読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-0268 投稿

更新

OpenAI Signs AI Deployment Deals with Abu Dhabi Sovereign Wealth Fund and National Oil Company Ahead of Reported Mega-Round

トピック · OpenAI阿布扎比协议重大
2026-11-02 21:48 GMT+8

According to Semafor, OpenAI CEO Sam Altman signed four major agreements in Abu Dhabi on November 2, 2026, to deploy artificial intelligence products across government, the oil sector, and sovereign wealth management. These deals come as the ChatGPT maker is reportedly looking to lock down a mega funding round ahead of an initial public offering (IPO).

The largest of the four agreements is with the Abu Dhabi National Oil Company (ADNOC). Signed with G42 subsidiary Inception42, this multi-year deal will deploy OpenAI’s Frontier platform for building and managing AI agents. The system will link decisions across oilfield development, production, refining, and trading, running on G42 infrastructure within the United Arab Emirates. No financial terms were disclosed. Additionally, the Abu Dhabi government is working with OpenAI to make some of its policies readable by AI agents.

A separate agreement involves Mubadala, Abu Dhabi’s sovereign wealth fund, which will apply OpenAI’s models to analyze potential deals. A consortium of UAE backers, including Mubadala-backed MGX, is expected to be the largest investor in OpenAI’s latest funding round and will get an option to lease new computing capacity in the region, the Financial Times reported.

Altman’s appearance at ADIPEC also highlighted a critical constraint facing the AI industry: energy. Asked onstage about his biggest challenge, Altman gave a one-word answer: “Energy.” He noted that while AI models will become more efficient and cheaper to deploy, the power needed to run them could keep costs high and slow down growth.

ソース:semafor.com

リサーチ

Semiconductor Analyst Richard Windsor: Kioxia Guidance Miss Stems from Long-Term Contract Pricing and Revenue Recognition Timing, Not a Signal of Storage Cycle Peak

トピック · 铠侠FQ3指引争议重大
2026-11-02 14:27 GMT+8

Following Kioxia's release of its Fiscal Q2 2027 results, its guidance for the next quarter disappointed the market: operating profit expectations were 14.9% below consensus, and revenue guidance was 11.6% lower, causing the stock to fall 17% on the day. The market immediately raised two concerns—demand is softening, and AI deployment is beginning to flatten. Independent semiconductor industry analyst Richard Windsor offered a different judgment in his blog: he believes this reflects Kioxia's own issues rather than a signal that the storage cycle has peaked.

Windsor's explanation revolves around three mechanisms. First is pricing deceleration: Kioxia expects prices to rise by 3-4% quarter-over-quarter in FQ3, far below the double-digit growth seen in the first two quarters. This appears to suggest supply and demand are moving toward balance, but it contradicts the company's statement that it remains sold out for all of calendar 2027 and that 50% of calendar 2028 capacity is already locked into long-term contracts. Windsor argues this is precisely the structural cost of long-term contracts—the industry trades off short-term pricing upside in exchange for longer-term guarantees of volume and pricing.

Second is management's proactive suppression of price hikes. In an interview, CEO Hiroo Ota stated that prices had risen enough and that further increases would hurt demand in the end-device market, which accounts for roughly half of Kioxia's revenue. He instructed sales teams not to push for overly aggressive price increases. Windsor points out that SanDisk also expects modest price increases in Q4 but guides to double-digit revenue growth due to volume gains, indicating that Kioxia's pricing restraint is not a universal industry trend.

Third is revenue recognition timing. Some orders in FQ2 were fulfilled early (Windsor speculates this may be related to higher quotes), while contracts expected to fall in FQ3 have been pushed to FQ4, resulting in lower FQ3 figures but a corresponding rebound in FQ4. These contracts are typically data center contracts, which take longer to fill and face risks of delays given that over 45% of US data centers encounter permitting, power, or political delays.

As evidence against a systemic peak, Windsor notes: both Kioxia and SanDisk state that capacity is sold out for calendar 2027; data center capital expenditure expectations remain rising rather than falling; DRAM manufacturers are unaffected; and Samsung and SK Hynix are completely unmoved by Kioxia's sharp share price decline. The tightness in NAND is far less severe than in DRAM, so there is no direct spillover effect.

