LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-10-22168 posts

Atualizar

NVIDIA and Nokia Advance AI-RAN: T-Mobile Completes First Commercial Call, Pilots Launch Across Europe

Tópico · 英伟达AI-RANMaterial
2026-10-21 18:11 GMT+8

At the GTC conference in Berlin on October 21, 2026, US chipmaker NVIDIA announced an expanded partnership with Finnish telecom equipment vendor Nokia and several global operators to bring accelerated computing directly into the Radio Access Network (RAN). This initiative aims to transform traditional telecom networks into distributed intelligence platforms, allowing AI models and applications to run closer to users and businesses at the cell site level.

As a landmark achievement of this announcement, US operator T-Mobile confirmed it had completed the world's first AI-RAN call on its commercial network in Bellevue, Washington. The test utilized NVIDIA's Aerial RAN Computer (ARC) servers and Nokia's CUDA-accelerated anyRAN software, successfully connecting to an existing 5G Standalone core and a commercial smartphone. Meanwhile, UK-based BT, Germany's Deutsche Telekom, and Finland's Elisa announced plans for AI-RAN pilots on their live networks, with BT and Deutsche Telekom explicitly targeting field trials in 2027.

To support dense compute requirements at high-capacity sites, NVIDIA also introduced two new ARC configurations—MGX 2U and MGX 4U. Powered by RTX PRO 6000 Blackwell Server Edition GPUs and offered by partners like Dell and Supermicro, these servers are designed primarily for centralized RAN deployments. Additionally, nine operators across Europe, the Middle East, and Africa (EMEA), including CRA and Etisalat, are building or expanding NVIDIA-powered network infrastructure.

Beyond hardware deployment, NVIDIA highlighted that its open model Nemotron is being integrated into network automation software by Nokia and Cisco to build agents capable of autonomously managing fixed, radio, core, and transport networks. These moves indicate NVIDIA's attempt to extend its data center compute advantage to the telecom edge, reshaping the infrastructure form of next-generation mobile networks through combined software and hardware solutions.

Fontes:NVIDIA Newsroom

Pesquisa

Swedish Legal AI Firm Legora Raises $550M Series E at $8.8B Valuation

Tópico · Legora融资动态Material
2026-10-21 21:32 GMT+8

Legora, a legal technology company headquartered in Stockholm, Sweden, announced on October 21, 2026, that it has closed a $550 million Series E funding round at a post-money valuation of $8.8 billion. The round was led by hedge fund Atreides Management, with new investors including Baron Capital, Capital Research, Fidelity Management & Research Company, and Glade Brook Capital Partners joining existing backers. This valuation represents an increase of approximately 57% compared to the $5.55 billion valuation established during its Series D round in April.

Legora provides artificial intelligence tools for drafting and reviewing documents, serving both law firms and corporate in-house legal teams. CEO Max Junestrand stated that the capital will fuel continued expansion in North America (with offices in New York, Denver, Chicago, Houston, San Francisco, Toronto, and Mexico City), as well as across Europe and the Asia-Pacific region. The company also plans to invest more heavily in its agentic operating system, known as Legora aOS. Junestrand emphasized that AI is reshaping the role of lawyers to focus more on judgment and advice, and Legora aims to lead this shift globally.

Regarding financial performance, Legora disclosed that its annual recurring revenue (ARR) grew from $1 million to $100 million in just 18 months, and surpassed $200 million less than six months later. CFO David Eckstein noted that the company prioritizes durable growth, requiring healthy gross margins and high retention rates. While specific terms regarding preferred rights or dilution were not fully detailed in the public announcement, the participation of long-term institutional investors signals market confidence in the maturity of vertical AI applications.

Fontes:City AM

Pesquisa

Musk says ASML and other suppliers are expediting shipments to Terafab, targeting volume chip production by Q3-Q4 2028

Material
2026-10-21 23:30 GMT+8

Elon Musk, CEO of Tesla and SpaceX, stated during a live broadcast of SpaceX's Q3 2026 earnings call on October 20 that Dutch lithography giant ASML and other critical semiconductor equipment suppliers are expediting shipments to the "Terafab" project. This joint venture between Tesla and SpaceX aims to build a vertically integrated chip manufacturing facility in Gibbons Creek, Texas.

