A viral model tier list lands the same week that a misread FT chart, AT&T's open-model migration, and a flipped Vercel token graph all point the same direction: the question is no longer simply who has the best model, but how you assemble a stack that routes the right task to the right one — and NVIDIA, quietly, is buying its way into the open half of that future.
The model war is no longer about who's best — it's about what belongs in your stack.
Frontier models have crossed a threshold where they can all do a lot, and usage volume has exploded — so individuals, teams, and enterprises are shifting from 'which model leads' to model efficiency and complete model architectures that route the right task to the right model. Theo's tier list is the fun version of that conversation; AT&T serving 40% of employee queries with open models, routers cutting coding costs 56%, and open tokens flipping to 62% of Vercel's gateway are the serious one.
Business Insider reports Hugging Face has engaged an investment bank to field offers at a $13 billion price — nearly 3x its 2023 round at $4.5 billion, which included Google, Amazon, Nvidia, Intel, and Salesforce. No deal has been reached yet, but the platform now hosts more than two million models, 1.5 million datasets, and 1.5 million AI apps.
As open models multiply, the coordination layer — helping developers find a model, judge whether it's safe, and put it into production — becomes harder to replace. Both Stripe's purchase of OpenRouter and the interest in Hugging Face look like bets that model fragmentation is the future, with commentators immediately floating NVIDIA as the buyer that makes the most sense.
Per Eric Newcomer, NVIDIA struck a $6 billion non-exclusive licensing deal for Poolside's technology plus a $1 billion equity investment at a $12 billion valuation, hiring over 100 Poolside engineers — the bulk of the team — to work on Nemotron. Poolside insists 'this is not an acquisition and it is not an acqui-hire': the founders stay, aiming to keep AGI from being 'a closed technology controlled by a few.'
The founders told shareholders they had a six-week window at the end of last year to raise $2 billion for a 40,000 GB300 cluster coming online in January. They didn't close in time, lost the cluster, and realized they'd run out of compute and capital as soon as next year — making NVIDIA the partner of necessity for a frontier open coding model.
Poolside for research talent, a stake in data-labeling startup Mercor (used for RL on the last two Nemotron models), and now a reported investment in Perplexity at a $30 billion valuation — a 50% markup — after initially floating a licensing-plus-hiring deal. Across all of these moves, NVIDIA is putting serious consideration into research, talent, training data, and the app layer as the open source frontier grows in importance.
The Information reports NVIDIA is notifying customers that top-end Grace Black and Vera Rubin chips will cost up to 17% more, applying even to chips ordered for delivery next year. A full 72-chip Vera Rubin rack is expected to hit $8 million, adding $5 billion to the cost of a gigawatt of compute — and cloud providers will 'almost certainly' pass it on. Blame the memory shortage, which looks set to stretch deep into next year or longer.
The largest offering of its kind in the Hong Kong market suggests China is pulling capital from every available source for AI — a departure from Alibaba's tight management of share supply. Michael Burry called Alibaba impressive as a disruptive force in the 'commodity low-cost LLM bloodbath,' but said he 'cannot bless share issuances' and expects return on invested capital to keep falling.
The robotics maker raised $900 million on the Shanghai Stock Exchange, debuted at a $9 billion market cap, and surged more than 460% on day one — a sign of strong appetite for China's embodied AI sector. Context matters, though: regulatory guardrails mechanically boost day-one performance, and this is the fourth Chinese IPO this year to rise more than 400% on debut.
— Dr. Dre, in a New York Times profile. In a New York Times profile, the rap legend said he's using AI extensively — a musical equivalent of brainstorming — and can't wait to see what happens next. Jimmy Iovine agreed ('when gifted people have AI, they're going to make better records') while noting AI companies have 'the worst public relations in the history of the world.' Both say plenty of producers are already using AI; they just won't admit it.
The AI entrepreneur and content creator put Fable 5 alone in S tier, GPT 5.6 Sol in A, and Kimi K3, DeepSeek V4 Flash, and GPT 5.6 Luna in B. Down in D: Anthropic's Opus 5 and Sonnet 5, GPT 5.6 Terra, and Composer 2.5. DeepSeek V4 Pro landed in F — and below F sat a dedicated 'Google tier' for Gemini 3.7 Flash and 3.1 Pro. Cue a ton of discussion.