Windsor concludes that the acceleration of this round of price increases will slow significantly in 2027, with uncertainty surrounding 2028, but that the current signal is noise rather than confirmation of a cycle top. However, Windsor himself notes that half of Kioxia's revenue comes from consumer electronics, a market that has already declined sharply due to rising storage costs, and he categorizes this risk as a 2028 issue rather than a current cycle-top signal. The boundary of this judgment lies in the following question: if Ota's pricing restraint is itself because end-user demand is already shrinking, whether this deterioration in the consumer sector will spread to the data center side over a longer period remains an unresolved issue.

ソース:radiofreemobile.com

リサーチ

AI Engineer Compiles RISC-V Processor Directly Into LLM Weights to Run DOOM Internally

トピック · Qwenputer项目クイックテイク
2026-11-02 21:30 GMT+8

AI engineer Divyansh Agrawal has released an experimental project called Qwenputer, successfully compiling and embedding a 32-bit RISC-V CPU directly into the weights of the Qwen 3.5 4B large language model. This innovation allows the model to use its own weights as memory (RAM) to execute processor arithmetic operations, C, and MicroPython code internally, without relying on any external runtime or tool calls, and even run the classic game DOOM at a very slow speed.

Unlike traditional model training, this built-in 'CPU' was not learned through machine learning but computed entirely by a compiler across 11 Transformer layers placed before Qwen's original layers. Agrawal claims the built-in CPU passes official riscv-tests, matches qemu programs, and has run 18 million instructions free-running without a single error. This design overcomes the common arithmetic weaknesses of LLMs by providing exact calculation capabilities. Additionally, the project uses DeltaNet state to provide the model with a filesystem supporting up to 512 files with persistent storage.

Due to expanded resource usage, Qwenputer is nearly twice the size of the underlying Qwen 3.5 4B model, reaching almost 19GB, thus requiring a workstation with at least 24GB of VRAM for testing. The project remains highly experimental, primarily exploring technical possibilities rather than solving specific production problems, though its MIT-licensed code is now available on GitHub.

ソース:Tom's Hardware

リサーチ

Marvell and Lumentum Announce All-Stock Merger Agreement Valued at $115 Billion

構造的
検証済み 2026-11-02 21:59 GMT+8

U.S. semiconductor firm Marvell Technology (NASDAQ: MRVL) and optical component supplier Lumentum Holdings (NASDAQ: LITE) announced on November 2, 2026, that they have entered into a definitive merger agreement. Under the terms, Marvell will acquire Lumentum in an all-stock transaction, where Lumentum shareholders will receive 4.500 shares of Marvell common stock for each share held. The transaction has an implied enterprise value of $115 billion for Lumentum, representing a premium of approximately 23.9% based on unaffected closing prices as of October 28, 2026.

The merger aims to address the shift in AI infrastructure from copper-based connectivity to optical technologies. Marvell brings strengths in electrical interconnects, network switching, and custom XPU, while Lumentum contributes expertise in high-performance lasers, optical transceivers, and optical circuit switching (OCS). The companies stated that the combined entity will offer a full-stack connectivity portfolio spanning CMOS, silicon photonics, compound semiconductors, and advanced packaging, expanding the total addressable market to $550 billion.

Financially, Marvell expects the transaction to be accretive to non-GAAP earnings per share in the first year post-close and to provide line of sight to more than $1 billion in annual run-rate cost synergies. Upon closing, current Marvell shareholders will own approximately 67% of the combined company, while Lumentum shareholders will own about 33%. Matt Murphy, Chairman and CEO of Marvell, will lead the combined company, and three members of Lumentum’s Board, including CEO Michael Hurlston, will join Marvell’s Board.

The transaction remains subject to regulatory approvals and shareholder votes, with an expected close in the first half of calendar 2027. This filing confirms details previously reported only as 'advanced talks' by media outlets, which had not disclosed specific terms or pricing.

