독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-1298 게시물

새로고침

AMD Announces Venice-X EPYC CPUs Returning 3D V-Cache to Data Center with 5.3 GHz Boost Clock for HPC Workloads

중대
2026-11-12 05:30 GMT+8

U.S. semiconductor company AMD has announced that its "Venice-X" series of EPYC server processors will launch in the second half of 2027. This marks the return of 3D V-Cache stacked cache technology to the data center after a four-year absence. Following the release of Genoa-X in 2023, the previous Turin generation did not feature an X-model with 3D V-Cache, temporarily removing the technology from the lineup.

The core breakthrough of Venice-X lies in unifying high frequency and large capacity cache, two characteristics that were previously difficult to combine. According to Tom's Hardware, Venice-X will feature 96 cores and 1152MB of L3 cache, matching the specifications of the previous-generation EPYC 9684X (Genoa-X), but with a maximum boost clock increased to 5.3GHz. In contrast, earlier X-series chips typically sacrificed frequency to accommodate the massive stacked cache; for example, Milan-X topped out at 3.5GHz, and Genoa-X at 3.7GHz. Venice-X achieves the highest frequency ceiling within the family for the first time, eliminating the need for users to choose between high-frequency chips and large-cache chips.

Beyond the frequency increase, Venice-X introduces PCIe 6.0 interfaces and relies on the SP7 platform to support a 16-channel memory configuration. Through MRDIMM technology, it achieves per-socket memory bandwidth of 1.6TB/s, nearly four times that of the previous generation. These improvements target High Performance Computing (HPC) workloads, particularly functional verification tasks in Electronic Design Automation (EDA).

Phil Steinke, an AMD Fellow, stated that AMD currently runs EDA workloads by mixing high-frequency CPUs and 3D V-Cache CPUs, as the high-frequency parts historically offered better clock speeds than the cache-heavy parts. Venice-X removes this dichotomy, allowing a single chip to possess both cache advantages and frequency advantages. According to Steinke, for certain logic simulation workloads, 3D V-Cache can improve performance by 40% and double instructions per clock (IPC).

Although the exact number of SKUs, base frequencies, and Thermal Design Power (TDP) have not been fully disclosed, the TDP is expected to be in the range of 400 to 600 watts. The launch of Venice-X signifies AMD's continued strategy of technical differentiation for specialized server loads in the Zen 6 architecture era.

출처:Tom's Hardware

리서치

United Airlines Blocks AI Agents From Booking Flights Citing Privacy and Regulatory Risks

주제 · 美联航AI代理禁令중대
2026-11-12 05:54 GMT+8

United Airlines has stated that it will not allow personal AI assistants to complete ticket bookings on its platform. The company cited privacy and regulatory concerns, noting that AI agents could expose customer information or make unintended purchases, and potentially skirt federal rules requiring airlines to have customers attest that their bags do not contain explosives.

According to The Wall Street Journal, United CEO Scott Kirby said at an event on Tuesday, "To be honest, I don't want to create another middleman." While agents can still show users United flights, they cannot execute the transaction on their behalf. This decision may limit the potential usefulness of newer personal AI assistants like Meta's Muse and OpenAI's Dots.

United is not alone in restricting unauthorized AI agents. eBay appears to have blocked agents from browsing its site, and several major banks have also placed restrictions on them. United suggested it could reconsider its decision based on future advancements, but emphasized the need to establish protections for travelers first.

출처:The Verge

리서치

SpaceX/xAI Reportedly Developing 'Kessel' Rack System: 84-Layer PCB Replaces NVLink Cables for Orbital Data Centers

주제 · Kessel중대
2026-11-11 23:03 GMT+8

According to a report by Tom's Hardware citing SemiAnalysis, SpaceX and xAI are developing a custom version of Nvidia's Vera Rubin NVL72 rack-scale AI system codenamed 'Kessel'. The system retains the fundamental configuration of 36 Vera CPUs, 72 Rubin GPUs, and nine NVLink switch trays but introduces a different solid-state interconnect architecture.

Unlike Nvidia's conventional arrangement using internal NVLink cable cartridges, 'Kessel' plans to position compute and switch trays back-to-back around a passive PCB midplane. The most distinctive feature is its 84-layer PCB midplane, which uses ultra-low-loss M10 copper-clad laminates (CCLs). This contrasts with Nvidia's Kyber architecture, which relied on a more complex 156-layer PCB. While reducing layer count could improve manufacturability, the advanced materials will likely increase costs.

