Reading

According to Crunchbase data, global travel tech startups raised $2.2 billion across 302 funding rounds in the first three quarters of 2026. This represents a significant drop from the $3.7 billion raised in 2025, with deal volume declining for a third consecutive year, indicating that investors are putting more money into fewer companies.

Despite a wave of startups promising to improve travel planning and booking with AI, this trend has not reversed the funding slowdown. For instance, Amsterdam-based hospitality software company Mews closed a $300 million Series D led by EQT Growth in January at a $2.5 billion valuation, and Chinese electric aircraft developer Aerofugia raised about $144 million in February. However, these large deals failed to boost overall sector investment heat.

Meanwhile, established travel companies are acquiring startups to enhance their AI capabilities. Amadeus acquired corporate travel AI assistant SkyLink in February, while Expedia bought AI trip-planning startup Layla in July. This M&A activity reflects industry consolidation rather than a recovery in independent startup funding capacity.

EMCOR Group (NYSE: EME), a US-based provider of electrical and mechanical construction services, reported record third-quarter results on October 29, 2026, driven primarily by demand for artificial intelligence data center construction. Quarterly revenue rose 28.8% year-over-year to $5.54 billion, including 25.5% organic growth. Adjusted diluted earnings per share increased 46.1% to $9.60, while the adjusted operating margin expanded by 110 basis points to 10.5%.

The Electrical Construction segment saw revenue increase 44.3% to $1.85 billion, and Mechanical Construction revenue climbed 43.1% to $2.55 billion. Together, these segments generated $4.4 billion in quarterly revenue with a combined operating margin of 12.6%. Network and communications work, largely consisting of data center projects, was a major growth driver: Electrical Construction revenue in this sector grew 68%, while Mechanical Construction revenue nearly doubled (up ~130%), aided by expanded cooling requirements on data center sites.

Remaining performance obligations (RPOs), a key indicator of future revenue visibility, reached a record $18.57 billion at quarter-end, up 47% from a year earlier. Notably, data center RPOs totaled $8.74 billion, more than double the prior-year level and up 13% sequentially. Chief Financial Officer Jason Nalbandian noted that most of the RPO increase was organic despite strong current revenue growth. CEO Guzzi stated that EMCOR primarily serves large hyperscalers and their support companies, rather than speculative 'neocloud' markets.

Based on continued demand, EMCOR raised its full-year 2026 outlook, now expecting revenue between $20.6 billion and $21 billion and adjusted diluted EPS between $34.50 and $35.50. The board also authorized an additional $500 million for common-stock repurchases. Through September 30, the company had returned $510 million to shareholders via dividends and buybacks.

The Linux 7.4 kernel has officially merged the AMD PerfOpt feature, which allows AMD integrated graphics (iGPUs) to bypass the Input-Output Memory Management Unit (IOMMU) when directly accessing system memory. This low-level driver optimization significantly reduces memory access latency, delivering substantial performance gains for local AI inference tasks that rely on shared system memory.

According to benchmarks published by Phoronix, enabling PerfOpt results in an 18% to 23% performance increase for AI and Large Language Model (LLM) workloads on AMD integrated graphics. The feature is enabled by default for Radeon integrated graphics systems using the AMDGPU kernel graphics driver. Users can verify if the feature is active on Linux 7.4 and above by checking the dmesg kernel output for the string "PerfOpt".

The update also includes other IOMMU-related improvements, such as Broadcom BCM2712 support for Raspberry Pi 5 display and multimedia devices, and a new para-virtualized IOMMU driver contributed by Microsoft for use with Hyper-V.

Robotic vision has established a relatively unified system of sensors and data, but tactile sensing remains an un-converged weak link in embodied intelligence systems. At the tactile session of IROS 2026, Li Yunchu, a professor at Columbia University (United States), pointed out that different laboratories often use fully self-developed, mutually incompatible tactile sensors. These sensors collectively face the dilemma of struggling to balance lifespan, flexibility, multi-dimensional information perception, resolution, cost, and consistency simultaneously. This makes tactile data collection, algorithm research, and reproducibility across different systems significantly more difficult.

