The AI Model Tier List

The AI Model Tier List

The story everyone will repeat from this episode is the tier list. Skip it. The data that matters is Vercel's: closed-model token share fell from 72% to 38% in two months, with open models climbing to 62%. AT&T is running 60-70% of employee queries on open weights while holding frontier spend flat. The single-model era is over for anyone running AI at scale. What replaces it is a router and a spreadsheet.

The core decision is whether to keep routing everything to a frontier model (GPT-5-class, Fable 5, Opus) or build a layer that sends most traffic to open weights and reserves expensive tokens for tasks that need them. The forcing function is Nvidia's chip price hikes, up to 17% on Vera Rubin and Grace Blackwell, passed downstream by cloud providers. Your inference bill is going up in 2026 either way.

On the Vercel data: the skeptic's read is that Vercel's user base skews toward cost-sensitive developers building coding agents, exactly the population most eager to dump frontier tokens the moment an open model gets close. That's not AT&T's help desk. Mark Austin's "2% quality degradation with 56% cost savings" figure also deserves scrutiny: a 2% average across a task mix can hide a 30% failure rate on the 5% of queries that actually matter. Know your tail before you trust that number.

AT&T is the stronger data point precisely because it's a regulated 100,000-person organization. It runs Nemotron, Meta, and Google open models on-prem and found it often cheaper than renting cloud compute. That's an accounting line, not an ideology.

The useful frame comes from Christian Catalini's three buckets: commodity open generalists, closed frontier generalists, and a premium middle of open weights plus your own fine-tuning and proprietary context. Gavin Baker's version of the same idea: closed frontier tokens end up as 60-90% of economic value but only 15-25% of token volume. The valuable work concentrates in a thin slice of frontier calls. Everything else commoditizes. The money is in owning the routing decision and the context you feed it.

On Nvidia: read the price hikes and the acquisitions together. A 72-chip Vera Rubin rack hits $8M. Nvidia just paid roughly $6B to license Poolside's technology and hire most of its engineers to staff Nemotron, invested in Mercor for RLHF data, and is reportedly circling Hugging Face for distribution. Three moves that only cohere if Nvidia intends to ship competitive open models. The mechanism is simple: open inference burns identical compute per token regardless of who made the model, so Nvidia profits from open-model adoption through GPU sales even if the models carry zero margin. Giving away strong open weights is pure demand-generation for its silicon. The concentration risk is real, a chip vendor that also owns the models and the distribution layer is worth watching, but Nvidia's incentives are aligned with more open-model usage, full stop.

What to actually build: a router in front of the workflows you already run. OpenRouter exists and Stripe just paid $7B for it, so this is infrastructure. Tag tasks: summarization, HR lookups, and support drafting go to open weights; code generation and review stay frontier. That mirrors AT&T's split exactly. Avoid the middle tier, models like Terra that aren't cheap enough to be the default or capable enough to be the specialist. And read every model's data-retention terms before evaluating capability. Fable 5's low enterprise adoption has nothing to do with quality: a 30-day data retention requirement is killing its deals. Your procurement constraint is now a model-selection input.

On-prem is cheaper only if your utilization is high. AT&T at 100,000 employees keeps those GPUs busy. Most teams don't. Build the eval on your own long-tail queries first, then let the router's fallback rules trigger on that eval, not on a vendor's blended percentage.

The call: by Nvidia's Q2 FY2027 earnings (expected late August 2026), Nvidia will have either closed or publicly walked away from the Hugging Face acquisition, and Nemotron will be positioned as a direct answer to DeepSeek and Moonshot. The $6B Poolside spend and hiring pattern make the intent explicit. The Hugging Face piece is the coin flip: a $13B deal invites regulatory scrutiny that the Poolside structure was deliberately designed to avoid. Medium confidence.

The story everyone will repeat from this episode is the tier list. That's the wrong thing to fixate on. The real content is the Vercel data: closed-model token share fell from 72% to 38% in two months, with open models climbing to 62%. And AT&T serving 60-70% of employee queries on open weights while holding frontier spend flat. The single-model era is over for anyone running AI at scale. What replaces it is a stack, a router, and a spreadsheet.

Let me frame what's actually being decided here for a team shipping AI into production.

The decision: Do you keep routing everything to a frontier model (GPT-5-6 Sol, Fable 5, Opus), or do you build a routing layer that sends most traffic to open weights and reserves the expensive tokens for the tasks that need them? This is a Type 2 decision at the task level (easy to reverse, try it on one workflow) and a Type 1 at the infrastructure level (a router plus on-prem hosting is a real commitment). The forcing function is Nvidia's chip price hikes, up to 17% on Vera Rubin and Grace Blackwell, passed downstream by cloud providers. Your inference bill is going up in 2026 whether you act or not.

---

The Skeptic

Two months of Vercel token data is a fad, not a trend. Vercel's user base skews toward developers building cost-sensitive coding agents, which is exactly the population most eager to dump frontier tokens the moment an open model gets close. That is not AT&T's help desk, and it is not your brand-safety classifier under load. The 56% cost cut with 2% quality drop sounds clean until you ask who measured the 2% and on what. Mark Austin gave a number; nobody showed the eval. And "2% quality degradation" averaged across a task mix hides the tail, where a 2% average can be a 30% failure on the 5% of queries that actually matter. For a PM: cheap models are fine until the one query that mattered breaks, and averages won't tell you which one.

