--- title: "🎙️ AI Daily Podcast: In-house Chip Jalapeño, DeepMind Talent Exodus, Model Distillation Battle" description: "Hello everyone, welcome to \"Mark's Tech Insights,\" I'm Mark. Today we have some truly exciting topics to talk about, from OpenAI and Broadcom's joint launch of their first self-developed inference chip, to the exodus of key figures from Google DeepMind shaking up stock prices, and the undercurrents between the US and China in AI technology. This episode is absolutely packed with information—are you ready to stay ahead of the latest AI trends with me?" canonical_url: "https://blog.markkulab.net/en/tech-news/ai-daily-podcast-2026-06-28" author: "Mark Ku" author_url: "https://blog.markkulab.net/en/author/mark-ku" site: "Mark Ku's Tech Notes" date_published: "2026-06-28 10:00:00 +0800" category: "What's New in Tech" tags: ["ai-daily", "podcast", "tech-news", "OpenAI", "Broadcom", "Jalapeño", "DeepMind", "Anthropic", "模型蒸餾", "主權AI", "AI創投"] language: "en" license: "CC BY 4.0" license_url: "https://creativecommons.org/licenses/by/4.0/" attribution: "when reusing or quoting, credit the author and link back to the original" --- # 🎙️ AI Daily Podcast: In-house Chip Jalapeño, DeepMind Talent Exodus, Model Distillation Battle ## Introduction Hello everyone, and welcome to "Mark's Tech Insights." I'm Mark. Today we have an incredibly exciting lineup of topics to discuss—from OpenAI and Broadcom joining forces to launch their first custom-designed inference chip, to a major talent exodus at Google DeepMind that shook stock prices, and the simmering undercurrents of AI competition between the US and China. The sheer volume of information today is absolutely mind-blowing. Are you ready to dive in and stay ahead of the latest AI trends with me? --- ## Today's Top Stories ### 1. OpenAI Partners with Broadcom to Unveil First Custom AI Inference Chip "Jalapeño" - **Source**: Tom's Hardware (） - **Summary**: Collaborating with chip giant Broadcom, OpenAI has developed its first custom ASIC chip designed specifically for AI inference, codenamed "Jalapeño," in just nine months. What makes this particularly remarkable is that they used their own AI models to accelerate the design process. Initial data shows that this chip's inference cost per token is roughly 50% cheaper than NVIDIA's GPUs, with full mass production expected by early 2028. - **Taiwan Perspective**: This is undoubtedly a massive boon for Taiwan's semiconductor supply chain, especially TSMC's advanced nodes and advanced packaging capacity. It also signals that the trend toward custom ASICs is now irreversible. - **Key Discussion Points**: - With OpenAI using its own models to design chips, does this mean the era of "AI designing AI" has officially arrived? - Will NVIDIA's dominance in the inference market be shaken as tech giants increasingly develop their own custom ASICs? - **Script Suggestion**: Folks, this chip codenamed "Jalapeño" is incredibly fascinating. It not only represents OpenAI's push to break free from its heavy reliance on NVIDIA, but it also showcases the power of "using AI to design chips"—pulling it off in just nine months. For Taiwan, this translates to massive business opportunities across both foundry services and silicon IP design. Although mass production won't start until 2028, this is already a wake-up call for NVIDIA. In the future compute market, custom ASICs will undoubtedly be the key battleground. --- ### 2. Anthropic Accuses Alibaba of Large-Scale "Model Distillation Attack" Using Over 20,000 Fake Accounts - **Source**: CNBC (） - **Summary**: Anthropic has filed a formal complaint with the US Senate, accusing Alibaba's Qwen AI lab of orchestrating an unprecedented "distillation attack." Alibaba allegedly used up to 25,000 fake API accounts to generate nearly 29 million interactions with Claude in just over a month, presumably to "steal" Claude's reasoning and knowledge to train Alibaba's own models. - **Taiwan Perspective**: Many development teams in Taiwan also frequently use APIs for fine-tuning. However, this incident highlights how critical intellectual property protection and API anomaly detection will become in the future. - **Key Discussion Points**: - While "model distillation" is common in academia, how should we draw the legal and ethical lines when it comes to scraping data at scale using fake accounts? - How will this incident intensify the friction between the US and China over AI patents and security safeguards? - **Script Suggestion**: This time, Anthropic went straight to the US Senate, accusing Alibaba of using 25,000 fake accounts to harvest Claude's conversational data. This has really thrust "model distillation" into the spotlight. To put it simply, it's like copying someone else's study notes to quickly boost your own grades. While technically clever, this kind of gray-area behavior is bound to trigger much stricter API access controls in the future. Developers in Taiwan might also face more rigorous identity verification when calling overseas APIs down the road. --- ### 3. Google DeepMind Core Talent Exodus Triggers Sharp Drop in Alphabet Stock - **Source**: Fortune (） - **Summary**: Google DeepMind is facing a severe talent crisis. Within a span of just six days, four top scientists—including Nobel laureate John Jumper and Transformer co-author Noam Shazeer—defected to Anthropic and OpenAI. This exodus directly triggered a 6% to 10% drop in Google's parent company Alphabet's stock price, prompting investors to question whether the tech giants' massive $450+ billion AI capital expenditure this year will yield a viable return. - **Taiwan Perspective**: Top talent remains the scarcest resource in the AI domain. This massive shift in human capital could directly reshape the competitive landscape of mainstream LLMs over the next few years. - **Key Discussion Points**: - Why would top scholars like John Jumper choose to leave Google? Is it due to the research environment or commercialization pressure? - Will the massive capital expenditures of the Big Four tech giants risk a bubble burst due to talent loss and unclear returns? - **Script Suggestion**: Google suffered a massive blow this time, failing to retain even John Jumper, the Nobel laureate who co-created AlphaFold, who jumped ship to a competitor. This didn't just bruise Alphabet's stock price; it also exposed how corporate R&D structures in large companies might be stifling top scientists. When these minds leave, they take their core technical intuition with them. It's a stark reminder that winning the AI race isn't just about how many tens of thousands of GPUs you buy—it's about whether you can retain the brilliant minds capable of changing the world. --- ### 4. National Security Concerns: US Government Forces Anthropic to Restrict Latest Flagship Model Outputs - **Source**: Dentro.de/AI (）與 The White House (） - **Summary**: Citing national security concerns, the US government took the rare step of forcing Anthropic to restrict access to its newly released flagship models, Fable 5 and Mythos 5. The White House subsequently issued a new executive order that significantly tightens national security oversight on frontier AI models while still promoting AI innovation. This move has sparked concerns among international leaders, including the Prime Minister of Canada, regarding over-reliance on a handful of US-based AI providers. - **Taiwan Perspective**: When adopting geopolitically sensitive AI technologies, Taiwan's government and enterprises must accelerate the development of "sovereign AI" or localized models to avoid the risk of sudden technology cutoffs due to international regulatory shifts. - **Key Discussion Points**: - What kind of impact will direct government intervention in private-sector model deployment have on the pace of AI innovation? - Will "sovereign AI," which many countries are now advocating, become the mainstream strategy for non-US nations following this event? - **Script Suggestion**: The US government stepped in directly this time, blocking Anthropic from freely deploying its most advanced Fable 5 and Mythos 5 models—something that is truly unprecedented. It shows that the government now views ultra-large AI models as strategic assets on par with nuclear weapons and advanced semiconductor nodes. This is a major wake-up call for Taiwan. If our businesses rely entirely on US API services, what happens if we get cut off one day due to geopolitical or national security reasons? This proves once again that developing Taiwan's own Traditional Chinese sovereign models is absolutely imperative. --- ### 5. June Sees Most Intense Model Release Wave in History, with Open-Source and Lightweight Models Reaching New Cost-Performance Heights - **Source**: Presenc AI (） - **Summary**: June 2026 has become the most fiercely competitive month for model releases this year, with top labs in both the US and China making moves simultaneously—including China's Qwen 3.7 and DeepSeek V4.1, and Western models