The strategy proposed by Li Yunchu's team is not to pursue more complex sensors, but to deliberately choose a sufficiently simple sensing mechanism. FlexiTac is an open-source flexible piezoresistive tactile sensor featuring a layered structure of Flexible Printed Circuits (FPC) and force-sensitive films. The team has also open-sourced the complete Bill of Materials (BOM), detailed assembly instructions, and software for reading data. Its core design logic is: because the sensing mechanism is simple, simulation implementation only requires building a spring-damper system and calibrating 3 hyperparameters to achieve high-fidelity alignment between simulated tactile signals and real-world signals.

This simplification brings chain reactions. In terms of assembly, once materials are prepared, it takes only 3 minutes to complete; in terms of form, the sensor is like a sheet of paper that can be bent, cut, and wrapped around various irregular curved surfaces, adapting to gloves, fingertips, or robot skin. Regarding consistency, Li Yunchu cited an example: his student Binghao made four sensors in the summer of 2024 to record videos. A year and a half later, using this same set of sensors, a model trained on data collected a year ago still worked perfectly. Output signals also showed good consistency across different batches, addressing a common pain point for many tactile sensors containing rubber materials.

Simulation adaptability is the key hub of this strategy. Li Yunchu's approach involves first collecting random interaction data in the real environment, then replaying the same actions in the simulation environment to collect two sets of tactile data and plot histograms. The calibration goal is to align these two distributions. He emphasized: the more precisely simulation aligns with reality, the less randomization is needed in simulation, making sim-to-real transfer more accurate. The same framework has been validated in laboratory dual-arm assembly tasks and on NVIDIA humanoid robots.

In scaling real-world data, FlexiTac fills a gap in the UMI handheld teleoperation device. The original UMI only recorded precise pose tracking information, ignoring the tactile feedback provided by internal mechanical springs to human operators. By integrating FlexiTac, the tactile feedback felt by human operators can be indirectly recorded as sensor output signals, allowing collected data to contain both visual and tactile information synchronized in time.

Japanese electronic component manufacturer Murata Manufacturing announced on October 30, 2026, during its second-quarter earnings briefing, that it plans to invest approximately ¥1 trillion in its multilayer ceramic capacitor (MLCC) business between fiscal years 2026 and 2030 to meet growing demand from AI data centers and other sectors. President Norio Nakajima described this as the biggest opportunity since the company's founding.

Under the new plan, Murata will increase its annual MLCC production capacity growth rate to approximately 20% for fiscal years 2027 and 2028, up from the previously planned 10%, aiming to raise total capacity by about 60% by fiscal year 2030. Of the ¥1 trillion investment, roughly ¥350 billion is allocated for building construction and ¥650 billion for production equipment. Murata currently holds about 40% of the global MLCC market share.

The expansion primarily targets high-capacity, miniaturized, and high-voltage MLCCs used in AI servers and optical communications. Nakajima noted that while AI data center demand might plateau around 2030, subsequent adoption in humanoid robots, autonomous vehicles, and drones will continue to drive market growth, making the long-term risk of such large-scale investment manageable.

Concurrently released financial results showed Murata's sales for April–September 2026 reached ¥1.0907 trillion, up 20.8% year-on-year, with net profit rising 40.6% to ¥186.1 billion. The company raised its full-year guidance for fiscal 2027 (ending March 2027) to ¥2.26 trillion in sales and ¥396 billion in net profit.

A deadly explosion at Intel's Fab 28 near Kiryat Gat, Israel, occurred earlier this month, killing two workers and seriously injuring another. The incident forced a complete two-week shutdown of the facility. Intel CFO David Zinsner stated during an earnings call that production is expected to restart by the end of this week, but the disruption will cost the company approximately $600 million in lost revenue across the fourth quarter of 2026 and the first quarter of 2027.

Fab 28 is a key manufacturing site for chips produced using Intel's 10nm-class process technologies, including Intel 7 (formerly 10nm Enhanced SuperFin). While Intel's latest Xeon 6 processors use compute dies fabricated on Intel 3, their I/O dies are still manufactured on Intel 7. Additionally, previous-generation Xeon 5 Emerald Rapids and Xeon 4 Sapphire Rapids CPUs continue to be sold and are made on Intel 7 technology. Zinsner noted that the disruption primarily affects 10nm-class server products and is expected to reduce Q4 non-GAAP gross margins by approximately 2%.

Although Intel has other fabs capable of 10nm production, such as Fab 12, Fab 22, Fab 32, and Fab 42 in Arizona, the company cannot fully offset the output loss from Fab 28 due to strong existing demand for server processors. Transferring production between facilities involves high costs and time, and some product lines may lack specific manufacturing qualifications at other sites. Beyond lost production time, Intel faces potential write-offs for wafers undergoing fabrication during the accident, with most impacts concentrated in Q4 2026.

