LectureRechercheRadarCadre d'investissement
Connexion / Inscription
ToutTechnologieProduitsInvestissementBourseAutres

Lecture

2026-10-28154 publications

Actualiser

Im Jae-hwan, Former Chairman of the Korea EdTech Industry Association: Educational Sovereign AI Is Not About Building Walls, But Synergy in Governance and Industrial Ecosystem

Sujet · 韩国教育主权AIMatériel
2026-10-28 15:00 GMT+8

The application of AI in South Korea's public primary and secondary education has stalled, while universities and corporate training programs rapidly absorb artificial intelligence technologies. In an expert commentary published on October 28, 2026, in The Electronic News (Electronic Daily), Im Jae-hwan, CEO of EdTech firm Ubion and former chairman of the Korea EdTech Industry Association, pointed out that this stagnation is not due to a lack of technical capability. Rather, it results from a tangle of social conflicts, emotional resistance, and feelings of marginalization within the EdTech industry accumulated during past pushes for AI digital textbooks, which have prevented the formation of a symbiotic ecosystem between the public education sector and the industry.

Im believes that breaking through this destructive cycle of pro-and-con debates requires a new path centered on 'educational sovereign AI,' proposed at the national level. He explicitly distinguishes this concept from technological nationalism: educational sovereign AI is listed as a mandatory national task alongside defense, health, and manufacturing. Its core logic lies in protecting national educational sovereignty and data while balancingly managing the benefits and risks brought by AI, rather than simply erecting walls against technology.

He proposes distinguishing between necessary and sufficient conditions for educational sovereign AI. At the necessary condition level, an independent AI education data center and management governance architecture must be built to securely protect learning data and support public education models. However, he emphasizes that the true sufficient conditions do not lie in infrastructure itself, but in the correct application of AI in classrooms—specifically, the organic interlocking of innovative AI pedagogy, teachers equipped with AI capabilities, and a private EdTech ecosystem supporting this process. While recent moves by South Korean education authorities to emphasize foundational literacy skills such as reading and writing as prerequisites for AI literacy represent a correct diagnosis, Im considers this merely a minimum condition.

At the institutional level, Im calls for a comprehensive revision of the 'E-Learning Industry Development Act,' drafted over twenty years ago, or its transformation into an 'EdTech Industry Promotion Act.' He notes that the National Assembly of South Korea has already submitted amendments to fully transition the E-Learning Industry Development Act to the AI era. The core content involves changing the law's name to the 'EdTech Industry Promotion Act' and expanding the definition of EdTech from online learning to cover a complete industrial ecosystem encompassing software, content, data, platforms, and hardware. He also recommends establishing a dedicated EdTech Revitalization Institute and an EdTech Support Center as governance bodies to drive public-private cooperation.

Im sets the goals of educational sovereign AI in two directions: quality innovation in public education and the acquisition of global competitive advantage. He particularly emphasizes that true educational sovereign AI does not build walls for oneself; instead, it uses South Korea's proven models to help global partner countries establish their own educational sovereignty, forming a symbiotic ecosystem, while simultaneously serving as a driving force for K-EdTech exports.

Sources :미래를 보는 창 - 전자신문

Recherche

SK Hynix Says DRAM and HBM Capacity Effectively Sold Out Through 2029, Announces ~100 Trillion Won Shareholder Return Plan

Sujet · SK海力士Q3利润暴增Matériel
2026-10-28 10:22 GMT+8

South Korean semiconductor maker SK Hynix disclosed during its third-quarter earnings conference call on October 28, 2026, that its DRAM and High Bandwidth Memory (HBM) capacity for the next three years, through 2029, is effectively sold out. This assessment combines volumes secured under existing Long-Term Agreements (LTAs) with demand currently being specifically negotiated with customers. The statement directly counters recent market concerns regarding an "AI peak" and potential slowdowns in HBM demand.

Financial results showed Q3 revenue reaching 104.6154 trillion won, up 327.9% year-over-year, and operating profit hitting 83.3338 trillion won, a 632.1% increase, with an operating margin exceeding 80%. This marks the first time the company's quarterly revenue surpassed 100 trillion won. On the demand side, SK Hynix expanded its LTA-signed customer base from 10 to 14, including major global cloud service providers (CSPs). Secured LTA volumes account for over 50% of total production capacity for 2027 and 2028, with non-LTA capacity to be flexibly adjusted based on market needs.

