 26, 2026
Blue Owl’s Erik Bissonnette and Anaplan CEO Charlie Gottdiener said AI is reshaping enterprise software, but Anaplan’s value remains its “system of decision” with auditable, deterministic outputs for CFOs. Its calculation engine provides 100% repeatable answers and can model complex, real-time variables such as supply chain disruptions, oil price shocks, and tariffs. Anaplan serves half of the Fortune 50, including Nvidia, OpenAI, Microsoft, and Google. It views AI providers as partners and acquired Syrup Tech for retail forecasting. Enterprises want AI embedded in proven platforms at predictable cost.
🔗 Source: Summary based on View Source from blueowl.com | Found on Sep 24, 2026
Artificial intelligence remains a promising investment theme, but the article says rising capital intensity, increasing leverage, and greater competition warrant caution. Following the recent Russell reconstitution, AI infrastructure beneficiaries make up about 36% of the Russell 1000 Growth Index, and hyperscalers add another 22%, for nearly 60% total AI exposure. By contrast, the S&P 600 Small Cap Index has about 5% direct AI infrastructure exposure and no hyperscalers. Outside the U.S., the MSCI Emerging Markets Index has about 36% AI infrastructure exposure, while the MSCI EAFE Value Index has almost none. The Asset Allocation Committee is overweight U.S. small caps and non-U.S. value equities.
🔗 Source: Summary based on View Source from troweprice.com | Found on Sep 23, 2026
Industries with low risk of functional displacement and clear AI tailwinds include metals and mining, software, and hyperscalers. Because AI is resource-intensive, copper mining and energy production in commodity-rich countries may benefit from rising demand, while media, entertainment, and online platforms face risks to discovery and advertising models from AI-driven search and agentic interfaces. Financial-services issuers are in between, with brokerages, asset managers, and exchanges facing partial erosion of informational advantages, though licensing, regulation, and scale remain barriers. For sovereigns, the framework focuses on structural leverage in AI supply chains, with Asian emerging markets showing the strongest indicators and some EMs exposed through power, thermal management, and data-center components.
🔗 Source: Summary based on View Source from alliancebernstein.com | Found on Sep 27, 2026
David Victor of UC San Diego and the Brookings Institution said AI data centers are driving steep U.S. electricity demand growth, approaching 5% of total power demand and about 4% to 5% of total emissions, similar to airlines. He said many data centers turn to natural-gas-fired generators because power must be obtained quickly, though some companies also seek renewables, batteries, and small modular reactors. He added that speculative interconnection requests are complicating grid planning, while data centers are also pushing for more responsive electricity use and, in his view, do not create a major water-consumption problem.
🔗 Source: Summary based on View Source from dimensional.com | Found on Sep 23, 2026
Ramp data show AI adoption broadening but expenditure remaining highly concentrated. About 10% of software-spending businesses now pay a GPU vendor, versus under 4% two years earlier, driven mainly by model serving and inference, whose share rose from 2.4% to 8.7%. Neocloud use increased modestly, while wholesale GPU capacity remained negligible. Most companies therefore consume AI as cancellable, usage-based software rather than reserving infrastructure. The top 10% of customers generate 99.5% of model-serving spend and 99% of neocloud spending; the bottom 90% contribute almost nothing. Unlike SaaS or CRM, AI costs track runtime, leaving infrastructure reliant on heavy users today.
🔗 Source: Summary based on View Source from apollo.com | Found on Sep 25, 2026
Goldman Sachs Research says S&P 500 earnings growth should decelerate, not collapse, and forecasts the index will rise to 8,700 over 12 months from 7,764 on September 21. The forward P/E has fallen from 23 times a year ago to 19 times, matching its 10-year average, while the cyclically adjusted P/E is near a record high. S&P 500 EPS are forecast to grow about 11% in both 2027 ($415) and 2028 ($460). The firm says almost half of 2026 EPS growth comes from AI investment, with hyperscaler capex at $800 billion this year, expected to reach $1.2 trillion in 2027 and $1.4 trillion in 2028.
🔗 Source: Summary based on View Source from goldmansachs.com | Found on Sep 24, 2026
Morgan Stanley’s Ariana Salvatore said AI safety scrutiny is more likely to boost compute spend and CapEx than slow it, as labs add safety-monitoring infrastructure. She sees little chance of a comprehensive AI regulation bill or sweeping open-weight rules, citing presidential opposition, weak legislative vehicles, and the need for a high-salience incident. Data-center pushback is mainly about environmental and affordability issues, not safety, and she expects AI CapEx to remain constructive this year and next. Midterms may affect sentiment, but policy direction depends more on event salience than government control. Investors should favor inference compute bottlenecks, cybersecurity leaders, and AI adopters.
