LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-03100 posts

Atualizar

VC Exit Capability Depends on Specific Bankers, Not Institutional Brand: Broken Underwriter Relationships Reduce IPO Exits by 9.5%

Material
2026-11-03 00:43 GMT+8

Whether a venture capital (VC) fund generates profits is typically thought to depend on two things: the ability to select good companies and the ability to help those companies grow. Tuft and Yimfor, in their 2026 study published in the Review of Finance, propose a third factor: the ability to take companies public. More specifically, this depends on the relationships maintained by specific individuals between VCs and underwriters.

The core finding of the study is that when an existing relationship between a VC and an investment bank underwriter is severed due to a merger or closure, the VC’s IPO exit rate drops by approximately 9.5%, and fund return multiples decline by about 7.8%. These figures come from a difference-in-differences analysis comparing the performance of affected VCs with those retaining their underwriter relationships before and after the shock.

The key to the identification strategy lies in exogeneity. Tuft and Yimfor use underwriter mergers and closures as exogenous shocks to existing business relationships; these events are driven by consolidation activities within the banking industry itself, not by choices made by VCs. The authors compare affected VCs with those retaining their underwriting connections, observing changes in exit performance before and after the event, thereby providing causal evidence rather than merely correlational data.

What truly changes the understanding is the mechanism analysis. The authors find that when mergers lead to high employee turnover, the damage to a VC’s exit capability from broken relationships is more severe; when the original team remains post-merger, the damage is smaller. This implies that the valuable asset is not the institutional label itself, but the trust, familiarity, and tacit knowledge accumulated through repeated cooperation with specific individuals. The authors also examined alternative explanations such as underwriter industry specialization, finding that these factors could not fully explain the results; the evidence points to relationship-specific human capital as a critical determinant of access to public markets.

This vulnerability is not uniformly distributed. Smaller VCs and funds nearing their exit phase suffer the most significant shocks. Large institutions possess broader networks, stronger reputations, and more alternatives; smaller VCs may struggle to find new trusted partners, while managers near the end of a fund’s life lack the time to wait for relationship rebuilding. Additionally, the authors note that underwriters intervene in portfolio companies as deal facilitators or late-stage investors in approximately 15% of cases before the IPO, indicating that the role of relationships extends far beyond the moment of listing.

The practical implication for VC due diligence is that historical returns may not fully reflect the importance or fragility of exit infrastructure. A manager’s strong performance may partly rely on relationships with a specific underwriter maintained by particular bankers. If these bankers leave or their institution is merged, future exit capabilities may decline significantly. When evaluating managers, dimensions worth investigating include the concentration of underwriter relationships, personnel stability, and the manager’s track record in handling changes among key service providers.

Fontes:alphaarchitect.com

Pesquisa

Google warns of 'LLMjacking': Hackers steal AI keys to rack up $100k cloud bills in minutes

Tópico · LLM劫持攻击Material
2026-11-03 23:25 GMT+8

Cybercriminals are stealing AI API keys and cloud credentials to hijack expensive large language model (LLM) services, according to a new report from Google’s Threat Intelligence Group (GTIG). The practice, termed "LLMjacking," involves attackers using stolen access to run AI workloads, create proxy services, and rack up massive bills for victims.

Unlike cryptojacking, where mining is relatively slow, LLMjacking costs can escalate extremely quickly. GTIG noted that one threat actor, identified as UNC6993, generated additional API keys within victim cloud projects after acquiring initial access, "quickly accumulating over $100,000 in billing charges." Another actor subscribed victim accounts to Pay-As-You-Go agreements and launched high-volume inference streams, bypassing rate limits by continuously creating and deleting short-lived service accounts.

The report highlights an emerging division of labor in this cybercrime ecosystem. While some groups focus on establishing initial access through exposed endpoints or stolen identity tokens, others known as "LLM Mixers" manage downstream monetization infrastructure. These operators act as routing gateways, proxying, aggregating, and reselling unauthorized LLM access to underground subscribers.

Google urges organizations to increase visibility into their LLM platforms, prioritize detection of modifications to LLM accounts, and enforce the principle of least privilege to mitigate these risks.