ソース:sec.gov

リサーチ

US Government Refuses to Release AI Model Safety Testing Framework and 'Trusted Partner' List

トピック · 美国AI安全测试透明度重大
2026-11-02 21:16 GMT+8

According to a report by German tech outlet heise online, the US government has refused to fully disclose the rules governing its voluntary safety testing of high-performance artificial intelligence (AI) models. This decision follows a Freedom of Information Act (FOIA) lawsuit filed by the civil rights organization Protect Democracy. While the Office of the National Cyber Director (ONCD) at the White House submitted some documents in response, key materials—including the full draft of the testing framework, accompanying documents, and the complete list of "Trusted Partners"—remain classified due to national security interests and protection of confidential business information.

The dispute stems from Executive Order 14409, signed by President Donald Trump in June 2026. The order established a voluntary mechanism allowing developers like Anthropic and OpenAI to grant the government access to their frontier models up to 30 days before planned release. Although the executive order explicitly states that this does not create a mandatory licensing or pre-approval requirement, the specific criteria for implementation have not been made public.

Citing documents obtained through the lawsuit, The Information reported that the actual framework likely consists of a seven-page document whose full content remains unpublished. The released materials primarily consist of emails from the ONCD and the Department of Commerce to AI labs, revealing which organizations were designated as "Trusted Partners" with early access, but leaving the selection standards and the full partner list undisclosed.

This lack of transparency raises concerns about industry fairness and the validity of safety assessments. With the benchmarking process kept secret, external research institutions cannot independently reproduce test results, and smaller AI providers struggle to align their security measures with publicly known standards. Furthermore, "Trusted Partners" with early access may enjoy asymmetric knowledge and market advantages.

Vivian Dong, Director at Legal Advocates for Safe Science and Technology, stated that it is unacceptable for the federal government to ask the public to trust a review framework whose contents are known only to the government and select AI companies. Her organization intends to continue fighting for transparency in court. The White House did not respond to requests for comment.

ソース:heise online

リサーチ

Report: DeepSeek Raises 105.5 Billion Yuan in Second Funding Round, State Capital Share Jumps to 27%

トピック · DeepSeek融资上市重大
2026-11-02 16:30 GMT+8

According to reports from China News Jingwei and other media outlets on November 1, details of AI startup DeepSeek's second funding round have emerged. The round raised 105.5 billion yuan (approximately $15.8 billion USD), attracting nearly 30 investors.

The investor list shows that Tencent Holdings and Contemporary Amperex Technology Co. Limited (CATL) remain the two major backers. Tencent contributed 10 billion yuan in both rounds; CATL and its chairman Zeng Yuqun contributed 5 billion yuan and 10 billion yuan respectively across the two rounds. Other industrial investors include ChangXin Memory Technologies, SMIC, vivo, Meituan-WR, Geely Auto, Midea Group, Luxshare Precision, Trip.com, and Xiaohongshu, with some individual investments ranging between 1 billion and 3 billion yuan.

Notably, state capital accounted for approximately 27% of this round's funding, a significant increase from 2% in the first round. Cumulatively, tech companies contributed about 42% and state investors about 19%. Foreign media reported that rising prices for Nvidia high-end chips and concerns about an AI industry bubble were reasons for DeepSeek doubling its fundraising target mid-process.

ソース:leiphone.com

リサーチ

Tesla Sees Strong Rebound in Europe as Swedish Registrations Surge 248%

トピック · 特斯拉校车区超速调查重大
2026-11-02 18:36 GMT+8

U.S. electric vehicle maker Tesla is showing continued signs of recovery in its European business. According to data cited by Reuters from local automotive authorities and industry platforms, new Tesla vehicle registrations surged 248% year-over-year in Sweden, 131% in Norway, and 34% in Denmark in October 2026. This marks a sharp improvement from September, when registrations in Norway and Denmark increased by just 2.2% and 2.9%, respectively.

