Manchi | 晚点聊 LateTalk Summary

Podcast: 晚点聊 LateTalk | Guest: Lu Zhang (张璐), Founding Partner of Fusion Fund

1. Key Themes

The Vertical Integration Arms Race in AI Infrastructure

The central thesis of the episode is that the winners in AI will be those who own the full stack — from chip design to deployment. Google's Gemini rapid improvement over OpenAI is attributed directly to its vertically integrated system (TPU → compute → model → data). Musk's TerraFab plan follows the same logic, and NVIDIA is now aggressively moving in this direction.

"Google Gemini's performance improvement speed is very fast. One of its biggest advantages is that it has its own full-stack technology system — from its own chips (TPU) to compute to AI models to data. Within this entire system, optimizing models and deploying compute becomes much easier." — Lu Zhang [00:02:56]

"Jensen is very clear that NVIDIA is not a chip or GPU company anymore — it wants to be a full-stack AI infrastructure company." — Lu Zhang [00:24:26]

The Shift from Training to Inference as the New Revenue Engine

The compute spend balance is shifting dramatically from one-time training runs to continuous inference, especially as AI agents become persistent and always-active. This is the underlying reason NVIDIA is projecting $1 trillion in data center revenue by 2027.

"Training is essentially a one-time upfront expenditure — you use many cards at once to support its cost and consumption. But in the future, inference will represent an ever-growing share of costs, because we use inference more and more... especially after we deploy agents broadly, the consumption at the inference level becomes a persistent, ongoing expenditure." — Lu Zhang [00:26:54]

"Years ago I was talking with a Microsoft AI expert. At that time maybe 70-80% was training, 10-20% inference. Now it's roughly 50/50. In the future it may flip — maybe 20-30% training, but 60-70% inference. And that inference is continuous." — Lu Zhang [00:27:52]

Enterprise AI (B2B) is the Overlooked Gold Rush

While consumer AI gets all the headlines, the real commercial opportunity right now is enterprise B2B AI — specifically vertical small language models deployed on-premise in regulated industries. Traditional enterprises are accelerating AI spend dramatically, and they actively prefer startups over big tech.

"We have a network of about 45 CTOs from Global 1000 companies. This year we universally hear that their budgets are getting larger and larger. The biggest one told me he has $12 billion in budget dedicated to AI-related acquisitions, technology integration, contracts, and strategic partnerships." — Lu Zhang [00:51:44]

"Large enterprises in traditional industries actually don't want to share their core internal data with big tech companies. They don't want to upload massive amounts of core data to the cloud. So in this context, they're more willing to work with startups — which actually accelerates vertical AI application iteration." — Lu Zhang [00:54:40]

2. Contrarian Perspectives

Space Data Centers Are Really About Escaping Government Regulation, Not Energy

While the stated rationale for Musk's space data center is energy (solar) and cooling, Lu Zhang argues the deeper strategic motivation is regulatory freedom — no government can regulate what's in orbit. This framing recontextualizes TerraFab as a governance play, not just an engineering one.

"He wants to build data centers in space. Actually, a very important underlying point is that he doesn't want any government to regulate it. In space, there's no regulation — he has tremendous freedom." — Lu Zhang [00:07:15]

"He's not just becoming the primary supporter of the space economy — he could also become the rule-setter for the entire future space economy. Becoming the rule-setter would give him even greater freedom." — Lu Zhang [00:15:04]

TPU Will Not Meaningfully Threaten NVIDIA's GPU Dominance Outside Google's Walls

Despite capital market narratives about TPU vs. GPU competition, Lu Zhang argues TPU's advantages are almost entirely locked inside Google's ecosystem and cannot be replicated by third parties — making external TPU adoption structurally disadvantaged.

"If I'm a third-party company using Google's TPU, I don't have Google's entire system. I can't achieve the same level of system optimization. So both performance and cost will take a discount... Google's own training cost is about one-third of OpenAI's — precisely because of excellent system optimization on top of TPU. But third parties can't achieve that." — Lu Zhang [00:46:23]

Space is the Natural Home for Humanoid Robotics — Not Earth Factories

The common assumption is that humanoid robots will first prove themselves in earthbound manufacturing. Lu Zhang argues the opposite: space is where humanoid robots make undeniably more economic sense today, because the cost of keeping humans alive in space far exceeds robot maintenance costs. Earth still has cheap enough human labor to make the ROI case harder.

"If you're building a space factory, even with relatively low launch costs, if you send humans to space you still have to create an ecosystem for them to survive, healthily and safely — which requires enormous technical cost and upfront investment. But if you send a robot, the maintenance and upkeep costs are much more manageable. That's why I say space factories are naturally AI-native and robotics-native." — Lu Zhang [00:11:38]

Grok's $20B Acquisition Price Was Not About the Chip — It Was About Completing a Story

The market and many observers evaluated the Grok acquisition on chip performance alone and found it overpriced. Lu Zhang reframes it: NVIDIA paid for narrative completeness and ecosystem gap-filling, not for silicon.

