Day 1 · R1 Industry Applications

Six talks on real-world generative AI adoption across banking, cybersecurity, consulting, telecom, laptop hardware, and industrial IoT.

Day 1 · Fri 27 Sep 2024Room 1 — Industry Applications13:00–17:00Chou Che-wei (周哲維), Assistant Professor, Feng Chia University

At a glance

SpeakerTalkIn one sentence
Huang Shih-chen (黃仕鎮), Director of Data Science, E.SUN BankGenerative AI applications in E.SUN BankE.SUN Bank shared its three-stage evolution of generative AI from embedded applications to autonomous agents, and the multimodal architecture design of its Genie platform.
Yang Cheng-lin (楊政霖), Director of Data Science, CyCraftLLMs in Cybersecurity Products: The Good, the Bad, and the UglyA cybersecurity company shared the benefits of adopting LLMs, four major concerns (privacy / alignment / domain knowledge / localization), and the security risks of RAG frameworks being attacked.
Wen Shao-chun (溫紹群), Senior Executive Vice President, Technology & Transformation, DeloitteUnleashing AI Empowerment: a Game-Changing Ecosystem of Shared InnovationFrom an industry consultant's perspective, discussed macro challenges of AI adoption such as organizational governance, data governance, and energy, and shared consulting case studies in smart logistics / lighthouse factories / smart cities.
Hu Te-min (胡德民), CIO and Executive Vice President, Far EasTone Telecommunications (FET)AI-Driven Future: FET's Digital Transformation JourneySummarized four common challenges enterprises face when adopting generative AI, and pledged on stage that AI adoption would not cost employees their jobs.
Kao Shih-fang (高士方), Senior R&D Deputy Director, MSI (Micro-Star International)AI ERA of Computing – Empowering the future with AIUsed the high-end GPU/NPU compute power of gaming laptops to build the on-premises AI assistant "AI Artist," and expanded into AI servers and AMR (autonomous mobile robot) product lines.
Lin Mao-chang (林茂昌), Chairman, NexAIoT (Nexcom Intelligent Systems)AI for All: Evolving from AI Supply Chain to Value ChainArgued that Taiwan should use open standards like EtherCAT to connect edge AIoT, and use on-premises LLMs linked to enterprise databases to build a "smart war room."

Talks

6 of 6 talks

1Generative AI applications in E.SUN BankHuang Shih-chen (黃仕鎮), Director of Data Science, E.SUN Bank

E.SUN Bank shared its three-stage evolution path of generative AI from embedded applications to autonomous agents, and the multimodal architecture design of its Genie platform.

Key points

  • Three-stage AI adoption architecture: AI Embedded (credit pricing, home price estimation, financial capacity assessment, fraud prevention, business opportunity discovery) → AI Copilot (system development, accompanying wealth advisors, training, intelligent customer service) → AI Agent (intelligent customer service, personalized marketing), describing AI applications' evolution path from "embedded in existing processes" to "assistive role" to "autonomous agent."
  • GENIE platform: E.SUN Bank's internal generative AI (OA intelligent assistant) platform, piloted in Q4 2023, officially launched in early 2024, with GENIE 2.0 released in September 2024.
  • Genie platform architecture: multimodal (Text / Image / Voice / HTML / PDF / Video), multi-channel, multi-model, multi-function (meeting minutes, regulatory comparison, customer service, etc.). Public reports confirm that GENIE 2.0's multimodal architecture added voice recordings and images (1.0 was text-only), while also expanding from desktop-only to mobile device availability (multi-channel); technically it uses RAG (Retrieval-Augmented Generation) to retrieve information from a knowledge base to aid generation, and has an NLP Gateway that detects and filters sensitive data before output as a last line of defense for security. The specific meaning of "multi-model" (whether it refers to multiple LLMs running in parallel or switching between them) — no more detailed public explanation could be found (unverified).
  • LLM GUI design logic: relevant data is retrieved first, then generative AI (GAI) compiles and outputs it.
  • Reason for building in-house: banking is a licensed industry and customer data is its lifeline, so the bank chose to build its own platform rather than directly connecting to external major-vendor LLMs.
  • GENIE's initial adoption rate was dismal: only 17% unique users, with a daily usage rate of about 2%; subsequent analysis identified three types of problem users (tried it but were dissatisfied, didn't know how to use it, or were banned by their department for cost reasons). Improvement measures included adding features and communication channels, providing a template library, and strengthening training (specifically targeting managers, on the reasoning that "if managers don't use it, neither will their subordinates"). After GENIE 2.0 launched, monthly active users grew 200% and actual usage grew 6-fold (GENIE 2.0 launched in September 2024, nearly concurrent with this talk, so it is uncertain whether this growth data had been fully disclosed at that point).
  • Specific application cases: a meeting minutes system (automatically removing filler words, extracting topics, flagging action items, analyzing speaking ratio and content quality); "Golden Secretary" (an AI assistant for branch staff, after launch reducing branch inquiry calls to headquarters by 30%); a credit report review system (using AI to compare historical rejection reasons, helping relationship managers automatically check report completeness before submission); an HR assistant; and automatic compilation of financial news into a morning briefing, among others.
  • E.SUN Financial Holdings separately partnered with IBM to establish an AI governance framework; reports have also mentioned that E.SUN Financial Holdings built its own "in-house GAI development framework" to support scaling of its business.