The design is aimed at both terrestrial and orbital deployments. Furthermore, SpaceX/xAI reportedly plans to shift its procurement model, acquiring AI accelerators, CPUs, switches, and networking components directly from Nvidia while sourcing other components independently. This allows SpaceX to select its own PCB manufacturers and optimize hardware for its infrastructure, though it assumes greater responsibility for system design, validation, and supply-chain management.

출처:Tom's Hardware

리서치

Poste Italiane: Why Valuation Fails When Analysts Cannot Find Comparable Companies

주제 · 意大利邮政估值争议중대
2026-11-04

In November 2026, Calum Holt, an investment analyst at the Monks fund under Baillie Gifford, published an investment thesis on Poste Italiane. His core argument is that the market undervalues this company not because of flaws in its fundamentals, but because analysts cannot find truly comparable companies. Consequently, they default to placing it in the peer group of Italian banks, pricing it using the average profit multiples of those banks.

This judgment is supported by a specific chain of evidence. In 2025, Poste Italiane’s net interest income—the classic engine of banking profits—contributed only one-fifth of its revenue. Its income sources are a mix of insurance, digital payments, logistics, and energy resale. Parcel revenue more than doubled between 2019 and 2025; the digital payments business has grown at an annualized rate of 22% over the past five years, currently generating €1.2 billion in annual sales. Since CEO Matteo Del Fante took office in 2017, although revenue growth has been modest at 3%, net profit per employee has grown at an annualized rate of 18%—a financial characteristic typical of the transformation from an inefficient state-owned enterprise to a high-profit platform.

Monks argues that the common thread among these businesses is not bank interest margins, but a trust network built over 150 years of brand heritage, combined with shared infrastructure. Customers can send parcels, pay utility bills, obtain passport photos, open savings accounts, buy insurance, and apply for loans all within the same branch network. The cross-selling success rate of the newly launched super-app 'P' is five times higher than that of old channels. In Monks’ view, the growth opportunities driven by e-commerce, digital payments, and the app are still in their early stages, with profits potentially growing at a rate exceeding 10% over the next decade.

However, the most critical aspect for readers to scrutinize is precisely the weakest link in this argument. The premise that there are 'no comparable companies' itself invites skepticism: every company is unique in some sense, and uniqueness does not automatically equate to being undervalued. If the growth logic were truly so clear, why did the market fail to recognize it earlier? One possibility is that analyst inertia indeed caused systematic mispricing; another is that as a holding party, Monks has a motive to describe its portfolio company as a misunderstood unique asset, thereby providing narrative support for its own buying decisions. Furthermore, the article mentions that if Poste Italiane’s acquisition of telecom operator Telecom Italia succeeds, it will add a telecommunications network business—a condition that has not yet materialized and should not be treated as an realized valuation catalyst.

The true value of this thesis lies not in offering a buy or sell conclusion, but in demonstrating a failure mode of valuation methodology: when an analyst’s peer-group tool encounters a genuinely hybrid enterprise, they retreat to the closest but incorrect classification. The resulting discount is not a result of fundamental judgment, but a product of the classification framework. Readers can use this question to examine other companies categorized into specific industries, investigating whether their discounts stem from fundamental assessments or from the classification inertia that arises when analysts cannot find better reference points.

출처:bailliegifford.com

리서치

London Startup Tano Raises $4M Seed Round Led by Cherry Ventures as AI Agents Handle 400K Monthly Creator Conversations

주제 · Tano간단한 요점
2026-11-11 22:00 GMT+8

London-based startup Tano has announced the closure of a $4 million seed round led by Cherry Ventures. This brings the company's total funding to $5 million, following an earlier approximately $800,000 round led by Seedcamp.

Founded in 2025 by former Encord employee Will Caplan and Sagar Shah, Tano utilizes AI agents to automate tasks such as negotiating rates, sending products, and paying creators in influencer marketing campaigns. According to CEO Will Caplan, the platform's AI agents currently conduct around 400,000 conversations with creators each month.

The new capital will be allocated to computing resources, engineering, and hiring human creator strategists. While highly automated, Tano maintains a human-in-the-loop system where strategists intervene when the AI's confidence levels are low.

출처:Tech Funding News

리서치

Meta Infrastructure Chief Santosh Janardhan to Step Down; Ex-Amazon Exec Dave Brown to Replace Him

중대
2026-11-12 04:12 GMT+8

U.S. tech giant Meta confirmed on November 11, 2026, that Santosh Janardhan, the company’s longtime head of infrastructure and co-head of engineering, will be stepping down from his role. Janardhan joined Meta in 2009 and was instrumental in the firm’s growth into a global social media and technology powerhouse. According to TechCrunch, he will be replaced by Dave Brown, who joined Meta earlier this year as vice president of infrastructure. Brown previously spent 19 years at Amazon, focusing on compute, cloud, and AI.

