Manchi | 晚点聊 LateTalk Summary

1. Key Themes

The Inference Revolution is Cerebras's Real Opportunity

Cerebras was originally built for training, but its actual market breakthrough is in inference. The shift from training to inference as the dominant compute workload — driven by AI Agents and applications — is what unlocked Cerebras's value proposition.

"When your inference becomes very very large, then Cerebras's advantages and strengths all stand out. So I think Cerebras, rather than fighting over training, should put all its energy into doing the inference workflow well." [00:57:55]

Scaling Law Was Born at Baidu's US Research Lab Before Anyone Systematized It

The Deep Speech 2 paper (2015), with Dario Amodei as first author, contained the embryonic insight that became Scaling Law — bigger models, more data, more compute = better models. This empirical intuition predated the formal mathematical articulation by years.

"If you trace back, the birth and germination of Scaling Law was the Deep Speech paper from back then... When the model is larger, when there is more data, when training runs longer, and when the computing system is stronger, the model's performance will continuously improve. I want to call this an empirical intuition." [00:30:10]

Baidu's US Research Lab Was a Remarkable but Forgotten Talent Incubator

Baidu AI Lab at its peak had 250+ people and was arguably the densest concentration of AI talent in the world — producing founders of Inflection AI, Anthropic (Dario interned there), and early OpenAI researchers. Geopolitics ultimately destroyed its potential as an investment platform.

"The talent density at that time was truly shocking. Greg also said last week when we caught up that it was an era of gods fighting gods. The talent density, including at Google DeepMind, had never been seen before." [00:27:23]

Non-Consensus Investment Windows Are Now Measured in Weeks, Not Years

The investing landscape has fundamentally compressed. What used to be multi-year windows of non-consensus opportunity (Cerebras took 10 years from investment to IPO) now closes in one to two months. This is forcing even top VCs to shift toward late-stage "bet on winners" strategies.

"The window of non-consensus to consensus is really too short. Short to the point where you haven't reacted yet and it's already formed consensus... Now VCs have evolved into: I need to raise a large fund to invest in these companies that have already emerged." [00:29:31]

Geopolitics Destroyed What Could Have Been the Most Consequential AI Fund in History

Baidu was building an independent AI fund with OpenAI, Databricks, Anthropic, and Scale AI all on the deal list. LP fear of geopolitical risk caused the fund to collapse before it launched.

"At that time because of geopolitical reasons, many LPs who originally wanted to participate in this fund all backed out. So in the end this fund never got off the ground. I think if this fund had been built, Baidu might now be the world's best shareholder — a shareholder in all those Frontier Labs." [00:01:46]

Wafer-Scale Architecture: A Fundamentally Different Bet on Compute Topology

Cerebras's core insight is not just "bigger chip" but eliminating the communication overhead between chips entirely — compute, memory, and networking all on one silicon die. This is architecturally opposite to NVIDIA's multi-chip cluster approach.

"Cerebras is a huge brain, trying to let computation and memory both happen on the same silicon die... The advantage of GPU is its mature ecosystem and strong generality, while Cerebras's advantage is letting inference, especially AI workloads related to inference, reduce distributed communication and memory movement." [00:21:32]

Sam Altman's Early Cerebras Investment Reveals a Long-Held Strategic Vision

Sam Altman personally invested in Cerebras in 2016, just months after founding OpenAI — showing that supply chain diversification and compute independence were strategic priorities from day one of the frontier AI era.

"Sam Altman is actually Cerebras's earliest investor. Sam Altman invested in this company already in 2016, before Baidu invested in 2017. This shows that when Sam Altman was establishing OpenAI, he also predicted some of the same things — that when the AI model gets very large in the future, you cannot rely only on NVIDIA's chips." [00:11:52]

Andrew Feldman's Non-Engineer Founder Advantage

The Cerebras founder is not a chip engineer by training, yet built the most credentialed semiconductor startup (80+ people, ~70 PhDs). His edge was product definition, team organization, client understanding, and first-principles risk articulation — arguably more valuable at the system level than pure engineering expertise.

