The short version

In late 2025, Databricks CEO Ali Ghodsi delivered the sharpest bear-case verdict in the entire model-layer debate: large language models are already a commodity: like gas stations, you just compare prices, and users can switch in a day. Frontier labs, in his telling, end up as fab-like companies: very valuable, but interchangeable. So the money flows to three places instead — enterprises' own data, applications, and governance and security. What makes the prophecy better than most slogans is that it can be checked: it supplies three gauges of its own. First, are enterprises using multiple vendors side by side? Second, are model-maker margins being squeezed toward foundry-style margins? Third, is any single lab persistently pulling away with the revenue? This piece does two things. One: it reads those three gauges as of July 2026 — and all three needles point the opposite way from the prophecy, though each caveat is a real methodology gap — selection bias, unaudited figures, leaderboard noise — not a formality. Two: it takes his TSMC analogy seriously and checks real foundry economics — and what comes back is sharper than any rebuttal: TSMC holds a near-monopoly in advanced nodes and has raised prices four years running; "interchangeable" is only true in mature nodes. Read correctly, the analogy doesn't predict wholesale commoditization of models — it predicts the tiered map: a top tier that oligopolizes, a commodity base, and a line between them that keeps moving. The analogy also has a break point: TSMC doesn't design its own chips, and model labs do. That break point caps the "money flows to the apps" leg of the prophecy — and it is the new judgment this piece adds.

What the prophecy is betting: one gas-station analogy, three checkable wagers

Start with the exact words. On the BG2 podcast in December 2025 — with Arvind Jain, CEO of enterprise-search company Glean, chiming in agreement — Ghodsi said: "I think the LLM is a commodity. People are not saying that, but it is a commodity. Like you can get gas from this gas station, you can get gas from that gas station... Just compare price." (BG2 Pod / Altimeter) He then supplied the endgame: frontier labs are "going to be kind of like these fab-like companies... TSMC is very valuable. But... they're interchangeable... People just switch LLMs like in one day." And if the model layer can't hold the money, where does it go? He named three layers: proprietary enterprise data, applications close to the workflow, and — once agents (AI programs that execute multi-step tasks on their own) are running everywhere — the governance and security layer that becomes unavoidable. In his words: "what is special is the data that you have... the governance security layer is going to be super super important... I do think most of the value will accrue to the apps."

This judgment earns a column not because it is provocative but because it directly decides a class of real-money choices. If models really are a commodity, then betting on a single lab is wrong, and the right position is the neutral multi-model platform — the model supermarket in the cloud, the routing layer, the governance layer. If lock-in is real, the right move is the opposite: bind early to the strongest lab. Which is also why it matters who is saying this. Databricks, having abandoned frontier-model development of its own, has rebuilt as exactly that multi-model distribution-plus-governance layer: a five-year agreement with Anthropic, a multi-year deal with OpenAI in the hundred-million-dollar range, both sold under its own governance gateway (Databricks press release). If the commodity thesis holds, his company sits at the exact center of where the value flows. That does not make him wrong — he backed the judgment with the real-money act of quitting frontier development — but it is the judgment of a position-holder, not a neutral observer, so it should be read here with that discount applied.

The prophecy names its own gauges, which is what puts it above most slogans. Gauge one: are enterprises genuinely running multiple vendors and switching often, the buyer behavior a commodity should produce? Gauge two: are model-maker margins compressing toward foundry-style margins — the seller P&L a commodity should produce? Gauge three is a counter-indicator: if one lab keeps taking the revenue curve for itself, the commodity thesis weakens. Here are the three readings as of July 2026.

Audit stop one: more people are switching models; the money isn't switching vendors

Start with the side that supports the prophecy, because the evidence there is real. By a16z's January 2026 enterprise survey, 81% of enterprises now run three or more model families in testing or production — 68% a year earlier (a16z). Perplexity's enterprise telemetry cuts finer, into behavior: 53% of users who actively pick their model switched at least once within a working day, and its fifty largest enterprise accounts averaged 30 models (Perplexity). On OpenRouter, the model-routing platform where developers congregate, US models' share of tokens fell from about seventy percent to about thirty within a year (Don't Worry About the Vase). At the API layer, "switch LLMs in one day" is not rhetoric; it is daily practice.

