AMD, Starcloud, Coatue..10 Hot Takes From The Biggest Names in AI

1. Key Themes

The Death of the Keyboard and Mouse as the Primary Human-Computer Interface

Max Cook of Coatue argues we are in the middle of a fundamental shift in how humans interact with computers — from structured GUI inputs to natural language and voice. He draws a historical arc from mainframe → PCs → desktop internet (mouse/keyboard) → mobile (touchscreen) → now agentic AI (natural language). Coatue investors are already "whispering into their collar" to give models better context.

"It seems pretty clear at this point, I think there's a good stat that like the average person has 60 to 80 apps on their phone. They use 11 apps daily on average and 30 apps in any given month. Like it seems quite clear that we're going from a world in which you have 60 to 80 apps to an agentic AI world where you've got a lot fewer than that. Two, three, four apps where everything else plugs in as an API or an MCP." 00:01:28 — Max Cook

"We've always conformed ourselves to computers like with that by inputting specific data with a keyboard and a mouse. And now we're finally entering the era where computers have to conform to us." 00:02:14 — Max Cook, quoting Greg Brockman

Open Source Model Adoption Is Growing Faster Than the Market Realizes — and Will Cross 50% of Enterprise Token Share by End of Year

Multiple speakers across the episode converge on the same signal: enterprise open source model usage is accelerating rapidly. Matan Grinberg of Factory has the data to back it — from under 1% to 10% open model token share within months.

"At the beginning of the year, we saw enterprises were using less than 1% open models, like less than 1% of their tokens were going to open models. Then in around March, it kind of crossed that 1% threshold. And then in May, it crossed the 10% threshold. So it's growing pretty quickly because of the kind of cost optimization." 00:21:35 — Matan Grinberg

"By the end of the year, I think we'll probably get to at least in the enterprise, we'll probably cross the 50% threshold by the end of this calendar year." 00:22:25 — Matan Grinberg

The Open vs. Closed Model Debate Is a False Binary — The Real Question Is What Happens When Frontier Labs Offer Fine-Tuning as a Product

Max Cook reframes the entire open/closed debate. The real disruption is not which model wins, but whether Anthropic and OpenAI offer model routing and fine-tuning of their older, smaller models as a product — which would compress the open source advantage.

"I don't think anyone's asking the question of what happens when Anthropic and open AI decide we can offer that as a product. I don't know that they're going to do that, but it seems plausible that they would see this opportunity to take their older, cheaper, more efficient, smaller models and sell that as a product." 00:08:14 — Max Cook

Agentic AI Is Transforming End-to-End Workflows at a Speed Even Industry Insiders Didn't Anticipate

Mark Papermaster, CTO of AMD, makes the point from a hardware perspective — agentic workflows are so intensive that the CPU-to-GPU ratio is approaching 1:1, a ratio nobody predicted. AMD itself is using thousands of sub-agents in chip design.

"None of us anticipated how fast these agentic workflows would just take off. The last six months have just been amazing. And it's just a harbinger of what's yet to come." 00:15:29 — Mark Papermaster

"We have thousands of sub-agents just doing incredibly complex chip design, speeding how fast we can get the next generation to market." 00:15:56 — Mark Papermaster

Foundation Models Must Be Built Along Three Axes — Capability, Efficiency, and Substrate — Not Just Intelligence at All Costs

Ramin Hasani of Liquid AI argues the entire AI industry has been optimizing on a single axis (maximum intelligence) and is now hitting the wall of energy, chip supply, memory supply chain, and geopolitical risk. The sustainable path requires co-evolving capability, cost efficiency, and deployment substrate simultaneously.

"All foundation model companies from 2019, when pre-training started panning out, they started working on one axis, you know. The axis was maximizing intelligence at all costs. Now, we are at a place where we see that, look, efficiency is not an afterthought. Energy is not abundant." 00:34:56 — Ramin Hasani

"Where does this intelligence system go? You know, so if you think about AI, AI is majorly like getting hosted in data centers, but you could also bring intelligence on phones, on laptops, on airplanes, on cars, you know, like you can actually host AI at any different substrate." 00:36:48 — Ramin Hasani

AI Is Moving to the Physical World — On-Device and Edge Intelligence Is a Major Emerging Theme

Liquid AI is powering in-car intelligence for Mercedes-Benz and building models cheap enough to run on Raspberry Pis. StarCloud is building orbital data centers. The physical world deployment of AI — not just cloud — is a major structural trend being discussed by multiple participants.