"If you just ask me purely looking at chip capability, it might not be worth the acquisition price. But its value to NVIDIA is not just a chip product — its value to NVIDIA is completing the entire ecosystem story and assembling all capabilities across every dimension of the ecosystem." — Lu Zhang [00:38:34]

Small Teams Serving as Acquisition Targets is the Most Reliable Path to Returns in AI Infra

Rather than building to IPO, the most capital-efficient AI strategy right now is build-to-acquire in under two years. Five of Fusion Fund's portfolio companies were acquired last year, four in under two years, at 10-20x returns. Grok: ~200 people, ~$20B acquisition = $100M per person.

"Last year we had five companies acquired. Four of them in under two years. The returns generated were 10x, 20x... This kind of fast-exit mechanism also promotes talent and capital mobility." — Lu Zhang [00:57:36]

3. Companies Identified

Fusion Fund Portfolio Company (AI Traffic Management for Satellites) AI system for managing satellite traffic and collision avoidance; also enables satellite data trading Mentioned as an early real-world example of AI applied to growing space infrastructure needs.

"This company uses AI for traffic management — because more and more satellites are being launched and they frequently collide, causing losses and creating space debris... In the process of traffic management, they can also enable satellite data trading." — Lu Zhang [00:12:07]

Fusion Fund Portfolio Company (Lunar Refueling Station) Robotic system that extracts water from lunar soil, then separates hydrogen and oxygen as rocket fuel Mentioned as a concrete near-term space economy startup that eliminates the need for rockets to carry full fuel loads.

"This company builds a refueling station in space — a robotic system that can extract water from lunar soil, then extract hydrogen and oxygen from the water. These are fuels used in rocket launches and space flight. Future rockets wouldn't need to carry as much fuel — they could refuel at the Moon." — Lu Zhang [00:13:06]

Star Cloud U.S. startup providing space computing services, based in Washington D.C. Mentioned as a direct analog to Musk's TerraFab space compute vision; reportedly in talks with the lunar refueling company.

"There's a company called Star Cloud — it's also doing space computing services. And it's based in Washington, not California. It's recently been in talks with the company I mentioned that does the lunar refueling station." — Lu Zhang [00:20:27]

Labton AI AI infrastructure company, acquired by NVIDIA; became the NVIDIA DGX Laptop Platform Mentioned to illustrate NVIDIA's speed and intent in AI infrastructure acquisitions.

"After Labton was integrated, it became the DGX Laptop Platform. Now if you go to Laptop's website, you'll be redirected directly to an NVIDIA page. This is a very important deployment for NVIDIA's future GPU cloud buildout." — Lu Zhang [00:35:41]

Nexus Flow AI infrastructure company co-founded by Jiao Jiantao, former Berkeley AI professor; acquired by NVIDIA Mentioned alongside Labton as evidence of NVIDIA's aggressive AI infra acquisition strategy.

"Another company is Nexus Flow, co-founded by Jiao Jiantao, a well-known AI scientist and professor from Berkeley. After we invested, it was acquired by NVIDIA." — Lu Zhang [00:35:11]

Grok (acquired by NVIDIA) Founded 2016; built LPU architecture optimized for low-latency, high-throughput inference; acquired for ~$20B Extensively discussed as the key acquisition enabling NVIDIA's full-stack inference and the Vera Rubin platform.

"Grok didn't go in the direction of optimizing GPUs or patching GPU problems — it redesigned the inference computation path from scratch, emphasizing low latency and high token throughput." — Lu Zhang [00:37:09]

Madra Medical robotics company; automates processes in pharmaceutical and life sciences industries Mentioned as a fast-growing Fusion Fund portfolio company in medical robotics.

"One is doing robotics automation for the traditional pharma and life sciences industry — this company is developing very fast. It's called Madra." — Lu Zhang [01:01:03]

Fusion Fund Portfolio (Micro/Nano Robotics in Medicine) Nanoscale robot systems for medical applications Mentioned as a highly promising forward-looking investment area.

"Another is doing micro-robots — nano-robots or microscale robot applications in medicine. This is also a direction I'm very bullish on." — Lu Zhang [01:01:03]

JP Morgan Chase Largest U.S. bank; Chief AI Officer is in Fusion Fund's CTO network Mentioned as example of traditional financial sector accelerating AI integration at scale.

"JP Morgan Chase — America's largest bank. Their Chief AI Officer is also in our network. This most traditional large American bank is iterating and integrating AI technology with great precision and speed." — Lu Zhang [00:56:08]

Koch Industries (KDT — Koch Disruptive Technologies) Private U.S. industrial conglomerate; runs KDT division specifically to bring new AI solutions to its portfolio companies Mentioned as a surprising fast-mover in enterprise AI adoption with very short partnership cycles.

"Koch Industries has a department called KDT — Koch Disruptive Technologies — specifically to bring in new AI solutions to all the large companies they control. Their collaboration cycle can be as fast as one to two months." — Lu Zhang [00:57:08]

4. People Identified

Lu Zhang (张璐) Founding Partner, Fusion Fund; investor in SpaceX, multiple NVIDIA-acquired AI infra companies Primary guest; expert synthesizer of AI infrastructure, space economy, and enterprise AI investment trends.

