A running analysis of AI's winners and losers — covering infrastructure, company outlooks, key predictions, and the strongest counterarguments.
Value flows to physical chokepoints and demand aggregators — not model builders or traditional software vendors.
The binding constraints on AI are physical: fabs, memory, power — not algorithmic. $700B+ in annual hyperscaler CapEx is the investment phase of a multi-decade build-out. A chip shortage ~2029 is near-certainty.
Google's ads get more targeted. Meta's feed gets sharper. Shopify's merchants get more tools. The "AI startup kills incumbent" narrative is mostly a startup funding mechanism, not serious analysis.
AI increases supply of content, code, and services. Whoever controls the demand side — users — gets stronger. Commoditized supply empowers aggregators above all others.
HBM memory and TSMC fab capacity are the real binding constraints — not model quality or software capability.
HBM demand for AI is crowding out consumer electronics across the entire industry simultaneously.
Sony — considering delaying PS6 to 2028–2029.
Samsung — reviewing contracts quarterly instead of annually.
Chinese OEMs — cut 2026 shipment targets by up to 20%.
Nintendo — making customers supply their own SD cards.
Valve — delisted Steam Deck in the US.
Massive chip shortage ~2029 due to TSMC's conservative capacity expansion relative to AI demand.
Microsoft exits first-party console hardware entirely. Xbox becomes a publishing and Game Pass brand.
Memory prices squeeze consumer electronics for years — phones, PCs, and consoles all simultaneously.
Government demands control of AI regardless of who funded its development. Inevitable, not normative.
The AI-native business model — the "feed" equivalent — has not yet been invented. OpenAI's banner ads are primitive.
OpenAI must find product-market fit in 2026 or face commoditization. Shallow weekly usage is not a moat.
Netflix eventually acquires both Paramount and Warner Bros. — possibly simultaneously — within the decade.
AI shopping benefits long-tail merchants (Shopify, Etsy) more than Amazon, by surfacing niche products agents can discover.
PE opportunity in beaten-down SaaS — buy at low multiples, restructure for profitability, run as cash generators.
EA's long-term play is live sports rights — a single platform for watching, playing, betting, and fantasy.
Agentic inference will become the largest compute market — and it won't look like today's GPU clusters. Memory hierarchy (DRAM, SSDs, databases) beats raw bandwidth; CPU speed for tool use matters more than GPU speed; latency is irrelevant without a human in the loop. Nvidia's dominance is training + answer inference, not agentic.
His framework is internally consistent — but these counterpoints deserve serious weight.
IBM embraced the PC. Newspapers launched websites. The sustaining phase is real and temporary. The analysis itself admits the AI-native business model hasn't been invented. You can't call the game at halftime.
WorldCom and Global Crossing had real technology and real revenue too. Cisco lost 90% despite being the "picks and shovels" play. When CapEx cycles crack, supply-chain winners get hit too.
Gross margins expanded from ~53% (2022) to ~58–60% (2026). Supply constraint is deliberate. The thesis can't have TSMC be both admirably rational and problematically insufficient.
Better answers mean fewer follow-up queries. AI Mode costs 5–10x more per query. Agents won't click ads. Antitrust stripping default status puts $40–60B in annual revenue at risk.
Apple Intelligence runs on-device. NPUs with 40+ TOPS are shipping at scale. Round-trip latency (50–200ms) makes real-time agents unusable in many scenarios. Privacy regulation pushes local.
"AI capabilities will continue to improve on a predictable trajectory."
This is treated as axiomatic. S-curves are only identifiable in retrospect. If improvement stalls: $700B+ CapEx becomes the largest misallocation in history, SaaS recovers, and supply-chain plays crash. This is the scenario this framework fails to model.