A Meta representative stated that Janardhan will stay on for a transition period and move into an advisory role for “Meta Compute,” an initiative launched earlier this year to scale the company’s data center investments and strategy amid intensifying competition in the AI race. While internal memos did not specify the reasons for Janardhan’s departure, reports indicate he cited the need to “pass the baton.” This leadership change occurs against the backdrop of exploding infrastructure spending at Meta, raising questions about how the new executive team will manage the pace and direction of its massive AI compute expansion.

출처:TechCrunch

리서치

Nvidia Donates OpenShell AI Agent Security Runtime to CNCF

주제 · OpenShell중대
2026-11-12 03:13 GMT+8

Nvidia announced this week at KubeCon + CloudNativeCon North America in Salt Lake City that it has donated OpenShell, its open-source runtime for securing and governing AI agents, to the Cloud Native Computing Foundation (CNCF). The project joins the CNCF as a Sandbox project, the foundation’s entry-level tier for early-stage projects. The CNCF formally accepted the project in late September.

Originally launched at Nvidia’s GTC conference in March, OpenShell runs AI agents inside a sandbox and enforces policies on which files, networks, credentials, and model endpoints those agents can access. Ali Golshan, Nvidia’s senior director of AI software, stated that choosing the CNCF was driven by the view that “Securing and creating trust around AI is really infrastructure.” He noted that the CNCF hosts mature infrastructure components like Kubernetes, aligning with OpenShell’s target users—platform teams already managing these cloud-native stacks.

This decision highlights the evolving landscape of AI infrastructure governance. While the Linux Foundation’s Agentic AI Foundation (AAIF) hosts protocols like MCP and A2A, Nvidia determined that OpenShell fit better within the CNCF’s cloud-native environment. Golshan explained that many necessary infrastructure pieces already existed in the CNCF, avoiding the need to build them from scratch. Currently, Red Hat, Docker, and Canonical have joined Nvidia as core contributors, with approximately 380 organizations contributing according to the CNCF project page.

Although OpenShell has entered the CNCF, its security governance is still being refined. Golshan mentioned that the team is working on establishing an independent security committee to handle issues like vulnerability disclosures. Jonathan Bryce, executive director of the CNCF, described neutral governance as “rocket fuel” for integration and adoption, helping to attract deeper community engagement.

출처:The New Stack

리서치

Marks Memo: AI's Concentrated Dependence on Stock and Bond Markets Is Eroding Diversification Assumptions

주제 · 马克斯AI集中风险论구조적
2026-11-11 16:00 GMT+8

In a memo published on November 11, 2026, Howard Marks presents an argument distinct from common discussions on AI investment: the issue is not whether AI itself is overvalued, but rather that AI has simultaneously become the single point of dependence for both the stock market and the bond market. This degree of concentration is unprecedented in history, and what it erodes is precisely the fundamental premise upon which investors rely to manage risk.

On the equity side, Marks cites J.P. Morgan data indicating that since January 2024, 42 AI-related companies have contributed approximately two-thirds of the price gains in the S&P 500. FactSet reports further show that Alphabet and Amazon alone generated combined book gains of $151 billion in the second quarter from equity investments in Anthropic and SpaceX (which now includes xAI), making them primary contributors to S&P 500 earnings. For 14 consecutive quarters, the "Magnificent Seven"—predominantly hyperscale cloud service providers—have seen their earnings growth outpace that of the rest of the S&P 500 constituents.

On the credit side, according to Goldman Sachs, global debt issuance related to AI this year has approached $600 billion, with approximately $400 billion flowing into the U.S. investment-grade bond market. This represents more than one-fifth of the total supply in that market, raising the share of AI-related bonds in the Bloomberg USD Investment Grade Corporate Index to approximately 12%. Marks points out that the AI industry may require an average of $1 trillion in new capital annually in the future, with hundreds of billions of that amount likely to be raised through debt financing.

This trend toward concentration stands in sharp contrast to the principle of diversification. Marks cites Michael Milken’s pioneering of the sub-investment-grade debt market in 1977-78 as an example, noting that the market accepted risk on the premise of sufficient diversification to ensure that a few defaults would not undermine overall portfolio performance. However, current stock and bond markets are highly dependent on a single factor: the success of AI.