In a podcast interview in late October 2026, Thai Duong, founder of cybersecurity firm Calif, put forward a judgment that sounds simple but carries profound consequences: if you train an AI to be good at writing code, you simultaneously train it to be good at hacking. This is not a bug that can be fixed; it is a structural property of the capability.

Duong’s reasoning chain is as follows. AI companies make money by selling tokens, and the current largest use of tokens is programming. A truly proficient programming AI must understand code, understand technical systems, be able to debug, and control computers—and Duong argues that this is precisely the definition of hacking. You cannot allow an AI to reach commercial levels of proficiency in programming while keeping it blind to system vulnerabilities.

This judgment is supported by concrete operational evidence. Duong worked at Google for twelve years, responsible for security testing before products like Chrome were released. He and the entire industry assumed that Chrome’s security investments had yielded results—not only eliminating vulnerabilities but also implementing mitigations that made them difficult to exploit. However, when they directed AI at Chrome, they discovered approximately one thousand vulnerabilities, ten times the expected amount. Duong said this surprised him greatly: either security engineers were underperforming, or the volume of code was so vast that manual auditing could never cover it all.

In contrast, the process of discovering PHP vulnerabilities illustrated the differing roles of AI across stages. Calif secured funding and tokens through OpenAI’s Patch the Planet program. Engineer Ryan Osteen found the first vulnerability just one or two days after launching the effort on July 30. But what made the situation urgent was not the discovery itself—the PHP Foundation had identified the same vulnerability back in 2024—but rather the exploitation. The foundation had chosen not to fix it because the repair cost was too high and there was no publicly available exploit method at the time. AI’s key contribution lay in developing the exploit code: engineers provided the AI with design directions for how to leverage the flaw, and within a day, the AI produced a working exploit, proving that remote attackers could compromise websites without any specialized knowledge.

Duong also described a collapsing time window. In the past, it typically took weeks or even months for attackers to develop usable exploits after a patch was released, giving users ample time to upgrade. Now, the moment an open-source project’s fix is published on GitHub, the information is already exposed. Attackers can monitor repositories and use AI to convert patches into exploits within hours. Duong calls this a 'minus-one-day vulnerability': by the time the patch reaches users, the exploit is already ready.

Regarding changes in attack speed, Duong supplemented this with a case involving a banking client. After the Calif team entered a client’s network, it usually took about a month to horizontally penetrate the entire system. An attacker’s AI, however, entered the network on Monday and was everywhere by Tuesday. Duong described this as cancer spreading at machine speed; existing detection systems cannot keep up—you detect the intrusion, but there are already a hundred attack points in the system, and shutting down operations is not an option for a bank.

Canadian large language model unicorn Cohere announced on October 29 that it has completed the first close of its Series E funding round. The round targets a maximum raise of CAD 7 billion (approximately RMB 33.4 billion). Cohere stated on its official X account that if the full fundraising target is reached, this transaction will become the largest single private funding round in Canadian startup history.

The round is led by German retail giant Schwarz Group, with continued participation from existing investors including Nvidia and Prosperity7 Ventures (a subsidiary of Saudi Aramco), as well as new investor the Government of Canada. Disclosed figures indicate that Cohere's valuation was approximately CAD 9.8 billion in September last year. If the current raise reaches its target, the company's valuation is expected to hit CAD 35.7 billion, representing a 3.6-fold increase within one year.

Beyond capital injection, Cohere unveiled strategic moves aimed at expanding its business boundaries. The company plans to advance two AI company acquisitions within a year: signing a letter of intent to acquire US-based AI cybersecurity startup AISLE to launch Cohere Cyber, an enterprise security product for automated vulnerability discovery and repair; and signing a merger agreement with German AI company Aleph Alpha to integrate model development capabilities and access European government and enterprise clients. Both transactions are currently pending completion.

Additionally, according to an agreement published by the Office of the Prime Minister of Canada on October 29, the Canadian government intends to invest up to CAD 1.6 billion (approximately RMB 7.6 billion) in Cohere in phases. As conditions for this investment, Cohere has committed to creating at least 1,000 technical jobs in Canada by 2030, investing at least CAD 1.7 billion in local computing infrastructure, and retaining its headquarters and registration in Canada.