"Andrew could not only make the vision clear, he could explain each risk point one by one from first principles. For investors, investing in deep tech is not just investing in a grand vision — we need to know exactly what risk we are taking on." [00:37:56]

Physical AI's "Aha Moment" Will Arrive Earlier Than Expected

Drawing on the lesson that language model generalization surprised even the researchers who built the foundation, Zhou Nan argues physical AI will similarly exceed timeline expectations — especially because it has lower accuracy requirements than autonomous driving.

"Based on my experience with language models, I believe the aha moment of Physical AI may arrive earlier than autonomous driving, earlier than we imagine. Although it still has many problems now — data bottlenecks, hardware-software integration bottlenecks — I believe this aha moment will come sooner than we think." [01:31:26]

The NVIDIA Monopoly That Researchers Tried to Prevent Still Happened

The entire origin of the Cerebras investment thesis was to avoid NVIDIA dominance. A decade later, NVIDIA has formed a de facto AI compute monopoly in public markets, and the researchers' fear was realized.

"Ten years ago, some researchers wanted to avoid this kind of over-dependence on a single company — it still happened. Will this state continue? Where might the future disruption begin from which crack? This will be a topic we repeatedly discuss." [01:38:07]

2. Contrarian Perspectives

The Best Deep Tech Founders Are Often Not Technical Domain Experts

Conventional wisdom says chip company founders must be engineers. Cerebras's Andrew Feldman is not a chip engineer — he came from the business/product side. Yet he built the most technically credentialed AI chip startup, precisely because his strengths (client empathy, team assembly, conviction articulation, ecosystem relationships) were what a system-level company actually needed.

"For a normal chip company, basically the founders are all engineers. But for a company at the system level like Cerebras, I think Andrew's strengths at that time were in product definition, his ability to organize teams, understanding customers, and his ability to persist in the long term." [00:35:32]

NVIDIA's GPU Was Never Optimal for AI — It Just Won by Default

NVIDIA's GPU was designed for gaming and graphics, with 96% of die area not optimized for deep learning. It became the AI standard not because it was the best architecture for the task, but because Andrew Ng publicly endorsed it and no better alternative was ready in time.

"NVIDIA at that time — ten years ago — was not the chip most friendly to deep learning, because 96% of its die area was not specially optimized for deep learning. It was more invented for gaming, for graphics." [00:07:05]

Great Investments Require Non-Consensus Timing — But That Window No Longer Exists at Early Stage

Most VCs believe they should move fast on conviction. Zhou Nan goes further: the window between non-consensus and consensus has collapsed to weeks, making traditional early-stage venture in AI nearly impossible. The rational response — which top VCs are now making — is to raise large late-stage funds to back confirmed winners. This contradicts the traditional VC model at its core.

"The window from non-consensus to consensus is really too short. Short to the point where you haven't reacted yet and it's already formed consensus. So I think now for early stage Vertical AI, the test is very hard." [01:28:04]

Anthropic Was Almost a Baidu-Backed Company — and the Early Anthropic Team Couldn't Raise

When key OpenAI researchers called Zhou Nan in summer 2020 to say they were leaving to found what became Anthropic, the honest advice was: you won't be able to raise money for this. The mainstream VC market in 2020 did not believe in frontier AI as an investment.

"I said what advice can I give you? The same as back then — you need to raise a lot of money to buy compute. I said do you think you can raise it? They said it seems quite hard, not sure if we can. I said in that case this is a big challenge — you need to find people who will give you money to buy compute, and mainstream VCs won't invest in this." [01:07:39]

Physical AI Has Lower Accuracy Requirements Than Autonomous Driving — Making It More Investable Now

The prevailing view is that physical AI (robotics) is harder than autonomous driving. Zhou Nan inverts this: autonomous driving requires 99.9% accuracy or people die; physical AI has more tolerance for error, which means the aha moment can arrive before perfection is achieved.

"Physical AI — I think its requirement for task completion and accuracy will be slightly more tolerant than autonomous driving... The generalization ability of the model may be faster than you want. When your model gets larger and larger, its rate of improvement will present a steep curve." [01:32:24]

3. Companies Identified

Cerebras Systems

AI compute company using Wafer-Scale Engine architecture — one entire silicon wafer becomes a single AI compute engine with 84 chips seamlessly interconnected. Provides full hardware-software stack including chips, servers, cooling, power, and compiler software. Recently IPO'd at ~$500B market cap (briefly near $1T). Has a $20B+ contract with OpenAI.

