ICYMI: AMD, Coatue… | Sourcery Newsletter Summary

Sourcery Newsletter | Molly O'Shea | RAISE Summit, Paris

1. Key Themes

Open Source Models Are Winning the Enterprise Token War — Faster Than Anyone Expected

Factory's data shows enterprise open-model token share went from <1% in January to 10% in May, with a predicted 50%+ by year-end. This isn't a philosophical debate — it's a measured, accelerating market shift with real customer data behind it.

"At the beginning of the year, we saw enterprises were using less than 1% of their tokens were going to open models. Then in around March, it crossed that 1% threshold, and then in May, it crossed the 10% threshold. By the end of the year, at least in the enterprise, we'll probably cross the 50% threshold." — Matan Grinberg, CEO, Factory

Agentic AI Is Fundamentally Reshaping the Compute Stack

The rise of agentic workflows isn't just a software story — it's a hardware story. AMD's CTO reports that agentic workloads are driving CPU demand back toward parity with GPUs, a significant shift in how AI infrastructure gets built and purchased.

"AMD, we got ahead of it. We've worked to getting our CPUs and GPUs ready for this, and now with these agentic workflows, you actually need both. In fact, the ratio of CPU to GPU is becoming one to one." — Mark Papermaster, CTO, AMD

"Every day I start with agents. They run my day... what's really cool is how we do our 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." — Mark Papermaster, CTO, AMD

Inference Economics Are Broken — and "Premium Inference" Is the Fix

Running the largest frontier models at full precision is economically unsustainable at scale. SambaNova's $1B raise is predicated on solving the inference cost/performance squeeze by enabling high-accuracy, high-speed output that lets service providers actually charge a premium margin.

"The world of inference is running into an economics problem. And 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, these trillion parameter models, at their full precision at incredible speeds." — Rodrigo Liang, CEO, SambaNova

"When you're running faster with the high accuracy models, you can charge more at a lower cost, allowing their service providers to generate more margins for their services." — Rodrigo Liang, CEO, SambaNova

Physical AI: Intelligence Is Leaving the Data Center

The next frontier isn't bigger models — it's AI that runs on constrained hardware in the real world. Liquid AI's framework explicitly treats deployment substrate (cars, phones, aircraft, edge devices) as a first-class design constraint, not an afterthought.

"The third axis I would say is substrate. Where does this intelligence system go? AI is majorly getting hosted in data centers, but you could also bring intelligence on phones, on laptops, on airplanes, on cars." — Ramin Hasani, CEO, Liquid AI

"We're building foundation models that are so cheap that you can bring them on Raspberry Pis. So you can host them on any kind of device that is on the planet." — Ramin Hasani, CEO, Liquid AI

The HCI Layer Is Collapsing — From 80 Apps to 3, and From Keyboards to Language

Coatue's sector head argues we are mid-transition between a desktop-internet interaction paradigm and an agentic one, with profound implications for every software layer above the model.

"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. And in that world, it would only make sense that we reinvent how we interact with compute. It's not the keyboard and the mouse. It's probably natural language." — Max Cook, Sector Head, Coatue

2. Contrarian Perspectives

Frontier Labs Will Eat the Open Source Opportunity Themselves

The prevailing framing is "open vs. closed." Coatue's Max Cook raises a question almost no one is asking: What happens when OpenAI and Anthropic simply productize their older, smaller models as a direct competitor to open-source offerings? This would fundamentally disrupt the current open-source tailwind narrative.

"I don't think anyone's asking the question of what happens when Anthropic & OpenAI decide we can offer that as a product. It seems plausible that they would see this opportunity to take their older, cheaper, more efficient, smaller models & sell that as a product." — Max Cook, Sector Head, Coatue

Most of AI Is Just Glorified Data Science — Model Complexity Is Overrated

General Intuition's Pim de Witte pushes back on the industry's obsession with model architecture, arguing that data quality is the actual moat. He also called Yann LeCun "underrated" — a contrarian take given LeCun's skepticism of LLMs as the path to AGI.