Tech, products & figures

GENIE / GENIE 2.0
E.SUN Bank's internal generative AI (OA intelligent assistant) platform
RAG (Retrieval-Augmented Generation)
GENIE's knowledge-base retrieval technology
NLP Gateway
Sensitive-data detection and filtering mechanism applied before output
Golden Secretary, Credit Report Review System, HR Assistant
Various internal AI applications at E.SUN Bank

Notable quotes

"It can't be thrown outside the wall, nor can external AI be used arbitrarily" — explaining why E.SUN Bank chose to build its own generative AI platform rather than directly connect to external major-vendor LLMs.

Q&A

  • No Q&A session.

Fact-check notes

2LLMs in Cybersecurity Products: The Good, the Bad, and the UglyYang Cheng-lin (楊政霖), Director of Data Science, CyCraft, PhD earned in the United Kingdom

A cybersecurity company shared the practical benefits of adopting LLMs, four reasons for fine-tuning their own model, and the security risks of RAG frameworks potentially being attacked.

Key points

  • Opening live poll: asked the audience by show of hands whether their companies had adopted or were considering adopting a language model — about 60–70% of the audience raised their hands.
  • Cited a Gartner report: the five most promising Generative AI applications currently (editing, chatbots, translation, summarization, etc.) are all related to language models.
  • Pain point before adoption: a single security incident at a cybersecurity company often involves hundreds of distinct malicious events, each requiring a security expert to review, write a report, and give remediation advice individually — extremely labor-intensive.
  • Malicious command-line obfuscation case: a Windows command that looked legitimate but was abused by hackers for obfuscation took the company's best security expert a full 40 minutes to decode its true intent; after adopting their in-house security LLM, an analyst can interpret it with one click, cutting the time to 3 seconds (speaker's own account, no external corroboration).
  • Threat intelligence aggregation problem: threat intelligence in different languages (Chinese/English/Japanese, etc.) and different formats (PDF, images, tables converted to images) is hard to integrate; after adopting multimodal AI to unify the format, combined with text-to-SQL, analysts can directly ask in natural language "which APT groups have been active recently" without hand-writing SQL.
  • Used an in-house embedding model to perform semantic-level correlation analysis of malicious command lines (malicious commands with different structures but the same semantic meaning can be grouped together); the related research has reportedly been accepted by EMNLP (the exact paper title/link could not be verified — unverified).
  • An internal chatbot handles repetitive administrative questions (such as common company-related FAQs), saving customer-service/HR time.
  • Why fine-tune their own model rather than directly using OpenAI/Anthropic: four considerations were raised.
    • Data Privacy: the privacy policy of OpenAI's web version (chat.openai.com) allows use of user input content; if an employee asks ChatGPT about company code, that is effectively leaking information to the vendor; also cited the example that OpenAI's GPTs feature was tricked via prompt injection into revealing its original data source download link the day after launch.
    • Model Alignment: general-purpose LLMs are trained to be "polite and not do evil," so when faced with the malicious command-line interpretation that security analysis requires, they simply refuse to answer (e.g., "This command looks bad, please don't run it"), which does not meet security personnel's need to have it "clearly explain what it is actually doing."