🔗 Source: Summary based on View Source from morganstanley.com | Found on Sep 24, 2026
OpenAI has introduced GPT‑6 Sol and GPT‑6 Luna, lower-cost models trained with techniques used for GPT‑6 Astra. API prices are cut 50% from promotional GPT‑5.6 levels: Sol costs $2/$10 per million input/output tokens, and Luna $0.10/$0.50. OpenAI says both improve professional workflows, factuality, coding and computer use, with benchmarks positioning Sol near or above competitors at substantially lower task cost. Better prompt caching offers 90% discounts on cached input reads, with dashboards and controls designed to preserve reuse. Sol and Luna also inherit Astra’s improved communication and alignment. They are available today in ChatGPT Work, Codex and the API now.
🔗 Source: Summary based on View Source from openai.com | Found on Sep 23, 2026
Anthropic introduced Claude Opus 5.5, the first model in its Claude 5.5 family, saying it performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5. It was tested by external evaluators including Frontier Design and METR, and scored best to date on the company’s automated behavioral audit. Opus 5.5 improved performance on coding, long migrations, and knowledge work, including a 680,000-line code migration completed in under a day. It is available on all platforms, with safeguards for cybersecurity, biology, and distillation.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 23, 2026
xAI announced Grok 4.7 on September 21, 2026, describing it as its most capable model for coding and knowledge work. Compared with Grok 4.6, it uses a new larger base model, was trained with a longer reinforcement learning run on harder, hours-long tasks, and is better at self-verification, longer context, documents, presentations, and the Grok Bot harness. It improved on GDPval and AA Briefcase and is competitive with frontier models. On safety, it uses a new safeguard stack, tops LatchBio at 62.4%, and allows only 3.3% of risky dual-use prompts through on HackerBench v0.3. It is available in Cursor, Grok Build, the API, and other platforms, starting at $2/$6 per million tokens.
🔗 Source: Summary based on View Source from x.ai | Found on Sep 22, 2026
Microsoft announced a new Copilot with three capabilities: Home, Code and Autopilot. Home combines Chat and Cowork and includes Office in Copilot, bringing Word, Excel and PowerPoint into the experience. Code lets users build solutions in a sandboxed environment using the same underlying technology as GitHub Copilot. Autopilot, previously called Scout, is a persistent agent that keeps working when users are away. Home and Code will roll out in Frontier in the coming weeks, and Autopilot will expand to private preview at month end. Microsoft also introduced new FinOps for AI capabilities and said Today will enter private preview in October.
🔗 Source: Summary based on View Source from blogs.microsoft.com | Found on Sep 25, 2026
Meta VR Glasses, announced by Facebook company on September 23, 2026, will go on sale in Spring 2027 for $1,299.99 USD. The device weighs about 100 grams and uses a two-part system: the glasses handle sensors and display, while a pocket- or bag-clippable puck handles compute, battery, and storage. It features a 5K Infinite Display, Dolby Vision, Dolby Atmos, Meta AI controls, and Qualcomm’s Snapdragon Reality Elite processor. The puck battery lasts up to three hours of continuous high-resolution media playback and supports 45W fast charging. The device is IMAX Enhanced certified and will support media, gaming, work, and video calling.
🔗 Source: Summary based on View Source from about.fb.com | Found on Sep 25, 2026
Facebook announced Petal on September 21, 2026, as the first subsea cable with petabit capacity to be deployed across an ocean. The cable will connect the US and France over 7,000 km, double the capacity of today’s most advanced transoceanic cables, and use multi-core fiber technology at transoceanic distances. Petal is designed to deliver 1 petabit per second, roughly enough for 75% of the world’s population to stream music at once. Meta is developing it with NEC and Sumitomo Electric Industries, and working with Orange to land it on France’s Atlantic coast. It is expected in service in 2029.
🔗 Source: Summary based on View Source from about.fb.com | Found on Sep 22, 2026
Anthropic formed a life sciences research group and lab in spring 2026 to use Claude for fundamental biology research, including DNA database searches and lab testing. In one program, Claude searched for 21 hours with roughly 950 agents and 210 million tokens, then identified a previously uncharacterized bacteriophage enzyme system called array-associated reverse transcriptses (ART). ART is based on a reverse transcriptase and includes a partner gene and a long array of evenly spaced DNA repeats, resembling CRISPR. Early experiments found the ART array is expressed as distinct short RNAs.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 23, 2026
Palo Alto Networks announced Unit 42 Continuous Frontier AI Defense on September 22, 2026, an always-on, agentic security service powered by Anthropic’s Mythos models and OpenAI’s GPT cyber models. It uses proprietary multi-model harnesses and offensive security experts to discover vulnerabilities, validate exploitability, and accelerate remediation across applications, identities, cloud infrastructure, and network assets. Internally, continuous Mythos-based scanning produced over a year of traditional penetration testing results in three weeks, found 3.2 times more high and critical vulnerabilities per product, and cut mean time to remediate by 51%. The service is available worldwide on an annual subscription basis.