"We're building foundation models that are so cheap that you can bring them on Raspberry Pis, for example. So you can host them on any kind of device that is on the planet." 00:32:47 — Ramin Hasani

"We are powering like in-car intelligence for Mercedes-Benz. Foundation models getting inside the car and making your car basically smarter." 00:34:09 — Ramin Hasani

The Inference Economics Problem Is Creating a New Category: Premium Inference

SambaNova's entire Series F thesis is built on solving inference economics — running trillion-parameter models at full precision faster and cheaper than GPUs, enabling service providers to charge more while spending less. This is a new framing of the inference layer.

"What we're able to do with premium inference is drive the performance of the largest models up. So you can actually deliver high quality models with these trillion parameter models at their full precision at incredible speeds. And when you're running faster with high accuracy models, you can charge more at a lower cost, allowing your service providers to generate more margins for their services." 00:39:59 — Rodrigo Liang

SpaceX's IPO Will Be Viewed as the Most Undervalued in History — Space Is the Infrastructure Layer for the Next Century

Philip Johnston of StarCloud makes a sweeping bull case for SpaceX, arguing they own the equivalent of the railroads for every space industry that follows — asteroid mining, lunar resources, comms, orbital compute.

"My hottest take right now is that the SpaceX IPO is going to be viewed historically as the most undervalued IPO of all time. I think they will tear through 10 trillion within, you know, certainly within the next couple of years." 00:24:06 — Philip Johnston

Most AI Is Downstream of Good Data — The Industry Is Overcomplicating What Is Fundamentally a Data Science Problem

Pim de Witte of General Intuition delivers the bluntest take at the event: most of what is being called AI is glorified data science, and model quality is primarily a function of data quality.

"Most of AI is just glorified data science. Good models mostly are downstream from good data and simple things." 00:13:31 — Pim de Witte

The Interaction Model Layer — Real-Time Multimodal AI Interfaces — Is Underloved and Will Sit on Top of Every Intelligence Stack

Max Cook flags that the Thinking Machines interaction model demo from May went underloved, and that real-time multimodal interfaces (voice, image, zero-latency) will become the universal layer sitting above every model and every app.

"To me that Thinking Machines interaction model demo was like a bit of a peek into the future. It's you're going to interact with compute at all times. It's got to be have zero low latency back and forth. It can ask to be able to input images, understand what you're saying in real time, respond to you. And so I watched that and said, of course, this is like going to sit on top of every intelligence layer, every model, every app." 00:03:55 — Max Cook

---

2. Contrarian Perspectives

Open Source Will Take 50%+ of Enterprise Token Share by End of 2025 — Against the Current Data Showing Decline

The counterintuitive data point is that enterprise LLM spend on open source actually dropped from 19% to 11% last year — yet Matan Grinberg predicts a 50% open source token share by year end, and his own platform data shows 0% → 10% in five months.

"Enterprise LLM spend was 19% open source last year and it's 11% this year. So it's dropped. Yet you've got businesses like Decagon telling you that 90% of their workloads are done via open source." 00:07:15 — Max Cook

"I think BigToken has been spewing some propaganda against Chinese models, which is how they're labeling the open models. And I think in fact these open models are incredibly performant, incredibly cheap, incredibly fast." 00:21:07 — Matan Grinberg

Open AI Is About to Introduce Almost a Billion People to Agents for the First Time — The Super App Moment Changes Everything

Most people are discussing the gradual adoption of agents. Max Cook argues a discrete discontinuity is coming: OpenAI's super app will onboard nearly a billion weekly active users to agents simultaneously, which will force an entirely new device and interaction paradigm.

