Evergreen Capital
Technology Investor, Semiconductors & AI Infrastructure
NVIDIA's CUDA may become the easiest platform for AI agents to accelerate because it has the deepest, broadest and most mature substrate to work on. OpenAI used NVIDIA's programming stack to unlock ultra-fast performance, and NVIDIA says GPT-6 Astra Ultrafast runs up to 8x faster on its platform. x.com x.com
Recurrent depth favours NVIDIA because it shifts some of today's memory-limited workloads onto compute, unlocking extra serving capacity per node. That helps maximize system value despite Rubin Ultra's reported HBM de-spec on short supply. More useful work per node raises revenue per node and returns for OpenAI and others, funding more infrastructure, and such efficiencies have repeatedly been recycled into more users and bigger workloads, sustaining demand across the stack, including memory. x.com x.com
NVIDIA's advantage extends beyond GPUs into engineering work that improves customers' AI economics. Its work with Meta on Andromeda redesigned ad-retrieval models and pipelines, processing over 1 billion ads per second on one Hopper GPU server with 100 times less network traffic and 300 times fewer FLOPs per inference. Meta says further optimization is underway for Grace Blackwell. x.com
Enterprise agents widen NVIDIA's opportunity beyond inference because reasoning, tool use and company-data workflows need infrastructure and controls outside the model. Satya Nadella describes those needs; Vera CPUs, OpenShell and Sentry on BlueField are the NVIDIA read-through. x.com
OpenAI's Astra shows harness capabilities moving into the model itself, but top scores still depend on costly inference. Per François Chollet, Astra scores 66% on ARC-AGI-3 under ARC Prize's standard harness; its 98.6% came from OpenAI's custom conversation harness at about $360 a game. ARC's team found the model inventing its own symbolic world modeling, which should worry scaffolding vendors. The NVIDIA trade does not change, because that bill still lands on inference. x.com
Nvidia can capture more AI infrastructure spending by supplying Vera CPUs and networking alongside GPUs. Oracle outlined an initial supply of 150,000 Vera CPUs, and OpenAI says it is putting them into production for agent tool execution; reducing CPU waits could improve cost per task and GPU-memory use. Benefits from independently scaling CPUs and GPUs still need validation in utilization and completion times. x.com
Nvidia is winning fast inference: OpenAI shipped its flagship speed tier, GPT-6 Astra Ultrafast, on Nvidia GPUs, gaining a second source and eroding the exclusivity Cerebras was priced on, though Cerebras is still faster and capacity likely played a role. Operators keep extending GPU lives, Amazon to 5-6 years, Microsoft and Google to 6, Meta to 5.5, because old chips keep earning on inference. More lifetime revenue per chip justifies customers' next Nvidia order. x.com x.com x.com x.com
Nvidia's case is a lot weaker than two years ago, because large providers such as OpenAI and Meta are likely to design custom chips tuned to their own models. Nvidia wins only if workloads and models stay diverse. The claim that frontier labs are decoupling from Nvidia is doubtful, since OpenAI's recent efficiency gains likely came from Blackwell, but it holds for Google and Amazon. Nvidia also needs to own the open-source ecosystem, which is contradictory by definition. x.com x.com
AI software and hardware advances make GPUs and power more valuable, creating huge long-term tailwinds for compute providers and AI model and app businesses. Each new software generation increases a GPU's value; each hardware generation increases the value of a gigawatt of power. That favors Nvidia, Broadcom, Intel and AppLovin. x.com
More AI tokens mean more compute demand for Nvidia even as pricing power starts to slip. OpenRouter data via the FT show weekly token use up ~250-fold since January 2025; the FT links the surge to agents' long call chains. Bloomberg's breakdown of a 1 GW AI campus that could cost roughly $30B puts well over half the bill in chips and servers, most landing at Nvidia as GPU racks. Servers wear out in 5-6 years, making AI capex a recurring chip refresh. x.com x.com x.com x.com x.com x.com x.com
Nvidia's push into AI models and agent software looks weak even though its GPUs and racks are great. Its Nemotron model is mediocre and its LangChain-based agent stack won't make it. If Nemotron were capable of frontier performance it would be there already, and Nvidia doesn't seem to know how to get there. x.com x.com
Nvidia's system components can reduce the idle time that erodes returns on expensive AI chips. CPUs, accelerator interconnects and server networking help operators such as Nebius and IREN keep machines doing useful work; solving the next bottleneck could win Nvidia more of the data-center equipment budget even when a rival chip looks appealing alone. Buyers must compare the cost of completed work across the whole system. x.com