Founded in 2019, Cohere provides privately deployable large models, AI agents, and search tools primarily for enterprises and governments. The company disclosed that its Annual Recurring Revenue (ARR) grew 165% year-over-year as of the end of Q3 this year, with global customer count increasing 20-fold over the past year to reach 300 companies. These financing and acquisition plans signal Cohere's transition from a pure model provider to a comprehensive enterprise AI solution vendor covering cybersecurity, European markets, and domestic infrastructure.

The publicly verifiable shipment volume of dexterity hands in the industry last year was approximately 20,000 units. However, Liu Xin, founder of Boya Intelligence (China), observed that most of these dexterity hands end up performing with humanoid robots for a few shows and then being left gathering dust in corners. In his view, current products are polarized: low-degree-of-freedom dexterity hands are cheap and durable, costing only a few thousand yuan, but their operational capabilities are very limited, primarily used for humanoid robot performances; high-degree-of-freedom products offer better operational capabilities, with mainstream prices exceeding 100,000 yuan, but their lifespan may be only one or two weeks, or at most one or two months. They are often purchased and kept as ornaments because users dare not use them, breaking down after just one week to two or three months of use.

Liu Xin believes the industry needs to create a 'third market': bringing scientific-grade operational capabilities into practical applications at an acceptable cost to solve problems that past automation equipment could not handle. He cited a specific example—some small factories near Shenzhen need to inspect parts under high-magnification electron microscopes after production, flipping each side of every part. Traditional grippers certainly cannot perform such tasks. Yet such products do not exist today; the most advanced and high-end dexterity hands on the market still fall far short of human hand capabilities.

The technical premise supporting this judgment is the core contradiction in dexterity hand design: it must pack mechanical structures, motors, circuits, and sensors into extremely limited space while ensuring sufficient force and power density. The thinner the fingers, the more easily structural strength and load-bearing capacity are constrained; joint connection points may bear forces of only a few hundred grams to about one kilogram, leading to fractures after continuous use for a week. To make the hand small, gear modules must be small, making teeth smaller and requiring higher machining precision. At today's material and process levels, wear, deformation, or even fracture occur easily. A human thumb can exert ten kilograms of force, whereas mainstream nominal force ratings for dexterity hand thumbs are currently one kilogram, with higher-end models reaching two kilograms. However, there is a discrepancy between nominal force and actual applied force.

From this, Liu Xin proposed a timeline judgment different from the industry mainstream: models may reach an inflection point in capability before hardware does. He believes that given the current iteration speed of AI and models, an inflection point in capability may arrive sooner. At that time, people will realize that models are already highly capable, but there is no sufficiently reliable hardware to support those capabilities. He mentioned a recent event attracting attention in the industry: GPT-6 can now operate dexterity hands. This indicates that common sense and cognitive abilities accumulated during the pre-training phase of large models are extending into the physical world. Ultimately, what holds things back may be the body, not the mind.

The telecommunications industry has long been jokingly referred to as the 'plumbers' of digital infrastructure—providing connectivity for cloud computing and AI, yet struggling to occupy a higher position in the value chain. As generative AI and agentic AI enter daily use, edge AI inference is seen as a new opportunity that could change this positioning. But 'the edge' is not a single location; it is an infrastructure gradient ranging from hyperscale data centers to mobile device terminals, with vast differences in economic feasibility across different tiers.

David Duggal, CEO of EnterpriseWeb, calls this distinction a confusion between 'telco edge' and 'customer edge.' The former refers to wireless base station towers, access points, and network-adjacent facilities; the latter may be located on enterprise campuses, factories, or warehouses. This distinction is crucial because deploying high-performance AI inference directly at a cell tower is far more difficult than doing so in metro-level facilities. Duggal points out: GPUs require significant power, more power means more heat, and more heat means more supporting infrastructure—a Catch-22.

In contrast, the metro edge—including central offices, regional aggregation hubs, and other facilities operated by large telecom operators or neutral hosts—can support hundreds of nodes, aggregate data flows from the cloud and the far edge, and provide a more practical environment for high-density GPU deployments. US telecom operators also hold a less discussed asset: real estate. Central offices, network huts, and legacy switching facilities may not look like the future of AI infrastructure, but many can become useful sites for edge inference, particularly in a context where community resistance makes the development of large new greenfield data centers increasingly difficult.