Addressing speculation that a reduction in HBM stacking layers (e.g., from 12 to 8) might signal weakening demand, SK Hynix explained this as a "desperate measure" by customers to build more AI accelerators within limited supply, aiming to increase the number of available chips per capacity unit. The company emphasized that even accounting for such adjustments, current customer requests exceed its supply capability by more than double. Additionally, the next-generation HBM4 is expected to surpass the incumbent HBM3E in both shipment volume and revenue share in the fourth quarter.

Regarding capital allocation, SK Hynix announced a shareholder return plan totaling approximately 100 trillion won for fiscal year 2026. This includes over 30 trillion won in cash dividends and over 60 trillion won in share buybacks and cancellations, fulfilling the commitment to return more than 50% of free cash flow (FCF) to shareholders. The board resolved to pay a Q3 cash dividend of 14 trillion won (20,000 won per share), which includes a special dividend of 10,000 won per share reflecting investment gains from Japanese memory chipmaker Kioxia. A previously announced 40 trillion won share buyback was completed last week and will soon undergo cancellation procedures.

Sources :zdnet.co.kr

Recherche

Africa's Largest-Ever IPO: Dangote Refinery Raises $1.6 Billion at Nearly $50 Billion Valuation

Sujet · 丹格特炼油厂IPOMatériel
2026-10-28 20:39 GMT+8

Dangote Refinery, owned by Nigerian billionaire Aliko Dangote, has completed its initial public offering (IPO), marking the largest such event in African history. According to Semafor, the deal closed on Tuesday, raising $1.6 billion and valuing the company at nearly $50 billion. The proceeds will support a more than $14 billion plan to expand capacity at the giant refinery in Lagos State.

A notable feature of this offering is its investor base, which was drawn primarily from within the African continent rather than traditional financial centers like London or Johannesburg. Dangote stated that this serves as a rare demonstration that Africa's savings can move across borders at the scale needed to fund its biggest companies. Multilateral lender Africa Finance Corporation estimates the continent's domestic savings at approximately $4 trillion. This event is widely seen as a test case for whether these domestic funds can deepen regional capital markets.

However, the deal poses a challenge to the Johannesburg Stock Exchange (JSE), which has long traded on its status as Africa's most liquid capital market, larger than the rest of the continent's bourses combined. Because the IPO largely bypassed the JSE, the exchange stated that local banks were "exploring JSE-listed instruments" to offer exposure to Dangote. Although the JSE noted in August that Dangote had "strong intent" to pursue a South African listing, the current strategy clearly prioritizes the mobilization of pan-African capital.

Sources :semafor.com

Recherche

Kodiak AI Signs First Commercial Contract for Public Highway Autonomous Trucks with DTL Transport

Sujet · Kodiak AIMatériel
2026-10-28 21:00 GMT+8

US-based autonomous trucking startup Kodiak AI has announced its first binding commercial contract for driverless heavy-duty trucks operating on public highways. Under the agreement, Kodiak will deliver 75 self-driving trucks to mid-market freight carrier DTL Transport starting in 2028. The deal comes just ahead of Kodiak's planned milestone in December to remove human safety operators from the cab and launch fully driverless highway services.

This contract marks a significant pivot in Kodiak's business model. Previously, the company primarily operated under a "Trucking-as-a-Service" (TaaS) model, where Kodiak owned and operated the trucks to haul goods for other companies. The new deal with DTL Transport utilizes a "Driver-as-a-Service" (DaaS) model, where the customer owns and operates the vehicles while Kodiak licenses its technology for a per-mile fee. DTL Transport, a California-based carrier serving grocery chains, plans to deploy these trucks on interstate routes including Dallas to El Paso and Dallas to Atlanta.

Kodiak founder and CEO Don Burnette told TechCrunch in an exclusive interview that partnering with mid-market carriers allows for faster execution and easier fleet integration, calling it a turning point for the self-driving industry. While Burnette stated the deal would provide a revenue boost, the specific contract value remains undisclosed. Based on Kodiak's 2025 investor presentation, which forecasts annual recurring revenue of $150,000 to $190,000 per truck, the 75-unit order could potentially generate approximately $11.2 million in annual recurring revenue once fully delivered.