Musk indicated that the company targets volume chip production at the primary facility by the third or fourth quarter of 2028, but emphasized that this timeline is "highly aspirational" and carries "big error bars." For context, he noted that xAI and SpaceX plan to achieve 13 gigawatts of AI computing capacity by the end of 2027, including a massive training cluster of 440,000 NVIDIA "Vera Rubin" GPUs.

Terafab aims to manufacture chips equivalent to 1 terawatt (TW) of AI computing power per year, covering logic, memory, and advanced packaging. The facility will utilize Intel's 14A manufacturing process technology and ASML's extreme ultraviolet (EUV) lithography systems. While Musk claims suppliers are cooperating to accelerate deliveries, industry analysts remain skeptical about the feasibility of such massive capacity expansion and timelines, noting that total global AI-optimized data center capacity is currently estimated at roughly 44 GW.

Fontes:Tom's Hardware

Pesquisa

New York Fed Research: Home Insurance Deductibles Leave 29% of Expected Losses to Households, with Financially Constrained Families Exposed to the Greatest Risk

Material
2026-10-05 19:00 GMT+8

A long-standing issue in homeowner’s insurance is that insurers cannot fully observe whether homeowners are properly maintaining their properties, creating a problem known as moral hazard. Deductibles and coverage limits are tools insurers use to preserve homeowners’ incentives to protect their property. In research published on October 5, 2026, in Liberty Street Economics, Hyeyoon Jung, a financial research economist in the Research and Statistics Group of the Federal Reserve Bank of New York, and Jaehoon (Kyle) Jung, an assistant professor at the NYU Stern School of Business, quantified the actual distribution of costs associated with this mechanism using millions of homeowner’s insurance contracts.

Their core finding presents a paradox: the direct cost to insurers of managing moral hazard is relatively small, but contract design leaves substantial residual risk with households. Specifically, deductibles and coverage limits expose households to 29% of expected losses. This means that although households are insured, nearly 30% of expected losses must still be borne by them in the event of a disaster.

Crucially, this 29% of residual risk is not distributed evenly. Contract-level estimates show that policyholders with lower FICO credit scores face higher risk premia, higher costs associated with moral hazard, and retain more uninsured exposure. The authors note that households with limited financial resources purchase insurance partly because they value transferring property risk to insurers; however, their contracts also leave them with more residual risk. In other words, the households least able to bear disaster-related losses are those whose insurance contracts leave them most exposed to loss risk.

The authors explicitly emphasize that these patterns represent correlations rather than causal estimates. FICO scores may also reflect factors other than financial constraints. Nevertheless, this distribution pattern provides informative insights into where residual risk concentrates.

To validate the reliability of the model estimates, the authors conducted two external checks. The first test examines the logic of "skin in the game": a lower loan-to-value ratio implies that homeowners have invested more of their own equity in the property, giving them stronger incentives to maintain it and reduce avoidable losses. If the model is indeed estimating moral hazard, the estimated cost of moral hazard for households with lower loan-to-value ratios should be lower. The results align with this intuition, even after controlling for borrower and property characteristics. The second check utilizes differences in state insurance regulation: some states require insurers to conduct home inspections or verification, which reduces information asymmetry regarding how the property has been managed. The estimated costs of moral hazard were also lower in these states. The authors clarify that these checks do not establish causal effects of leverage or regulation but provide external evidence that the direction of the model estimates is consistent with the underlying economic mechanisms.

Fontes:Liberty Street Economics

Pesquisa

Dust Co-Founder Stanislas Polu: Models Will Change, but the AI Work Traces Accumulated by Enterprises Are the Enduring Assets

Tópico · Dust智能体痕迹资产Material
2026-10-21

Stanislas Polu is the Co-Founder and CTO of the AI collaboration platform Dust. In an interview with Snowflake in October 2026, he offered a key insight: Large language models have made intelligence abundant, but abundance itself creates new problems—how can enterprises preserve their unique identity amidst constantly changing sources of intelligence?