"Most of AI is just glorified data science." — Pim de Witte, CEO, General Intuition

"Good models mostly are downstream from good data and simple things." — Pim de Witte, CEO, General Intuition

SpaceX's IPO Will Be the Most Undervalued of All Time

Against a backdrop where most tech IPOs are debated as overvalued, the CEO of Starcloud makes a maximalist bull case for SpaceX, projecting a $10 trillion valuation trajectory and framing their launch cost advantage as a tollbooth on every future space industry — asteroid mining, lunar resources, comms, and orbital compute.

"My hottest take right now is that the SpaceX IPO is gonna be viewed historically as the most undervalued IPO of all time. I think they will tear through 10 trillion within the next couple of years, and then have almost unlimited time for what they're building." — Philip Johnston, CEO, Starcloud

"They own what will be by far the most cost-effective launch vehicle, and that opens up every industry in space that will be possible beyond that." — Philip Johnston, CEO, Starcloud

3. Companies Identified

Factory | AI coding/enterprise software platform | Used as the primary data source for open vs. closed model token share trends in enterprise | "At the beginning of the year, we saw enterprises were using less than 1% of their tokens were going to open models... in May, it crossed the 10% threshold."

SambaNova | AI chip and inference infrastructure company | Announced $1B Series F at $11B valuation; positioned as the solution to the inference economics problem with their SN50 chip | "Our series F is a $1 billion raise at 11 billion valuation. The round was led by General Atlantic."

Liquid AI | Foundation model company focused on efficiency and physical AI | Powers in-car intelligence for Mercedes-Benz; building models deployable on Raspberry Pi-class hardware | "We're bringing the cost of tokens to zero."

Black Forest Labs | Open-weight image generation model company | Cited as a case study for the open-source accessibility argument | "Having accessible AI models is key."

General Intuition | AI company training multi-action models beyond text | Makes the case that data quality beats model architecture complexity | "Good models mostly are downstream from good data and simple things."

Starcloud | Orbital data center company | Building AI compute infrastructure in space; CEO is a SpaceX bull | "The SpaceX IPO is gonna be viewed historically as the most undervalued IPO of all time."

SpaceX | Aerospace/launch vehicle company | Identified as a generational investment opportunity due to launch cost monopoly | "They own what will be by far the most cost-effective launch vehicle."

AMD | Semiconductor company | Positioned as an agentic AI infrastructure beneficiary; CTO described using thousands of sub-agents internally for chip design | "The ratio of CPU to GPU is becoming one to one."

Coatue | Technology-focused hedge fund/growth investor | Represented by sector head Max Cook; offered the HCI and open-vs-closed frameworks | "The keyboard and mouse are slowly dying."

Good Future Media | AI-powered media/clipping agency | Described as producing thousands of short-form clips monthly for distribution | "Clipping is the new ad."

RAISE Summit | Europe's high-profile AI summit | Venue and context for all interviews; framed as the leading European AI gathering | "We are all in our tech bubble, token maxing on a daily basis, but it's good to have a bit of history in our sight."

4. People Identified

Matan Grinberg | CEO, Factory | Cited with the most concrete, data-backed prediction in the piece — the 90% open-model token share forecast, grounded in Factory's own enterprise usage data | "In the next 12 months, 90% of tokens will be going to open models."

Mark Papermaster | CTO, AMD | Described both the enterprise transformation from agentic AI and AMD's internal use case (thousands of chip-design sub-agents); key voice on the compute stack shift | "People don't even understand how transformational agentic AI is."

Rodrigo Liang | CEO, SambaNova | Announced the $1B Series F and articulated the "premium inference" market thesis | "The world of inference is running into an economics problem."

Ramin Hasani | CEO, Liquid AI | Articulated a three-axis framework (intelligence, efficiency, substrate) for next-generation foundation models; company powers Mercedes-Benz in-car AI | "Efficiency is not an afterthought."

Robin Rombach | CEO, Black Forest Labs | Co-founder of the open-weight image generation lab; argued that accessibility — not the open weights vs. open source distinction — is what matters | "Fear-mongering around AI models is only going to lead to them being more closed, and that's ultimately going to lead to a slowdown of all AI progress."