    • Domain Knowledge: general-purpose LLMs lack knowledge of cybersecurity terminology (such as "juicy potato," a password-theft tool commonly used by hackers), and can only respond "no relevant knowledge."
    • Localization: the proportion of Traditional Chinese in open-source LLM training corpora is extremely low (based on BLOOM's training data, roughly only four in a million), causing models asked in Traditional Chinese to answer with Simplified Chinese wording (e.g. "计算机" instead of "電腦") — a major taboo for public-sector clients.
  • Hidden cost example: the same query that costs about 2,500 tokens with GPT-3.5/4 balloons to 15,000 tokens with OpenAI's o1 model because of its embedded reasoning tokens (the speaker cited "tests actually done by people abroad" without naming a source; it is publicly known that o1's reasoning tokens substantially raise overall token consumption and billing, with input/output pricing about 5–6 times that of GPT-4o — see Vellum's analysis "Analysis: OpenAI o1 vs GPT-4o vs Claude 3.5 Sonnet" — but the specific "2,500 to 15,000 tokens" figures the speaker cited could not be traced back to an original test; speaker's own account). Also cited a case of a Japanese security expert complaining on Twitter: after adopting a major vendor's Security Co-pilot service, their bill suddenly spiked to 2 million yen because the vendor unilaterally raised the price of a certain feature threefold (speaker's own account, no corresponding tweet or report could be found, no external corroboration; however, Microsoft Security Copilot's "hourly billing, decoupled from actual usage" Security Compute Units pricing model is itself a publicly known point of controversy that can lead to bills exceeding expectations — see Redress's "Security Copilot Pricing: The SCU Dial" — but the specific figures the speaker mentioned, "2 million yen" and "tripled," could not be verified).
  • Security risks of self-hosted RAG frameworks (such as LangChain): cited research presented at Black Hat Asia in early 2024, which penetration-tested 51 "framework + LLM provider" combinations (such as LangChain+Llama, LlamaIndex+Anthropic); 17 had vulnerabilities, of which 16 could achieve RCE (remote code execution) and 1 could achieve SQL injection. Of the 16 RCE cases, 14 could establish a reverse tunnel (a persistent connection back to an external host), and 4 of those could obtain a reverse shell directly. The attack principle is that frameworks (such as LangChain) fail to properly filter LLM response content, automatically executing malicious instructions embedded in the response. Verification suggests this research is Tong Liu et al.'s "Demystifying RCE Vulnerabilities in LLM-Integrated Apps" (full paper at arXiv:2309.02926, formally published at ACM CCS 2024), whose paper tested 51 potentially affected applications with 17 successful attacks (16 RCE, 1 SQL injection) — figures that exactly match what the speaker stated; the same team also presented the same technique, "LLM4Shell," at Black Hat Asia 2024, matching the timing the speaker described as "earlier this year." The detailed reverse-tunnel/reverse-shell figures (14/4) could not be individually confirmed in the public abstract and are recorded here as the speaker stated them.
  • Key takeaways:
    • Don't blindly believe LLMs are all-powerful; adoption should start from small, non-critical functions (Start from small).
    • If fine-tuning an in-house model, be sure to evaluate the four dimensions above — privacy / alignment / domain knowledge / localization — plus security risk, for five dimensions in total.
    • Self-hosting an on-premises model does not mean zero vendor lock-in or zero cost; teams must assess whether they have enough security talent to deploy and retrain it.
    • Many enterprises "buy for the sake of buying" — equipment gets purchased but nobody knows how to use it, just to use up budget and pad performance metrics.