Kodiak went public in September 2025, but its shares have fallen 87% from their opening-day price. This transaction supports the company's strategic goal of becoming an "asset-light" technology supplier, a path also pursued by competitor Aurora Innovation. Currently, Kodiak conducts pilot operations for TaaS customers in Texas and provides off-road DaaS services to Atlas Energy Solutions in the Permian Basin.

Sources :TechCrunch

Recherche

Radio Free Mobile Founder Richard Windsor: Long-Term Storage Contracts Cannot Stop the Cycle Reversal

Sujet · 存储长期协议与周期博弈Matériel
2026-10-28 14:27 GMT+8

SK Hynix delivered a stellar performance report for Q3 2026: revenue reached KRW 104.6 trillion and earnings per share were KRW 86,300, both exceeding market expectations. The company once again declared that the memory cycle has ended, forecasting that demand will exceed its supply capacity until 2030, with capacity for 2027 already sold out. Independent technology analyst and Radio Free Mobile founder Richard Windsor agrees that the outlook for the next twelve months remains strong, but he believes the industry's judgment that the cycle has ended will come back to haunt these companies around 2028.

Windsor's core argument revolves around long-term agreements. The storage industry is leveraging its current market dominance to migrate customers to long-term contracts with relatively fixed prices and volumes; SK Hynix has already secured 50% of its production capacity for 2027 and 2028 under these agreements. He argues that these contracts can indeed be executed as long as DRAM prices remain flat or stable, but the problem arises at the moment the cycle reverses.

His reasoning mechanism is as follows: once the industry enters a state of oversupply (which he believes is inevitable), someone will inevitably break ranks, compete for customers by offering lower storage prices, and customers will subsequently breach their contracts. The key constraint here lies in the market structure—the storage market is fairly concentrated, and the last thing one wants to do in a commodity market is to start suing customers for breach of contract. Therefore, these contracts are largely ineffective at preventing a price crash.

Windsor also points out that this turning point is still some distance away. AI models that "think" longer to improve results and agents all require substantially more memory than the use cases that came before them, meaning that storage demand is likely to continue outpacing the growth rate of any other component in the industry. He suspects that SK Hynix is underestimating growth expectations for 2027 in the short term, but the fragility of long-term agreements will be exposed when oversupply truly arrives.

At the end of the article, Windsor disclosed that he holds a position in Samsung Electronics, stating that he remains comfortable with it but acknowledging that the stock often tests his patience.

Sources :radiofreemobile.com

Recherche

Nvidia's Vera CPU Goes Live as Standalone Instance on Oracle Cloud, Benchmarks Show Bandwidth Lead

Sujet · 英伟达Vera CPUMatériel
2026-10-28 19:00 GMT+8

Nvidia and Oracle announced on Tuesday at Oracle AI World that Nvidia’s Vera Rubin NVL72 rack and Vera CPUs are now generally available on Oracle Cloud Infrastructure (OCI) Superclusters. Crucially, the announcement highlights "standalone" Vera instances, meaning pure CPU deployments not tied to GPU compute. This marks a strategic expansion for Nvidia, pushing its Arm-based Vera CPU into the general-purpose server market previously dominated by x86 architectures from Intel and AMD.

While Google Cloud recently announced Vera Rubin availability, Oracle's move to offer standalone Vera instances sends a different signal. Hyperscalers typically rely on x86 for their CPU fleets, with Arm options often coming from third parties like Ampere. By placing Vera alongside x86 competition regardless of GPU attachment, Nvidia is directly challenging the status quo, particularly as agentic AI workloads grow.

Technology outlet Tom's Hardware conducted independent benchmarking on a local Vera server, revealing significant advantages in memory bandwidth. Using the Stream benchmark, Vera leveraged LPDDR5X-9600 memory and a distributed mesh cache to more than double the bandwidth available on a dual Intel Xeon 6980P system equipped with MRDIMMs. This high bandwidth is critical for agentic AI loops that involve large database queries or extensive document processing.

Vera also demonstrated strong Python performance, which underpins many agentic frameworks like CrewAI and LangGraph. Tests showed that Nvidia’s second-generation Scalable Coherency Fabric (SCF) reduced data movement overhead in real-world workloads such as SQLite database queries.