"Cerebras is in a very disruptive architecture outside of NVIDIA — it redesigned the entire architecture for AI training and inference... It put one entire wafer into a huge AI training engine, with 84 chips seamlessly interconnected on that wafer." [00:04:17]

NVIDIA

Dominant AI compute provider. Its moat is not just chips but the CUDA ecosystem, developer community, networking, software, customer trust, and supply chain.

"NVIDIA's moat is not just these chips, but its CUDA, plus its developer ecosystem, networking, systems, software, customer trust and supply chain — this entire series of ecosystem is its strongest moat." [00:08:01]

OpenAI

Frontier AI lab, signed a $20B+ compute contract with Cerebras in January 2026. Sam Altman personally invested in Cerebras in 2016. GPT-3 was in post-training in summer 2020.

"OpenAI — I think the logic is very clear. Any model company, I think, needs to have diversified supply. You cannot just depend on this one chip provider." [00:13:18]

Anthropic

Frontier AI lab founded by former Baidu US Lab and OpenAI researchers (including Dario Amodei). Currently valued at ~$1 trillion range. Zhou Nan describes investing in Anthropic now as a "no-brainer."

"Anthropic — I think investing in Anthropic is a no-brainer. If this company becomes a trillion-dollar company, I think that would not be surprising." [01:30:29]

Databricks

Data warehouse/engine company. Was on Baidu's fund deal list in 2018. Zhou Nan sourced it very early but the fund never closed.

"I also sourced Databricks very early — it was on our deal list. Because at that time I had a thesis to invest in Data and Data Engine." [01:01:17]

Scale AI

Data annotation company founded by Alexander Wang in 2016. Was on Baidu's fund deal list.

"Even Scale AI was on my shortlist, because Alexander Wang came out in 2016 to do it, and Data Annotation was also one of my investment theses." [01:02:45]

Graphcore

Alternative AI chip company. Was considered alongside Cerebras; eventually acquired by SoftBank for $500M in 2024.

"Graphcore was acquired by SoftBank in 2024 for $500 million... The conclusion at that time was it could improve speed and efficiency, but its effect was not as good as Cerebras, and its architecture was also not as disruptive as Cerebras." [00:33:04]

Wave Computing

Early AI chip company with similar philosophy to Cerebras. Was eliminated first from Zhou Nan's consideration due to team issues. No longer exists.

"This company — its architecture and concept were quite similar to Cerebras, but its team had problems, so this company was the first one I eliminated." [00:34:03]

G42

UAE-based technology company. Was both an investor and early major customer of Cerebras, creating single-customer concentration risk.

"At that time there was a single-customer risk — G42 was both its investor and its customer." [00:55:27]

Coatue (Code2)

Major technology investment fund. Was an early Cerebras investor. Thomas LaFont (Founding Partner) was the person who introduced Zhou Nan to Cerebras. LaFont had a background as a semiconductor research analyst.

"Thomas LaFont — he is Coatue's Founding Partner, and at the same time he spent a long time researching semiconductors. So he was very excited about the pitch I was giving, then started telling me he had invested in a very disruptive company called Cerebras." [00:48:47]

Benchmark

Early Cerebras investor and board member. Recently raised a ~$2B fund that closed within 24 hours.

"Benchmark just today announced raising a two-billion-dollar fund — they had a 24-hour window, immediately closed, raising one to two billion in one day." [01:29:01]

Foundation Capital

Cerebras board member.

"The three board members are Benchmark, Foundation Capital, and Eclipse." [00:54:30]

Eclipse

Cerebras board member. [00:54:30]

Groq

AI inference chip company (LPU architecture). Acquired by NVIDIA in December for $2B. Now integrated into NVIDIA's compute platform as an inference-focused solution.

"NVIDIA acquired Groq — LPU. When Groq was raising in 2019-2020, I saw Groq and thought this company's thinking was quite similar to Cerebras." [00:23:01]

Fireworks AI

AI inference infrastructure company. Named by Zhou Nan as a category winner worth investing in now.