Tech, products & figures

In-house fine-tuned Traditional Chinese security LLM
Used to automatically generate incident overview reports and interpret malicious command lines
Multimodal AI + Text-to-SQL
Unifies threat intelligence formats from multiple sources (PDF/images/web pages) for natural-language querying
In-house embedding model
Performs semantic-level correlation analysis of malicious commands
RCE vulnerabilities in RAG frameworks such as LangChain
Known as "LLM4Shell," presented at Black Hat Asia 2024; the formal paper (arXiv:2309.02926) found vulnerabilities in 17 of 51 tested framework combinations

Notable quotes

"Your budget is not your budget" — describing how, after enterprises adopt language models, token-based billing (especially for models with reasoning tokens) often brings unexpected hidden costs.
"This is a lamb on the chopping block, at the mercy of others" — describing how, if an enterprise has not clearly negotiated pricing and price-increase notification terms with its LLM/cloud service provider, its bargaining power is completely unequal.

Q&A

  • No Q&A session.

Fact-check notes

3Unleashing AI Empowerment: a Game-Changing Ecosystem of Shared InnovationWen Shao-chun (溫紹群), Senior Executive Vice President and Head of Digital Transformation Services, Technology & Transformation, Deloitte (勤業眾信聯合會計師事務所)

From an industry consultant's perspective, discussed the organizational/data governance and macroeconomic challenges of generative AI adoption, and shared several smart manufacturing consulting case studies.

Key points

  • Opening viewpoint: most Taiwanese enterprises' investment in generative AI does not focus on developing the underlying large language model itself (foreign startups' investment share in this area is as high as 50–60%), but rather on "using large language models for specific industry applications"; this is related to Taiwan's PhD/master's-degree population structure, and in the long run, investment in more generally applicable underlying models has "higher long-term expected value."
  • Cautioned that the economic outlook for 2025 is expected to be tighter, with several financial industry executives privately noting that next year's AI budgets would be more conservative; suggested that technical feasibility testing remains necessary, but that when seeking budget, one should use language that CFOs and CHROs understand (cash flow, workforce/inventory management efficiency) rather than purely technical language.
  • Cited a retail application scenario as an example: if customer foot traffic (unstructured image data) could be combined with deli/fresh food POS sales data analysis, it could in theory automatically detect slow-moving items around 3–4pm each day and automatically apply a 20% discount promotion (the speaker gave this as an example of a potential application scenario, without specifying whether it was already an implemented real-world case); also pointed out that unstructured data may account for more than 50% of enterprise data, and mostly just "sits there, unanalyzed."
  • Financial industry case: mentioned a recent public news case of a financial institution spending approximately NT$8–10 billion on a core system transformation (institution not named); even for highly digitalized financial institutions, extracting data from legacy systems for further analysis remains very time-consuming and costly.
  • Pointed out the long-standing real-world adoption difficulties of RPA (robotic process automation): if enterprises do not first "centralize" at the organizational/process level (consolidating the same work scattered across departments A/B/C into a shared service), adopting technology alone is unlikely to truly save significant labor costs.
  • Suggested AI/data governance framework: enterprises should establish governance mechanisms across dimensions such as "strategy, organization, risk management, assessment methods, data ethics, trustworthy AI, and security compliance"; non-financial industries might cover this with two or three people, while financial industry firms often need five or six units jointly overseeing it.
  • Deloitte's actual consulting case studies:
    • Assisted Chunghwa Post in planning the Taoyuan A7 Smart Logistics Park (AIoT warehouse digitalization across 5 buildings) as planning consultant — the A7 Smart Logistics Park itself is a real, publicly known project, but no public information could confirm Deloitte's role as planning consultant for this case; this part is the speaker's own account, unverified externally.
    • Assisted an unnamed semiconductor manufacturer in competing for and maintaining WEF "Lighthouse Factory" certification (about 130 worldwide, this company being the only semiconductor company selected), currently assisting it toward becoming a "sustainability lighthouse factory" — the WEF Global Lighthouse Network itself is a real, publicly known program, but the speaker did not name the manufacturer on stage, and this verification effort could not find public reporting confirming which specific semiconductor manufacturer or Deloitte's consulting role — unverified (during verification, only confirmed that Taiwan's selected lighthouse factories are mostly panel makers such as AUO and Innolux; no corresponding semiconductor manufacturer case was found).
    • In 2022, assisted New Taipei City in competing for the top ranking in a global smart city evaluation (Smart City Top 1) — verified that the New Taipei City Government was indeed awarded the ICF (Intelligent Community Forum) Smart City Top 1 award in 2022, and public information also mentions that Deloitte's digital transformation services team participated in assisting — this part is corroborated by external reporting.
  • Raised the need for impartial third-party verification of "robo-advisors / explainable AI" in the financial industry: financial regulators require AI to be free of bias, but the speaker argued that "precision" inherently carries a bias tendency (finding the most valuable customer segment); how to balance "precision" against the social responsibility of "financial inclusion" is a governance question each company must think through for itself; also mentioned that foreign financial institutions have already created "Chief AI Ethics Officer" positions.
  • Taiwan energy-and-AI linkage issue: Taipower plans to "balance power supply across regions," no longer transmitting power from the south to the north or from central Taiwan to the north, meaning that without local power plants, it will be difficult to build AI data centers locally; when the share of renewable energy rises to 20%, power supply becomes intermittent, requiring energy storage, virtual power grids, and demand forecasting/dispatch to balance supply and demand — this will become a hot topic.
  • Distinguished between "playing with models" and "building a platform": going from 0 to 1 is "playing with models," while going from 1 to 100 (scaling up to hundreds of people using different bots/platforms) enters the realm of data integration and governance issues, which is the real key to whether AI can truly scale.
  • Cautioned companies looking to monetize data or even plan an IPO to pay close attention to data security and privacy issues, because stock exchanges/regulators will strictly scrutinize whether related risks are properly controlled during review.

