LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-10-22168 publicaciones

Actualizar

Symantec Discovers 'LLMjacking': AI Agent Autonomously Hijacks Unprotected Models for Cyberattacks

Tema · LLM劫持攻击Material
2026-10-22 19:30 GMT+8

Security firm Symantec has discovered a new attack method dubbed "LLMjacking," in which malicious AI agents hijack unprotected large language models (LLMs) on the internet and use them to launch cyberattacks autonomously. According to Symantec's analysis, the agent can execute up to 100,000 steps without human involvement, while the owners of the hijacked models are left stuck with the compute bill for these attacks.

The discovery stems from a Symantec honeypot system designed to mimic open AI services and lure attackers. Researchers intercepted the full working prompt of the agent in summer 2026, revealing its architecture. The attack process involves scanning for more targets, exploiting flaws in common server software, stealing databases and credentials, and installing cryptominers in the background. To bypass safety rules in commercial models, attackers employed a five-layer jailbreak technique, including assigning fake identities to the model, redefining malicious actions as "normal developer work," and suppressing the model's safety warnings.

Symantec notes that this marks a shift in how attackers use AI: rather than merely manipulating public chatbots, AI is now built into the core of offensive tools, actively seeking out models to power attacks. While individual techniques are not new, combining them into a self-directed agent makes large-scale attacks much easier to execute. Organizations running AI services should assume that unprotected endpoints will be found and probed shortly after going live. It remains unclear whether this jailbreak method fully defeats current commercial models' defenses, but the case confirms the potential threat of autonomous AI agents in the real world.

Fuentes:The Decoder

Investigación

SpaceX and Tesla Consider Merger; Conglomerate Discount May Weigh on Overall Valuation

Tema · 特斯拉SpaceX合并Material
2026-10-22 17:28 GMT+8

The potential merger between SpaceX and Tesla is currently in the formal consideration phase by both parties and has not yet resulted in a deal. However, this possibility raises a specific valuation question: if two companies spanning aerospace, electric vehicles, AI, robotics, and other fields were to merge into a single entity, how would investors price it?

First, consider the breadth of the combined entity's businesses. On the SpaceX side, this includes space launch services for Falcon 9 and Dragon, Starlink broadband internet, the Starshield government satellite system, Starlink mobile carrier operations, and Musk's grandest ambition of colonizing Mars. The Tesla side covers electric vehicle product lines such as the Model 3, Model Y, Cybertruck, and Roadster; Robotaxi autonomous driving; Optimus humanoid robots; Powerwall and Megapack energy storage systems; frontier AI models (Grok); AI infrastructure (Colossus); AI coding tools (Cursor); social media (X); and payment applications (X Money).

When a company operates across so many industries, the market does not always grant a trust premium. There is a specific term in finance for this phenomenon: the conglomerate discount. This means that a diversified company's stock price is lower than the sum of the valuations of its individual business segments. Investors fear the company may be too complex to manage, or that consolidated financial statements could obscure true operating conditions, or that strong business units might subsidize weaker ones.

A widely cited 1995 academic paper found that from 1986 to 1991, diversified companies in the sample exhibited a discount of approximately 13% to 15%. However, a 2024 McKinsey report offered a correction to this conclusion: the discount does not occur automatically but is linked to financial transparency and business performance. In other words, if a diversified company demonstrates strong performance across its segments and provides ample disclosure, the market may not apply a discount.

Returning to the specific scenario of SpaceX and Tesla. Imagine an investor who is bullish on Starlink's prospects but skeptical about humanoid robots. In a merged entity, expensive robotics projects could consume capital that was intended for Starlink; conversely, cash flow generated by Starlink might be used to subsidize underperforming divisions. This is precisely why large, sprawling conglomerates are often less favored by the market than pure-play companies.

Business Insider notes at the end of the article that a deal ultimately comes down to selling a vision to investors, and arguably no one is better at doing that than Elon Musk. He now needs to prove that SpaceX and Tesla were meant to be together from the start.

Fuentes:Business Insider

Investigación

Independent Tech Analyst Richard Windsor: AI New Cloud Valuation Anchored to Unbuilt Data Centers; Confidence Reversal Can Become a Vicious Cycle

Tema · AI基建融资缺口Material
2026-10-16 16:11 GMT+8

Richard Windsor, founder of independent tech research firm Radio Free Mobile, put forward a core judgment regarding AI infrastructure financing: in financial markets, confidence itself is a mechanism. It does not matter whether Jensen Huang’s comments about CoreWeave are correct or not; what matters is that by making them, he may have restored enough confidence to allow those expected outcomes to come true.