"There are a few I like, like Fireworks, and like Fil, and like BaseTen — some infra companies that have formed an early flywheel effect." [01:30:58]

BaseTen

AI model serving infrastructure company. Named as an infrastructure winner. [01:30:58]

Agan (likely "Ngen" or similar — garbled)

Inference optimization startup. Founded by a student of MIT Professor Song Han. Acquired by NVIDIA in March of the relevant year, just months after founding. Zhou Nan discovered it in August of the prior year when it was just starting.

"There is a company called Agan, it was acquired by NVIDIA in March of this year. This is a very typical inference optimization startup. I discovered this company last August when it had just been established — introduced through a good friend of MIT Professor Song Han, whose student is the founder." [01:23:40]

Kunlun Chip (昆仑芯)

Baidu's AI chip spinout. Named as one of Baidu's lasting AI hardware contributions, connected to Zhou Nan's work period at Baidu.

"Kunlun Chip — at that time when I was investing in Cerebras, I also communicated with the person in charge of Kunlun Chip. I think at that time he already had the initial intention of founding Kunlun Chip." [01:15:54]

Pony.ai (小马智行)

Autonomous driving company. Named as a Baidu US Lab spinout that Zhou Nan wished he had invested in.

"If you invest in everyone who came out of Baidu US Lab, including like Pony.ai, and of course Anthropic can also count..." [01:17:23]

Inflection AI

AI company. Named as having co-founders who came from Baidu US Lab.

"Some of the researchers — they are Inflection co-founders, Adapt co-founders, and some also became founding members of Meta's FAIR lab, not to mention OpenAI and Anthropic." [00:46:25]

Adapt AI

AI company. Named as having co-founders from Baidu US Lab. [00:46:25]

Coreweave

Neo-cloud GPU provider. Named as an example of NVIDIA-aligned cloud infrastructure.

"NVIDIA has its own — like Coreweave that kind of thing, and close cooperation with Neo Cloud like Coreweave. These Neo Clouds also have a strong pain point — their compute is not enough." [00:17:39]

Ninus (likely "Manus" or similar AI Agent company — garbled)

General purpose AI Agent. Named on Zhou Nan's February 2025 investment memo as a category to watch.

"I listed General Purpose AI Agent, at that time I listed Ninus and GenSpark." [01:27:34]

GenSpark

AI Agent platform. Named on Zhou Nan's February 2025 investment recommendation list. [01:27:34]

Fil / FIL

Multi-modal AI infrastructure company. Named as a current category winner in infra.

"Like multi-modal Infra companies, like Fil." [01:27:35]

4. People Identified

Zhou Nan (周南)

Investor at Qualcomm Ventures. Former Baidu US AI Lab investment team member. Made the Cerebras investment as his first deal. Has 10 years of AI investing experience spanning the full arc from 1.0 to 3.0. Identified Databricks, Scale AI, OpenAI, Anthropic early — but was unable to invest due to fund failure.

"I started my AI investment career ten years ago at Baidu's US Research Institute, and I witnessed the entire AI 1.0 era take off." [00:02:22]

Andrew Feldman

CEO and co-founder of Cerebras Systems. Serial entrepreneur; previously founded SeaMicro (acquired by AMD). Not a chip engineer by training — his strengths are product definition, team building, and customer understanding. Was in his 40s when he founded Cerebras. Spent 4 weeks doing 2-hour daily sessions to walk Zhou Nan through every risk point.

"Andrew is a very outstanding entrepreneur — he can not only make the vision clear, but can explain each risk point one by one from first principles." [00:37:26]

Andrew Ng (吴恩达)

Former Chief Scientist of Baidu and head of Baidu US AI Lab. Credited with publicly establishing that GPUs could be used for AI model training, giving NVIDIA crucial early endorsement from the AI research community.

"Andrew Ng was probably the first to announce to the world that GPUs can be used very well to train AI models as a compute system. So I think there is a very important connection between Andrew Ng and NVIDIA." [00:26:27]

Dario Amodei

Co-founder and CEO of Anthropic. First author on the Deep Speech 2 paper. Was recruited into Baidu US Lab by Greg (not through traditional hiring) despite having a math, physics, and biology background — not computer science or AI. Called Zhou Nan in summer 2020 to discuss leaving OpenAI.

