Charles Wang

Terence Tao told me something that is both clarifying and unsettling about large language models. The mathematics underlying today’s LLMs is not especially exotic. At its core, training and inference mostly involve linear algebra, matrix multiplication, and some calculus. This is material a competent undergraduate could learn. In that sense, there is very little mystery about how these systems are constructed or how they run. And yet the real mystery begins there. What we do not understand well is why these models perform so impressively on certain tasks while failing unexpectedly on others. Even more striking, we lack reliable principles that allow us to predict this behavior in advance. Progress in the field remains largely empirical. Researchers scale models, change datasets, run experiments, and observe what emerges. Part of the difficulty lies in the nature of the data itself. Pure randomness is mathematically tractable. Perfectly structured systems are also tractable. But natural language, like most real-world phenomena, lives in an intermediate regime. And we humans hate that liminal space! It is neither noise nor order but a mixture of both. The mathematics for this middle ground remains comparatively underdeveloped. So we find ourselves in a peculiar position. We understand the machinery, yet we cannot reliably explain its capabilities. We can describe the mechanisms that produce these systems, but we cannot predict when new abilities will appear or how performance will vary across tasks. That tension, between relatively simple mathematical tools and highly unpredictable behavior, is the central puzzle of modern AI. (Video link in comments)

I just finished going through BofA’s research on $NBIS. Here are the key competitive advantages they highlight: ⬇️ 1) Differentiated virtualization layer and distributed compute fabric Nebius’ most important differentiator is its virtualization layer, which allows GPUs across multiple data centers to function as a single unified cluster. This matters a lot more than it might seem at first. AI workloads are no longer just about raw compute, they’re about how efficiently that compute can be assembled, scaled, and allocated across geographies. With power constraints, GPU shortages, and long data center build timelines, the ability to pool distributed resources into one system becomes a major advantage. It allows for: - Higher utilization - Faster deployment - Better ability to serve variable demand GPU innovation is moving faster than infrastructure expansion, so software layers like this become increasingly valuable over time. Importantly, Nebius appears to be the only player currently enabling global GPU orchestration. Most neoclouds are limited to virtualization within a single data center, and hyperscalers like Oracle or Microsoft don’t yet offer this type of GPU-specific orchestration across locations. That makes this a potentially durable and hard-to-replicate edge. 2) Full stack AIaaS cloud offering creates defensible long-term moat On top of its infrastructure, Nebius has built a full AI cloud platform that simplifies how companies build, train, and deploy models. The key value proposition here is abstraction: customers don’t need large internal AI/ML teams to operate complex workloads. As AI moves from experimentation to production, this becomes increasingly important. Then there’s Token Factory, which is a big part of the differentiation: - Enables large-scale inference and model serving out of the box - Supports 60+ models (including open-source) - Allows easy fine-tuning and deployment - Offers predictable, transparent pricing ($/token) - Claims significant cost savings vs proprietary models - Provides enterprise-grade governance and security - Uses an OpenAI-compatible API (lower switching friction) From Nebius’ perspective, this does two important things: - Improves utilization of idle compute - Introduces a more recurring, usage-based revenue stream It also increases platform stickiness, especially for enterprises that want a turnkey AI solution without building internal infrastructure. Long term, if inference becomes the dominant workload (which is the current direction), Token Factory could become a very meaningful moat. 3) Leadership team with proven hyperscale execution and deep technical DNA Nebius is led by a team that previously built and operated Yandex’s infrastructure across multiple verticals (search, cloud, autonomous driving, etc.). That experience is highly relevant: - Managing large-scale data centers - Operating multi-generation hardware fleets - Building distributed systems at scale What stands out is not just the technical capability, but the execution track record. Yandex was able to outperform global players like Google and Uber in its home market. In their view, this combination of deep technical expertise and proven execution is one of Nebius’ strongest assets.

Over the next 5 years, NVIDIA will ship ~100,000 AI factory PODs across 4 chip generations. Each generation is 3-14x more powerful than the last. The annual cadence is relentless: • Vera Rubin (2026): 60 exaflops/pod • Rubin Ultra (2027): ~220 exaflops/pod • Feynman (2028): ~600 exaflops/pod • Feynman Ultra (2029): ~1,200 exaflops/pod Blended across the full ramp, that’s ~50 ZETTAFLOPS of aggregate compute deployed globally. 100M+ GPUs. Billions of dies. System ASPs climb each generation — more HBM, denser packaging, liquid cooling at scale. If demand holds, it is realistic that NVIDIA will generate $30 TRILLION in revenue with gross profit margins of~75%! Thats more profit than U.S. GDP.

The biggest AI companies won't sell AI. Every technology cycle produces generational companies: - light alloys brought Boeing, - the assembly line brought Ford, - solid-state electronics brought Texas Instruments, - fabless enabled Nvidia. AI is that next enabling layer, and we expect great companies to emerge redefining how we move, produce energy, and live longer, healthier lives. Right now, most startups are focused on building AI itself. The next wave will leverage it under the hood and build products - from airplanes to bridges to computers - that don't even resemble what's commonplace today.

AI coding agents can now deliver one-shot custom apps straight to your phone. It’s the beginning of the end for the iPhone’s dominance.