eric vishria

Specialized inference providers like Fireworks achieve order-of-magnitude better price/performance than AWS/Azure/GCP on identical hardware by focusing exclusively on model serving optimization, capturing cloud margins while running on top.

Just as AWS/Azure/GCP plus Cloudflare created a 40/30/20 oligopoly with room for $100B niche players, AI infrastructure (foundation models, neoclouds, inference specialists like Fireworks, custom silicon) will support many large winners because no single vendor can scale fast enough to consume the entire market; zero-sum thinking misses the market's true size.

Traditional product management (specify requirements, engineers build) fails when model capabilities jump unpredictably; winners must deeply understand the 'jagged edge' of what models can/can't do, obsessively iterate, and bridge that edge to customer problems — technical taste and customer empathy now matter more than rigid roadmaps.

Winning AI applications (Cursor, Sierra) embrace building disposable 'sand castles' — products architected around current model jagged edges, with teams that re-evaluate every assumption every 4 weeks as capabilities shift, inverting traditional product management.

Legacy SaaS companies executing old playbooks (high gross margins, sticky databases) are worth 3x revenue because AI makes database migration trivial, favors zero-to-infinite scaling and low cost over lock-in; the market has compressed multiples from 30x to 6x despite 4x growth, and CEOs must invert from hill-climbing to rebuilding for AI-native criteria.

AI workloads are fundamentally limited by core-to-core communication, not compute. Cerebras' wafer-scale approach maximizes the three known hardware levers (cores, communication, memory proximity), while a new CPU architecture opportunity emerges for LLM-generated code execution.

Only three knobs exist to accelerate deep learning: more cores, faster core-to-core communication, memory closer to compute. Cerebras maximizes all three via wafer-scale. Next wave: LLMs generate code that runs on legacy CPUs burdened by backward compatibility; a clean-sheet CPU for AI-generated code could create a sixth major compute category.

Zero-sum thinking dominates AI investing (one lab wins, one chip wins), but cloud history shows oligopolies plus specialized $100B winners at every layer. The market is large enough for CSPs, neoclouds, inference specialists, chip startups, edge, and data center models all to succeed simultaneously.