| |
|
| |
Event Overview
Since launching the series in May 2026, weve successfully hosted five exclusive, closed-door Top AI Researchers Roundtables, fostering meaningful connections & in-depth conversations among researchers & leaders from the worlds foremost AI labs & research institutions. Participants have included:
Ed Chi VP of Research at Google DeepMind
15 researchers from OpenAI
5 researchers from Anthropic
26 researchers from Google DeepMind
6 researchers from xAI
7 researchers from NVIDIA
2 Stanford professors
Xuedong Huang CTO, Zoom
Yuandong Tian Co-founder of Recursive (Raising $650M)
Were excited to continue this tradition with our6th Top AI Researchers Roundtable, taking place onNovember 15, from 3:00 PM to 5:00 PM, at a private residence in Los Altos Hills.
This invitation-only gathering will once again bring together leading researchers from OpenAI, Google DeepMind, Anthropic, xAI, & other frontier AI labs & research institutions for an intimate, closed-door discussion exploring some of the most important questions shaping the future of artificial intelligence.
This will be a fully closed-door, off-the-record conversation focused on the future of AI systemsfrom frontier research breakthroughs & emerging technical challenges to the next wave of opportunities that could shape the industry over the coming years.
Following the roundtable, guests will continue the discussion over a private BBQ reception with fellow researchers, founders, investors, & industry leaders.
Core Discussion Topics
AI progress depends not only on better models, but also on how effectively we build & use the infrastructure behind them.
From GPU utilization & changing compute demands to hardware diversity, AI-assisted research, & scaling constraints, the next wave of breakthroughs raises fundamental questions about efficiency & system design.
This roundtable brings together leading AI researchers to explore five critical questions shaping the future of AI infrastructure.
Rather than formal presentations, the discussion will focus on open questions, technical challenges, differing perspectives, & the breakthroughs needed to advance frontier AI.
QUESTION 1 THE UTILIZATION GAP
Labs are committing to unprecedented amounts of compute. How much of it actually translates into learning?
Anthropics IPO disclosures reveal at least $518 billion in planned infrastructure spending over the coming decade, much of it tied to long-term commitments that are difficult to cancel. But contracted capacity is not the same as useful training compute. Microsofts MAI-Base-2 report offers a concrete example: a 2.8-trillion-parameter MoE model trained on 16,000 GB200 GPUs achieved an average MFU of 21%, with goodput of 89%. The gap between theoretical hardware capacity & actual training performance remains substantial.
Follow-up questions
Where does the remaining compute go todayand which sources of inefficiency are hardest to eliminate?
As RL & post-training become more important, is MFU still the right metric, or do we need a better measure of how efficiently compute produces learning?
If you could choose between 20% more compute & 20% better utilization, which would make a bigger difference to model development?
QUESTION 2 THE CHANGING COMPUTE MIX
How will the rise of reasoning models & AI agents change what AI infrastructure needs to look like?
AI workloads are becoming more diverse. Training a model, running extensive reasoning at inference time, & deploying agents that execute code or interact with external tools can place very different demands on computing systems. Recent infrastructure announcements, including Oracles plans for NVIDIA Vera CPU cloud services aimed at agentic workloads, highlight how the supporting infrastructure may evolve beyond GPUs alone.
Follow-up questions
How will the balance between pretraining, post-training, & inference evolve as models become more capable?
As agents spend more time using tools, running code, & interacting with environments, where do the main computing bottlenecks shift?
Will future AI systems require fundamentally different infrastructure, or can existing systems adapt to these new workloads?
QUESTION 3 ONE MODEL, MANY CHIPS
How much does the hardware shape the way we build AI models?
Anthropic has described using multiple accelerator platformsincluding Nvidia GPUs, Google TPUs, AWS Trainium, & AMD hardwareto meet its compute needs. This growing diversity raises an important research question: different chips offer different trade-offs in memory, performance, & software support, but moving workloads across platforms can require significant engineering effort.
Follow-up questions
How difficult is it to move a frontier training workload across different chip architectures, & what are the biggest challenges?
To what extent should model architecture be optimized for the hardware, rather than designed independently of it?
Could greater hardware diversity lead to meaningful new approaches to model design, or will software compatibility remain the main challenge?
QUESTION 4 AI RESEARCH ON AI
Can AI meaningfully accelerate the development of the next generation of AI?
AI systems are increasingly helping researchers write code, analyze results, generate hypotheses, & run experiments. Anthropic has publicly described the growing role of Claude in its internal AI research & engineering work. If AI systems become more capable research collaborators, they could change not only the cost of experimentation but also the pace at which new models are developed.
