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

Interaction Models and the Decline of Traditional Interfaces

• Max Cook (Cotu) identifies the keyboard and mouse as "slowly dying," driven by a historical shift from mainframes to mobile to an agentic AI future.
• Data indicates the average user interacts with only 11 apps daily out of 60–80 installed, signaling a transition to an "agentic" ecosystem where fewer apps serve as the central interface.
• Future interaction will likely rely on natural language and real-time audio, exemplified by investors at Cotu whispering into collars to provide models with better context.
• Cook references a Greg Brockman concept that the industry is shifting from humans conforming to computers (via keyboards) to computers conforming to humans (via natural speech).
• The "Thinking Machines" reaction model demo is cited as a precursor to zero-latency, multi-modal interaction (speech-to-text, image input, real-time response) that will sit atop all intelligence layers.
• No consensus exists on the final hardware format, with speculation ranging from wearables and neural links to entirely new device categories.
• OpenAI's upcoming "super app" is expected to introduce agents to nearly 1 billion weekly active users, fundamentally altering mobile interaction expectations.

The Open Source vs. Frontier Model Ecosystem

• Max Cook argues the open-source vs. closed (frontier) debate is a false dichotomy; 90% of Decagon's production workloads run on open-source models due to latency, fine-tuning needs, and cost.
• Enterprise LLM spend on open-source models dropped from 19% last year to 11% this year, yet high-scale production workflows increasingly favor open-source solutions.
• The end state will likely involve "model routing," where frontier models handle complex discovery tasks while fine-tuned open-source models handle production-grade workflows.
• Matan (Factory) predicts that by the end of the current calendar year, 50% of enterprise tokens will be routed to open models, up from less than 1% at the start of the year.
• Factory observed a 10% threshold for open-source token usage in May, driven by cost optimization and the realization that not all tasks require frontier-level intelligence.
• Matan notes that while frontier models will remain valuable, their token share in the overall compute ecosystem is shrinking as "droid" agents automate software engineering.
• Philip Johnston (StarCloud) anticipates the SpaceX IPO will be viewed historically as the most undervalued IPO, potentially exceeding a $10 trillion valuation within two years due to dominance in launch vehicle costs.
• Robin (Black Forest Labs) posits that open innovation is fundamental to safety and progress, warning that fear-mongering leads to restrictive closed models that slow AI development.
• Philip Johnston emphasizes that space-based computing is becoming viable, with StarCloud planning to house up to 10 megawatts of compute per Starship launch starting in August.

Infrastructure, Hardware, and Enterprise Adoption

• AMD CTO Mark notes that agentic AI workflows are shifting the CPU-to-GPU ratio from traditional imbalances to a near 1:1 requirement.
• AMD reports that thousands of "sub-agents" are currently being utilized for chip design, significantly accelerating next-generation product cycles.
• Ramin (Liquid AI) proposes a "three-axis" foundation model framework: maximizing intelligence, prioritizing efficiency/cost, and optimizing the substrate (e.g., edge devices vs. data centers).
• Liquid AI is targeting the physical AI space by making foundation models cheap enough to run on resource-constrained devices like Raspberry Pis and in-car systems (e.g., Mercedes-Benz).
• Rodrigo (SambaNova) announced a $1 billion Series F raise at an $11 billion valuation, led by General Atlantic, to accelerate supply chains for their new SM50 cloud-scale RDU.
• The SambaNova SM50 chip is designed for "premium inference," delivering high-precision, trillion-parameter model performance at lower costs and power consumption than GPUs.
• StarCloud (Philip Johnston) confirms chip-agnosticism, having already flown ARM chips, ARM GPUs, and partnering with Cerebras and SambaNova for space-based compute.
• Ramin observes that 85% of current AI activity is R&D on code, suggesting enterprises are waking up to the necessity of treating AI as a deployed product rather than just a research tool.

European Tech Landscape and Market Dynamics

• The RAISE Summit is positioned as a convergence point for European and Silicon Valley tech, with attendees discussing sovereign AI, power constraints, and grid bottlenecks.
• Robin notes that while European tech is improving, it lacks the "nice AI" reputation of the region's historical landmarks, requiring a mindset shift toward pushing innovation locally.
• Matan highlights that European banks and enterprises are beginning to adopt "droids" for software engineering automation, indicating a shift from US-centric urgency to global adoption.
• Chris (Good Future Media) asserts that "clipping" long-form content into short videos is the new advertising standard, with the first three seconds acting as the critical hook for virality.
• Chris reveals that high-quality clipping services (involving human psychology and visual hooks) cost approximately $200 per clip, contrasting with cheaper AI-only services.
• Chris cites successful viral clips featuring Chamath and David Friedberg (5 million views on TikTok regarding the Ukraine famine) as benchmarks for the strategy.

Funding and Strategic Investments

• SambaNova secured a $1 billion Series F at an $11 billion valuation, with investors including General Atlantic, Sequoia Ventures, T. Rowe Price, and Capital Group.
• Capital is being directed toward supply chain acceleration to ensure global delivery of racks for the new SM50 inference hardware.
• StarCloud is constructing a new campus for "StarCloud 3," a vehicle compatible with Starship capable of massive in-orbit compute capacity.
• SambaNova's SM50 targets the "economics problem" of inference by allowing service providers to generate higher margins through faster, high-accuracy model delivery.