But move the lens from "how many models are in use" to "did the money move," and the picture inverts. Menlo Ventures' enterprise survey measures what happens after a vendor is chosen: enterprises stay. Only 11% of teams switched model vendors in the past year; 66% upgraded to newer models inside their existing vendor (Menlo Ventures). Another 2026 enterprise survey supplies an uglier number: 89% of enterprises believe they could switch vendors, but among those that actually attempted a migration, 58% hit failure or unexpected difficulty (Parallels survey, as relayed). Both numbers need discounting before use. Migrations that get attempted skew toward hard cases, so the 58% carries selection bias; and whether an 11% annual switch rate is high or low for enterprise software at large has no reliable baseline to compare against — it shows the money moves slowly, not, by itself, that the money can't move. Still, between believing you can switch and actually switching sits a full layer of engineering reality: prompts need retuning, evals need rerunning, fine-tuned assets don't come along. And vendors are actively growing that cost: 15% to 30% discounts for annual commitments of $500K and up, multi-year tiered pricing, millions of dollars in compute credits for startups that sign long: textbook switching-cost manufacturing, every move of it.

The two data sets don't contradict each other; they measure two different layers of reality: adding a model is cheap; switching a vendor is expensive. Developers rotate through three models a day at the routing layer while procurement signs the annual contract with the same vendor. The weak version of the commodity thesis holds: multi-model use is now the norm. The strong version fails on buyer behavior: zero switching cost has not cashed out into zero stickiness or zero pricing power. Needle one points against the prophecy, with a caveat: routing-layer tooling keeps maturing, gateways supporting a thousand-plus models are already off-the-shelf commodities, and today's engineering friction is a current condition, not a law of physics.

Audit stop two: margins and concentration, both needles running the wrong way

The second gauge is the seller's P&L, and it needs one sentence of provenance first: Ghodsi's own words gave the foundry analogy and no margin number — "model-maker margins converge toward foundry margins" is how this piece translated the prophecy into a testable claim. But on the analogy's own logic — capital-intensive, interchangeable, scale-driven — the only direction it can predict is margins getting flattened. What happened is the reverse. Anthropic's inference-serving gross margin, per internal and analyst accounts, climbed from 38% in 2024 to around 65% in 2026, and per Normal Technology's relay of Wall Street Journal reporting, the company is entering its first profitable quarter since founding (AI Snake Oil / Normal Technology). The rule this series set in the previous pricing-power column applies immediately here: these margin figures are all unaudited, and the same company currently has three margin books in circulation — overall gross margin, inference-serving gross margin, compute-cost ratio — with incompatible definitions; listing documents are the court of final appeal. But the direction is unambiguous: the needle is climbing, not being crushed. And the climb has a destination with some dark comedy in it: it is approaching, from below, TSMC's 67.7% — which, as the next section shows, is a monopolist's gross margin, not a commodity's. Flagship prices are rising too: OpenAI more than doubled its flagship API tier (GPT-5 at $1.25 per million input tokens and $10 per million output became $5 and $30 with GPT-5.5), and Anthropic opened a Fable 5 tier priced above Opus ($10 and $50). Even Jain — the man who seconded the commodity thesis in that room — said on another podcast seven months later that "in the last 6 to 9 months... every model actually increase[d] their per token price" (20VC). A commodity's price trajectory is supposed to be monotonically down; this is an anomalous reading reported by the commodity camp's own instrument. In fairness, Jain has a self-consistent account: the top tier is what's rising, while ninety percent of enterprise use cases sink to cheap models. The account holds together internally. But look at what it pays to hold together: it abandons "the model layer as a whole is a commodity" and replaces it with a tiered version, where the top has pricing power and only the base is a commodity. That tiered version is precisely what this piece is about to build out of foundry economics.