Jon Alexander, Senior Vice President of Products at Akamai's Cloud Technology Group, supplements this judgment from a workload perspective. He believes physical AI (factory and robotics scenarios) is one of today's most latency-sensitive use cases. 'You can't wait 10 seconds to send data all the way around to the other side of the world and back; everything in the factory would grind to a halt,' he says. However, for broader commercial AI workloads that do not require instant on-device intervention, metro-level edge infrastructure may offer the best balance of performance and cost. His specific criterion is: if GPU-enabled metro nodes can achieve response times of 5 to 10 milliseconds, 'that is great for pretty much every compute experience that exists today, meaning good utilization and good economics.'

Luis Fernandes, Senior Research Manager for Cloud and Data Centers at IDC, offers a reminder regarding the popular narrative of 'latency-driven edge deployment.' Research from IDC's 'Edge Observatory for the Digital Decade' project, conducted in collaboration with the European Commission, found that while latency is indeed an issue in very specific cases and regions, in most instances it is not the primary driver for edge deployment. 'Workload type is a determining factor,' Fernandes says. Real-time workloads require instant responses shaped by local context and timing; analytical workloads often do not, and data can be staged or preprocessed at the edge before being sent back to the center.

US enterprise software giant Salesforce announced on October 29, 2026, that it has signed a definitive agreement to acquire Rune Technologies, a defense AI startup. Kendall Collins, President and CEO of Salesforce Missionforce and Government Cloud, described this as the company's "most significant national security-focused acquisition." This deal is the third in a series of October acquisitions by Salesforce, following deals for analytics platform Loops and customer engagement platform Braze.

Rune Technologies was founded in 2024 by David Tuttle and Peter Goldsborough, both former leaders in command-and-control at defense tech firm Anduril. The company's core product, TyrOS, is a super intelligence (SI) platform for tactical edge logistics designed to operate in Denied, Degraded, Intermittent, or Limited (DDIL) connectivity environments. Already deployed with the US Army and US Marine Corps, the platform connects enterprise system data with real-time operational data, enabling frontline units to forecast supply needs, optimize resources, and suggest transport routes even when disconnected, syncing automatically once connectivity returns.

Salesforce will integrate Rune into its Missionforce portfolio, aiming to complete a three-layer architecture spanning headquarters, strategic planning, and the operational edge. Missionforce currently serves as the system of record for the first two layers across the US military and allied forces. By acquiring Rune, Salesforce seeks to bridge the gap between frontline decisions and enterprise records. This move aligns with the US Department of Defense's shift since March 2025 toward prioritizing commercial software pathways, and builds on Salesforce's existing 10-year, $5.6 billion ceiling contract with the US Army.

While Salesforce did not disclose the acquisition price, the deal reflects a broader defensive strategy by traditional SaaS companies facing competition from AI-native solutions. According to Reuters, Salesforce acquired AI agent platform Fin for $3.6 billion earlier this year and Informatica for $8 billion last year. Investors have expressed concern that rapid AI adoption could erode demand for legacy software, prompting Salesforce to aggressively acquire specific AI capabilities to maintain its market position.

Korean document AI startup KDL (한국딥러닝) and IBK Industrial Bank have completed a proof-of-concept (PoC) for an AI agent designed to automate voice phishing victim relief processes. According to Platum, the test conducted between April and August 2026 resulted in an 88% reduction in document processing time.

In traditional workflows, bank employees manually reviewed faxed documents such as relief applications and ID copies, then entered data into systems. This delay hindered rapid responses like payment freezes before funds could be transferred out. KDL's 'DEEP Agent for Finance,' built on a Vision-Language Model (VLM), goes beyond simple OCR by understanding document context to classify files, extract information, and infer crime types.

Data provided by KDL indicates the system achieved 100% accuracy on printed text and over 97% on handwritten text, successfully handling low-quality IDs and mixed handwriting that previous AI-OCR systems struggled with. Compared to existing AI-OCR, accuracy improved by up to 20.4 percentage points. The system uses a confidence scoring mechanism: high-confidence results are processed automatically, while low-confidence items are routed to human reviewers for final judgment.

The project won an excellence award in the PoC category of the '2026 Public Institution AI Innovation Challenge' hosted by South Korea's Ministry of Economy and Finance and Ministry of Science and ICT. IBK Industrial Bank plans to deploy the system in actual operations by the first half of 2027. KDL intends to expand this solution to other financial and public institutions.