Tech, products & figures

Deloitte "AI and Data Operations" managed services
Sponsor advertising content, not a focus of the speaker's own presentation
Chunghwa Post Taoyuan A7 Smart Logistics Park
The project itself is publicly reported; Deloitte's consulting role could not be verified with an external source
WEF Global Lighthouse Network certification case
The program itself is publicly known to exist; the speaker did not name the manufacturer, and neither the specific company nor Deloitte's consulting role could be verified
2022 New Taipei City Global Smart City Top 1 consulting case

Notable quotes

"In the age of AI, you have to know how to tell a story" — emphasizing that when reporting AI projects to CXOs, if you cannot articulate a concrete application scenario, decision-makers won't understand it or buy into it.
"This dilemma is like your wife and your mother both falling into the sea — which do you save first" — describing the dilemma between "precision marketing/risk control" and the social responsibility of "financial inclusion."

Q&A

  • No Q&A session.

Fact-check notes

  • Wen Shao-chun's current title is Senior Executive Vice President at Deloitte (Head of Digital Transformation Services).Sources：清華人故事館—溫紹群高雄市政府簡報資料：溫紹群資深執行副總經理
  • The speaker's mention of a financial institution spending "NT$8–10 billion" on core system transformation cites a public news case but does not name the institution; the specific company name was not further verified.
  • When the speaker mentioned the "lighthouse factory" semiconductor client, they did not explicitly name the company, so it is recorded as-is as "an unnamed semiconductor manufacturer" without forcing an identification; verification could only confirm that the WEF Global Lighthouse Network program itself exists, and could not find public reporting on the corresponding semiconductor manufacturer or Deloitte's consulting role.Sources：WEF Global Lighthouse Network
  • Regarding the speaker's mention of assisting Chunghwa Post in planning the A7 Smart Logistics Park, the park itself is a publicly known project, but Deloitte's role as planning consultant could not be corroborated by external reporting — this is the speaker's own account.Sources：中華郵政「智慧物流園區」開工｜數位時代
4AI-Driven Future: FET's Digital Transformation JourneyHu Te-min (胡德民), CIO and Executive Vice President, Far EasTone Telecommunications (FET)

FET's CIO summarized the four major challenges enterprises commonly encounter when adopting generative AI, and pledged on stage that AI adoption would not cost employees their jobs.