However, these tests were performed on local hardware, not OCI infrastructure, so actual cloud performance may vary due to virtualization overhead. Oracle stated it will proceed cautiously with this multi-gigawatt deployment, also offering HGX NVL8 clusters attached to dual Intel Xeon 6 hosts. The event underscores a fundamental shift in server CPU selection criteria as AI workloads move beyond simple matrix math to complex logic and data retrieval.

Sources :Tom's Hardware

Recherche

Cisco Announces MCP Servers and Native ThousandEyes Integration for Network Platforms at Partner Summit

Sujet · 思科开放网络转型Matériel
2026-10-28 20:33 GMT+8

U.S. networking giant Cisco announced at its 2026 Partner Summit in Vancouver that it is expanding its network platforms to support AI-assisted operations. Key updates include the release of standardized Model Context Protocol (MCP) servers for platforms such as Meraki, Catalyst Center, and the ISE network access control solution, as well as the native integration of ThousandEyes network testing features into supported Wi-Fi 6E and Wi-Fi 7 access points.

The new MCP servers allow external AI tools and agents to securely access configuration and telemetry data from Cisco's network platforms. This move aims to simplify automated error analysis, troubleshooting, and reporting for customers and partners, making it easier to embed network data into enterprise-owned AI agents and workflows. The servers are documented on Cisco's DevNet developer portal and available for download on GitHub.

Additionally, Cisco announced the direct integration of ThousandEyes functionality into supported wireless access points. This enables key network services and applications to be tested automatically and continuously in the background, even when no clients are connected to the Wi-Fi. Previously, ThousandEyes agents typically required separate hardware or ran as containers on switches. The feature requires specific Cisco Networking Essentials or Advantage licenses.

To address potential security risks associated with AI agents, Cisco also introduced 'Privileged Infrastructure Access' in collaboration with Teleport. This feature allows administrators and AI agents to obtain time-bound and task-specific access rights, using short-lived certificates instead of permanent privileged access, and records all sessions to ensure control over critical infrastructure is not compromised.

Sources :iX Magazin

Recherche

Legal AI Company Legora Says Internal NDA Review Capacity Share Dropped from Approximately 20% to Below 1%

Sujet · Legora融资动态Prise rapide
2026-10-28 17:30 GMT+8

A structural dilemma for legal teams is that there is never enough time, leading to high-value work that should have been completed simply not getting done. Max Junestrand, co-founder and CEO of Stockholm-based Swedish legal AI company Legora, cited this as the company's starting point in a case study released by OpenAI: junior lawyers were asked to review an overwhelming number of contracts overnight, able to focus on only a small fraction, while the rest were either skipped or handled roughly.

Legora's legal team designed a specific mechanism for reviewing NDAs (Non-Disclosure Agreements). Emma Kadri Bergström, the General Counsel, explained that her team first analyzed all NDAs the company had historically signed to extract a review playbook defining which term parameters fall within acceptable ranges. When new NDAs arrive, an AI agent performs an initial screening based on this playbook, flagging terms that exceed acceptable parameters for lawyers to make final judgments. Bergström estimated that before this workflow was introduced, reviewing these tedious documents consumed about 20% of the legal team's capacity; now, that figure has dropped to below 1%.

The key to this mechanism is not that AI replaces lawyers, but that it changes the scope of the review. With limited capacity, teams could only deeply review the most critical few contracts, skipping or quickly scanning the rest. The AI's initial screening allows a large volume of files that would previously have gone unexamined to be processed, thereby expanding the exposure of risk clauses. Alex Fortescue-Webb, Global Head of Legal Engineering at Legora, used due diligence as an example: traditional reviews are constrained by time and cost, forcing teams to concentrate on the most valuable contracts; the Tabular Review workflow can analyze thousands of contracts across multiple questions in minutes, allowing lawyers to apply detailed scrutiny to a much broader range.

The engineering team's approach follows a similar logic of parallel exploration. CTO Jacob Lauritzen stated that coding agents have changed the speed at which new ideas are developed: in the past, the team would research one possible solution, commit engineering time, and then find out if the direction was correct; now, they can explore 20 solutions simultaneously and select the best one within days. In troubleshooting, coding agents can search through millions of lines of logs and track multiple possible explanations simultaneously, sometimes compressing the time from bug report to fix completion to approximately five minutes.