like GPT-5.6 and Gemini 3.2. The standout is DeepSeek V4 Flash, which has captured intense market interest by offering highly competitive performance at an ultra-low price of just around €0.09 per million tokens. - **Taiwan Perspective**: The explosion of ultra-low-cost, lightweight models is a massive boon for Taiwan's SMEs and independent developers, significantly lowering the barrier to entry for building AI applications. - **Key Discussion Points**: - Will this extreme price war drive the AI API market toward razor-thin margins? - In an era where so many models coexist, how should enterprises choose the best model portfolio for their specific business scenarios? - **Script Suggestion**: Have you guys noticed that the AI world has felt like a carnival this month? Models from both the US and China are dropping one after another. But what surprised me the most are models like DeepSeek V4 Flash that push prices to the absolute limit—costing only about 3 NTD per million tokens. This means that when we develop AI applications in the future, compute costs will be practically negligible. Future competition won't be about who has the biggest model, but who can integrate AI into daily workflows at the lowest cost and fastest speed. For Taiwan's many software startups, this is an incredible entry point. --- ### 6. Global AI VC Funding Tops $300 Billion in H1 2026, with Nearly 90% of Capital Flowing to the US - **Source**: Digital Applied (） - **Summary**: According to the latest reports, global AI startup funding reached a staggering $319 billion in the first half of 2026, with a whopping 88% of that capital flowing to US-headquartered companies, particularly OpenAI and Anthropic. Funding in Q1 alone surpassed the total for all of 2025, and the average Series A round for AI startups has climbed to over $50 million—roughly 30% higher than for non-AI enterprises. - **Taiwan Perspective**: Faced with the extreme concentration of nearly 90% of global capital flowing to the US, Taiwanese VCs and startups must seek out differentiated vertical application scenarios or leverage our hardware advantages to carve out a niche in the global funding landscape. - **Key Discussion Points**: - Will the hyper-concentration of AI funding in a handful of US giants lead to market monopolies and stifle innovation in other regions? - With Series A rounds averaging over $50 million, how much of a valuation bubble is baked into these numbers? - **Script Suggestion**: I was absolutely blown away when I saw these numbers. Out of over $300 billion in funding, nearly 90% went straight into the pockets of US companies—especially cash-guzzling giants like OpenAI and Anthropic. This means global capital is betting that the future of AI infrastructure will be dominated by a select few US players. For Taiwanese startups, going head-to-head in this funding war is practically impossible. However, our advantage lies in combining these powerful models with our strengths in hardware manufacturing, healthcare, manufacturing, and other vertical domains to build highly practical, real-world applications. That is our key to breaking through. --- ## Conclusion Alright, that wraps up today's exciting lineup on "Mark's Tech Insights." From the hard power of custom silicon to the soft power struggles of talent retention and national security regulations, the pace of change in the AI industry is truly breathtaking. I hope today's episode helped you make sense of it all. If you enjoyed our show, don't forget to subscribe and share it with your friends who are interested in AI. See you next time, bye! --- --- ## About this article and its author Originally published on [Mark Ku's Tech Notes](https://blog.markkulab.net/en/tech-news/ai-daily-podcast-2026-06-28) License: [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) — when reusing or quoting, credit the author and link back to the original ### About the author **[Mark Ku](https://blog.markkulab.net/en/author/mark-ku)** — Software Solution Provider - 10+ years senior software engineer, now an AI Builder - Focused on large-platform architecture — North-American e-commerce, AI SaaS subscription billing - Combining AI Agents and automation to build evolvable product foundations ### Free tools built by the author All of these are free to use: - [Free PDF Sign Tool](https://blog.markkulab.net/en/tools/pdf-sign): Online PDF sign tool — draw, type, or upload a signature, then drag, resize, and download. Everything runs in your browser; nothing is