Fontes:TechRadar

Pesquisa

a16z: AI Makes Health Q&A Cheaper, Value Will Flow to Companies That Can Schedule Exams and Prescriptions

Tópico · a16z健康执行层报告Material
2026-11-03 23:06 GMT+8

In the 'Healthcare Home' section of a16z's longevity track report, a core judgment is presented: AI is making the interpretation of health information cheap, while true commercial value will migrate to companies that can execute actions. The report states: 'As AI makes answers and interpretation cheaper, value will migrate to the companies that can figure out the action layer.'

The logic chain behind this judgment is as follows: In the past, you needed a doctor to understand lab results; now, ChatGPT can explain your metrics in plain English. But when interpretation is no longer the bottleneck, the real difficulty shifts to execution—scheduling the next test, booking a doctor's appointment, ensuring the patient receives prescriptions, and completing follow-up care. The report argues that asking an AI a health question is useful, but turning that question into action is much harder.

Around this execution layer, a16z depicts the endgame form of a 'Healthcare Home': a single app that knows your entire medical history and exam records, automatically schedules annual labs and imaging, connects to your wearable devices, uses AI to help you understand what is happening in your body, and directly connects you with online or offline doctors for prescriptions and surgical referrals when treatment is needed. Currently, various players start from different baselines—Function Health has longitudinal lab and imaging data, Oura and WHOOP are already embedded in daily habits, General Medicine and Counsel start from clinical capabilities, Ferry Health understands fragmented care paths under insurance coverage, while OpenAI, Anthropic, Amazon, and Apple stand on existing massive user bases.

The report argues that the complexity of the healthcare system plays a dual role in this track: it makes building a 'Healthcare Home' difficult, but once built, it is also hard to replace. a16z expects multiple large companies to emerge in this field rather than a winner-takes-all outcome, because 'Almost everyone could use this kind of help.'

Survey data supports this argument. 54% of U.S. consumers are already using wearable devices, and 83% of users check their data daily, yet 55% of users say this data has never prompted them to change behavior or visit a doctor. Information is available, but action has not followed—this is the specific manifestation of the execution layer gap.

It should be noted that this is a venture capital firm's outlook on the track. Authors Daisy Wolf and Eva Steinman explicitly state at the end of the piece, 'If you're building one, we'd love to hear from you,' indicating they are actively seeking investment targets. Their survey data comes from a self-collected sample of 500 U.S. consumers, and the conclusions align with the institution's interests. The report itself includes a disclaimer stating that the content does not constitute investment advice.

For readers, the actual incremental value of this report lies in advancing the 'AI + Healthcare' discussion from 'who can answer health questions' to 'who can turn questions into action.' If you are evaluating companies in this track, a useful screening criterion is: Is it providing information (interpretation, advice, education) or executing actions (appointments, ordering tests, follow-ups, prescription fulfillment)? The former will be rapidly eroded by general AI assistants, while the latter, requiring integration with real-world healthcare infrastructure, presents more durable barriers.

Fontes:a16z.news

Pesquisa

OpenAI Hires Former US Army Secretary Dan Driscoll to Pursue Defense Contracts and Shape Military AI Rules

Material
2026-11-03 23:03 GMT+8

OpenAI has hired Dan Driscoll, the former civilian head of the US Army, as its new "Senior Counselor for National Security Modernization." According to The Wall Street Journal, the move is read as a clear signal that OpenAI wants closer ties with the military, a bigger share of government contracts, and a seat at the table when rules for AI use in the armed forces and intelligence agencies get written.

As the 26th Secretary of the Army, Driscoll oversaw nearly one million soldiers and more than 265,000 civilian employees. During his tenure, he drove AI modernization across the service branch, including through Project ARIA, an initiative that brought AI tools to budgeting, logistics, and tactical model deployment. He submitted his resignation in late August 2026, reportedly after months of friction with Defense Secretary Pete Hegseth.

Beyond his Pentagon connections, Driscoll could open doors on Capitol Hill. A close ally of Vice President JD Vance, he maintains strong relationships with both parties in Congress. That bipartisan network could prove especially useful if the balance of power shifts after the current midterm elections.