Marks further questions the predictability of AI. Citing an assessment by Claude, he notes that the overall predictability score for AI is only 3 out of 10, with core variables of interest to investors—profit size, distribution patterns, and speed of realization—scoring as low as 2. He specifically mentions that although frontier AI labs such as OpenAI and Anthropic are regarded as industry leaders, their predictability scores are comparable to early-stage venture capital, falling into the category of 'practically unknowable.'

Marks believes this concentrated dependence constitutes a 'fault line' running through portfolios, with potential impacts far exceeding the AI sector itself. He reminds investors that in areas of extreme uncertainty, accepting 'that you do not know everything' is necessary. He concludes by quoting Warren Buffett from 1979: The future is never clear, and uncertainty is actually the friend of the long-term value buyer.

출처:Oaktree Capital

리서치

Ukrainian Drone Maker Reports 6-10x Fiber Optic Price Hikes Due to AI Data Center Demand

주제 · AI数据中心光纤短缺중대
2026-11-11 19:00 GMT+8

According to Tom's Hardware, a research and development representative for Ukrainian drone manufacturer BlueBird Tech told Ars Technica that prices for fiber optic cables have risen approximately six to ten times in the past couple of years as AI data center demand absorbs more of the global supply. This shortage is hindering Ukraine's ability to scale production of fiber-optic guided drones.

The representative noted that while both Russia and Ukraine are estimated to produce around 50,000 fiber optic drones monthly, Ukraine acquired only 72 miles of fiber optic cable from Chinese suppliers in August 2025, whereas Russia imported 328,000 miles during the same month. It was reported that Russia secured 37 million miles of fiber throughout 2025, accounting for over 10% of global production. Additionally, some Chinese companies were claimed to have refused to sell fiber optic cables to Ukraine.

There is an estimated 111-million-mile shortfall in global fiber optic cable supply. Beyond material scarcity, trade barriers and supply chain concentration exacerbate the issue. While there has been talk of producing fiber optic cables domestically in Ukraine, such manufacturing requires specially trained personnel, and the currently limited supply makes it difficult to support the training and quality control needed before large-scale automation can be implemented.

출처:Tom's Hardware

리서치

Theory Ventures Partner Tomasz Tunguz: The AI Industry Chain Gross Margin Waterfall — Applications Capture 40%, Neocloud Post-Depreciation Margin Is Only $0 to $2 per Watt-Year

주제 · AI产业链毛利瀑布중대
2026-11-11 08:00 GMT+8

The AI industry chain faces a fundamental measurement difficulty: utilities sell kilowatt-hours, neoclouds sell GPU hours, inference providers bill by token, and application layers price based on completed tasks. Each of the five layers quotes in different units, making it impossible to directly compare who earns more. In an analysis published on November 11, Tomasz Tunguz, partner at Theory Ventures and former Google product manager, proposed a normalization solution: convert all layers to dollars per watt-year ($/w/yr), which represents how many dollars of revenue are generated annually for every watt of continuous power. This denominator brings the entire chain onto the same ruler.

After normalization, Tomasz Tunguz assembled a gross margin waterfall using CoreWeave’s Q3 2026 earnings report, Anthropic’s S-1/A filing, disclosures from Core Scientific with the U.S. Securities and Exchange Commission (SEC), and CBRE data center leasing data. Assuming an application layer gross margin of 40%, the distribution of the total gross profit pool is as follows: Power 3%, Data Center 5%, Neocloud 28% (before GPU depreciation), Inference 24%, and Applications 40%. The three layers of Applications (40%), Neocloud (28%), and Inference Providers (24%) collectively capture 92% of the total gross margin—this figure is before GPU depreciation.

However, this waterfall has a critical accounting boundary. The Neocloud figure of 28% is pre-depreciation. The capital expenditure for GPU hardware—$35 to $50 per watt, depreciated over six years—adds $6 to $8 per watt in annual capital costs, a portion of which effectively belongs to Nvidia (which has a hardware gross margin of approximately 75%). CoreWeave’s depreciation and amortization in Q3 amounted to $1.89 billion, representing 52% of its revenue. After deducting GPU depreciation, Neocloud’s gross margin drops to $0 to $2 per watt-year under the current blended rate of $9 per watt-year. In other words, while Neocloud appears to earn significantly, most of that gross margin is consumed by GPU hardware costs.