🔗 Source: Summary based on View Source from paloaltonetworks.com | Found on Sep 24, 2026
At Dreamforce in San Francisco, Salesforce said it is adapting to an AI-focused enterprise landscape by making its apps and data accessible to agents through tools such as Claude, Slack, ChatGPT, and Agentforce Coworker. CEO Marc Benioff said customers want different pricing models, including per user, per agent, consumption, usage, transaction-outcome, and business outcome pricing, and Salesforce is giving sales teams flexibility to negotiate. The company left guidance unchanged, including a fiscal 2030 revenue target of at least $63 billion. Barclays kept a buy rating and $276 price target, while William Blair said consumption or outcome-based pricing may be most popular.
🔗 Source: Summary based on View Source from salesforce.com | Found on Sep 25, 2026
IBM introduced Forward Deployed Units, or FDUs, small teams of three or four specialists working with 10-20 AI agents to build and deploy AI systems directly with clients. Gartner said only 22% of organizations had scaled AI across multiple business units or adopted an AI-first approach, while 85% of functional leaders planned to increase AI spending in 2026. IBM’s Institute for Business Value reported an average AI ROI of 17% for enterprises in 2026. IBM said an India bank data modernization project that might have needed about 300 people for 12 to 15 months was completed in less than nine months with a fraction of that team.
🔗 Source: Summary based on View Source from ibm.com | Found on Sep 23, 2026
Google Private AI Compute Team published a technical update on September 23, 2026, describing a new Private AI Compute architecture for persistent, cross-device AI memory with on-device privacy standards. The update says Google will bring private, server-side memory to the platform to preserve long-term continuity across devices while maintaining strict privacy. It describes a persistent memory layer as a secure digital vault in the cloud, with information sealed in encrypted storage and cryptographic keys held only on personal devices. An authenticated, end-to-end encrypted channel connects the device to a protected cloud secure enclave that temporarily decrypts data, handles the request, saves new context, and re-encrypts it immediately.
🔗 Source: Summary based on View Source from deepmind.google | Found on Sep 24, 2026
The article says analysts in security operations centers may paste sensitive incident data, such as internal hostnames, service account names, file paths, IP ranges, detection rules and live-incident indicators, into personal chatbots to speed up work. It warns this shadow AI can leak information to unreviewed outside services and can also lead analysts to inherit AI mistakes. It cites research that employees enter sensitive data into AI tools about once every three working days, nearly half of generative AI users use personal accounts, and heavy shadow AI use adds about $670,000 to breach costs. It argues bans fail and approved, managed AI channels work better.
🔗 Source: Summary based on View Source from ibm.com | Found on Sep 25, 2026
Microsoft announced Integrated Security Operations Center (ISOC) in Microsoft Defender on September 23, 2026, as a foundation for agentic security that combines SIEM and threat protection in one system. The company said ISOC gives people and agents a shared foundation to see, understand, and act across the environment without operating separate systems. Microsoft also said it introduced its end-to-end cyber stack and Project Perception in July 2026. ISOC is available in preview today and is intended to support continuous defense, helping practitioners investigate, automate, manage incidents, understand threats, and take action using shared context and controls.
🔗 Source: Summary based on View Source from microsoft.com | Found on Sep 24, 2026
Meta said AI glasses need cloud-based processing for stateful, personal, and advanced features such as translation, transcription, contextual search, and long-term recall. In 2026, it described Private Processing, its confidential-computing infrastructure for AI workloads, which runs models inside confidential virtual machines so even Meta cannot access user data. The system uses TEEs, remote attestation, anonymous credentials, encrypted storage, and public transparency ledgers, and it is built around five requirements: hardware isolation, fail-closed guarantees, public verifiability, non-targetability, and encrypted storage. Meta said it had introduced Private Processing for WhatsApp and the Meta AI app in 2025.
🔗 Source: Summary based on View Source from engineering.fb.com | Found on Sep 24, 2026
Physicist and science writer Matt von Hippel challenged AI companies to solve a frontier problem in theoretical particle physics: the six-particle amplitude in planar N=4 super Yang-Mills at nine loops. In late August, Anthropic’s Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude Science with Fable 5.1 to compute it, costing about $1,000–$2,000 for an end user and about $100 of that on a bootstrap run using 96 CPUs for a week. Lance Dixon independently validated the result. Soon after, Song He’s group reported they had also gotten most of the result, with some AI assistance based on GPT-6.
