🎙️ AI 日報 Podcast — 模型輕量化、算力軍備賽、實體 AI

Opening

(Intro music)

Mark: Hey everyone, welcome to Mark's Tech Insights! I'm Mark. Today's AI news is "five layers of fire and ice." On one hand, AI models are racing toward extreme lightweighting and high efficiency. On the other, the compute arms race and capital spending behind them is bigger than ever. Today we'll dig into these contradictory and fascinating trends, and see how AI is getting both "lighter" and "heavier" at the same time.

Today's Top Stories

1. PrismML Unveils 1-bit LLM "Bonsai": 8x Faster, 80% Less Energy

Source: HPCwire / AIwire (link)
Summary: A startup just emerging from stealth, PrismML, announced a family of 1-bit LLMs called Bonsai. With extreme weight compression, they dramatically reduce memory footprint, latency, and energy consumption — claiming 8x speed on existing hardware and a 75-80% drop in energy use.
Taiwan Angle: This is great news for Taiwan's IC design firms and end-device makers. If 1-bit models become a trend, chips optimized for this architecture will be the key to winning Edge AI.
Discussion Points:
- Will 1-bit compression sacrifice model accuracy or "reasoning ability"?
- What's the impact on GPU giants like NVIDIA, who focus on high-precision compute?
- Which end applications (phones, cars, IoT) will explode first because of this?
Talking Points:

Mark: Hey, today's first story is super cool! A company called PrismML dropped a bombshell — a 1-bit model called Bonsai. "1-bit" means they compress the model parameters from complex floating-point numbers down to just 0/1 switches — extreme compression. That makes the model run 8x faster while using only a quarter of the power. What does that mean? Soon the AI assistant on your phone might not need to call the cloud at all to do real-time, complex conversations — and barely use any battery. For Taiwan's hardware industry, this path is a must-watch. Whoever first builds chips best suited to running 1-bit models will be in pole position for the next Edge AI wave. But the thing I'm most curious about is — with this much compression, will the AI get dumber? That's the test the field will run next.

2. Broadcom Expands Chip Deals with Google and Anthropic, Covering 3.5 GW of Compute

Source: CNBC (link)
Summary: Broadcom signed bigger orders to manufacture next-generation custom AI chips (ASICs) for Google and Anthropic. The Anthropic portion alone will require around 3.5 gigawatts of compute starting in 2027. It shows top AI companies are designing their own dedicated chips and building independent supply chains to lock in compute.
Taiwan Angle: This re-confirms that the "custom chip" trend is unstoppable. The end fab might still be TSMC, but for Taiwan's IC design firms it's both a warning and a chance — they need to push more aggressively into high-value, custom AI chip design services.
Discussion Points:
- 3.5 gigawatts — what does that mean? Why don't top AI companies just buy NVIDIA GPUs?
- Does this mean the AI chip market is moving from "NVIDIA dominates" to "GPU + various ASICs"?
- With Google (TPU) and Anthropic both designing in-house chips, what pressure does that put on OpenAI?
Talking Points:

Mark: We just talked about models getting lighter — now look at how absurdly heavy compute is getting. Broadcom is making custom AI chips for Google and Anthropic, and Anthropic's compute alone is 3.5 gigawatts! Do you know what that means? That's roughly the output of 2-3 nuclear reactors — just to train one company's AI models! The strategic point is they don't want to be held hostage by NVIDIA anymore. General-purpose GPUs are great, but power and cost are too high. They want chips tailored to their own models — like building a supercar designed specifically to run Claude. The takeaway for Taiwan: big customers want more than fabs — they want "design partners." TSMC is sitting comfortably as foundry king, but the design step in front of it is where you climb the value chain.