Follow-up questions
Which parts of AI research are already benefiting most from AI assistance, & which remain difficult to automate?
What would it take for AI-assisted research to produce a sustained acceleration in model development?
How would we know whether AI research is genuinely accelerating, rather than simply becoming cheaper or faster to execute?
QUESTION 5 SCALING UNDER CONSTRAINTS
If more compute becomes harder to obtain, what will drive the next wave of AI progress?
Frontier AI infrastructure increasingly depends on more than chips. Power availability, networking, memory, & construction timelines all affect how quickly new capacity can come online. As major labs commit to increasingly large infrastructure projects, a key research question is whether future progress will depend primarily on scaling hardware or on finding more efficient ways to use it.
Follow-up questions
If access to more compute becomes constrained, how would that change research priorities: larger models, more training, more inference-time compute, or more efficient algorithms?
How much can algorithmic progress compensate for limits in hardware & power availability?
Could training across geographically separated clusters become a practical way to expand capacity, or would the coordination costs outweigh the benefits?
Additional Questions
1. CPU Requirements for Million-Agent Systems
How much CPU compute is required to support one million concurrently active AI agents?
As agentic workflows become increasingly complex & task horizons extend from minutes to hours, days, or even weeks, how significantly could CPU requirements per agent increase? What are the primary scaling bottlenecks, & how might the CPU-to-agent ratio evolve over the next three to five years?
2. x86 vs. ARM: Which Architecture Will Dominate Agentic Computing?
As AI agents become a major driver of CPU demand, will x86 or ARM emerge as the dominant architecture for agentic workloads?
How do x86 & ARM compare in terms of performance, power efficiency, cost, virtualization, & scalability when running agent virtual machines & orchestrating complex workflows?
What architectural advantages or limitations does each offer for large-scale agent deployment? Could agentic computing fundamentally reshape the competitive landscape between x86 & ARM?
3. Distributed Inference Across Geographically Separated Data Centers
Can today's AI infrastructure efficiently distribute inference workloads across multiple geographically separated data centers?
As power availability increasingly constrains data center expansion, could geographically distributed inference become a practical solution for utilizing fragmented compute capacity?
What are the key technical challenges involving network latency, bandwidth, synchronization, KV-cache management, & workload scheduling? Under what conditions is cross-data-center inference economically & technically viable today, & could it become a mainstream architecture?
4. KV-Cache Offloading: Can SSDs Significantly Reduce Inference Memory Costs?
How far can KV-cache offloading to SSDs go in reducing the growing memory costs of AI inference, particularly for long-running agents?
What percentage of KV-cache data can realistically be offloaded from DRAM or HBM to SSD storage without significantly degrading inference performance?
How do latency, bandwidth, cache reuse patterns, & task duration affect the trade-offs? As agentic workloads generate increasingly long contexts, could SSD-based KV-cache architectures fundamentally change the economics of inference infrastructure?
5. The Future of AI Inference Hardware: Which Decoder Architecture Will Win?
Among emerging inference hardware architectures from Groq, Cerebras, Positron, Etched, d-Matrix, & SambaNova, which is best positioned for the next generation of AI inference?
How do these approaches compare in terms of decoding throughput, time per output token, memory bandwidth, power efficiency, scalability, programmability, & total cost of ownership?
What are the fundamental architectural trade-offs between specialized inference accelerators & more flexible computing platforms?
As models evolve toward longer contexts, mixture-of-experts architectures, multimodal reasoning, & increasingly complex agentic workloads, which architecture is most likely to achieve a sustainable competitive advantageand why?
6. Next-Generation Optical Interconnects: Will Coherent Lite Become Mainstream?
As data center interconnect speeds reach 2.4 Tbps & beyond, will Coherent Lite emerge as a mainstream optical interconnect technology for scale-out AI infrastructure?
How does Coherent Lite compare with conventional intensity-modulation/direct-detection (IM-DD) solutions in terms of power consumption, cost, reach, bandwidth density, latency, & implementation complexity?
At what bandwidth & distance thresholds does coherent technology become more attractive? Could Coherent Lite become a foundational technology for next-generation AI clusters, or will alternative optical interconnect architectures remain dominant?
7. Foundation Models for Industrial Robotics: How Close Are We to Scalable Deployment?
What is the current frontier of foundation models for industrial robotics, & how close are we to achieving reliable, economically viable deployment at scale?
Is the dominant training paradigm still based on simulation & human demonstration videos for pretraining, followed by task-specific post-training & some form of in-context adaptation? Or are fundamentally new training approaches emerging?