"You've got open AI coming out with a super app for the first time. That's going to introduce almost a billion weekly active users to agents for the first time. It's going to change how people interact with their phone and change how people want to interact with their phone. And we're going to need something new for that." 00:05:25 — Max Cook

Orbital Data Centers Are a Real Infrastructure Play — Not Science Fiction

Philip Johnston is building data centers in space and was quoted in SpaceX's S1. He's moving into a new campus in August to manufacture StarCloud 3 vehicles capable of fitting 10 megawatts of compute per Starship launch, and is in active conversations with SambaNova and Cerebras about flying their chips.

"We can potentially fit up to 10 megawatts of compute capacity per Starship launch." 00:25:59 — Philip Johnston

Intelligence Maximization as the Sole Axis of AI Development Is a Dead End — Efficiency and Substrate Must Be Co-Designed from the Start

Against the dominant industry assumption that scaling wins, Ramin Hasani argues the energy, chip supply, and memory supply chain constraints mean intelligence-at-all-costs is structurally unsustainable. Companies retrofitting efficiency after the fact (distillation) are already behind.

"If you have maximized intelligence at all costs, that axis alone is not going to get you to the place that you want to go." 00:35:26 — Ramin Hasani

The Frontier Model vs. Open Source Debate Is Missing the Key Variable: What Frontier Labs Do Next

Most analysis treats OpenAI and Anthropic as purely closed and assumes open source grows at their expense. Max Cook argues both could offer fine-tuning and model routing of their older models, potentially collapsing the open source advantage entirely.

"It seems plausible that they would see this opportunity to take their older, cheaper, more efficient, smaller models and sell that as a product. So that's like the question I don't think is being asked enough when we consider what is the end state compute ecosystem look like." 00:08:14 — Max Cook

---

3. Companies Identified

Coatue

Large multi-stage investment firm. Cited as an early mover on the AI compute-to-app thesis; their analysts are literally whispering into microphones to interact with AI models during the workday, demonstrating real internal adoption.

"We've got a number of investors that are whispering into their collar half the day. Because that's how to give models better context, better prompting and ultimately like produces much better work output." 00:02:44 — Max Cook

Factory (Droids)

AI software engineering automation platform. Cited for having real enterprise data on open vs. closed model token share transitions, and for actively routing customers to the best model (open or closed) dynamically for each task.

"The droids are in Europe. They've recently made the jump across the pond. A lot of the largest banks and enterprises in Europe are now automating their software engineering with droids and with factory." 00:20:36 — Matan Grinberg

StarCloud

Orbital data center company. Building space-based compute infrastructure on top of SpaceX's Starship, moving into a major new manufacturing campus in August, in discussions with SambaNova and Cerebras about flying their chips. Quoted in SpaceX's S1.

"We are about to move into a much, much larger facility. We're building a kind of campus to set up a huge manufacturing line for what we're calling StarCloud 3, which is like the vehicle which will fit on Starship. So we can potentially fit up to 10 megawatts of compute capacity per Starship launch." 00:25:53 — Philip Johnston

SambaNova Systems

AI inference chip company. Just announced $1 billion Series F at $11 billion valuation led by General Atlantic. Building the SN50 RDU chip for premium inference — running trillion-parameter models at full precision faster and cheaper than GPUs. StarCloud is in talks to fly their chip in orbit.

"We're announcing the first close, our series F is a $1 billion raise at $11 billion valuation." 00:38:55 — Rodrigo Liang

Black Forest Labs

Multimodal visual AI company based in Germany, building visual models for content creation and physical AI. Strong open source advocate; Robin Rombach is a founder and argues open innovation is fundamentally good for safety and accessibility.

"We are building multimodal visual models for content creation and for now physical AI." 00:17:04 — Robin Rombach

Liquid AI

Foundation model company building ultra-efficient models deployable on edge devices and automotive systems. Powering in-car intelligence for Mercedes-Benz. Argues for a three-axis framework (capability, efficiency, substrate) vs. the industry's one-axis approach.

"We're building foundation models that are so cheap that you can bring them on Raspberry Pis, for example. So you can host them on any kind of device that is on the planet." 00:32:47 — Ramin Hasani

General Intuition

AI research company building models that can predict a wide range of actions, not just text — a more general approach than standard LLMs. Pim de Witte argues most AI quality is simply downstream of good data.