Windsor traced the specific chain of wobbles in confidence. First, Anthropic: the company held a pre-IPO investor meeting at which it appeared that Q3 2026 revenues were lighter than expected. Going into the anticipated IPO, expectations were that Anthropic would increase revenues by 10x in 2026 such that the annual run rate by the end of the year would be over $100 billion. However, it looks like revenue in Q3 2026 was around $16.8 billion, which, taking into account the growth so far in 2026, leads one to estimate that the ARR for Q3 2026 is closer to $73 billion. This implies that growth is slowing more quickly than expected, meaning that everyone is starting to wonder about demand.

Second, the United States government: its unwillingness to cut the fiscal deficit is causing the yield on U.S. Treasury bonds to rise. When governments spend more than they earn, it tends to be inflationary, which means that savers become more concerned with the declining purchasing power of fiat currency from inflation and demand a higher return to compensate. Government bonds are the foundation from which all other costs of capital are based, meaning that the cost to issue bonds and the return that investors demand from equity all rise.

Windsor argues that the neo-clouds are hit the hardest because their valuations remain predicted on data centers that they have not yet built or paid for. Although they all have massive order books, once market confidence wobbles, financing costs rise, repayment capacity drops, and confidence deteriorates further, forming a vicious cycle. He believes the end result could be a collapse much like Lehman Brothers in 2008.

Jensen Huang’s intervention—Nvidia owns 10% of CoreWeave and has agreed to purchase any unsold capacity—with the statement “CoreWeave’s strength is improving every day,” temporarily restored some confidence. But Windsor makes it clear whether this is enough remains to be seen.

Windsor points out that the current economic environment for selling AI compute capacity is superb, with returns of 50% to 85% per year on offer, but you can only earn this if you have capacity already built to sell. However, customer concentration is still very high, with OpenAI and Anthropic being responsible for most of the deals being struck, meaning that if something goes wrong with one of them, it could trigger a widespread collapse. The upcoming Q3 2026 reporting season will be a key test.

Windsor disclosed at the end of the article that he holds a position in Meta Platforms, believing its valuation is relatively low, with a target price of around $770. He expressed his hope that Meta will announce through its newly formed enterprise business unit that it will start selling compute capacity to the market.

Fuentes:radiofreemobile.com

Investigación

European Commission Seems Determined to Undermine EU-Inc Initiative; Founders and Investors Warn European Entrepreneurs Will Lose Confidence

Material
2026-10-22 18:30 GMT+8

Eileen O’Mara, Vice-Chair of fintech firm Stripe, notes that many people perceive the European Union (EU) as a single market of 27 member states where one can form a company and operate freely anywhere. This is not the case.

She cites data showing that revenues of AI companies in the United States are 30 times larger than their European equivalents. Europe has only 20% non-bank capital compared to over 60% in the US. This divergence becomes apparent starting at the Series A funding stage.

In March of this year, the European Commission proposed the EU-Inc scheme, with the core objective of allowing entrepreneurs to register a limited liability company online for operation across all member states, aiming to complete the process within 48 hours.

O’Mara expressed concern that the proposal might be diluted into ineffectiveness during negotiations among individual countries.

Last week, a group of founders and investors published an open letter titled 'Dear Europe, why are you killing EU-Inc?'. They stated that compromises have rendered the proposal toothless, leaving it with merely the function of fast incorporation. The letter calls on political leaders to intervene to rescue it.

O’Mara pointed out that AI companies are globalizing faster than any previous cohort in the SaaS sector—the median AI company in the Forbes 50 list entered 59 countries within one year. The open letter warns: If European leaders cannot even reach consensus on a proposal like EU-Inc, European entrepreneurs will lose not just a proposal, but confidence in Europe’s ability to meet the challenges of our time.

Fuentes:diginomica

Investigación

OpenAI Claims GPT-5.6 Sol Helped Uncover Critical PHP Flaws Affecting 70% of Websites

Material
2026-10-22 18:30 GMT+8

According to a report by Tom's Hardware, OpenAI announced on October 21, 2026, that researchers participating in its "Patch the Planet" cybersecurity initiative used the GPT-5.6 Sol model to identify critical security vulnerabilities in PHP, an open-source programming language powering approximately 70% of the internet.

The vulnerabilities, tracked as CVE-2026-109618 and CVE-2026-109619, involve flaws in PHP's memory management and error handling. OpenAI claims that while human developers had been aware of the underlying code issues since 2024, they were unable to determine an exploitation method or a fix until AI-assisted researchers demonstrated practical exploits within two weeks.