Regarding organizational adoption, the case study lists one lesson: when more than 75% of leaders actively use Legora products, the median usage across the entire organization is three times higher. Lauritzen also noted that the rapid iteration of model capabilities forces him to reconsider the product roadmap every two weeks.

It must be pointed out that all the above data—the shift in capacity from 20% to 1%, the 90% acceleration in contract turnaround, the savings of over $20 million in external litigation and due diligence fees for a single Fortune 100 client, and the return of 9.5 hours per lawyer per week—come from self-statements by Legora management in the case study published by OpenAI. The 20% baseline is an estimate by the General Counsel rather than a system measurement; the $20 million figure does not specify the time frame or calculation method; and the 90% acceleration applies only to one unnamed client. As a success story written by a model provider for a paying customer, positive selection bias is a structural feature. The value of this case lies in providing a description of specific workflow mechanisms, rather than proving any universal performance conclusions.

Sources :OpenAI

Recherche

UK Circuit Board Design Startup Circuit Mind Raises $25M, Launches AI Schematic Verification Engine

Sujet · Circuit MindPrise rapide
2026-10-28 20:00 GMT+8

UK-based circuit board design startup Circuit Mind Ltd. announced on October 28, 2026, that it has raised $25 million in new funding to extend its artificial intelligence platform across more stages of the hardware design process. The company also launched a new tool called the Agentic Verification Engine (AVE), which uses AI agents to automatically hunt for errors in circuit schematics.

Circuit Mind's core business involves automating much of the manual work behind printed circuit board (PCB) design. According to the company, engineers can spend more than a year on a new piece of hardware, partly reading datasheets that run to thousands of pages, with mistakes often remaining hidden until a physical prototype is built and tested. Circuit Mind argues that general-purpose large language models are not up to the task on their own because they have never been trained on professionally designed circuits. Its platform layers AI agents over a patented, physics-based engine that tests each design decision against electrical limits.

The newly launched AVE hands design review work to a swarm of AI agents. Working from component datasheets and any review checklists loaded by customers, these agents trace every connection and bus in the circuit to find mismatches with pin tables or operating voltages. Nothing reaches an engineer until the system has critiqued its own findings.

Regarding customer results, Circuit Mind cites a public report from Los Alamos National Laboratory in New Mexico, stating that lab engineers finished a medium-difficulty board in just over four hours using its design application, compared to an estimated 60 to 80 hours by hand. Daniel Gheno, Chief Technology Officer for Energy Management at Schneider Electric SE, said early evaluations point to a cut of more than 90% in the effort spent exploring electronic architectures. Charlie Eastberg, Director of Hardware Engineering at Emerson Electric Co., confirmed that both applications sped up his team's move from system architecture to schematic.

The $25 million consists of two rounds: Norwest Venture Partners led an $18.5 million Series A round with participation from new investor Lightscape Partners and existing backers Third Prime, Differential Ventures, and Blackhorn Ventures. The remainder comes from a previously unannounced seed round led by Third Prime. Dave Zilberman, a general partner at Norwest, will join Circuit Mind's board. Including this new funding, investors have put a total of $33.4 million into the business. Circuit Mind plans to use the funds to push the platform into further stages of the design lifecycle, including testing and firmware, and to establish a U.S. headquarters to serve North American customers.

Sources :SiliconANGLE

Recherche

Seoul's First Conversational AI Voice-Recognition Bus Stop Launches in Cheongnyangni

Sujet · BICNS首尔AI公交站Prise rapide
2026-10-28 16:23 GMT+8

Seoul's Dongdaemun District has launched its first "conversational" AI voice-recognition smart bus stop at the Cheongnyangni Station Bus Transfer Center. The kiosk was installed on October 22, 2026, by South Korean data and AI technology company BICNS (비아이씨엔에스), which announced the start of pilot operations on October 28. This deployment follows a partnership agreement signed with the Dongdaemun District government in January 2026.

Unlike traditional touchscreens or mobile apps, the kiosk allows passengers to ask questions in natural language. For instance, a user can simply ask, "When does the bus to City Hall arrive?", and the AI will provide arrival times, the fastest route, and transfer information via both voice and screen. It also offers guidance on nearby facilities such as subway stations, convenience stores, and pharmacies. Supporting 12 languages including Korean, English, Chinese, Japanese, and Vietnamese, the system aims to assist digitally vulnerable groups, including the elderly, people with disabilities, and foreign tourists.

