AI model layer — open source & commoditization (new) Model layer commoditizes ▼ / infrastructure wins ▲

2026-AUG-04 — Dan Niles (CNBC): the 90% token-cost cut from open source plus automatic model routing by Azure / Google Cloud pushes margin out of the model layer and into the platforms and semis beneath it; frontier labs keep the hard tasks ("design a new database") and lose the routine volume ("you don't need them for what's three plus three… you don't need a Ferrari to go to the corner store to get milk"). Corollary governance point from the Anthropic–Figma board episode (Anthropic stepped off Figma's board and "like a week later, they launched a competing product"): enterprises "have to keep their data proprietary," itself a tailwind for self-hosted / open-weight deployment. Consequence for the private labs: "Anthropic and OpenAI become more commoditized over time." 2026-08-17 — independent corroboration of the model-layer dependency from strategists with no stake in it: "you could argue the hyperscalers have a real business and there are real moats around it, but it's a lot more capital intensive. The Anthropic's or Open AI's of the world, that's where it's most questionable… they're totally negative cash flow," and "the Chinese are competing with them." The systemic number: "something like 70% of the AI hyperscaler revenue is from just those two companies. So if this is the big if, if Open AI and Anthropic ever get in trouble, the ecosystem is in trouble" — Eisman: "I think it's undeniable." Same map he built solo on the Aug 14 Weekly Wrap, reached from separate client work. (Trennert on Eisman Ep 73, Aug 17) 2026-AUG-17 — Jay Singh (David Lin Report): "the models themselves will be commodities and I think everyone understands that now" — which is why the only way to capture the pie is first-mover compute, producing the mad dash and the hardware frenzy. 2026-AUG-23 — Jay Singh (SSR call, 2026-AUG-23): open-weight local models now match cloud frontier models on 70-80% of routine enterprise tasks (he discounts the headline 89% figure as "an optimized ensemble calculation… statistically inflated" assuming perfect routing). Zhipu AI's stealth OX Alpha (likely GLM-6) is "absolutely mogging every frontier model in cyber benchmarks" and could be "another deep seek moment" — constrained only by compute. Meanwhile Anthropic's flagship Fable 5 has plateaued at ~11% of total corporate spend on Anthropic tools as clients route to cheaper tiers (Opus 5) and open weights, which "challenges the assumption that technical capability leadership automatically translates into revenue" ahead of an October IPO. Goldman's Jim Covello argues the winners are the hyperscalers, who get their capacity filled. Kedrosky (Meb Faber #648, Aug 28): model performance hit its maximum year-over-year inflection almost four years ago (2022–23); on composite indices — not the promoted benchmarks, which models ingest ("no more than if you had seen the SAT before you'd done it") — gains have flatlined over the last six months, from 10–12% a year to "one or two% at most." Cross-vendor variance has collapsed: "very little difference between a frontier model from Anthropic and a frontier model from Qwen or from DeepSeek" in practical composite terms, and in his blind Pepsi-Coke tests behind a harness "inevitably no one can tell the difference." What creates the illusion of progress is the harness (Claude Code, the codexes) — "the models are the bratty kids and the harnesses are Julie Andrews." Conclusion on capital: "the game is almost over in terms of pretending that you can justify multi-billion dollar training runs" — "the most successful frontier AI company will be the first one to stop pretending they can train new AI models." Convergence also has a data cause: "the median data nugget inside of a large language model is a 37-year-old male on Reddit," so models amplify herding rather than break it. Steve Eisman (New Money clip, 2026-AUG-15): the sharpest statement of the no-moat case yet — "the debate has really shifted because there just don't seem to be any moats, or at best, the moats are shallow. Enterprises are switching between models and using cheaper open-source Chinese models in order to control costs. The future for these large LLM providers is very questionable. The Chinese models are much cheaper, and this could eventually cause a price war." The tell he points at is customer behaviour, not benchmark scores: switching costs low enough that a procurement decision can move a workload is the definition of an absent moat. Host exhibit (Artificial Analysis, not Eisman): DeepSeek V4 Flash ~3¢ per test vs Moonshot's Kimi K3 ~86¢, GPT-5.6 $1.86 and Claude Fable 5 $3.15. (2026-AUG-28, The David Lin Report): asked why the substitution isn't visible yet — "I'm hearing that people are starting. Takes time. But if that does happen, there would be a price war." (2026-SEP-03, Dan Niles, Excess Returns) He re-states and extends the chain — "you don't need a Ferrari to go to the corner store to get milk… people aren't going to be using Anthropic to go summarize their emails" — so "the value is going to move from the model providers to the infrastructure providers and then ultimately the companies sitting on top of that that create great businesses," the way stranded dot-com bandwidth produced Google, Amazon, Facebook and Netflix. Within the model layer he is long Anthropic and short OpenAI on evidence rather than taste: Anthropic "turned profitable in Q2" while OpenAI "lost even more money in Q2 relative to Q1," and Anthropic on enterprise plus Google on consumer "squeeze OpenAI between them." 2026-SEP-07 (Jay Singh): per OpenRouter, Chinese open-weight models are ~70% of usage versus US LLMs, and SiliconData's usage-weighted token-expenditure index fell below $1 per million tokens for the first time, from $2.04 at end-May — −50% in three months. He supplies his own caveat (OpenRouter over-indexes model-switching and smaller developers). The mark-to-market: Anthropic's estimated valuation fell from ~$1.4tn to ~$900bn and the IPO slipped again to mid-October, “as Chinese models gain share.” 