Bottleneck Capital
Long/Short Investor, Technology, Media & Telecom
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
The market should pay less for Nvidia's moat, though its rental stream need not disappear. OpenAI's Jalapeño shows agents can do kernel and assembly work, cutting one CUDA toll, while NCCL, cuDNN, TensorRT and Nvidia-heavy training data still bias agent code toward Nvidia. Some demand rests on Nvidia's own balance sheet, including up to $105B guaranteed for OpenAI's Ohio data center, and per Ed Zitron's read, five customers held 70% of receivables at the end of Q2, up from 56%. x.com x.com x.com
Low Anthropic spending relative to customer expenses points to an enterprise AI adoption gap that leaves Nvidia reliant on intermediaries able to build their own chips. Anthropic's S-1 says over half the Fortune 500 were $100k+ annual customers as of September, yet spending among its publicly traded customers averaged under 0.1% of expenses in Q3 2026. That measures one provider and cannot show whether AI deployments succeed; Nvidia needs more enterprises buying its compute directly. x.com
Cerebras should dominate the enormous fast-inference market because agentic workflows often run steps sequentially, making each step's speed critical. Its newly announced Oracle partnership brings its ultra-fast inference to Oracle Cloud Infrastructure. Cerebras delivers inference speeds up to 13x faster than Nvidia GPUs, and its chips should eventually reach every major AI cloud. 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
NVIDIA faces rival chips from every Mag 7 customer, which has an incentive to cut GPU spending and improve margins on its AI buildout. If they can make a GPU 85% as good for 50% less, it is worth using instead of NVIDIA's; the shift will take time. NVIDIA makes a GPU for $5,000 and sells it to those customers for $50,000, leaving them an incentive to save 90% by building their own. x.com x.com x.com x.com
Nvidia's AI profits are real, but its compute demand depends heavily on fragile venture funding. Per Chamath, 1% of firms account for 90% of AI spending, and 72 cents of each dollar comes from unprofitable VC-backed firms. Venture money passes through startups to Nvidia and cloud suppliers; if the next raises fail, concentrated orders could contract together. x.com x.com x.com
Nvidia retains pricing power if its full system keeps reducing customers' cost per effective AI task. Strong AI demand alone does not ensure profits across the chain: rising chip, memory and optical costs burden buyers, while tighter financing and slow collections threaten operators. Large customers developing their own chips could pressure Nvidia's margins even as the market grows. x.com
Neoclouds such as Nebius, CoreWeave, IREN and TeraWulf remain long-term winners because cheaper, better AI models drive more compute use, not less. Per early testing by @Signal_65, GPT-6 Astra completed 279 of 280 enterprise tasks with zero fabrication, cutting weighted errors about 51% versus Claude Fable 5.1 at lower cost. Enterprises are deploying far more agents, pushing the upper end of 2028 capex estimates for the five biggest hyperscalers toward $1 trillion. x.com
AI agents' productivity over human labor should lift GPU demand and rates for neoclouds. The bottom looks in for CoreWeave, Nebius, Boost Run, Sharon AI and IREN. Gavin Baker says token prices can fall as revenue per GPU rises; paraphrasing Elon, even $38 per GPU-hour can pay off. CoreWeave's five-year fixed contracts may lag current rates; Boost Run has less capacity locked in. If SpaceX takes more growth, slower cash burn could rerate neoclouds, though smaller customers may pay less. x.com … Show more
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
Google's TPUv7 inference efficiency supports the bullish case for Alphabet. Per SemiAnalysis' first TPUv7 FP8 vLLM benchmark submission, Google's TPUs delivered 67% more AI inference performance per dollar than Nvidia's GB300 NVL72, even with Nvidia's system using FP4 precision. x.com x.com