"We train models that can predict lots of different actions where text is just a subset, which is a much more general approach in LLMs." 00:13:31 — Pim de Witte

Decagon

AI customer service / enterprise automation company. Cited by Max Cook as running 90% of production workloads on open source models due to latency and fine-tuning requirements — a leading indicator of production-grade enterprise behavior.

"I think 90% of Decagon's workloads run on open source models. And that is because they are at kind of production scale. They need low latency. They need to be able to fine tune a model and use a small model for that low latency." 00:06:44 — Max Cook

Replit

Developer platform. Cited by Max Cook as a forward-looking voice on natural language as the new interface for compute, via Michele (likely Michele Catasta, President of Replit).

"Michele from Replit was talking about how natural language is how people communicate. That's clearly the direction we're going in." 00:02:14 — Max Cook

Databricks

Data and AI platform. Cited as a public software company benefiting from growth in AI app development, and Ian Stoica (co-founder) named as a panelist at RAISE.

"Snowflake, Databricks, Datadog, all of these businesses that are tied to growth in AI app production, AI app development are doing great." 00:06:07 — Max Cook

Snowflake

Cloud data platform. Cited as a beneficiary of AI app development growth.

"Snowflake, Databricks, Datadog, all of these businesses that are tied to growth in AI app production, AI app development are doing great." 00:06:07 — Max Cook

Datadog

Cloud monitoring and analytics. Cited alongside Snowflake and Databricks as benefiting from the AI development wave.

"Snowflake, Databricks, Datadog, all of these businesses that are tied to growth in AI app production, AI app development are doing great." 00:06:07 — Max Cook

MongoDB

Database platform. Cited multiple times in the episode as a beneficiary of AI app development; 75% of Fortune 100 runs on it. Sponsor of the episode.

"I think Mongo is going to be the exact same type of story." 00:06:07 — Max Cook

AssemblyAI

Speech-to-text and audio intelligence company. Cited by Max Cook as directly relevant to the natural language interaction paradigm he is describing.

"There's AssemblyAI working on it. There's a lot of a lot of exciting companies that are working in this space." 00:04:48 — Max Cook

Cerebras

AI inference chip company. Cited by Philip Johnston as a chip StarCloud hopes to fly in orbit; described as "an incredible inference chip."

"With Cerebras, they're an incredible inference chip. We hope to fly that chip." 00:25:25 — Philip Johnston

SpaceX

Launch vehicle and space infrastructure company. Philip Johnston makes a sweeping bull case — argues it is the equivalent of the railroads for every industry in space and that its IPO will be seen as the most undervalued in history.

"They own what will be by far the most cost effective launch vehicle. And then that opens up every industry in space that will be possible beyond that." 00:24:33 — Philip Johnston

Mercedes-Benz

Automotive OEM. Cited as a customer of Liquid AI for in-car AI intelligence — a flagship example of a traditional enterprise deploying foundation models in a physical product.

"We are powering like in-car intelligence for Mercedes-Benz. Foundation models getting inside the car and making your car basically smarter." 00:34:09 — Ramin Hasani

Good Future Media

Podcast clipping and social media agency. Founded by Chris Madden, generates thousands of clips per month for VCs, CEOs, and investors including the All In podcast besties and Harry Stebbings at 20VC.

"Now I have a huge agency. We get out thousands of clips a month." 00:27:43 — Chris Madden

Brex

Corporate card and expense management platform. Sponsor of the episode; cited as the finance platform used by OpenAI, Anthropic, Vercel, Granola, and Deepgram.

Neuralink

Brain-computer interface company. Mentioned by Max Cook as one possible endpoint for the new human-compute interaction paradigm beyond keyboards and mice.

"Obviously wearables is a place people are spending a lot of time. Neuralink is a place people are spending time." 00:05:25 — Max Cook

---

4. People Identified

Max Cook

Investor at Coatue. Identified for his clear-eyed historical framing of tech waves, his non-consensus take on the keyboard/mouse dying, and his incisive reframing of the open/closed model debate as missing the key variable of what frontier labs will do next.