The PHP Foundation has released patched versions for PHP 8.2.35, 8.3.36, 8.4.27, and 8.5.12, urging all sites using PHP to upgrade immediately. Older versions, including PHP 8.1 and below, will not receive official fixes.

This incident highlights OpenAI's "Defender's Window" concept, which argues that organizations should proactively use advanced AI models to strengthen defenses before malicious actors can exploit similar capabilities.

Fuentes:Tom's Hardware

Investigación

Radio Free Mobile Founder Richard Windsor: xAI Halts Grok 5 Training to Rent Out Compute; Spot Premium Expected to Fade by 2027

Material
2026-10-21 14:53 GMT+8

xAI has made a seemingly counterintuitive decision: pausing Grok 5 training to rent out compute capacity. In an analysis dated October 21, 2026, Richard Windsor, founder of Radio Free Mobile, stated that this decision was made approximately six to nine months ago, resulting in year-over-year growth in data center business revenue of seven times and quarter-over-quarter growth of three times. He attributes this to a simpler calculation: the return on renting out compute has already surpassed the potential benefits of Grok 5 training.

Windsor’s estimates for spot pricing are as follows: the spot price for compute capacity under short-term contracts is approximately $30 billion to $50 billion per GW, yielding annualized five-year returns of 45% to 85%. He explicitly states that these calculations are RFM's own. Based on this estimate, any company with idle capacity should drop what it is doing and sell its compute to the highest bidder—exactly what xAI is doing.

However, there is a key structural asymmetry here. Windsor points out that new cloud providers are almost unable to leverage this spot premium because their financing models require selling compute via multi-year contracts in advance to raise funds for building data centers. In other words, the spot market rewards companies that already possess capacity and can flexibly switch its use (such as xAI), rather than those building capacity and locked in by financing terms. The term market prices are much more moderate; Windsor believes long-term deals are in the range of $20 billion to $25 billion per GW.

High prices are the cure for high prices. Windsor estimates that in a normalized market, AI compute will fall back to approximately $20 billion per GW, leaving substantial returns for providers while imposing reasonable prices on renters. He expects xAI’s financial performance to slow significantly in 2027, as prices charged then will drop by more than 50%. Meanwhile, he mentions that about half of new data center construction faces delays, meaning supply-demand balance may take longer than expected to arrive.

Windsor disclosed in the article that he holds a position in Meta Platforms and believes Meta’s enterprise business creation aligns with this compute shortage logic. He also mentioned that Korean new cloud provider Elice dropped 30% on its first day of listing due to retail panic, subsequently recovering most of its losses. These observations point to the same judgment: currently, it is a window period for capacity holders, but the window will not remain open forever.

Fuentes:radiofreemobile.com

Investigación

OpenAI says ChatGPT advertisers top 100,000 as ad reach expands to 89 countries

Material
2026-10-22 16:57 GMT+8

OpenAI announced on October 21 (local time) that the number of advertisers using its chatbot ChatGPT has exceeded 100,000, while expanding ad delivery services from 63 to 89 countries. This move signals the company's accelerated effort to convert its free user base into an advertising revenue stream, diversifying a business model previously reliant on subscriptions.

According to ZDNet Korea, OpenAI began piloting ads for free users in the United States in February and has since pushed for international expansion, adding five countries including South Korea in August. Data released by the company in August indicated that the annualized revenue from ChatGPT's advertising business had reached $1 billion, approximately six months after launching the service in the U.S. It is important to note that annualized revenue is a projection based on current sales speed, not actual cumulative earnings.

Outlets such as Axios predict that OpenAI's total advertising revenue for this year could reach $2.5 billion, representing about 8% of the market's expected $30 billion total annual revenue. As competition among AI service providers shifts from user acquisition to monetization, the expansion of ChatGPT's ad platform is viewed as a significant diversion of demand from traditional search engines and internet platforms.

Fuentes:zdnet.co.kr

Investigación

Meta Runs Google TPUs in Production for First Time, Challenging Nvidia's Dominance

Material
2026-10-22 18:00 GMT+8

According to Tom's Hardware, Meta is now running Google's Tensor Processing Units (TPUs) in production within its data centers. This marks the first time Google's custom AI chips have been used by an external major customer for actual business workloads. The breakthrough is enabled by a newly released PyTorch-native version of the SGLang inference engine, developed through a close collaboration between Meta, Google, and RadixArk, the startup founded by SGLang's creators.