The hire builds on OpenAI's push into defense following rival Anthropic's ban. In February 2026, the Pentagon classified Anthropic as a "supply chain risk" after CEO Dario Amodei refused to drop restrictions on mass surveillance and autonomous weapons. That same day, OpenAI announced its own deal with the Pentagon to deploy its models on classified networks.

Fontes:The Decoder

Pesquisa

Apple CEO Cites 'Unprecedented' Memory Costs as Manufacturing Bill Jumps $33 Billion

Tópico · AI内存涨价冲击Material
2026-11-03 22:28 GMT+8

Apple CEO John Ternus described current memory price increases as "unprecedented" during the company's November 2 earnings call. According to Tom's Hardware, while Apple has mitigated some impact by using lower-priced inventory, cutting other costs, and raising prices on select products, its manufacturing bill has climbed by $33 billion in just a few months. This sharp rise highlights how artificial intelligence infrastructure demand is tightening global DRAM and NAND supply.

Major suppliers including Samsung, SK Hynix, and Micron have warned that memory shortages could persist until 2028 as AI data centers absorb production capacity. For Apple, this signifies a fundamental shift in the cost structure of its hardware lineup, encompassing iPhones, Macs, and iPads. Although no new broad price hikes have been announced, Ternus expects memory costs to rise further in the December quarter, potentially forcing the company to absorb higher component expenses or adjust pricing strategies during the critical holiday shopping season.

Analyst Dave Nicholson noted that unlike tech companies pouring hundreds of billions into AI infrastructure, Apple faces a more quantifiable challenge: the direct impact of rising component costs on margins is easier to measure than the long-term returns on massive AI capital expenditures. However, this measurable pressure is eroding profitability. If memory prices do not stabilize quickly, Apple may need to implement further price increases to maintain gross margins, impacting consumer affordability and its competitive position in the premium market.

Fontes:Tom's Hardware

Pesquisa

Barry Ritholtz: Why Polls Keep Failing — The Three-Layer Gap Between Attitude, Intention, and Behavior

Material
2026-11-03 20:30 GMT+8

In an article published on November 3, 2026, Barry Ritholtz presents a core judgment regarding prediction failure: pollsters mistakenly believe their core work is data analytics, but history suggests it is more akin to the work psychologists do when trying to predict future behavior. Failing to discern this distinction is why polling track records have fluctuated from decent (2024) to abysmal (2016, 2020). He states at the outset that his motivation for writing is that this topic fascinates him because it closely tracks prediction errors in markets and macroeconomic forecasts.

Ritholtz breaks down respondent error into three layers. The first is attitude: people do not truly know what they believe. Many individuals create unstable answers on the spot for policy issues they have never considered; these are known as "non-attitudes." Others use polls to send messages rather than report true preferences—what he calls expressive answering. Ambivalent voters who like one candidate’s personality but another’s policies give squishy answers that they themselves may later revise.

The second layer is intention: people declare with high confidence that they will do something, yet the cost of making that statement is zero. Ritholtz writes that we tend to say what we want to believe or what we think the listener wants to hear. Expressing support costs the respondent nothing; their commitments may be conditional or aspirational. Additionally, there is the bandwagon effect: some state their preference based on who they perceive will win, because people do not want to back a loser.

The third layer is behavior, which is the largest gap: self-reported voting intention significantly overstates actual turnout. Ritholtz cites examples such as the 2022 U.S. midterm elections, where a "red wave" failed to materialize, while 2024 saw the reverse. Low-propensity voters who show up only for a specific candidate are precisely the people likely-voter models often exclude, and the results of model adjustments have been decidedly mixed. More fundamentally, between answering and voting, factors such as weather, breaking news, geopolitics, the economy, and inflation can intervene.

Ritholtz explicitly defines the boundaries of his discussion in the article: he does not address whether pollsters reached a representative sample (he believes they did not) nor whether the questions themselves were biased (he believes it varies by situation). His argument is that the better pollsters understand respondent psychology, the more accurate their numbers will be. At the end of the article, he quotes the classic saying—"all models are wrong, but some are useful"—and points out that models assume the future looks like the past, an assumption that is particularly problematic during periods of disruption.