The tier list matters not because it's fun to debate (though it is) but because individuals and increasingly businesses aren't picking one model — they're building infrastructure that moves between models by task. There's even a whole category of router companies built for exactly this, the best known of which, OpenRouter, was just acquired by Stripe for $7 billion.
The Financial Times chart showing limited enterprise spend on Fable 5 was read as businesses consciously rejecting the price. But Fable 5 carries a 30-day data retention policy — a provisional safeguard from when the model came back online after being shut down by the government — and that alone is enough for a huge number of enterprises to say absolutely not. Note how aggressively OpenAI has been pushing zero-data-retention for frontier models this past week.
The chart comes from Ramp's token and spend management product, so it samples tech-forward companies whose users are predisposed to cost control — as Simon Smith put it, 'Fable simply isn't cost-effective for most tasks.' And the shock that enterprises haven't adopted a months-old model misses the glacial pace of enterprise IT: people are still using GPT 5.2 and other nine-month-old models because that's what their companies give them.
Per The Information, AT&T will hold OpenAI and Anthropic spending flat and slowly supplant it with open models — already serving 40% of employee queries across its ~100,000 staff, with a target of 60-70%. VP of Data Science Mark Austin says open models are as good or better than previous-generation frontier models, which were already up to the task; frontier models stay for advanced work like code generation.
The company is using NVIDIA's Nemotron plus open models from Meta and Google, routed by task. Switching to open models also let AT&T host part of the service in its own data center stocked with NVIDIA and AMD chips — often cheaper than renting cloud compute. Chinese models aren't in the mix yet, but the risk analysis is underway.
— Theo, AI entrepreneur and content creator, in his tier-list video. Theo's paradox: Fable is the only S-tier model — 'it knows more than any model I've interacted with,' the one whose code he wants to merge and the one he trusts to double-check everything else — yet if forced to pick, he'd keep Sol, his default. 'Fable is that genius at the company that no one wants to work with, but no one wants to fire because they're the smartest person there.'
On any given day, the host is jockeying between Fable 5 and GPT 5.6 Sol — for many tasks initiating in both, going back and forth a little, then deciding which one to hone in on. That usually, but not always, ends up being Fable.
Theo puts Luna — theoretically the least capable of the three GPT 5.6 models — two tiers above the middle Terra model. Luna is 'smart, fast, and good at a bunch of random stuff' and his most-called model by volume, for tasks like categorizing code and reading content (nothing irreversible). Terra? 'It makes sense on a pricing chart, but doesn't make sense in reality for me.'
As labs compete on deliberate performance-efficiency trade-offs rather than just the state of the art, many models will end up neither frontier enough to justify a premium nor efficient enough for high-volume tasks. And price intuitions can mislead: Theo notes Kimi K3, despite being open weight, actually costs slightly more than Sol on extra high, because Sol does more with fewer tokens.
A vocal strand of the discourse is tired of model connoisseurship — comparing it to arguing over favorite colors or Pokémon — and even Theo concedes a tier list isn't the best way to compare models given the many axes: task capability, cost, token efficiency, speed. 'Whether you're using expensive best-in-class stuff like Fable or surprisingly cheap and effective stuff like DeepSeek V4 Flash, it's kind of hard to go wrong.'
Vercel CEO Guillermo Rauch showed that closed-model tokens went from roughly 72% of AI gateway usage in late June to 38% two months later, with open models rising from 28% to 62%. The data skews toward developers, but for some it shows exactly where the winds are blowing.
Investor Gavin Baker reads the Vercel data as open source taking share from OpenAI and Anthropic even as both accelerated in July — meaning total token demand grew even faster. His prediction: closed frontier tokens end up as 60-90% of economic value but only 15-25% of tokens. And since an open source token costs just as much compute to produce, nothing about open source inference is free — which is bullish for AI infrastructure.
MIT's Christian Catalini sees three spend categories: cheap generalists (commodity open weights), state-of-the-art generalists (closed-lab tokens), and in the middle, state-of-the-art specialists that combine open weights with enterprise proprietary context. It's the thesis behind Microsoft Foundry, which lets companies post-train on their own data atop MAI models — though how common that becomes across enterprises is genuinely debatable.
Increasingly, the job is understanding where different models fit for different reasons — and even slow-moving, single-ecosystem enterprises should set up environments where small groups can test approaches and hunt for these new efficiencies. Whether the days of getting excited about state-of-the-art releases are over, we'll find out soon: chatter says Fable 5.1 is coming shortly, while OpenAI's Astra has slipped to September.