Previously, TPUs could only be accessed via Google's proprietary JAX framework or the older PyTorch/XLA bridge, limiting their adoption in the mainstream PyTorch ecosystem. The new TorchTPU engine allows developers to deploy PyTorch workloads directly to TPUs without rewriting code. Zongwei Zhou, Meta's PyTorch-TPU lead, stated that the engine provides Meta with "a remarkably mature, PyTorch-native serving stack."

This development poses a potential threat to Nvidia, which has built a strong moat around its CUDA software ecosystem. Meta, as a primary supporter of the PyTorch framework, shifting to TPUs signals that large tech companies are actively seeking to break free from single-hardware constraints. Jana van Greunen, Director of PyTorch Engineering at Meta, noted that Meta "cannot be constrained by a single hardware architecture" and that TorchTPU enables them to deploy PyTorch workloads to Google TPUs.

It is worth noting that this is not Meta's first attempt to use TPUs. Google describes PyTorch/XLA as a "predecessor" to the new TorchTPU, and SemiAnalysis reported last year that Meta research teams used TPUs through this bridge from 2020 to 2023 before canceling those contracts. The current deployment is based on a multi-year rental agreement, though specific scale and locations remain undisclosed.

Regarding performance, Inferact, a startup founded by vLLM's creators, reported that a tuned vLLM setup serving Moonshot AI's Kimi K3 model on seventh-generation Ironwood TPUs achieved over 700 tokens per second for a single user, compared to 452 tokens per second on Nvidia GB200s. However, these figures come from vendor and partner internal tests and have not been independently verified.

Fuentes:Tom's Hardware

Investigación

US plans to restart three retired nuclear plants for AI data centers with $4.4 billion federal loans

Material
2026-10-22 00:30 GMT+8

According to TechRadar, the United States is reportedly gearing up to restart three previously closed nuclear power plants specifically to keep up with the energy demands caused by AI data centers. Together, these three facilities are expected to add 2.25 gigawatts (GW) of electricity to the grid, enough to power two million homes and businesses.

The U.S. Department of Energy is a major supporter of these projects, providing up to $4.4 billion in federal loans. Microsoft and Google have already signed agreements to purchase electricity from two of those facilities, with power plant restarts slated for between 2027 and 2029. These moves likely mark the first commercial nuclear reactors to be restarted after being formally decommissioned, driven by surging demand for reliable and low-carbon electricity.

Specifically, the Palisades Nuclear Plant in Michigan, which closed in 2022, has electricity purchase agreements requiring it to return to service by March 2027. The Crane Clean Energy Center in Pennsylvania is also targeting a 2027 online date, while the Duane Arnold Energy Center in Iowa is looking for a 2029 go-live date.

This initiative aligns with the Trump administration's goal to quadruple US nuclear generating capacity from around 100GW in 2024 to 400GW by 2050. However, restarting aging reactors has drawn criticism over safety concerns, especially in a world of heightened geopolitical activity and natural disasters.

Fuentes:TechRadar

Investigación

Why Backend Infrastructure Is Becoming More Valuable as AI-Generated Apps Multiply: Supabase’s Embedding Strategy and Operational Moats

Tema · SupabaseMaterial
2026-10-08 19:00 GMT+8

AI tools are enabling more people to generate applications. In June 2026, during its Series F funding round, Supabase announced that its annual database creation volume grew by 600%, with over 60% generated through AI tools. It has nearly 10 million developer users and more than 250,000 enterprise customers. Third-party firms Sacra and Latka estimated its annual recurring revenue (ARR) at $170 million in May 2026. These figures reveal a counterintuitive fact: while AI lowers the barrier to creating applications, it does not reduce the demand for backend infrastructure—in fact, it creates more demand.

Supabase packages the components needed for application backends into a single product: PostgreSQL databases, user authentication, file storage, real-time data push, APIs, and edge functions. Its strategy is not to reinvent the database but to directly use existing open-source components—Postgres for the database, PostgREST for the API, and Deno for edge functions—and then develop the missing parts itself, such as the management interface, real-time engine, and connection pooling. Founder Paul Copplestone started Supabase when he was CTO at a startup in Singapore; frustrated by Firestore's slow query speeds and lack of real-time capabilities, he had to build a custom real-time engine on top of Postgres. He stated in 2022 that while Postgres is scalable, developers still struggle to use it effectively.