Polu points out that current enterprise adoption of AI is largely fragmented, with individuals adopting one assistant or model at a time. This creates three forms of dependency: reliance on a single model provider, loss of control over the data and operational traces generated by agents, and insufficient visibility into costs and behaviors. He emphasizes that agent work traces record how work was completed—including instructions, context, model and tool choices, outputs, and human feedback. Over time, this record may become one of an enterprise's most valuable AI assets, helping them evaluate workflows and develop systems tailored to their specific operations. Enterprises should not default to abandoning this asset.

Polu views agent work traces as assets that enterprises can understand, improve, and retain. In Dust’s design, when an enterprise switches models or providers, it can preserve the accumulated context and learning around its workflows without having to rebuild its entire operational layer.

Polu believes it is a mistake to view AI as a race with a single permanent winner. Models will continue to evolve in capability, cost, latency, and specialization. Enterprises should not be forced to predict winners or rebuild their operational layers every time the frontier shifts. In his view, the enduring asset is what the enterprise builds around intelligence—how the organization operates, what it knows, what decisions it has made, and where human judgment plays a role. This layer should belong to the enterprise itself, not to whichever model happens to be leading at the moment.

Fontes:snowflake.com

Pesquisa

CoreWeave stock sinks 29% as AI infrastructure debt fears mount

Tópico · AI基建债务风险Material
2026-10-22 05:12 GMT+8

U.S.-based AI cloud computing company CoreWeave is facing a crisis of confidence on Wall Street. As of October 21, 2026, its stock has fallen 29% this month, with market capitalization shrinking from a peak of approximately $90 billion to around $34 billion. More concerning for investors are its rising debt costs: yields on CoreWeave's unsecured bonds have spiked to about 15%, and the spread on its five-year credit default swaps (CDS) has widened by 164% over the past year, reaching nearly 1,067 basis points (10.67%).

This trend is not isolated but reflects broader challenges across the 'neocloud' sector—emerging cloud providers specializing in AI compute rental. Shares of peers Nebius and IREN dropped 27% and 32% respectively this month. In Australia, data center operator NextDC plunged 40% after slashing its annual earnings forecast. UK-based neocloud Nscale, backed by Nvidia, also faces scrutiny over its planned IPO amid market volatility.

The core of market panic lies in CoreWeave's massive debt burden and uncertain future financing. The company has approximately $20.7 billion in debt maturing through 2030, with another $14.9 billion thereafter. Rising interest rates significantly increase refinancing costs. According to CNBC reports, the high-yield bond market has effectively paused trading until after the U.S. midterm elections in two weeks, due to political pushback against data centers and permitting issues. Fixed income investors want regulatory requirements substantially resolved before committing funds.

Although CEO Michael Intrator emphasized in an internal memo that the company maintains diverse funding options—including infrastructure-backed borrowing, corporate debt, equity, and customer prepayments—analysts expect third-quarter free cash flow to deteriorate to negative $10.1 billion, a record low. Barclays credit analyst Andrew Keches noted that sentiment toward AI bonds is 'decidedly negative,' shifting rapidly day-to-day.

Notably, some institutions remain bullish on CoreWeave's fundamentals. Goldman Sachs, JPMorgan Chase, and Barclays issued positive notes this week, highlighting its GPU-backed loan structure and revenue stability from investment-grade customers like Microsoft and Meta. Goldman argues that CoreWeave ranks in the top quartile or even decile of neoclouds regarding execution capability and performance. However, the market is reassessing whether technical narratives alone can sustain valuations in capital-intensive AI buildouts; financial sustainability and accessible financing channels have become critical considerations.