For investors, the transfer value of this framework lies in the ability to determine which layer—attitude, intention, or behavior—a given piece of stated data belongs to when faced with any data based on statements. Ritholtz himself only briefly noted the parallel relationship between polling errors and those in market and macroeconomic forecasts with a single sentence, without elaborating on specific investment applications in this article. Voting and investment decisions face different incentive structures—voters express opinions at zero cost, while investors involve real financial risk—so this analogy is heuristic rather than causal.

Fontes:The Big Picture

Pesquisa

Morgan Stanley Raises Palantir Price Target to $233, Sees 65% Revenue Growth in 2027

Tópico · Palantir估值争议Material
2026-11-03 17:35 GMT+8

U.S. investment bank Morgan Stanley released a research report on November 3, 2026, raising its 12-month stock price target for AI software company Palantir Technologies Inc. (NASDAQ: PLTR) from $205 to $233, while maintaining an "Equalweight" rating. The adjustment follows Palantir's strong third-quarter earnings report, with the bank citing accelerating revenue growth, expanding margins, and rising demand for Sovereign AI solutions as key drivers.

According to the report, Palantir's Q3 revenue grew 96% year-over-year, and operating margins expanded by nearly 1,100 basis points to 61.7%. U.S. commercial revenue increased by 148%, while U.S. government revenue growth accelerated to 96%. Morgan Stanley analysts suggested that Sovereign AI could be another potential inflection point for the company, as the technology allows customers to train AI models using their own operational data and ontology. Additionally, the firm revised its future financial outlook upward: it increased the 2027 revenue growth forecast from 45% to 65% and raised the operating margin estimate for that year from 61% to 63.5%.

Despite analyst optimism regarding fundamentals, market reactions indicated continued investor concern over high valuations. Following the news, Palantir shares dropped 6% in pre-market trading, extending weekly losses of 10%. This divergence suggests that while institutions are confident in long-term growth prospects and new opportunities in Sovereign AI, current price levels may have already reflected some expectations, or tolerance for valuation premiums is decreasing.

Fontes:Finbold

Pesquisa

Warcraft Logs Changes API and Introduces Paid Commercial Keys Amid Flood of AI-Generated Addons

Tópico · 魔兽日志API调整Visão rápida
2026-11-03 22:00 GMT+8

According to 404 Media, as the release of 'World of Warcraft: Forever' approaches, the game community has been overrun with AI-generated addons. These plugins, often created through 'vibe-coding,' are frequently low-quality and rip off the work of long-time developers.

In response, the prominent combat data analysis platform Warcraft Logs is changing its API strategy. CEO Sildri Silentcrest stated that while the API was previously used mostly by hobbyists, they now receive at least one commercial request per day, many of which commercialize without approval.

Silentcrest noted that the new API changes aim to better support applications that help players dive deeper into their data (such as death analysis tools), while cracking down on 'scrape and show basic statistics' apps that pull large amounts of data from the site to display surface-level info elsewhere. Additionally, Warcraft Logs is introducing a paid commercial key so that developers with commercial ambitions can support the foundation for their projects.

This shift highlights the governance challenges posed by AI-assisted programming in specific vertical domains. Veteran addon creator plusmouse reported that AI-generated addons often copy unique designs (like nameplate styles) with high code similarity, forcing them to file copyright complaints with Curseforge. Some developers have made their GitHub repositories private to prevent LLMs from using their code.

Fontes:404 Media

Pesquisa

OpenAI Introduces Pando: A Git Backend Rebuilt for AI Agent Traffic, Cutting Push Latency by ~80%

Material
2026-11-02 23:30 GMT+8

On November 2, 2026, OpenAI announced the launch of Pando, a new Git server backend now powering all development in its internal monorepo. The system was designed to address the infrastructure bottlenecks caused by the massive parallelization of work enabled by AI coding agents like Codex. According to OpenAI's engineering team, Pando has reduced median push latency by nearly 80% and currently delivers close to 1 petabyte of Git data per day.

Traditional Git servers typically enforce repository-wide consistency and generate unique packfiles for each client, which becomes a bottleneck under thousands of concurrent pushes. Pando abandons these assumptions: it employs 'causal consistency' instead of global snapshots, allowing unrelated branches to update independently and eliminating write contention across the repository. Additionally, it stores each pushed packfile as an immutable object and reuses these files for subsequent fetch requests, significantly reducing server-side computation.