The strategic choice worth focusing on is distribution. In December 2025, Supabase began offering white-label services to AI building tools: Lovable, Bolt.new, and Baidu’s MeDo can prepare and manage the backend on behalf of their users. Another path involves Vercel’s v0 and Figma Make, where users build on their own Supabase projects. In its Series F announcement, Supabase noted that its platform-facing white-label products were its fastest-growing business line in the past year. This means that as AI tool users create applications, the backend is already in place, eliminating the need to search for and configure infrastructure separately.

However, creation volume does not equal revenue. Supabase’s pricing model combines organizational subscriptions with compute resource usage per project. Free projects stop if unused for a week. Paid projects charge based on reserved compute size and runtime, rather than access frequency. In an interview with First Round Review, Copplestone explained that sales commissions are calculated only on the increased usage after a customer is onboarded; in a TechCrunch interview, he mentioned that the company turned down multi-million-dollar contracts requiring custom product features to maintain a unified product experience for all developers. The key logic here is: free entry points allow applications to grow, and revenue comes from the continuous operations once applications are truly running.

The article concludes by pointing out two structural risks. First, dependence on AI tool partners: if Lovable switches to another backend solution, Supabase will lose this growth channel. Second, the replicability of open-source code: the main repository uses the Apache 2.0 license. CTO Ant Wilson wrote in 2022 that when scale is sufficient, large cloud vendors could offer equivalent products leveraging their distribution capabilities, similar to the situation faced by Elastic and MongoDB. Additionally, AI-generated apps themselves pose security configuration issues—CVE-2025-48757 revealed that projects generated by Lovable lacked row-level security settings, indicating that ease of creation does not equate to ease of secure operation.

Fuentes:note(ノート)

Investigación

OpenAI Korea Head: South Korea Ranks Third Globally in Weekly Active Users for 'Dots' Agent

Resumen rápido
2026-10-22 16:55 GMT+8

Kim Kyung-hoon, General Manager of OpenAI Korea, revealed on October 22, 2026, at the "OpenAI DevDay Exchange: Seoul" event that South Korea ranks third globally in weekly active users for OpenAI's artificial intelligence agent "Dots." He also noted that despite not having the largest population, South Korea maintains a top-five position in ChatGPT Pro usage scale.

"Dots" is an AI agent capable of autonomously executing tasks across applications and websites based on user-defined goals, continuing to work on dedicated cloud computers even when users are away. Kim emphasized that feedback from Korean developers and users has been crucial for OpenAI's product iterations, citing optimizations for Linux systems, multi-folder project handling, and mobile usage as responses to local demand. Additionally, OpenAI improved its voice model "GPT Live 1.1" by adding Korean voice options named "Haetsal" (Sunshine) and "Haneul" (Sky) to enhance the localized experience.

Fuentes:zdnet.co.kr

Investigación

Alpha Exchange Selection: 400 Capital CIO Chris Hentemann on AI Infrastructure Financing and Credit Risk Structuring

Tema · AI基建融资缺口Material
2026-10-09 16:00 GMT+8

The Alpha Exchange podcast released a conversation with Chris Hentemann, Founder and Chief Investment Officer of 400 Capital Management, on October 9, 2026. Based on the official episode description, the discussion centers on Hentemann's framework for credit investing and his views on the current market environment. 400 Capital currently manages approximately $8 billion, focusing on asset-based and securitized credit.

The description indicates that Hentemann will reflect on his early career at First Boston, Salomon Brothers, and Bank of America, discussing how leverage, insufficient risk capital, and misaligned incentives led to unsustainable bank balance sheet expansion prior to the Global Financial Crisis. These experiences form the foundation of his investment philosophy at 400 Capital, which starts with raw, unlevered returns, seeks strong underlying asset selection, and looks for structures providing asymmetric or convex outcomes.

Furthermore, the episode delves into Significant Risk Transfer (SRT), explaining how regulatory changes have created opportunities for private investors to assume credit risks that banks increasingly want off their balance sheets. Hentemann plans to walk through a recent transaction involving OceanFirst and a seasoned residential mortgage portfolio to illustrate the underwriting and structuring logic behind such investments.

Regarding the current AI infrastructure boom, the description notes that Hentemann shares 400 Capital's perspective: the firm is cautious on data center real estate itself but sees opportunity in financing tangible assets such as power equipment and turbines. The discussion concludes with topics on risk management, dry powder reserves, and the value of being a thoughtful second mover in this point of the cycle.

Fuentes:alphaexchange.simplecast.com

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión

In summary, Supabase’s case reveals a transferable insight: in an era where AI lowers creation barriers, the value anchor for infrastructure companies shifts from technology itself to the combination of distribution entry points and continuous operational capabilities, which together support the growth of recurring revenue.