Fontes:CNBC

Pesquisa

PC Watch tests 64GB RTX Spark laptop: runs 27B LLM locally, but bandwidth limits speed

Tópico · 英伟达RTX SparkVisão rápida
2026-10-22 05:04 GMT+8

Japanese tech outlet PC Watch has published detailed benchmark results for the ASUS ProArt P16 laptop, equipped with the NVIDIA RTX Spark N1X chip and 64GB of unified memory. The tests aim to evaluate the practical capabilities and limitations of running local artificial intelligence workloads on mobile hardware.

For Large Language Model (LLM) inference, the test used the Qwen3.8-27B model (Q4_K_M quantization), achieving an average generation speed of approximately 32 tokens per second. This rate supports real-time text interaction and code assistance. However, when the input context length increased to 65,536 tokens, the time to first token reached about 102 seconds, indicating a significant latency bottleneck for long-document processing.

In image and video generation tasks, the device required approximately 14 to 40 seconds to generate a 1024x1024 pixel image and about 69 seconds to produce a 5-second video using ComfyUI workflows. While these times allow creators to iterate on visual concepts on-site, they lag behind high-end desktop GPUs.

The primary cause of this performance gap is memory bandwidth. The RTX Spark N1X offers a maximum memory bandwidth of approximately 300GB/s, compared to 1,792GB/s for the NVIDIA GeForce RTX 5090 desktop GPU used as a comparison. Although RTX Spark provides larger model hosting capacity through shared CPU/GPU unified memory, its data throughput rate constrains the execution speed of compute-intensive tasks.

Additionally, the tests noted that not all 64GB of memory is available for GPU acceleration, as the operating system and applications require resources. In Windows settings, the memory allocation for the GPU was adjusted to a cap of 40GB. This underscores the need to distinguish between 'total memory capacity' and 'effective VRAM equivalent' when planning local AI deployments.

Fontes:PC Watch

Pesquisa

IBM Releases Granite Speech 5.0: Open-Source Voice Models Improve Accuracy and Speaker Attribution

Material
2026-10-21 22:58 GMT+8

IBM released the Granite Speech 5.0 series on October 21, 2026, comprising the new granite-speech-5.0-2b model and updated versions of granite-speech-5.0-470m-turboctc and granite-speech-4.1-2b-plus. This collection targets enterprise speech workloads, ranging from high-accuracy transcription and translation to ultra-fast edge processing and complex multi-speaker recordings. All models are released under the Apache 2.0 license with native support for frameworks like Transformers and vLLM.

The flagship granite-speech-5.0-2b succeeds the previous 4.1 version at the same parameter count. It demonstrates improved English short-form Automatic Speech Recognition (ASR) accuracy, with enhanced robustness against noise, reverberation, and distant-microphone speech. The model supports transcription and translation in English, French, German, Spanish, Portuguese, and Japanese, introducing language ID prompting to steer multilingual tasks via prompts. It also supports keyword biasing for up to 2,000 terms, improving recognition of product names, people, and acronyms. According to the Open ASR Leaderboard, it ranks second among top open-weight ASR models with an average English Word Error Rate (WER) of 5.06%.

For high-throughput and on-device deployment, the updated granite-speech-5.0-470m-turboctc retains its 470M parameter footprint while adding multilingual support (English, Japanese, Spanish, German, French, Italian, Portuguese) and native punctuation and capitalization. By compiling a single Conformer block and reusing its graph, inference speed increased to 19,870 RTFx (Real-Time Factor), approximately 1.6x faster than the initial launch rate of ~12,600 RTFx, maintaining the same accuracy level. This allows for higher concurrency or lower power consumption on equivalent hardware.

Targeting meetings and support calls, granite-speech-4.1-2b-plus features optimized weights for better speaker attribution and timestamps without changing the architecture or API. Data shows a 7% reduction in Word Diarization Error Rate (WDER) and a 7% drop in speaker-attributed WER (cpWER). Average Timestamp Shift (AAS) is cut by more than half, enabling tighter timing for captions and search. Long-form speech WER decreased by 24%, and errors on biased keywords fell by 45% with 20-term lists. These improvements enhance utility in long-duration, multi-speaker recording scenarios.