The rebound was supported by several factors. Rising fuel prices, driven by global energy market instability amid the US-Iran conflict and disruptions to Strait of Hormuz supplies, have made electric vehicles more attractive to consumers. Additionally, government incentives and easier year-over-year comparisons with weaker sales periods last year contributed to the significant percentage gains. However, analysts note that the high growth rates are partly due to low base effects, and sustainability remains to be seen.

Tesla CEO Elon Musk stated on a recent earnings call that demand for their vehicles is growing rapidly and "exceeding our production by a significant margin." In the third quarter, Tesla delivered 486,532 vehicles and reported its largest passenger-vehicle order backlog in four years. Despite this, Tesla's net income fell 43% in Q3 as the company continued heavy investment in future businesses such as AI and robotics.

Investors are closely watching upcoming registration figures from Germany and the UK, Europe's two largest car markets, to confirm if this rebound trend holds. If major markets validate the recovery, the market may re-rate Tesla from 'stabilizing' to 'growing again.' Meanwhile, competition remains intense, particularly from Chinese manufacturer BYD, which continues to expand its presence in Europe.

ソース:invezz.com

リサーチ

OpenAI cuts private funding target to $20 billion at $1.25 trillion valuation

トピック · OpenAI估值与融资重大
2026-11-02 21:09 GMT+8

According to reports by CNBC and the Financial Times, US artificial intelligence company OpenAI has lowered its target for its ongoing private funding round to as much as $20 billion, down from the previously sought $30 billion. The company maintains a valuation target of $1.25 trillion. The round is being led by investors from the United Arab Emirates and venture firm Coatue.

The reduction was first reported by the Financial Times, which noted that OpenAI gave investors until October 30 to commit, with some choosing to reduce their allocations. Reuters, citing the FT, reported that the first close is expected to bring in less than $20 billion. CNBC confirmed the new target on October 31, noting that the round remains open and final numbers could still change.

This adjustment comes amid a broader sell-off in technology stocks driven by concerns over the return on massive AI capital expenditures. Intel dropped 25% after weak guidance, and Nvidia fell 4%, while Microsoft, Alphabet, and Amazon reported capital spending below analyst expectations. In this context, OpenAI’s decision to lower its target suggests investor hesitation about committing large sums before seeing how public markets value rival Anthropic, which filed its prospectus on October 26 for a Nasdaq listing under the ticker ANT.

OpenAI’s annualized revenue was approximately $50 billion at the end of September, according to Bloomberg, with projections reaching at least $70 billion by the end of 2026. At a $1.25 trillion valuation, this implies a price-to-sales ratio of roughly 25 times current revenue run-rate. The company confidentially filed its own IPO prospectus with the U.S. Securities and Exchange Commission in June but has not disclosed a debut date.

ソース:Tech Funding News

リサーチ

Microsoft VP Says Windows 11 Speed Test Prototype Built by Agentic AI in One Day

重大
2026-11-02 21:02 GMT+8

Marcus Ash, Microsoft's VP of Windows Design and Research, recently revealed on X that the updated built-in internet speed test in the Windows 11 preview was largely designed and built using an "agentic AI workflow." Ash noted that after receiving user feedback, the team had a working prototype "within a day," which was then integrated into Windows by the connectivity team. He emphasized that a human guided the process "in their spare time" and highlighted this as a way to accelerate value delivery without producing "AI slop."

However, the statement has generated controversy online. Critics argue that given recent complaints about bugs in Windows 11 updates, particularly since the 24H2 release, promoting AI-driven speed raises concerns about insufficient quality assurance. Comments on Reddit suggest that users are less opposed to AI usage itself but are worried about "woefully insufficient human oversight and even more woefully insufficient QA," which they believe leads to disastrous bugs reaching end users.

The disclosure also touches on sensitive issues regarding AI's impact on software jobs. By describing the feature as mainly implemented by an agent, including its design, with humans merely guiding it in their spare time, the narrative suggests a marginalization of traditional engineering roles. This timing coincides with Microsoft's significant workforce reductions and voluntary redundancy offers, making the promotion of AI efficiency appear tone-deaf to some observers. While Microsoft has openly acknowledged using LLMs in development, this specific case study intensifies the debate on responsibility and quality in AI-assisted coding.