To accommodate high-frequency interactions from AI agents, Pando also optimizes ref advertisement. In repositories with millions of branches, a full ref list can exceed 100 MiB. Pando advertises only the trunk branch as a base for new feature branches, enabling clients to reuse packs based on shared trunk history. While this approach causes repeat or stacked branch pushes to resend non-trunk history (with p99 pack sizes under 3 MiB), it substantially improves overall throughput.

Pando was designed by a single engineer and implemented by teams of Codex agents over 10 weeks, written in Rust. It has been serving production traffic since April 2026. Although Pando itself is not open source, OpenAI states that it has contributed several client-side improvements (such as parallel pack indexing and explicit haves support) back to the upstream Git project and is exploring ways to share the technology more broadly.

Fontes:OpenAI

Pesquisa

Apple Confirms Siri AI Cloud Features Will Have Usage Limits, Paid Access Planned

Material
2026-11-03 22:11 GMT+8

Apple CFO Kevan Parekh stated during the Q4 2026 earnings call that as Siri AI increasingly relies on Private Cloud Compute for complex tasks, server-side inference costs have become a primary consideration. Parekh noted that while on-device execution is highly efficient with minimal additional cost, cloud inference incurs significant expenses.

To address this cost structure, Apple plans to implement daily usage limits for Siri AI's cloud features. According to previous Apple documentation, lightweight tasks will remain free for all users, but advanced functions requiring AFM 3 Cloud models (such as intelligent image editing and Image Playground) will count towards these limits. Users exceeding the limits will need to pay for expanded access through iCloud+ subscription tiers in the future.

Specific pricing structures and limit thresholds have not yet been announced. Parekh emphasized that since Siri AI was only released in mid-September with iOS 27, Apple is still in the early stages of evaluating the costs of its hybrid computing model. This strategic shift indicates Apple is transitioning its core AI assistant from a hardware selling point into a service product with recurring revenue potential.

Fontes:TechRadar

Pesquisa

Google Releases Gemini 4 Carbon, Topping Intelligence and Low-Hallucination Benchmarks

Material
2026-11-03 21:00 GMT+8

Last week, Google released its latest AI model, Gemini 4 Carbon. According to data from third-party evaluator Artificial Analysis, the model ranked first in the latest Intelligence Index V5 benchmark suite with a score of 59, surpassing Anthropic's Claude Opus 5.5 by one point.

Beyond general intelligence, Gemini 4 Carbon also topped Artificial Analysis' new AA-Omniscience v1.1 test with a score of 54. This test specifically tracks model hallucination rates and knowledge recall. Although this score remains far from perfect, indicating that even frontier models have limitations in information accuracy, it represents a significant reduction in hallucinations compared to competitors like Claude Opus 5.5 (which scored 48).

In terms of cost, Gemini 4 Carbon has an operating cost of $7.76 per task, lower than Claude Opus 5.5's $10.04, while using 29% fewer output tokens. This 'cost-performance' strategy aims to meet the multiple needs of enterprises and government agencies for AI security, reliability, and cost control.

Currently, Google is prioritizing access to the model for cybersecurity defenders through its Fairwind program, with wider availability coming later. Google emphasized that the model underwent safety testing in collaboration with the U.S. government during the pre-release phase and performed well in preventing indirect prompt injection attacks.

Fontes:Tom's Hardware

Pesquisa

OnlyCFO Author: AI May Redefine ARR and Cost Presentation; Annual Plans May Not Hold for a Full Year

Tópico · OnlyCFOMaterial
2026-11-03 20:39 GMT+8

The core function of annual planning is not merely the three financial statements, but rather establishing strategic consensus among the company’s internal teams and its board regarding the most important priorities for the coming year. The author of the OnlyCFO newsletter raised this point on November 3, 2026, noting that due to continuous disruption and evolving impacts brought by AI, the importance of such alignment has never been greater. His specific questions were: how the financial framework presented by CFOs to boards needs to adjust as AI changes revenue definitions and cost structures, and how long an annual plan itself can realistically remain valid.