2026-SEP-07 (RiskReversal — Nathan & Adami): Nathan — every frontier model “will be sitting on AWS or Azure… you're going to choose the cheapest one from a token perspective,” so a one-week leaderboard win moving a mega-cap 3.5% is “really stupid.” Adami closes with the bull case's own analogy turned against it — Jensen Huang's “most important thing since electricity,” and “electricity is a commodity.” The value accrues to the distribution layer (AMZN/MSFT/GOOGL), not the model maker; Meta lags OpenAI/Anthropic/Gemini and has no cloud business to recoup the spend. 2026-SEP-08 (WSJ AI & Business): capital keeps flowing to non-US challengers at the model layer — China's Moonshot AI valued at $50bn in its latest round (one of several Chinese labs challenging the leading US developers) and France's Mistral at more than $24bn. App Economy Insights (2026-SEP-11): Apple states the 'own the layer above the model' thesis at the device level. It 'does not necessarily need to win the foundation-model race,' can rely on outside models, and competes on device, OS and personal context. Monetization is indirect ('does not need a $20 monthly Siri subscription if AI helps sell more $1,200, $2,000, or even $3,000 devices'), in contrast to Meta's same-week $20/$100 Muse agent tiers. Steve Eisman (2026-SEP-11): OpenAI's CFO disclosed a price cut on GPT 5.6 Luna shortly after its July release, claiming a tenfold increase in usage - "I wonder if Anthropic will have to follow… Is this the harbinger of a price war?" 2026-SEP-14 (Carlson, via Colossus): The heads-I-win-tails-I-win frame. If models commoditize, value accrues to the complements (distribution, attention, personalization, commerce) and pure model sellers like Anthropic and OpenAI are worth "dramatically less". If not, a company within ~6 months of the frontier with its own model (Meta) is insulated, while Apple, with no in-house AI, stays "in his competitor's prison". Dan Niles (2026-SEP-15): "you don't need a Ferrari to go to the corner store to get milk" — ~90% of requests will go to open-weight models, ~10% to frontier models. Cost per token is down ~50% since end-May while tokens produced are up almost 4x. He expects few frontier winners (Google, Anthropic); OpenAI is "stuck between" them. Anthropic's run rate is ~$65B vs OpenAI ~$40B. Mike Taylor (Hedgeye, 2026-SEP-15): pricing is the next shoe — pharma (Bristol Myers, Lilly) internalizing AI, and select software (ServiceNow, Palantir) aggregating compute purchases for thousands of clients to push prices down. Too many eating the pie: Anthropic gets its raise done, OpenAI is the one 'they got to get rid of' (possibly via a bankruptcy that wipes out investors while the tech survives). McCullough adds high-grading: the first $1.5T of capacity earns better pricing than the next. Daniel (Eisman Ep 75, 2026-SEP-14): enterprise channel check — AI tokens +30–40% but costs −60%, routing only the most important queries to Anthropic/OpenAI and the rest to Chinese open-weight models: "how do we get to the returns on invested capital?" Sep-10 (Paulo Macro): a new DeepSeek release sits “literally right behind Astra. And 1.4% of the cost” — “these open models are almost there and they cost a fraction.” Peterffy, 2026-SEP-16: even if frontier models stopped progressing, "so many open-source models" adopted across companies would drive huge earnings and productivity gains. The benefit accrues to users, not model builders. CNBC Halftime (2026-SEP-16): Zuckerberg opposes Amodei and Altman's frontier-model slowdown and regulation call, siding with Jensen Huang; Meta is up ~18% in September. Harrington: "why are you telling us it's dangerous, but you're still building hand over fist?" Weiss: "disregard what they're saying. They're all advancing." Eisman (CNBC, 2026-SEP-17): the AI-safety alarm is "all nonsense" — "no evidence at all that AI has achieved artificial general intelligence… most of the evidence seems to point that it never will." What is really happening: token maxxing "is over," "the open weight models are taking big market share," and the labs "realize that there are no moats around their business whatsoever," so they are "trying to manufacture a crisis" that yields regulation they can shape "to create the duopoly that they want." To labs urging a slowdown: "really postpone your IPO." 2026-SEP-18 (Eisman): "business is potentially slowing because token maxing is ending and open weight models keep taking market share. There are no pricing moats in this business. Today I have the best LLM and tomorrow yours is better and cheaper" - while data-center and capital costs rise; the labs "see I believe a price war coming." Joseph Carlson (2026-SEP-21): Meta's Muse beat ChatGPT, Gemini and Anthropic to the consumer agent (#1 App Store, 4.9 stars) on distribution (3.6B DAU), UX and generous free usage rather than a model moat; he calls Google's lack of an agent despite owning Gmail/Docs/Maps "embarrassing" and predicts an OpenAI counter-launch during Meta Connect.