What breakthroughs are happening in robot learning, including vision-language-action (VLA) models, reinforcement learning, world models, & sim-to-real transfer?
How far have robotic foundation models advanced in terms of generalization, dexterity, adaptability, & autonomous task execution?
For robots deployed in real manufacturing environments, what task success rates are realistically achievable today? How do these systems perform in terms of reliability, human intervention requirements, return on investment, & total cost of ownership?
Looking ahead three to five years, what technical breakthroughs are still needed for general-purpose industrial robots to become commercially viable at scale?
Event Format
30 curated attendees only
Invitation-only
No media / no public recordings
Chatham House Rule-style discussion
Intimate, high-signal conversations among top researchers & operators
Why This Matters
The future of AI will not be shaped solely by public announcements or viral demos.
It will be shaped by small groups of people exchanging ideas, debating technical tradeoffs, & building trusted relationships behind the scenes.
The BFC AI Researcher Inner Circle is designed to create:
Deep technical & strategic dialogue
Cross-organization insight exchange
Authentic relationship-building among frontier researchers
A trusted environment for discussing sensitive ideas & future directions
About the Sponsors
We are where insight & AI shape the future.
Llama Ventures is a Silicon Valleybased early-stage venture firm investing in AI companies from Pre-Seed through Series A. With offices in Sunnyvale & San Francisco, the firm has invested in 60+ companies across AI infrastructure, enterprise software, healthcare, fintech, & deep-tech, & manages $300M+ in direct fund capital alongside a $600M+ fund of funds. Founded by entrepreneurs & technology leaders, Llama Ventures partners closely with founders, providing strategic support & access to a global network of operators, engineers, customers, & investors.
Apera AI | Silicon Valley-Based AI Gateway
One API. 200+ Leading AI Models. 35% Average Cost Savings.
Access the world's leading AI models through a single, reliable, cost-efficient gateway. Simplify AI integration, optimize performance, & reduce inference costs by an average of 35%.
TRUSTED IN PRODUCTION
Powering production AI workloads for 60+ customers.
Want to reduce your AI token costs? Feel free to book a 30-minute consultation with our team:
https://calendly.com/apera-ai-model-review/30min
OrphLux is turning data chaos into training truth, providing the data backbone that helps robotics, world models, & next-generation AI systems learn from the real world. We specialize in three core areas:
Embodied AI large-scale, real-world data collection & operation trajectories that teach robots & physical AI systems to perceive, reason, & act in dynamic environments.
Benchmark Data rigorous, high-quality evaluation datasets that test & validate generative AI & physical AI systems from development through deployment.
Expert Ecosystem a global network of domain specialists & PhD-level experts delivering precision annotation, instruction writing, & preference ranking at scale.
At scale, OrphLux operates 12 global delivery centers across Asia, North America, & South America, with access to more than 100,000 PhD & domain experts.
Viggle builds physics-native world foundation models using a novel 3D-native architecture for generating scenes, characters, & motion. JST-2 generates a 3D scene in about one second of model inference & supports real-time world expansion & editing. Our models support video creation, interactive games, & robotics applications such as generating & adapting 3D simulation environments. We train our JST models on diverse, controllable 3D data produced by our in-house data engine.
IterCSI.ai | Scientific Data Scientific Agents Autonomous Labs
IterCSI.ai builds scientific data infrastructure & AI agents to accelerate scientific discovery, with a long-term vision of autonomous labs.
Scientific data is our core strength. We turn scientific literature & experimental data into high-quality datasets for training & evaluating AI models. We are preparing to release what we believe will be the largest benchmark dataset for drug discovery, & we invite AI model developers to test their models on it.
Building on this data foundation, we are developing scientific agents that help researchers plan experiments, analyze results, & guide the next round of research. Over time, we aim to connect these agents with automated laboratories to create a continuous cycle of experimentation & discovery.
We welcome conversations with large language model companies about benchmark evaluations, scientific data partnerships, & agentic AI for science.
About Bay Area Founders Club
The Bay Area Founders Club (BFC), founded by Dr. Paul Fang in Silicon Valley in 2022 & managed by a dedicated group of Stanford alumni & students, is a thriving community of over 100,000 members, including more than 5,000 startups & over 1000 venture capital firms. With most startups in our network raising $1-10 million, & some exceeding $20 million, our mission is to empower creators who aspire to make a global impact. We provide the essential resources, support, & connections needed to transform visionary ideas into reality.
Unlock insider access & elevate your founder journey join BFC NOW.
Learn more: https://www.bayareafoundersclub.com
|
|
|
|
|
|
|
|