"Dario's ability to enter Baidu was actually a very important part of his career. Dario was recruited in by Greg. Greg DeAmends, after meeting Dario, thought this person is talented — his vision for AI is very deep, even though he had no AI background." [01:11:04]

Greg DeAmends (Greg D'Amens / Greg Diamos)

Key NVIDIA CUDA architect during his PhD; one of the central people who built CUDA's ecosystem. Was recruited to Baidu US Lab. Co-author on Deep Speech 2. Brought Dario Amodei into Baidu. Led hands-on due diligence of Cerebras's simulator. Later co-founded a company with Andrew Ng. Was training a 300-million parameter language model at Baidu that took 3+ months on NVIDIA GPUs — creating the direct pain point that motivated the Cerebras investment.

"Greg is the key dealer of NVIDIA's CUDA. Greg during his PhD was helping NVIDIA build its CUDA ecosystem. You can say without exaggeration that Greg was one of the most important people building CUDA for NVIDIA at that time." [00:39:21]

Sam Altman

CEO of OpenAI. Personally invested in Cerebras in 2016 — before OpenAI was even a year old. Had a mentoring relationship with Baidu's Lu Qi. Had friendly enough relations with Baidu that OpenAI was willing to accept Baidu investment circa 2017-18.

"Sam Altman invested in this company already in 2016 before Baidu invested in 2017. This shows that when Sam Altman was establishing OpenAI, he also predicted the same things." [00:11:52]

Thomas LaFont (Thomas Laffont)

Co-founding Partner of Coatue. Former semiconductor research analyst. Introduced Zhou Nan to Cerebras during a Baidu office visit. His semiconductor background allowed him to immediately understand the investment thesis.

"Thomas LaFont — he is Coatue's Founding Partner, and at the same time he spent a long time researching semiconductors. So he was very excited about the pitch I was giving, then started telling me he had invested in a very disruptive company called Cerebras." [00:48:47]

Robin Li (李彦宏)

Co-founder and CEO of Baidu. Was on the investment committee that approved the Cerebras deal in under two days. Had early vision in AI, establishing Baidu US Lab and participating in conversations around Geoffrey Hinton's lab.

"Robin had that much vision — establishing Baidu's US Research Institute so early, and at that time participating in advising about Geoffrey Hinton's lab." [00:49:43]

Lu Qi (陆琪 / Qi Lu)

Former Baidu COO/President. On the investment committee for Cerebras. Was described as Sam Altman's mentor, which partly explains OpenAI's early willingness to accept Baidu investment.

"Sam Altman's relationship with Baidu was still quite good at that time, because Lu Qi was apparently his mentor." [01:01:46]

Jennifer Lee

CFO of Baidu. Was Zhou Nan's direct boss and also on the Cerebras investment committee.

"At the investment committee were my boss Jennifer Lee, who was Baidu's CFO, as well as Lu Qi, and Robin himself." [00:46:25]

Song Han (韩松)

MIT Professor. His student founded the inference optimization startup that was acquired by NVIDIA. Also co-founded a chip company, DeepScale.

"Through a good friend of MIT Professor Song Han — his student is the founder." [01:23:40]

Andrew Ng / Manchi (曼琪)

Host of the LateTalk (晚点聊) podcast and interviewer for this episode.

Eric Schmidt

Early Anthropic investor. Former Google CEO. Named as one of the non-VC early backers of Anthropic.

"Anthropic's earliest investors were actually not VCs — one was Eric Schmidt, Google's former CEO." [01:05:40]

Demis Hassabis

Named as an early Anthropic supporter alongside Eric Schmidt.

"One was Eric Schmidt and another was Hassabis." [01:05:40]

5. Operating Insights

Run Structured Risk Decomposition Before Any Deep Tech Investment, Not Just High-Level Diligence

Zhou Nan's Cerebras diligence is a template: enumerate every failure mode (yield/良率, heat dissipation, power, compiler, customer adoption), calculate the worst-case scenario cost and time delay for each, verify that company cash flow can absorb the worst case, and only then assess whether the risk envelope is acceptable. He brought in Baidu AI researchers for compiler/software risk, hardware engineers from the autonomous driving team for physical risks, and Stanford professors for chip manufacturing questions.