"I think there's been a number of good discussions about this... we've always conformed ourselves to computers... and now we're finally entering the era where computers have to conform to us." 00:02:14 — Max Cook

Matan Grinberg

Co-founder/CEO of Factory. Identified for having unique proprietary data on open source model adoption in enterprise, and for making the most specific and falsifiable prediction in the episode (50% open source enterprise token share by end of 2025).

"At the beginning of the year, we saw enterprises were using less than 1% open models... Then in May, it crossed the 10% threshold." 00:21:35 — Matan Grinberg

Mark Papermaster

CTO of AMD. Identified for providing rare ground-level insight into how agentic AI is changing hardware ratios (CPU:GPU approaching 1:1) and for demonstrating that AMD is deploying thousands of sub-agents internally in chip design.

"We have thousands of sub-agents just doing incredibly complex chip design, speeding how fast we can get the next generation to market." 00:15:56 — Mark Papermaster

Philip Johnston

Founder/CEO of StarCloud. Identified for building a genuinely novel infrastructure category (orbital data centers), having a chip-agnostic approach, being cited in SpaceX's S1, and for the boldness of the SpaceX IPO valuation prediction.

"My hottest take right now is that the SpaceX IPO is going to be viewed historically as the most undervalued IPO of all time. I think they will tear through 10 trillion within, you know, certainly within the next couple of years." 00:24:06 — Philip Johnston

Rodrigo Liang

CEO of SambaNova Systems. Identified for closing a $1 billion Series F at $11 billion valuation and for clearly articulating the inference economics problem and SambaNova's solution via the SN50 chip.

"What we're able to do with premium inference is drive the performance of the largest models up. So you can actually deliver high quality models with these trillion parameter models at their full precision at incredible speeds." 00:39:59 — Rodrigo Liang

Ramin Hasani

CEO/co-founder of Liquid AI. Identified for the three-axis framework (capability, efficiency, substrate) which reframes how foundation models should be designed, and for having already deployed it in a real enterprise product at Mercedes-Benz.

"If you have maximized intelligence at all costs, that axis alone is not going to get you to the place that you want to go." 00:35:26 — Ramin Hasani

Robin Rombach

Co-founder of Black Forest Labs. Identified for being a strong and articulate advocate of open AI innovation as a safety and accessibility imperative, not just a business model choice.

"Open innovation is good for the world. And I think fear mongering around AI models is only going to lead to them being more closed and that's ultimately leading to slow down all AI progress." 00:17:26 — Robin Rombach

Pim de Witte

Founder of General Intuition. Identified for the blunt, contrarian reduction of the AI hype cycle: most AI quality is a data science problem, not an architecture problem.

"Most of AI is just glorified data science. Good models mostly are downstream from good data and simple things." 00:13:31 — Pim de Witte

Jesse Zhang

CEO of Decagon. Mentioned for writing a public post documenting that 90% of Decagon's production workloads run on open source models, cited by Max Cook as the clearest articulation of why open source wins at production scale.

"Jesse Zhang from Decagon posted recently on X about how 90% of Decagon's workloads run on open source models." 00:06:44 — Max Cook

Greg Brockman

Co-founder of OpenAI. Cited for a tweet about computers finally conforming to humans rather than humans conforming to computers — used by Max Cook as the intellectual anchor for the natural language interaction thesis.

"We've always conformed ourselves to computers like with that by inputting specific data with a keyboard and a mouse. And now we're finally entering the era where computers have to conform to us." 00:02:14 — Max Cook, quoting Greg Brockman

Ian Stoica

Co-founder of Databricks and co-founder of the LM Arena evaluation platform. Mentioned by Ramin Hasani as a legend in the field, with whom he is co-paneling at RAISE to discuss open source and enterprise AI value creation.

"I will be on a panel with Ian Stoica, who is like a legend of the field, a co-founder of Databricks and co-founder of like arena." 00:37:39 — Ramin Hasani

Chris Madden

Founder of Good Future Media. Identified for building a large clipping agency from scratch (started by clipping the All In podcast), generating thousands of clips per month at $200+ per clip, and articulating a clear framework for viral content (the first three seconds rule).