Max Cook | Sector Head, Coatue | Framed the HCI transition away from keyboards/mice and raised the underasked question about frontier labs commoditizing open-source's cost advantage | "I don't think anyone's asking the question of what happens when Anthropic & OpenAI decide we can offer that as a product."

Philip Johnston | CEO, Starcloud | Made the maximalist SpaceX IPO bull case; positioned Starcloud as building orbital compute on top of SpaceX's launch infrastructure | "They own what will be by far the most cost-effective launch vehicle."

Pim de Witte | CEO, General Intuition | Argued that LLM-centric AI is overhyped ("glorified data science") and that multi-action models are a more general approach | "Most of AI is just glorified data science."

Chris Madden | CEO, Good Future Media | Provided the tactical media playbook for short-form clip virality; rates $200/clip as the market standard | "Clipping is the new ad."

Henri Delahaye | CEO, RAISE Summit | Co-founder and host of the summit; framed the event's philosophy of grounding tech ambition in historical context | "It's good to have a bit of a history in our sight."

Yann LeCun | Chief AI Scientist, Meta | Mentioned (not interviewed) by Pim de Witte as "underrated" — an implicit endorsement of LeCun's skepticism of LLM-first approaches | Referenced as support for the "data over architecture" thesis.

5. Operating Insights

Instrument Your Model Portfolio by Task Complexity — Don't Default to Frontier

Factory's data reveals that defaulting to frontier models for all enterprise workloads is increasingly wasteful. Open models are "incredibly performant, incredibly cheap, incredibly fast" for routine tasks. Operators should audit their token spend and route simple, high-volume tasks to open models while reserving frontier models for discovery and high-stakes inference.

"Not every task you need Opus 4.8 or GPT 5.6 Ultra High, and many tasks day-to-day you do are very simple." — Matan Grinberg, CEO, Factory

The First Three Seconds Determine Everything in Short-Form Distribution

Good Future Media's playbook is concrete: the visual and spoken hook must convert within three seconds or the viewer leaves. At $200/clip and a minimum cadence of one clip per day (ideally two to three), this is a scalable, systematizable content operation — not a creative lottery.

"It's really the first three seconds. 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." — Chris Madden, CEO, Good Future Media

Use Agentic Workflows Internally Before You Sell Them Externally

AMD's CTO described using agents personally to manage his day and deploying thousands of sub-agents in chip design to accelerate time-to-market. For operators, this is a signal: internal agentic deployment is a live competitive advantage, not a future roadmap item.

"Every day I start with agents. They run my day, look at what's coming up. But what's really cool is how we do our 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." — Mark Papermaster, CTO, AMD

6. Overlooked Insights

Open vs. Closed Is the Wrong Debate — Workload Type Is What Actually Determines Model Choice

Max Cook makes a point that cuts through the binary framing dominating AI discourse: open source wins at production scale; frontier wins at discovery. The debate isn't about which paradigm "wins" — it's about matching model type to use-case stage. This is an underappreciated portfolio management insight for any enterprise AI buyer or investor evaluating model providers.

"Open source is good for production scale workflows, yet we're so early in this AI journey that there's not that many production grade workflows. And so if you are trying to discover a new use case, you want the smartest model you can possibly get, you use the frontier." — Max Cook, Sector Head, Coatue

Space Compute Is a Real Infrastructure Play, Not Sci-Fi

Philip Johnston's Starcloud isn't just making a SpaceX investing argument — he's building orbital data centers and discussing chips specifically designed for space compute. This "space as compute substrate" angle received only a brief timestamp mention ("Chips for Space Compute" at 25:13) but represents a nascent infrastructure investment category that sits at the intersection of the physical AI and launch economics theses.

"They own what will be by far the most cost-effective launch vehicle, and that opens up every industry in space that will be possible beyond that. So that's all of asteroid mining, lunar resource mining, all of the comms businesses that are gonna be built." — Philip Johnston, CEO, Starcloud