Fontes:huggingface.co

Pesquisa

Meta Earnings Reveal: AI App Muse Surpasses 10M Users, Shopping Orders Average Over $200

Tópico · Meta助手MuseMaterial
2026-10-22 05:14 GMT+8

US technology company Meta disclosed during its third-quarter earnings call on October 21, 2026, that its artificial intelligence application Muse has been adopted by more than 10 million users. This figure confirms estimates released earlier this week by market intelligence firm Sensor Tower.

Meta CEO Mark Zuckerberg stated that about 25% of people who set up Muse now use it daily, helping improve retention. He also shared that over a quarter of engaged users who were shopping via the AI agent had an average purchase size of over $200, a number surprising even to industry insiders. Zuckerberg noted that Plaid, the financial technology company connecting payment apps to bank accounts for Muse, told them they hadn't seen activity like this with their connector on any other service before.

Beyond shopping, Zuckerberg said other popular uses for Muse include personal productivity tasks, such as connecting the AI agent to users' email or calendars. He stressed that this indicates users trusted Muse's security setup, which includes a feature called "Muse Secure VM" designed to protect their data and privacy. He also mentioned that the forthcoming "Muse Confidential VM" will prevent even Meta from seeing users' content.

Now that Muse has achieved "product-market fit," Zuckerberg said Meta's focus will shift to scaling the AI agent, increasing revenue, and optimizing compute costs. Remarkably, the tech company is forecasting that Muse will eventually bring in more revenue per user than its "family of apps" users do today—meaning its users on Facebook, Instagram, WhatsApp, and Messenger. Currently, that figure is around "several hundreds of dollars annually," Zuckerberg said.

Meta intends to monetize Muse by keeping it free for most users but retaining a percentage of the commerce taking place on the platform. Power users wanting access to the most advanced AI models or more tokens will also have to pay. Zuckerberg noted that Meta believes Muse will be profitable after it scales. On the compute side, Meta is building out capacity to serve more than a billion people over the next few years. While Muse is very compute-intensive to run in the near term, the company expects to improve its cost efficiency by more than 10x in the next year, enabling profitability once revenue lines reach scale.

Fontes:TechCrunch

Pesquisa

IBM Beats Q3 Revenue and Earnings Estimates, Raises Full-Year Free Cash Flow Guidance

Tópico · IBM第三季度财报Material
2026-10-22 04:28 GMT+8

US technology giant IBM released its third-quarter earnings report on October 21, 2026, beating Wall Street consensus estimates for both revenue and adjusted earnings per share. The company posted revenue of $17.10 billion, up about 5% year-over-year, exceeding the LSEG consensus of $16.82 billion. Adjusted EPS came in at $2.95, surpassing the expected $2.86. Net income rose to $2.19 billion, or $2.30 per share, compared to $1.74 billion a year ago.

By segment, software was the primary growth driver, generating $7.80 billion in revenue, an 8% increase that topped analyst expectations of $7.66 billion. Infrastructure revenue reached $3.70 billion, up 4% and also above consensus. However, consulting services revenue remained flat at $5.34 billion, slightly missing the $5.37 billion estimate. Additionally, the adjusted gross margin of 58.5% fell short of the StreetAccount consensus of 59.1%, indicating some pressure on costs or product mix.

Regarding guidance, IBM raised its forecast for full-year free cash flow increase from 'approximately $1 billion' (as of July) to 'more than $1 billion,' while maintaining its constant-currency revenue growth outlook of 4% to 5% for 2026. CEO Arvind Krishna stated in a release that clients are deploying 'Super Intelligence' to transform entire operating models rather than just addressing narrow problems, reflecting deeper enterprise AI adoption.

Notably, IBM shares climbed in extended trading, but the stock has fallen 15% year-to-date as of Wednesday's close, significantly underperforming the S&P 500 index, which gained 13%. This divergence suggests that despite strong short-term results, investors remain cautious about IBM's long-term competitiveness in the traditional IT cycle.