Key points

  • Four major challenges of enterprise GAI adoption:
    • Ambiguity of application scenarios
    • Difficulty monetizing value
    • Lack of specialized talent
    • Insufficient effective data
  • Advocated "let professionals develop their expertise, and let us focus on our own expertise," emphasizing division of labor and collaboration rather than reinventing everything in-house.
  • Characterized generative AI as driving a "white-collar revolution," emphasizing that enterprise adoption needs to be "goal-oriented, planned, fast, and able to measure results."
  • In response to employees' concerns about AI, the speaker declared on stage that AI adoption would never affect employees' jobs, as a way of reassuring employees about AI adoption.
  • Background: FET established its "Transformation Office" in 2018, with "Big Person" (Big Data, AI, IoT) as its core technology focus; in 2022 the Transformation Office was merged with the IT department into the "Information & Digital Transformation Technology Group," dedicated to transforming the telecom business into a technology services company.
  • Representative cases: the 5G telemedicine platform (2022 results, covering 12 counties/cities and 35 health centers, cumulatively reaching 27,529 visits, including 21,122 video consultations and 6,407 remote consultations, with estimated reductions of NT$4.14 million in wasted medical costs, NT$7.55 million in patient costs, and 81.56 metric tons of carbon emissions); friDay Video rose to become Taiwan's third-ranked video platform after adopting Azure OpenAI's recommendation feature.
  • "Internal use, external application" model: first solving real problems inside FET, then packaging mature solutions into external enterprise offerings (AI Ready enterprise plans, AI talent training, GenAI workshops, Copilot adoption consulting services), extending application domains to smart cities, smart healthcare, and smart campuses.

Tech, products & figures

5G Telemedicine Platform
Covers 12 counties/cities and 35 health centers
friDay Video + Azure OpenAI recommendation feature
Rose to become Taiwan's third-ranked video platform
"Big Person" (Big Data / AI / IoT)
FET's core digital transformation technology focus

Notable quotes

"Let professionals develop their expertise, and let us focus on our own expertise."
Declared on stage that AI adoption would never affect employees' jobs.

Q&A

  • No Q&A session.

Fact-check notes

5AI ERA of Computing – Empowering the future with AIKao Shih-fang (高士方), Senior R&D Deputy Director, MSI (Micro-Star International)

MSI used the high-end GPU/NPU compute power of gaming laptops to build the on-premises AI assistant "AI Artist," and expanded into enterprise product lines for AI servers and AMR robots.