uploaded. - [VS Code Refactory](https://blog.markkulab.net/en/tools/refactory): Refactory is a VS Code refactoring extension: 34 actions plus a 37-rule code-smell inspection layer with a Code Health dashboard, across 18 languages, backed by 534 tests. It learns your repo's conventions: where interfaces live, where DI is registered, whether 'use client' belongs. It ranks files by git churn × complexity so you know what to fix first, and hands any smell to the Claude Code already on your machine. Free to use, and your source never leaves your computer. - [DB-Kit Database Manager](https://blog.markkulab.net/en/tools/db-kit): DB-Kit is a lightweight, cross-platform database manager built with Tauri + Rust + React. Manage MySQL, MariaDB, PostgreSQL, SQL Server, Oracle, SQLite, MongoDB, Redis, Kafka, Elasticsearch and RabbitMQ from one consistent interface: passwords encrypted in the OS keychain, SSH tunnels, full CRUD, a visual query builder, stacked multi-statement result sets, cross-connection data transfer and compare/sync, Excel / CSV import & export, visualized execution plans, ER diagrams, scheduled backups, SQL stress testing with p50–p99 latency percentiles, a 15-rule SQL review engine, Kafka message browsing with monitoring & alerts, a bilingual UI (Traditional Chinese / English), a built-in AI assistant (natural-language SQL, AI review and tuning advice) and the dbk CLI. Free and open source (MIT), with installers for Windows, macOS and Linux. - [VS Code Super Mermaid](https://blog.markkulab.net/en/tools/super-mermaid): Super Mermaid is a VS Code extension for beautiful Mermaid diagrams out of the box: auto-colored live preview, mouse pan & zoom, high-res PNG / SVG export, 21 templates and multiple themes. Free and open source (MIT). - [React Super Mermaid](https://blog.markkulab.net/en/tools/react-super-mermaid): react-super-mermaid is an open-source React component library: render beautiful Mermaid diagrams with a single , with built-in colorful / sketch themes, pan & zoom, in-diagram search, and high-res SVG / PNG export. Lightweight, SSR-safe, fully typed. Free and open source (MIT). - [Jira / Confluence Super Mermaid](https://blog.markkulab.net/en/tools/jira-super-mermaid): An Atlassian Forge app: write Mermaid syntax directly inside a Jira issue or a Confluence page and get flowcharts, sequence diagrams, state machines and Gantt charts. 11 diagram types, SVG / PNG export, light and dark themes, full CJK support. Runs on Atlassian: your diagrams live in your own site and the app calls no third-party service. Free, coming soon to the Atlassian Marketplace. - [Mermaid Live Preview](https://blog.markkulab.net/en/tools/mermaid-preview): Write Mermaid in your browser, see it render instantly, and share the whole diagram as a single link. No sign-up, nothing uploaded to a server, and mermaid.live share links work as-is. - [React Intl Phone Number](https://blog.markkulab.net/en/tools/react-intl-phone-number): react-intl-phone-number is an open-source React component: framework-agnostic and antd-free, with E.164 in/out, a searchable flag / country-code dropdown, configurable validation levels (strict / mobile-strict / loose), themeable CSS, and i18n — phone logic powered by google-libphonenumber. Lightweight and fully typed. Free and open source (MIT). - [Uptime Kuma Cluster](https://blog.markkulab.net/en/tools/uptime-kuma-cluster): Turn single-node Uptime Kuma into a highly available cluster: OpenResty + Lua smart load balancing, shared MariaDB state, health checks and automatic failover, plus cluster-management REST APIs. One Docker Compose command to start. Free and open source (MIT). - [Special Education](https://blog.markkulab.net/en/education): Learning materials crafted for special education students ### Daily podcasts - [Tech News](https://blog.markkulab.net/en/category/tech-news): Daily curated AI and tech trends. Catch the latest developments via audio summaries — covering AI applications, software architecture, DevOps, and engineering practice. — RSS: https://blog.markkulab.net/feed.xml - [AI Stock Chat](https://blog.markkulab.net/en/category/ai-stock-chat): Every trading day, an AI-analyzed take on the Taiwan stock market, delivered as a two-host conversation covering the session and the next-day outlook. — RSS: https://blog.markkulab.net/ai-stock-chat/feed.xml ### Subscribe to Newsletter [Subscribe to the newsletter](https://blog.markkulab.net/en/subscribe) — Be the first to know about new posts. No spam, unsubscribe anytime.