Fontes:CNBC

Pesquisa

Rob Arnott on Hidden Costs of Passive Investing and AI Winners

Material
2026-10-22 02:11 GMT+8

This episode features Rob Arnott, founder of Research Affiliates and chairman of Syzygy Asset Management, joining host Kai Wu to delve into structural issues in passive investing. Arnott points out that owning today’s market leaders is not necessarily the same as owning tomorrow’s biggest investment winners, especially amid rising S&P 500 concentration.

The conversation analyzes how index construction rules shape what investors actually own and why index additions can create hidden implementation costs not reflected in expense ratios. Arnott shares his view on considering buying richly valued IPOs before index funds are forced to buy them, discussing price pressures surrounding index inclusion for companies like Tesla and SpaceX.

Additionally, the episode covers the shift toward capital-intensive business models in AI, including GPU depreciation, economic obsolescence risks, and intangible assets missed by conventional book value. Arnott emphasizes the importance of testing investment ideas rather than chasing impressive backtests, and explores how AI customers could capture value beyond infrastructure builders.

Fontes:podcasters.spotify.com

Pesquisa

80,000 Hours Podcast: How Frontier AI Staff Should Reassess Work Priorities Amid Recursive Self-Improvement Risks

Tópico · 前沿AI安全与对齐Material
2026-10-22 00:11 GMT+8

The 80,000 Hours podcast released an episode titled "It’s time for frontier AI staff to reassess their work: 8 strategic updates" on October 21, 2026. Host Robert Wiblin issued a plea to all AI company staff to reassess whether their current role is their best opportunity for impact, suggesting they move if it is not.

Wiblin noted that many technical staff, including Anthropic co-founder Jack Clark, now believe it will become possible to fully hand over AI R&D to AIs themselves as early as next year, or in 2028 or 2029. If employees believe recursive self-improvement (RSI) could come that soon, it likely demands taking decisive steps to ensure their role either delays RSI or makes it less likely to blow up.

The episode suggested eight ways the strategic situation appears to have shifted: chain-of-thought monitoring probably stops working without a big push; it's unproductive to patch surface indications of deep structural problems; safety and alignment evals need a credibility revolution; helping with capabilities advances looks worse than before; there's ever less need to help with purely commercial work; 'pacing the frontier' is unambiguously necessary; staff have an ever-shrinking window to organise; and there are new opportunities for impact outside frontier companies.

Wiblin emphasized the importance of 'pacing the frontier', arguing that AI capabilities are truly dangerous without solving alignment. He suggested staff consider working on efforts to slow down AI development or joining organizations like Resolution, which focus on alignment research built to keep working through recursive self-improvement.

Fontes:80000hours.org

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

The study also finds that the financial condition of insurers affects risk allocation. Insurers with stronger financial constraints, measured by their risk-based capital ratios, tend to insure riskier properties and collect higher dollar risk premia. However, the risk premium as a share of total premiums is roughly flat across insurers with different levels of financial constraint. This suggests that higher premiums do not simply reflect higher profit margins but rather that these insurers are matched with riskier properties and different types of policyholders. The authors argue that endogenous matching among households, properties, and insurers must be considered when studying the property insurance market.

This research reveals a fundamental trade-off in homeowner’s insurance: the direct cost of managing moral hazard is small, but residual risk is not uniformly distributed—households with lower credit scores and homes facing severe tail risks retain more exposure. This quantification offers more specific insight than the general explanation that deductibles exist due to moral hazard; it not only confirms the existence of the mechanism but also measures its magnitude and distribution.

The authors note that these results raise a natural policy question: if deductibles and contract terms leave risk with households, especially those with fewer resources, why not simply require insurers to provide more complete coverage? The answer is not obvious, because the same contract features that leave households exposed also help preserve incentives to maintain and protect property. In forthcoming companion research, they will use models to directly test this trade-off.

It should be noted that Liberty Street Economics is a research blog of the Federal Reserve Bank of New York. The views expressed in the post are those of the authors and do not necessarily reflect the position of the Federal Reserve Bank of New York or the Federal Reserve System. This blog post is a popularized presentation of their formal paper; technical details and full estimates can be found in the authors' paper.