At the level of revenue metrics, the author believes that changes driven by AI may require updating how ARR is presented. He assesses that if a company’s AI products are entirely based on consumption-based pricing, and future growth bets are placed there, then this revenue stream must be separated out because gross margin profiles associated with different pricing models can vary significantly. He recommends splitting revenue streams by pricing model type and notes that if the definition of ARR changes, this must be clearly marked in the appendix. The author’s exact words were that the ARR definition will likely need to change or expand, rather than asserting that all companies must do so.

On the cost side, the author argues that AI spend is currently or will very soon become the single largest expense for a company. He suggests breaking it down for the board into two categories: the COGS portion affecting gross margins, and the Opex portion dispersed across Sales & Marketing, R&D, and G&A. He mentions that R&D is typically the largest category of AI operating expenses, including costs for AI tools, usage volumes, and personnel building AI products; AI costs related to free trials and proof-of-concepts (POCs) can grow rapidly. He also suggests showing total Claude and ChatGPT usage by department as a metric for AI penetration, but emphasizes that this must be immediately followed by an explanation of how much time and resources these usages actually save, rather than just reporting volume.

These changes point to the author’s core judgment: AI is changing too fast, and strictly adhering to a single annual plan for 12 months may no longer be realistic. He believes every company must be willing to completely tear up the plan when underlying assumptions are broken. Meanwhile, the plan should set guardrails—how costs should be adjusted if actual performance deviates from the plan—rather than blindly executing it. He acknowledges that every financial plan is wrong on day one; the difference lies only in how wrong it is.

The author also mentions several operational details supporting this judgment. Board members sit on multiple boards simultaneously and do not live inside the financial data like the executive team does; therefore, reporting should start with goals, first stating what is to be achieved over the next 12 to 18 months, and then describing tailwinds and headwinds. Boards truly care about the handful of key assumptions driving more than 80% of the model, not hundreds of minor ones. In headcount planning, if more than 70% of hiring is concentrated in the first half of the year, the board will get nervous—because if sales underperform, the company’s ability to adjust cash burn is very limited. These details indicate that when AI simultaneously changes revenue definitions and cost structures, the board needs to understand not a static number, but a set of assumptions that may become invalid within the year.

Fontes:onlycfo.io

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

The strongest counterargument to this study rests on the parallel trends assumption. Difference-in-differences methods require that, had no underwriter mergers or closures occurred, the exit trends of affected and unaffected VCs would have remained consistent. If VCs that happened to partner with a merging underwriter were systematically different in quality or scale, the estimated effect might mix in selection bias. The authors use mergers and closures as 'plausibly exogenous' shocks to construct identification, but this assumption itself cannot be fully verified.

Nevertheless, this study establishes a previously underestimated fact: a VC’s exit capability is a form of relational capital embedded in specific individuals, which can be destroyed by exogenous events, and whose loss directly reflects in LP returns. For institutions and individuals allocating private market capital, questions such as 'who does this manager know, how deep are these relationships, and are the key people still there,' are just as important to due diligence as 'what projects did this manager select.'

Of course, this judgment has a clear failure condition: If OpenAI or Apple directly integrates the ability to schedule exams and prescribe medications within their general assistants, and users do not need to download a dedicated 'Healthcare Home' app, then the window for specialized companies will be significantly compressed. The report itself acknowledges these giants as competitors but believes they are 'unlikely to prioritize' healthcare as a vertical scenario. Whether this assumption holds depends on future product priority choices by general AI platforms.

It should be noted that the author did not disclose his company’s specific size, industry, or AI revenue share. The premise of his described practices is that AI product revenue already constitutes a significant part of the company’s revenue structure, and AI spend is large enough to require separate breakdowns for the board. He himself mentioned that this calendar generally applies to larger, more predictable companies; smaller companies or those with lower predictability may need to compress timelines or revise multiple times. For traditional enterprises where AI investment is still in the experimental stage and contributes minimally to revenue, the applicability of these recommendations needs to be discounted.

The author also mentions an operational aspect: topics such as AI gross margins, total AI spend, and where funds are directed can all trigger strong disagreements on the board. He describes his own practice: if there have been past disputes or potential sensitivities in the plan, he communicates separately with board members before the formal meeting to gauge their positions; however, he explicitly states he does not always do this. This is not a universal recommendation, but a conditional practice based on the author’s personal experience.