To address the technical challenge of voice recognition in environments where vehicle noise often exceeds 80 decibels, BICNS implemented "noise-canceling beamforming" technology to filter out background noise and focus on the speaker's voice. The current pilot phase will verify recognition performance and user satisfaction in these real-world noisy conditions. If successful, the company plans to expand the installation to major bus stops across Dongdaemun District and subsequently to other public spaces like traditional markets and community centers.

Sources :미래를 보는 창 - 전자신문

Recherche

Japanese Chip Designer Socionext Returns to Profit in Q2, Raises Full-Year Guidance on Weak Yen

Sujet · Socionext业绩Prise rapide
2026-10-29 03:30 GMT+8

Japanese semiconductor design firm Socionext reported on October 28, 2026, that it returned to profitability in the second quarter of fiscal year 2026 (July–September). Revenue reached 60.4 billion yen, up 14.7% year-on-year; operating profit was 2.3 billion yen, flat year-on-year; and net income rose 31.3% to 2.1 billion yen.

The recovery was primarily driven by the clearance of delayed shipments. In the previous quarter (April–June), approximately 10 billion yen worth of automotive products were postponed due to a review of production and shipping systems, resulting in a loss. The successful delivery of these goods in Q2 restored black ink. Although the change in product mix lowered gross margins compared to the prior quarter, increased sales volume boosted total product gross profit by 4 billion yen.

By segment, while telecommunications revenue declined, growth in automotive and industrial sectors lifted product sales by 12.8% year-on-year to 50.2 billion yen. Non-Recurring Engineering (NRE) revenue, collected during the design phase, surged 28.4% to 10.2 billion yen, supported by demand in data center and network applications.

Citing a weaker yen against the dollar than initially anticipated, Socionext revised its full-year fiscal 2026 guidance upward. The company now expects annual revenue of 225 billion yen (up 4.7% from April forecasts), operating profit of 15.5 billion yen (up 10.7%), and net income of 10.5 billion yen (up 5%).

Sources :EE Times Japan

Recherche

SK Hynix Reports Sevenfold Jump in Q3 Profits as AI Boom Drives South Korean Growth

Sujet · SK海力士Q3利润暴增Matériel
2026-10-28 19:02 GMT+8

South Korean memory chip maker SK Hynix reported a sevenfold increase in its third-quarter profits compared to a year ago. This strong performance was primarily driven by surging demand for artificial intelligence infrastructure, marking the latest sign of a tech boom fueling nationwide economic growth in South Korea. According to Semafor, the earnings release coincided with official data showing South Korea's GDP growth beating analysts' expectations, prompting Goldman Sachs to raise its forecasts for expansion this year and next.

However, this prosperity driven by a single high-tech sector has sparked concerns about macroeconomic structural imbalances. While semiconductor exports performed exceptionally well, private consumption in South Korea declined. The Bank of Korea warned earlier this year of "semiconductor Dutch disease," referencing the 1960s gas boom in the Netherlands which strengthened the currency and crimped demand for goods from other sectors, suggesting that over-reliance on semiconductors could harm the health of other parts of the economy.

Sources :semafor.com

Recherche
Suivant →
Connexion / Inscription
LectureRechercheRadarCadre d'investissement

It is worth noting that Im's identity is that of an EdTech enterprise manager and former industry association chairman. His advocacy for industry-promoting legislation, the establishment of revitalization agencies, and ecosystem construction is directly related to the commercial interests of his sector. Furthermore, the binding of educational sovereignty with industrial exports in his argument raises questions about whether the goal of educational publicness should be subordinated to the logic of industrial growth.

The original article, titled regarding how to advance educational sovereign AI, fully presents Im's policy proposals. Its core contribution lies in disentangling sovereign AI from the popular understanding of technological protectionism, proposing a dual-layer condition framework of governance infrastructure and industrial ecosystem, and providing specific legislative and institutional pathways. The value of this argument does not lie in proving that a certain policy will inevitably succeed, but in revealing the institutional bottlenecks causing the stagnation of AI in South Korea's public education, as well as the specific visions within the industry for breakthrough paths.