🔗 Source: Summary based on View Source from anthropic.com | Found on Sep 27, 2026
The article says agent evaluation has shifted from scoring single function calls to scoring full tasks executed through tool use, because agents operate across multi-step environments and can recover from failures. It distinguishes step-level scoring, which checks whether each call was valid and useful, from end-to-end scoring, which checks whether the final state shows the task was completed. It also notes that benchmarks now vary by task complexity, statefulness, and methodology, and that executable verification is preferred when available. Public suites, domain-specific evaluations, and real environment state are presented as the best way to judge whether a model can complete work.
🔗 Source: Summary based on View Source from developer.nvidia.com | Found on Sep 22, 2026
Jev AI is TypeSafe’s hosted System One API for zero-shot typed decisions, supporting up to 64k tokens per request and up to 255 choice options. Laya is an open-weight decision model that can run locally and be fine-tuned; the evaluated checkpoint used 512 tokens per question. In JevBench v1.3.0, checked September 22, 2026, 52 systems were tested on 534 typed decisions, and Jev scored 74.4 overall (#1) versus Laya’s 54.4 (#33). Jev is recommended for managed zero-shot serving, while Laya is suited to open weights, data residency, multilingual routing, or fine-tuning.
🔗 Source: Summary based on View Source from huggingface.co | Found on Sep 27, 2026
The paper “Propose, Don’t Judge: An Anytime-Valid Referee for LLM Agents That Mine Investment Factors,” submitted on 22 Sep 2026 by Bo Qu, Mingguang Chen, and Licheng Wang, studies a governed self-evolution setup in which an agent may propose investment factors while a frozen statistical referee judges them. The referee scores candidates only on post-submission market outcomes and provides a false-discovery guarantee at every stopping time. In tests with a script, bandit, and language model, the frozen referee admitted 5–11 times fewer sub-threshold factors than leaky referees under a scripted proposer, but certified true factors waited about 500 trading days.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 24, 2026
Kemal Kirtac’s paper introduces MFAST, a Market-Friction-Aware Sentiment-to-Trading framework that converts timestamped financial text into auditable and reproducible trading decisions. The framework links Refinitiv News Analytics with CRSP equity data, limits primary out-of-sample evaluation to post-release news outside disclosed foundation-model data-freshness periods, and includes a public replication arm using open financial text and public price data. The results show decoder-only language models outperform encoder baselines and dictionary sentiment in classification, calibration, return prediction, and net portfolio performance, while diagnostics show trade-offs among accuracy, latency, memory, throughput, and inference cost.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 22, 2026
Mojtaba Abdolmaleki, Stefanus Jasin, and Boyu Wang presented a paper on designing agentic AI workflow portfolios under imperfect selection and compute cost. The study examines a portfolio-and-selector approach in which multiple workflow executions are run and the final answer is chosen after observing outputs, balancing improved accuracy against compute cost and distractors. The authors derive bounds, exact formulations, linear programming relaxations, randomized rounding procedures, and performance certificates, and also develop a finite-dimensional dual and ellipsoid method for large workflow classes. On ABCD, Schema-Guided Dialogue, and HotpotQA, portfolio optimization improved held-out selector accuracy by 3.1, 7.5, and 0.9 points, respectively.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 17, 2026
TimeEvo is a proposed failure-driven self-evolution method for time series agents that use external tools. The paper reports two problems in current setups: a 21-tool expert-curated library improves some tasks but hurts others, dropping anomaly accuracy across every tested backbone, and one round of generic self-revision changes 147 answers, breaks 56, and shifts the final score by less than one point. TimeEvo clusters diagnosed failures into capability gaps, plans measurements, synthesizes evidence-only tools, and admits candidates through a paired gate. Experiments on ten time series QA tasks and three backbones show improved accuracy on every task and backbone, starting from an empty library.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 24, 2026
The paper introduces market signal injection (MSI), an attack on LLM pricing agents that changes numerical formatting, competitor ordering, or qualitative market commentary without explicit instructions. The authors evaluate nine open-weight models in simulated Bertrand duopoly and triopoly markets and three proprietary models in duopoly markets. Sentiment-based attacks produce the largest behavioral shifts, affecting other firms as well as profits and consumer surplus. Susceptibility varies across model families, and larger models are not consistently more robust. Episode-held-out probes separate baseline from attacked activations in all eleven re-evaluated model-condition pairs, with linear AUC 1.00 and MLP AUC 0.93 to 0.99.
🔗 Source: Summary based on View Source from arxiv.org | Found on Sep 17, 2026