From this, Tomasz Tunguz draws the first implication: rising compute prices are testing who can transmit costs to end users. Anthropic’s S-1/A filing discloses that short-term burst compute has reached $13 per GPU-hour. On CoreWeave’s Q3 earnings call, CEO Michael Intrator confirmed that prices for new four-to-five-year contracts are shifting from $40 per megawatt-year toward $20 to $25 per megawatt-year—more than double CoreWeave’s current blended fleet rate of $9 per watt-year. If application layers can charge end users higher prices due to improved effectiveness, cost increases can cascade through the chain; if not, Neoclouds, inference providers, and applications must compete for a fixed pool of gross margin.

The second implication is that pricing power for different applications depends on their return on investment (ROI). Tomasz Tunguz points out that an AI that saves 30% in costs does not possess the same pricing power as one that saves 70%. Small models, fine-tuning, and routing techniques can help maintain margins, but unlike the SaaS era, different vertical software markets will support varying sizes of gross profit pools. This means the 40% gross margin in the AI application layer is not evenly distributed but highly differentiated.

출처:Tomasz Tunguz

리서치

US Corporate AI Spending Growth Slows to 1.65%, Customer Support Share Rises to 13.3%

주제 · Ramp AI指数중대
2026-11-12 03:35 GMT+8

According to the Ramp AI Index released by financial services provider Ramp, US corporate AI spending reached a new high in October 2026, but month-over-month growth slowed to just 1.65%. The index, based on payment data from more than 85,000 US companies, reflects API-level spending trends.

Spending structure remains extremely concentrated. The top 1% of companies by AI spending paid a median of $10,200 per employee per month, up 9% from the previous month. In contrast, a typical company spent only $16.32, an increase of less than $1. Ramp Chief Economist Ara Kharazian noted that this top tier—mostly venture-backed firms in tech, finance, and technical services—still drives 80% of revenue for OpenAI and Anthropic.

Departmental allocation shifted significantly. Software development teams accounted for 72% of all tracked AI spending, though their share has declined from summer highs. Meanwhile, customer support and success departments saw their share of AI spending climb from 0.8% in June to 13.3% in October, likely driven by AI chatbots in service roles. Kharazian analyzed that these use cases tend to favor cheaper, faster models over expensive flagship ones, meaning new customers may bring lower margins.

Although the average price per million processed text tokens dropped 21% last month and usage volume grew faster than spending, market concentration risks persist. Anthropic reached 44.8% adoption on the Ramp platform, while OpenAI held steady at 40.6%. It is worth noting that Ramp's data sample skews toward large, tech-savvy companies, potentially overstating overall market AI adoption rates.

출처:The Decoder

리서치

Instagram Launches Effect Builder for No-Code AI Story Effects

간단한 요점
2026-11-11 14:30 GMT+8

Meta's Instagram has launched the Effect Builder, a new web-based tool integrated into Edits that allows creators to generate custom AI effects for Stories using simple text descriptions. According to a report by German tech outlet t3n.de, the tool aims to lower technical barriers, enabling users without design or programming knowledge to create unique visual styles.

In terms of workflow, users describe their desired effect (such as "nostalgic film look" or "comic style") at edits.com/effectbuilder. The system generates a preview and supports real-time adjustments. Submitted effects undergo an Instagram safety review before becoming available in the community browser for other users.

A key mechanism of this feature is attribution linking: when other users apply a created effect to their Stories, the original creator's profile remains linked. This design intends to help creators break out of existing follower bases and gain new exposure through popular effects.

Currently, the Effect Builder is available only to users aged 18 and older in the United States, with creation and publishing free of charge. Meta has not announced a timeline for expansion to the European Union or other regions. If released in the EU, users would need to comply with local regulations regarding the labeling of AI-generated content.

출처:t3n.de

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

The third implication is that vertical integration is already absorbing adjacent layer margins. In 2025, CoreWeave acquired Weights & Biases ($1.03 billion), OpenPipe, Marimo, and Monolith AI ($185 million), moving into training, evaluation, and enterprise workflows. Nebius spent over $1 billion in 2026 acquiring four companies spanning search, model optimization, cold-start scaling, and orchestration to build its Token Factory. Owning multiple layers allows integrators to absorb price fluctuations that squeeze independent suppliers.

Tomasz Tunguz himself marked the boundaries of this model: the conversion to $/w/yr relies on modeling approximations rather than audited disclosures, involving parameters such as quarterly starting active power, facility PUE (1.2 to 1.45), rack power draw (150 to 180 kW), billing utilization (65%), and cache hit rates (87% to 90%), leading to significant error across workloads. The inference layer modeling is based on open-weight models rather than proprietary frontier lab APIs. The 40% application gross margin is an assumed input, not a measured result. This waterfall describes a static snapshot and does not predict which layer will collapse first or when.