"We've had five companies acquired last year, two of which were acquired by NVIDIA — both AI infrastructure companies." — Lu Zhang [00:23:56]

Jia Yangqing (贾扬清) Founder of Labton AI; prolific open-source contributor in AI; acquired by NVIDIA Called out specifically for excellence as an AI infrastructure founder.

"Labton's founder is Jia Yangqing — many people in the AI field have heard of him. He's also a very outstanding contributor to the open-source ecosystem." — Lu Zhang [00:35:11]

Jiao Jiantao (焦建涛) Co-founder of Nexus Flow; former AI scientist and professor at Berkeley; acquired by NVIDIA Called out as an outstanding researcher-turned-founder whose company was acquired under two years after founding.

"Jiao Jiantao, a well-known AI scientist and professor from Berkeley. He built this company — we invested, and then it was acquired by NVIDIA." — Lu Zhang [00:35:11]

Jensen Huang (黄仁勋) CEO of NVIDIA Referenced repeatedly for his strategic clarity in redefining NVIDIA as an AI factory, not a chip company, and projecting $1T+ data center revenue by 2027.

"Jensen's definition of NVIDIA now is not a chip or GPU company — it's a full-stack AI infrastructure company. The core concept is the Token Economy: how to support the rise of the entire token economy industry." — Lu Zhang [00:24:26]

Peter Steinberg Individual developer who built Open Claw (OpenClaw); subsequently joined OpenAI Cited as the archetype of the small-team-to-acquisition model that is the most reliable near-term AI startup path.

"A typical example is Peter Steinberg — as an individual developer, he built OpenClaw. After it went viral, he quickly joined OpenAI." — Host [01:03:30]

5. Operating Insights

Build AI Infra Startups for Acquisition, Not IPO — Two Years is the New Timeline

The data from Fusion Fund's portfolio is concrete: four of five acquired companies exited in under two years at 10-20x returns. Founders and investors should explicitly design companies to be acquired into large platform ecosystems (NVIDIA, Qualcomm, Google, Meta) rather than defaulting to the 7-10 year IPO path.

"Four of last year's five acquired companies were under two years old. The returns generated were 10x, 20x... This fast-exit mechanism promotes both talent and capital mobility." — Lu Zhang [00:57:36]

Enterprise AI Startups Should Target Regulated Industries First — They Can't Use Big Tech

The structural insight for B2B AI go-to-market: regulated industries (finance, healthcare, insurance) are systematically blocked from using hyperscaler AI due to data sensitivity. This creates an explicit opening for startups offering on-premise or private-cloud small language models. The sales cycle can compress to 3-4 months.

"Traditional industry large companies actually don't want to share core internal data with big tech companies... so they're more willing to work with startups, which actually accelerates vertical AI application fast iteration and development." — Lu Zhang [00:54:40]

"We see many financial companies and insurance companies completing integration of new AI technology in just three to four months." — Lu Zhang [00:52:13]

Founders Should Explicitly Decide: Serve the AI Data Center or Build the Data Center

Lu Zhang flags this as a strategic fork that startups in AI infrastructure must answer clearly before building. The capital requirements, competitive dynamics, and exit paths are fundamentally different.

"Startups need to think clearly: is it better to serve such a data center, or to build the data center yourself? That's something startups need to figure out — what is your real innovation opportunity?" — Lu Zhang [00:22:26]

6. Overlooked Insights

Satellite Data is an Undervalued AI Training Asset — And It's Becoming Tradeable

This was mentioned only briefly but is structurally significant. As satellite constellations proliferate, high-quality satellite data (mineral detection, wildfire prediction, weather) is becoming a new class of AI training data. A portfolio company is building the infrastructure to manage satellite traffic and trade this data as a byproduct. This is an entirely new data market forming, largely invisible to mainstream AI investors focused on language/text.

"Most satellite data is very high quality and very valuable — not only for space applications, but for solving problems on Earth. Companies are using satellite data for mineral detection, early wildfire detection, meteorological data for Earth-side applications." — Lu Zhang [00:12:36]

"In the process of AI traffic management, they can also facilitate satellite data trading." — Lu Zhang [00:12:07]

Musk's True Moat is 3D Real-World Physical Data — Nobody Else Has It

This was mentioned briefly toward the end and largely passed over, but it may be the most important long-term competitive insight in the entire episode. The next frontier of AI is world models requiring high-quality 3D real-world data. Every other tech company has video data at best. Musk uniquely has 3D sensor data from Tesla factories, Tesla vehicles, SpaceX facilities, and satellites — making his world model / physical AI potential categorically different, not just marginally better.

"Other tech companies can't compete with him on 3D real-world data capability. Most companies at best have video data. But Musk has 3D real-world data. If he can use this data to build his AI ecosystem — especially a world-model AI ecosystem — his ecosystem capability could be an entire order of magnitude above what we see from other tech companies today." — Lu Zhang [00:40:31]