"We calculated out all the worst-case scenarios — maybe it would take an extra six months, maybe cost an extra five to ten million dollars. Can the company's cash flow support going through the wafer fab process again? We at the time thought this risk was controllable." [00:42:43]

Use Simulator Validation as a Proxy for Real Hardware When the Chip Hasn't Taped Out Yet

Cerebras had no physical chip when Baidu invested — only a simulator. The diligence insight was to run Baidu's actual (world's largest at the time) language model on that simulator under explicit assumptions, generating directional results that were positive enough to proceed. This is a replicable method: find the organization with the most demanding real-world workload and run it on the simulator, making all assumptions explicit.

"Only Baidu could say — on its architecture, run such a model, and give a diligence result. The diligence result at that time was also very important for Cerebras, because in a certain sense you verified its idea, verified its technical concept." [00:40:46]

For Hardware Companies, Build Cloud Wrapper First to Eliminate Adoption Friction

The single biggest go-to-market barrier for hardware companies is the time and effort required for customers to deploy new hardware (modify data center, rewrite software stack, retrain team). Cerebras's strategic cloud platform eliminates this entirely — customers use an API, never touching the hardware layer. This is a generalizable playbook for any hardware company seeking faster enterprise adoption.

"Hardware companies have a big challenge called Adoption Friction. If you tell a customer to buy a whole new hardware set, deploy it to their data center, then modify their software stack — that cycle is very very long. Cerebras Cloud wraps the entire complex underlying system. The customer can directly use API, immediately plug in Cerebras's product." [00:16:10]

Write an Investment Memo in Two Languages, Distill to Core Risk/Return, and Set a Fast Decision Deadline

Zhou Nan submitted a ~10-page investment memo in both Chinese and English within four weeks of beginning diligence. The committee approved in under two days. The speed and bilingual format served two functions: it forced rigorous distillation of conviction and risk, and it respected the time of a senior investment committee whose approval would otherwise take months. The lesson: memo quality and format discipline are as important as the underlying analysis.

"I sent my deal memo of about ten pages — one Chinese draft and one English draft — to the investment committee after four weeks. In less than two days they approved, passed the investment." [00:46:52]

6. Overlooked Insights

The CPU Renaissance in Inference Is the Biggest Unpriced Opportunity in AI Compute Right Now

Zhou Nan briefly mentions — almost in passing — that new CPU architecture startups are already being founded to address inference workloads, specifically the sequential/non-parallelizable tasks in AI Agent workflows that GPUs handle poorly. He says he personally knows teams already working on this, and Intel and AMD have both seen significant stock appreciation reflecting this shift. This was mentioned in less than 60 seconds but is potentially a multi-hundred-billion-dollar structural shift.

"I also believe there may be a new CPU architecture appearing as an inference chip. Because in the current inference chip space, Cerebras is a company that has just broken out. There may also be some other companies based on CPU solutions coming out. I know some startups are already working on this — very early stage. I think this is also a point worth paying attention to." [01:34:49]

The host's closing commentary reinforces this: Intel and AMD have both had significant stock gains recently, which this thesis directly explains. This is a non-obvious investment angle that was barely surfaced in the conversation but sits at the intersection of the AI Agent boom, inference economics, and semiconductor architecture — a convergence that could produce the next generation of compute infrastructure winners.

The Anthropic Origin Story Reveals the Exact Moment When Mainstream VC Consensus Was Still Missing

Zhou Nan received the first call from future Anthropic founders in summer 2020, when GPT-3 was in post-training — two full years before launch. The founders themselves weren't sure they could raise. This is a precise data point showing that even the people building the most important AI company of the decade couldn't get conviction from the market. The implication for investors: the people closest to a breakthrough are often the last ones believed, and calls from credible technical insiders who are "not sure they can raise" are the highest-signal investment opportunities that exist.

"In summer 2020 they called me and said GPT-3 we're almost done training it... They said they wanted to come out and start a new company, and asked if I had any advice... They said it seems quite hard, not sure if we can raise. I said in that case this is a big challenge." [01:06:09]

This pattern — credible technical founders who can't raise — perfectly predicts Anthropic's subsequent trajectory. Identifying this signal type and acting on it is the single most valuable early-stage investing lesson in the entire episode, yet it was delivered as an anecdote of regret rather than a named framework.