ソース:TechRadar

リサーチ

EQT Global Chair Jean Salata: The Bottleneck for AI Data Centers Is Power Delivery, Not Contract Signing

トピック · AI数据中心电力交付瓶颈重大
2026-11-02 20:44 GMT+8

In AI infrastructure investment, there is an easily overlooked distinction: signing a power contract and actually delivering electricity to the server rack are two entirely different things. In an interview with Semafor, Jean Salata, global chair of EQT, used a specific term to label the former—'bragawatts' (bragging watts)—and used it to explain why this Nordic private equity giant, managing $400 billion in assets, has chosen a path distinct from most of its peers during the AI boom.

Salata’s core judgment is that power and energy assets are the true limiting factors in current AI construction. EQT had previously invested in power assets and data centers geared toward cloud migration; at the time, these investments were unrelated to AI. By 2026, however, these assets suddenly became the bottleneck for AI infrastructure—'We suddenly find ourselves holding $100 billion in energy assets, which are precisely the real limiting factor for AI right now.' EQT currently partners only with hyperscale cloud service providers (hyperscalers), with 92% of its data center contract revenue coming from such clients.

This contrasts sharply with the model relied upon by much of the current AI infrastructure financing. Salata points out that many people have signed massive contracts with anyone—they call them bragawatts—but signing something and delivering it on time, with power connected, are two very different things. He believes sophisticated investors understand this distinction, but many in public markets are investing in things they truly do not understand. This concern is well-founded: in the fourth quarter of 2025, investors withdrew more than $7 billion from the largest private credit funds, including those managed by Blackstone, Apollo, Blue Owl, and Ares; Blue Owl was even forced to raise its quarterly redemption cap from 5% to 17% and borrow money to meet payouts.

EQT’s overall stance on private credit is also noteworthy. Salata describes the credit business as a 'beta business,' arguing that its best-case scenario yields only an 8% return and is unrelated to EQT’s strengths in equity operational capabilities. EQT deliberately avoided chasing the path taken by US peers like Apollo and Blackstone, who expanded their asset management scale through private credit and insurance businesses.

However, Salata’s position comes with specific conditions. EQT was co-founded in 1994 by Sweden’s Wallenberg family, which remains the largest shareholder today. Its patient holding model resembles the long-hold structure that many private equity firms are attempting to transition toward. This means EQT can withstand longer asset realization cycles than typical ten-year closed-end funds. Salata himself acknowledges that the entire industry is making the same AI trade, and differentiation ultimately depends on execution rather than directional choice—'There are no patents in finance. We are all moving in the same direction, and it ultimately comes down to execution.'

For readers, this distinction has practical implications: when evaluating AI infrastructure investments, 'signed capacity' and 'delivered power' are two different dimensional metrics. Investors who own built-out power assets and have contracts with hyperscale cloud service providers face a risk structure different from those relying on credit exposure to the future revenue streams of new cloud companies. Salata used the term 'bragawatts' to name this execution gap, which has been obscured by market narratives.

ソース:semafor.com

リサーチ

Pillar Security Researchers: AI Coding Agent Sandboxes Isolate the Host, Not the Sandbox Itself

トピック · AI智能体沙箱隔离重大
2026-02-25

When local AI coding agents execute code on a user's machine, they typically operate with user-level privileges: access to the file system, network capabilities, and shell execution. In an analysis published in February 2026, Pillar Security researchers Eilon Cohen and Ariel Fogel summarized this dynamic with a core distinction: isolation technologies protect the host from code running inside the sandbox, but they do not protect the sandbox itself.

This means the sandbox is a containment measure, not a prevention measure. When trust is violated, the sandbox limits the blast radius, but it does not address the underlying trust issue. AI agents process content from sources that cannot be fully verified—model weights that execute code upon loading, package dependencies that include installation scripts, and user-provided repositories—each representing an injection vector. The sandbox is the last line of defense, not the first.