Key points

  • Core argument: because MSI gaming laptops are equipped with high-end discrete graphics cards (such as the RTX 40 series), their compute power far exceeds the NPU requirements defined by general AIPC specifications, so "a gaming laptop is the most powerful AIPC on Earth"; taking the RTX 4050 and above as an example, it can contribute over 250 TOPS of compute, with the 4090 exceeding 1000 TOPS (more than 5 times the NPU compute power of the time, such as Lunar Lake's 45 TOPS).
  • Product iteration timeline: launched the first-generation "AI Artist" text-to-image tool at Computex 2023, with on-premises computation paired with an NVIDIA TensorRT accelerator, generating images in just 1.5 seconds; the CES 2024 version added image-to-image, image-to-text, and fine-tuning functions; in the second half of 2024, it pivoted to language models, embedding an LLM into AI Artist, so users can ask in natural language "why can't I see myself in a video call," and AI Artist can directly diagnose and control hardware (such as automatically turning a disabled camera back on or adjusting screen brightness).
  • Adopted RAG (which the speaker called "AI Fire"): letting users import their own documents (examples given: the Six Codes of law, elementary school Chinese-language textbooks) to create customized Q&A or quiz-generation tools, with the goal of building "an AI hand that every industry and profession can use."
  • Rationale for the technical pivot: having observed the trend toward smaller LLMs (SLMs at scales such as 70B, 7B, or even 2B) and multimodal development, MSI shifted AI Artist's focus from pure text-to-image to language models; the challenge for small task-specific models lies in accuracy and architectural planning for the task direction, not inference speed (inference speed was already solved via the NVIDIA TensorRT accelerator).
  • Deepfake detection: cited the well-known AI-fake video of Tom Cruise playing guitar, and South Korea's virtual news anchor "Kim Joo-ha" (able to broadcast around the clock) as examples showing that generative AI can now pass as highly convincingly real; MSI is collaborating with Intel to research detecting deepfakes through minute physiological changes such as skin tone and blood vessels, but the current detection data is not yet precise enough — it remains an ongoing research direction with no mature product yet.
  • Enterprise product line strategy: the AI Server line (server products spanning L10–L11, 2U–4U chassis, supporting AMD/Intel platforms) plays the role of "the brain"; AMR (Autonomous Mobile Robots, whose technology lineage descends from robot vacuum cleaners and extends into factory automation) plays the role of "the hands and feet," equipped with vision sensors for autonomous mapping and decision-making; application scenarios cover retail restocking, warehouse heavy-item transport, and hospital environment disinfection and equipment delivery — tasks in settings unfriendly to humans.
  • Emphasized "on-premises" deployment for enterprise-grade AI: to meet the needs of the financial industry (where fund flows and customer information are legally prohibited from being on the cloud) and factory production lines (which require high real-time responsiveness), MSI's enterprise solutions focus on deploying compute power inside enterprise IT rather than the public cloud; MSI also has a joint development project with NVIDIA to bring cloud development experience and platforms on-premises.

Tech, products & figures

MSI AI Artist
MSI's on-premises AI assistant, whose functionality evolved through four stages: "text-to-image → image-to-image/image-to-text → language model + hardware control → RAG (AI Fire)"
NVIDIA TensorRT accelerator
Used for on-premises image generation and inference acceleration
MSI AI Server product line (L10–L11, 2U–4U, supporting AMD/Intel)
MSI AMR (Autonomous Mobile Robot)
Applied in retail, warehousing, and healthcare settings

Notable quotes

"A gaming laptop is the most suitable, and the most powerful, AIPC on the planet."

Q&A

  • No Q&A session.

Fact-check notes

  • The speaker's title per the official agenda is "Senior R&D Deputy Director, MSI" (Kao Shih-fang). A web search did not find a public page confirming details of Kao Shih-fang's title (unverified); per task requirements, the official agenda information is treated as the factual baseline.
6AI for All: Evolving from AI Supply Chain to Value ChainLin Mao-chang (林茂昌), Chairman, NexAIoT (Nexcom Intelligent Systems Co., Ltd.); Founder, Nexcom Group / NEXCOM

Argued that Taiwan should use open standards such as EtherCAT to connect edge AIoT, and use on-premises LLMs linked to enterprise databases to build a "smart war room" to accelerate manufacturing transformation.