3. Q1 2026 VC Funding Hits Record: $300B into Startups, 80% to AI

Source: Crunchbase News (link)
Summary: Global VC went insane in Q1 2026, with $300 B d e p l o y e d .80 300B deployed. 80% of that —$ 242B — flowed to AI companies. Four of the five largest investments in history happened this quarter, led by OpenAI, Anthropic, xAI, and Waymo.
Taiwan Angle: Most of the money concentrates in U.S. top players, but the spillover effect will be massive. Valuations and fundraising opportunities for Taiwan's AI startups will rise, and the war for top AI talent will intensify.
Discussion Points:
- Is this AI investment frenzy healthy growth, or a giant bubble?
- Money concentrates in a few giants — is that good or bad for the startup ecosystem?
- Beyond model companies, which underrated investment opportunities exist in the AI app or infra layers?
Talking Points:

Mark: Buckle up for this number. Just three months into the year, $242 b i l l i o n — o v e r N T$ 7 trillion — has been poured into AI companies. What does that mean? More than Taiwan's entire annual government budget! OpenAI alone took $122 billion — they're literally a wealthy country. This is the "money arms race" behind the "compute arms race" we keep talking about. Honestly, there's definitely some bubble in here, but it's also acting as a super catalyst — using money to compress what would happen in the next ten years into two or three. For Taiwan's founders and engineers, it means your market value is rising — but the competition has never been fiercer.

4. NVIDIA Releases New Physical AI Models — Robotics Goes Mainstream

Source: NVIDIA Newsroom (link 1, link 2)
Summary: NVIDIA is pushing AI from the digital world into the physical one. They released "Physical AI Models" designed for robotics and announced they're open-sourcing their "Physical AI Data Factory" blueprint, helping developers generate simulation data and reinforcement learning environments at scale to train robots.
Taiwan Angle: Taiwan is a hub for industrial computers, automation, and precision manufacturing. NVIDIA's full robot development stack — from simulation to training to deployment — is a super weapon for Foxconn, Delta Electronics, Quanta, Advantech, and others to accelerate smart factories and autonomous robotics.
Discussion Points:
- How are "Physical AI Models" fundamentally different from the LLMs we know?
- How important is open-sourcing the "data factory" blueprint for accelerating the whole robotics industry?
- When robots can understand the physical world and self-learn, how far are we from general-purpose humanoid robots?
Talking Points:

Mark: After all the cloud-based models and data talk, let's see how AI walks into the real world. NVIDIA isn't just selling chips this time — they're handing you a complete toolbox for "building a robot brain." They launched what they call "Physical AI" — AI that understands physics and can interact with environments. Even better, they're open-sourcing the "data factory" used to generate training data. It's like NVIDIA isn't just selling you flour and an oven — they're publishing a Michelin chef's recipe. For Taiwan's manufacturing sector, this is huge news. We've always been strong in hardware; this set of tools is like giving robots a "soul." Future factories may not just have robotic arms — they'll have AI workers that can judge and learn on their own.

5. Utah Sets a U.S. First: AI Systems Authorized to Renew Drug Prescriptions

Source: Crescendo AI / Latest AI News (link)
Summary: Utah passed legislation becoming the first U.S. state to formally grant AI systems the legal authority to renew drug prescriptions for patients. The intent is to ease physician shortages, but it's also sparking heated debate over medical liability, patient safety, and regulation.
Taiwan Angle: Taiwan's health insurance system and medical staff have been under sustained pressure for years. The Utah case is a valuable reference, but bringing it to Taiwan — the legal updates, public trust, and most critically, "who's responsible when something goes wrong" — will be far harder to solve than the technology itself.
Discussion Points:
- If the AI prescribes incorrectly, where does liability fall — the physician, hospital, or AI vendor?
- Which prescription types (chronic care vs. controlled substances) are appropriate for AI?
- Would you personally let an AI decide your medication? Where's the line of trust?
Talking Points:

Mark: I think this last story deserves serious thought. Utah actually passed a law allowing AI to renew your prescriptions. It sounds futuristic and a bit scary. The upside is obvious — free doctors from massive, repetitive admin work to focus on more complex cases. But there's a big problem: if the AI gets it wrong and someone gets hurt, who takes responsibility? The supervising physician? The AI vendor? In Taiwan, this would absolutely raise huge questions, and our NHIA and Ministry of Health are surely watching closely. The big shift is this: AI is no longer just an "assistive tool" — it's becoming an entity with "decision-making power." Once you cross that line, how society works can change in unexpected ways.

Closing

(Outro music)

Mark: Alright, today we went from ultralight 1-bit models to ultraheavy compute investment, then into Physical AI walking into real factories and pharmacies. You can feel that AI is expanding at a pace that's hard to imagine, in every dimension at once. It's an era full of opportunity and challenge. Thanks for listening — I'm Mark, see you next time on Tech Insights!