The Researcher

The interesting claim is Gavin Baker's: closed frontier tokens end up as 60-90% of economic value but only 15-25% of token volume. That is the whole game stated in one line. The valuable work concentrates in a thin slice of frontier calls, and the bulk commoditizes. Theo's own behavior proves it at the individual level, he says Fable 5 is the best model he trusts to double-check work, then picks Sol because it's fast and follows instructions. Best and most-used have split. Christian Catalini's three-bucket read is the useful frame: commodity open generalists, closed frontier generalists, and a premium middle of "state-of-the-art specialists," open weights plus your proprietary fine-tuning and context. For a PM: the money isn't in owning the smartest model, it's in owning the routing decision and the context you feed it.

The Open-Source Advocate

This is the episode where the open-weight case stopped being ideological and became an accounting line. AT&T runs Nemotron, Meta, and Google open models on-prem and found it "often cheaper than renting compute from cloud providers." That's a telecom, not a hobbyist. And Nvidia just told you where the puck is going: $6B for Poolside's tech plus a $1B equity stake, hiring most of the engineering team to staff Nemotron with the stated goal of beating DeepSeek and Moonshot on open models. Then Mercor for RLHF data, then a possible $13B Hugging Face acquisition to own distribution. The chip vendor is now building the open-model supply chain end to end. For a PM: the company that sells the shovels is now also giving away the gold, because it makes money either way as long as you keep buying compute.

The Compute Pragmatist

Read Nvidia's moves and the price hikes together and the strategy is obvious. A 72-chip Vera Rubin rack hits $8M, adding $5B to the cost of a gigawatt. Nvidia is pushing open models precisely because Baker is right: open-source inference burns identical compute per token. Lower model margins don't reduce GPU demand, they increase it. So Nvidia would love for you to run cheap open weights all day, as long as you run them on its silicon. On-prem gets more attractive as cloud markups stack on top of chip hikes, but "on-prem is cheaper" only holds if your utilization is high. AT&T at 100K employees keeps those GPUs hot. Your team probably doesn't. For a PM: buying the hardware only beats renting if you keep it busy, and most teams can't.

The Builder

What do I ship Tuesday? A router in front of the workflows I already run. OpenRouter exists, Stripe just paid $7B for it, so this is now infrastructure, not a science project. Start by tagging tasks: summarization, HR lookups, and customer-support drafting go to open weights, code generation and review stay frontier. That mirrors AT&T's split exactly, and it maps onto bid factors, the same multiplicative routing logic we already run for audience, geo, and device. The trap is the middle tier, Theo's "uncanny valley" models like Terra that aren't cheap enough to be the default or smart enough to be the specialist. Don't route to them. And watch the operational tax: Fable 5's low enterprise adoption isn't capability, it's the 30-day data retention requirement. A data-handling clause killed a frontier model's adoption. Your procurement constraint is a model-selection input now.

---

Where the council splits

The Skeptic and the Open-Source Advocate genuinely disagree on the Vercel number. One sees a self-selecting developer population fleeing to the cheapest thing that clears the bar; the other sees the leading edge of what every enterprise does once the router is built. AT&T is the tiebreaker, because it's a regulated 100K-person org, not a Vercel coding-agent user, and it landed in the same place: majority open, frontier reserved.

The second split is Nvidia's. The Compute Pragmatist says Nvidia wins no matter which model you pick, because every token burns its silicon. The Open-Source Advocate says a chip vendor that also owns the models and the distribution layer (Hugging Face) is a concentration risk the ecosystem should worry about. Both are right, which is the uncomfortable part.

What it hinges on

Whether the routed open model holds quality on YOUR task tail. AT&T's 2% number is a task-mix blend; your mix is different. The council leans hard toward building the router regardless, because even a conservative split (open weights for summarization and support, frontier for code) captures most of the savings at low risk, and the chip price hikes make doing nothing the expensive option.

Before committing: build the eval on your own long-tail queries first, then let the router's fallback rules trigger on that eval, not on a vendor's blended percentage. And read every model's data-retention terms before capability, because Fable 5 just showed a retention clause outweighs being "the best model."

---

Prediction: By Nvidia's Q2 FY2027 earnings call (expected late August 2026), Nvidia will have either closed or publicly walked away from the Hugging Face acquisition, and its Nemotron open models plus the Poolside talent will be positioned as a direct answer to DeepSeek and Moonshot, not as a side project.

Confidence: Medium. The $6B Poolside spend and hiring pattern make intent explicit, but a $13B deal could stall on antitrust or price.

Why: Nvidia has paid roughly $6B to license Poolside and hire most of its engineers explicitly to staff Nemotron, invested in Mercor for RLHF data (reinforcement learning from human feedback, the process that fine-tunes models to follow instructions), and is reportedly circling Hugging Face's distribution layer. Three moves that only cohere if Nvidia intends to ship competitive open models, not just sell chips. The mechanism is straightforward: open inference burns identical compute per token regardless of who made the model, so Nvidia profits from open-model adoption through GPU sales even if the models carry zero margin, which makes giving away strong open weights pure demand-generation for its silicon. The opposite outcome, Nvidia quietly abandoning the open-model push, is unlikely because it has already spent the money and hired the people. The Hugging Face piece remains the coin-flip: a $13B acquisition invites regulatory scrutiny that Poolside's structure was deliberately designed to avoid.

Revisit by 2026-11-30: We're right if Nvidia has closed or abandoned the Hugging Face deal AND publicly frames Nemotron as a DeepSeek or Moonshot competitor by Q3 FY2027 earnings. We're wrong if Nemotron remains an unmarketed internal effort with no open-model positioning and the Hugging Face talks are still open with no decision.