The authors analyzed fourteen sandbox solutions designed for coding agents, categorizing them into four isolation tiers. Containers share the host kernel, meaning a kernel vulnerability can breach the boundary; gVisor intercepts system calls to reduce the attack surface, but its own correctness becomes security-critical; hardware-virtualized microVMs provide an independent kernel for each sandbox, offering the strongest isolation but constrained by session limits and boot latency; and kernel capabilities (such as Landlock and Seatbelt) offer no escape path but are limited to local environments and cannot be deployed on multi-tenant platforms.

However, regardless of the tier chosen, a overlooked fact remains: a sandbox only isolates what you put inside it. Mount SSH keys, and they are accessible within the sandbox; grant full filesystem read access, and credential files can be read; return standard output to the large language model, and any printed content can be exfiltrated. Cursor already suffered credential exposure due to full filesystem read access in November 2025; even with writes restricted to the workspace, the agent could still read credential files using the cat command and output the content to stdout.

CVE-2026-22708, disclosed in January 2026, further exposed the structural flaws of static allowlists. Cursor’s allowlist mechanism validates commands line-by-line for safety, but shell built-in commands can modify environment variables without confirmation, turning subsequent allowed commands into attack vectors within a poisoned context. A command that appears safe in isolation is no longer safe once the surrounding environment has been manipulated.

From this, the authors propose a selection framework: the correct isolation tier depends on two questions—what are you isolating from, and what are you isolating with? If the threat is untrusted user-provided code, a microVM combined with violation detection is a reasonable choice; if the threat is third-party packages with installation scripts, defense-in-depth is more appropriate; if the threat is prompt injection via external content, network isolation combined with restricted filesystem read access is key. Conversely, if production credentials exist in environment variables, they are inside the sandbox; if full filesystem read access is granted, credential files are readable; if stdout is returned to the model, anything printed is exfiltrable.

ソース:pillar.security

リサーチ

Security Firm Trail of Bits Reports Using Codex to Cut Linux Patch Backporting Time from Days to Hours, Fixing 149 CVEs

重大
2026-11-02 20:00 GMT+8

US-based cybersecurity consulting firm Trail of Bits published a report on November 2, 2026, stating that it used OpenAI's Codex tool to reduce the human effort required for backporting security fixes to stable Linux versions from days to hours. By October 29, the project had helped deliver fixes for 149 distinct CVEs into stable Linux kernel releases, along with several fixes to Fedora and Ubuntu packages.

Traditionally, adapting security fixes from the latest codebase to older versions is a time-consuming and thankless task, often requiring a week or more of human effort per bug. The new workflow described by Trail of Bits involves AI agents performing activities such as identifying missing fixes, adapting patches, assembling validation evidence, and running tests, while human engineers review the patches and submit them to maintainers. According to the company's self-report, excluding initial setup, preparing and reviewing a straightforward backport now takes about 30 minutes of human attention; more involved cases take around three hours.

This development aims to address the long-standing tradeoff between "stability and security." Data shows that the vast majority of users (e.g., 95% of Ubuntu installations) rely on Long Term Support (LTS) versions, leaving many known vulnerabilities unpatched in stable branches. Trail of Bits notes that as AI-assisted vulnerability hunting becomes more capable, accelerating the security maintenance of these legacy systems has become increasingly urgent.

ソース:The Trail of Bits Blog

リサーチ
次へ →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

A microVM with mounted AWS credentials is not meaningfully more secure than a container with the same credentials. Choosing an isolation tier is an architectural decision, not a procurement decision. Its importance is secondary to the threat model it addresses and the trust boundaries it actually enforces.

It should be noted that Pillar Security is an AI security product company, and this analysis simultaneously serves to promote its SAIL 2.0 framework and Agentic Endpoint platform. The assessment that Claude Code had no known sandbox vulnerabilities as of the analysis date is a snapshot at a specific point in time and does not represent a permanent state.

For developers, actionable checks include: What credentials exist in your development environment? Does the agent have full filesystem read access? What is mounted into the execution environment by default? For security teams: Does the sandbox restrict read access, or only write access? Is there a mechanism to detect sandbox boundary violations? Who controls the sandbox configuration—the vendor, the platform team, or the end user?