"The three seconds need to grab the viewer in. If the spoken line isn't catching you, you need to have a visual hook on the screen that will tell the viewer what is the payoff for them." 00:28:27 — Chris Madden

Yann LeCun

Chief AI Scientist at Meta. Mentioned twice — once by Pim de Witte (prompted by friends, with no further comment) and once by Mark Papermaster as one of the bookends of the RAISE Summit speaker lineup (alongside the President of France).

"Jan LeCun is underrated." 00:14:16 — Pim de Witte

Henri Delahaye

Organizer of the RAISE Summit. Identified for building what is described as a rapidly maturing, high-quality European AI conference, now bringing together founders, operators, and investors from across the global AI ecosystem.

---

5. Operating Insights

Use Voice/Whisper Prompting to Get Better Outputs from AI Models Throughout the Workday

Coatue's analysts have discovered that giving models richer context through voice — literally whispering into a collar microphone — produces materially better work output than typed prompting. This is not a gimmick; it reflects the higher information density and naturalness of spoken language for context-setting.

"We've got a number of investors that are whispering into their collar half the day. Because that's how to give models better context, better prompting and ultimately like produces much better work output." 00:02:44 — Max Cook

Route AI Tasks Dynamically: Use Frontier Models for Discovery, Open/Fine-Tuned Models for Production

The optimal enterprise AI stack is not a single model choice. Use the smartest frontier model when exploring new use cases, then fine-tune or route to smaller, faster, cheaper open source models once the workflow is understood and production-grade. Factory does this dynamically for every task.

"If you know all of the parameters of the workflow you're trying to execute, you've thought about all the edge cases, you've used a frontier model to hone that, then it makes sense that maybe you're actually seeing better capability with an open source model that's fine tuned or the combination of a frontier model for your hardest tasks and an open source or previous version of a model for your kind of production grade workflows." 00:09:40 — Max Cook

For Viral Content, the First Three Seconds Are Everything — Lead With the Payoff, Not the Setup

Chris Madden's agency, which has generated 4-5 million view clips, has a single non-negotiable rule: the viewer must know their payoff in the first spoken line or via an on-screen visual hook. If it isn't there in three seconds, restart entirely.

"The three seconds need to grab the viewer in. If the spoken line isn't catching you, you need to have a visual hook on the screen that will tell the viewer what is the payoff for them? What's the value that the viewer is going to get from watching this clip? If you don't have that in the first line of the speaker or a hook on the screen, think, you know, restart." 00:28:27 — Chris Madden

---

6. Overlooked Insights

AMD's CPU-to-GPU Ratio Approaching 1:1 Is a Major Structural Signal Almost Nobody Is Tracking

Mark Papermaster drops a single sentence that is potentially one of the most important hardware demand signals in the episode: agentic workflows require CPU and GPU at close to a 1:1 ratio, which is dramatically different from pure GPU-centric inference workloads. This has profound implications for AMD's competitive positioning (they make both) relative to pure-play GPU companies, and for anyone doing infrastructure capacity planning or investing in the semiconductor supply chain.

"With these agentic workflows, you actually need both. In fact, the ratio of CPU to GPU is becoming like one-to-one." 00:15:29 — Mark Papermaster

Nobody in the conversation picked up on this or asked a follow-up question, but it implies that as agentic AI scales, CPU demand could surge alongside GPU demand in a way the market is not currently pricing — and AMD, as a producer of both, is uniquely positioned.

StarCloud Flying ARM GPUs on Its First Spacecraft Is Proof That Commodity Edge Chips Already Work in Orbit

Philip Johnston mentions almost in passing that StarCloud flew three ARM GPUs on their first spacecraft — something he notes "nobody knows about." This is a profound data point: ARM's commodity chip architecture, which powers almost every smartphone on earth, has already been validated in the space environment. This has major implications for the cost curve of orbital compute, since commodity chips are vastly cheaper than radiation-hardened space-grade components.

"We flew an ARM chip on our first spacecraft. We threw three ARM GPUs on our first spacecraft. Nobody knows about that, but we did." 00:25:25 — Philip Johnston

If commodity chips survive in orbit, the economics of space-based compute could be far more favorable than the market assumes, and the timeline to viable orbital data centers may be much shorter than anyone is modeling.