Key points

  • Core viewpoint: although the AI server supply chain has already been raking in profits (NVIDIA's gross margin about 75%, TSMC's about 53%, with midstream AI server contract manufacturers' gross margin only 6–8% but revenue scale dozens of times larger than traditional servers), the supply chain will eventually become saturated; downstream cloud service providers, meanwhile, face an "arms race" of massive investment (Microsoft/Google/Meta/Amazon averaging about USD 50 billion per year, potentially exceeding USD 100 billion within a few years) with unclear profit prospects — the speaker called this "lost in the depths of the clouds" (an allusion to a classical Chinese line meaning the outcome cannot be seen).
  • Argued that Taiwan is not suited to being an "AI Cloud Service Provider": demographics-related (internet-derived) big data has a "territorial" characteristic, and only countries with population scales like China (1.4 billion) or the United States (estimated at over 2 billion when counting the English-speaking world's potential market) can sustain cloud giants; by contrast, markets related to "physics" (material properties, mechanical properties, hardware functions, environmental/physiological monitoring) are unrelated to location or population and are suited to global operation — this is precisely Taiwan's golden opportunity to develop Edge AI/AIoT.
  • Specific technical approach: adopted EtherCAT (an industrial Ethernet real-time communication standard; the official standards body is the EtherCAT Technology Group) to build a controller that uniformly controls heterogeneous equipment such as CNC machines, general motion control units, laser engravers, AOI inspection machines, robotic arms, and AMRs; for closed-architecture brands (such as the "Big Four" robot manufacturers and some machine tool controllers), interoperability is achieved through gateways. This controller platform is called "iRPA2000," one of NexAIoT's official product lines.
  • The smart factory architecture is divided into four layers; the core monitoring system product is "iAT2000" (a cloud-intelligent monitoring system), one of NexAIoT's official product lines.
    • L1 Machine Equipment Layer: connects heterogeneous equipment using EtherCAT
    • L2 Connectivity Layer: aggregates OT-layer and IT-layer data, such as ERP/MRP/MES/CRM
    • L3 Data Application Layer: uses the "NextData" data hub to make smart manufacturing application development as fast as building a mobile app
    • L4 Enterprise War Room: real-time decision-making, which after connecting to generative AI is expected to achieve "predictive" responsiveness
  • Killer application argument: connecting a company's own NextData database to generative AI is more accurate and reliable than a generic ChatGPT trawling the whole internet and easily "speaking nonsense with a straight face," because the database is "bounded and has clearly defined content"; developing a vertical solution (such as predictive maintenance or energy conservation/carbon reduction) used to take two to three months, but now the enterprise war room can get an answer via one question-and-answer exchange "in under two minutes," and the solution is fully deployed on-premises, so sensitive corporate data does not leave the company, results are explainable, and there are no AI hallucination concerns.
  • Industry argument: cited a McKinsey forecast that the global IoT market could reach USD 5.5 trillion to 12.6 trillion by 2030; argued that if Taiwan's machinery/hardware contract manufacturing industry can move toward open standardization, it could replicate the path by which open ecosystems drove the PC/mobile phone industries' boom in their day, and even export "unmanned production line / unmanned factory" turnkey services abroad.
  • Open strategy: placed technologies such as iRPA2000, iAT2000, eSafe (an OT security platform), and DVOI (a smart imaging platform, which the speaker called "the eyes of the Internet of Things") on an open platform (tentatively noted as "物聯羽" [Wulian Yu], presumed to be NexAIoT's own proprietary AIoT platform name, though the official name could not be verified) for free download, with the goal of lowering the technical and cost barriers to AIoT innovation and entrepreneurship.

Tech, products & figures

EtherCAT
An industrial Ethernet real-time communication standard, used by NexAIoT to connect heterogeneous machinery equipment
iRPA2000
NexAIoT's next-generation robotics co-creation platform
iAT2000
NexAIoT's cloud-intelligent monitoring system (the core product of the enterprise war room)
NextData
NexAIoT's manufacturing data hub, designed to connect to generative AI for a question-and-answer war room
eSafe
An OT (operational technology) security platform
DVOI
A smart imaging platform
McKinsey's forecast for the global IoT market size by 2030 (USD 5.5 trillion to 12.6 trillion)

Notable quotes

"Compute power and national power — it seems like the greater the compute, the stronger the nation. Is that really true?...We just need enough compute to get by." — questioning whether the narrative of "racing for compute" fully applies to a small country in terms of population like Taiwan.
"If AI's cloud services are the heavenly dragon, then we at NexAIoT are the earthly tiger." — describing NexAIoT's self-positioning as the "on-the-ground collaborator" of AIoT/edge AI.

Q&A

  • No Q&A session.

Fact-check notes

  • The speaker's name and title match the official agenda: Lin Mao-chang (林茂昌), Chairman, NexAIoT (Nexcom Intelligent Systems); also Founder of Nexcom Group (NEXCOM), which he founded as Nexcom Computer in 1992.Sources：商業周刊—董事長林茂昌科技創新個案：新漢集團TEEIA—新漢智能系統
  • The name of the free-download platform mentioned near the end of the talk (tentatively noted as "物聯羽" [Wulian Yu]) could not be verified as its official name, so it is noted as-is.

中文版