The third gauge is the prophecy's own counter-indicator, and it is ringing loudest. Menlo's two time points show enterprise LLM spend share concentrating, not evening out: Anthropic went from 12% in 2023 to 32% in mid-2025 to 40% by December 2025, with the top three labs taking a combined 88% (Menlo, as relayed). In coding — the high-frequency workload that should commoditize first — a single lab holds roughly 54%, two and a half times the runner-up; Claude Code alone went from $500 million to $2.5 billion in annualized revenue in five months (Stratechery). One rebuttal has to be blocked in advance: share concentration by itself does not refute commoditization — commodity industries can be highly concentrated on scale economics alone, and TSMC is itself the example. What a real commodity market cannot produce is concentration, price increases, and rising margins all at the same time: a commodity leader can hold a big share, but it does not have the nerve to raise prices year after year. One definitional detail also needs flagging: Menlo measures spend share, not token share: usage is far more dispersed, with enormous token volume running on cheap models. Concentrated spend coexisting with dispersed usage is, once again, a tiered map.

This stop's verdict needs two caveats written into it, and they are both real. First, leaderboard leadership genuinely is rotating: in July 2026, GPT-5.6 reportedly pulled even with Claude Fable-5 on Code Arena, the developer-voted coding leaderboard (via X user @infwinston; the original post is not archived and no permanent link exists to verify it), and it's now widely said that no frontier model wears the crown for long. But leaderboard rotation and procurement stickiness are two different meters — one measures capability, the other measures money, and so far the money is not following the leaderboard. Second, the price war has genuinely opened at the base: Google is pushing in with Gemini 3.5 Flash at half to a third of comparable prices, and large customers are shifting from burning tokens freely to conserving them (CNBC); open-weight models discount closed flagships by 96% or more (DeepSeek). The commoditization mechanism is forming for real. It just isn't happening at the top of the stack. Where it happens is exactly the question the analogy, read properly, answers next.

Taking the analogy seriously: foundry economics is not on the side of "interchangeable"

"Labs will become companies like TSMC: very valuable, but interchangeable." That analogy is the skeleton of the whole prophecy — and the strange thing is that in seven months of debate, nobody on either side actually checked it. Is the real foundry business actually a "valuable but interchangeable" industry?

The answer that comes back is: precisely the opposite, and opposite in a structured way. In the first quarter of 2026, TSMC took 72.3% of the global foundry market; in advanced nodes — the newest two or three process generations — its share exceeds ninety percent. Its 2-nanometer capacity is booked through the second quarter of 2027, and the launch-customer list reads Apple, AMD, NVIDIA, MediaTek (the Taiwanese fabless chip designer) — not one of them hands its flagship product to Samsung or Intel (industry roundup). It is expected to raise 3-nanometer prices by up to 15% in the second half of 2026 — the fourth consecutive year of increases (TrendForce, as relayed). Its gross margin in the second quarter of 2026 was 67.7%, an all-time high — a figure from its SEC filings, and the only audited number in this section. Consecutive price increases plus customers with nowhere else to go: that is called pricing power, not a commodity. "Interchangeable" does exist in the foundry business, but only in mature nodes: the old processes anyone can run, where the price war ground on until 2027 brought the first increase in three years. The real shape of the foundry industry is: monopoly at the top, commodity at the base, and between them a line that only enormous capital expenditure can hold.

So take "labs are like TSMC" seriously, and what the analogy yields is not "the model layer fully commoditizes" but: the flagship tier oligopolizes and holds pricing power; the trailing tier commoditizes into a red ocean. Map that back onto the last two stops' readings — flagship prices rising, spend share concentrating at 40%, open-weight discounts at 96%, a price war at the base — and it is exactly the current shape of the model layer. It is also the same line this series has now measured twice from different angles: the coding column measured it along the task-difficulty axis — native integration earns a premium on hard tasks, easy tasks decompose into parts; the pricing-power column measured it along the price-tier axis — flagship list prices moving up, the budget tier burning in a price war. Read the analogy right, and the three maps are one map.

A methodological self-warning belongs here, because it too comes from the commodity camp. Analyst Benedict Evans is the independent second voice for "commoditization is the default endgame," but he is equally explicit that arguing by analogy has no predictive force — mobile, fiber, semiconductors, cloud: every analogy differs, and "it's like industry X" is not an argument. The warning cuts both ways: Ghodsi doesn't get to prove commoditization with "like TSMC," and this piece doesn't get to prove oligopoly with it either. What this piece does with the analogy is demote it — from proof to structural template: it proves nothing about what the model layer will do; it only supplies a set of checkable correspondences — and the checks, the gauge readings of the last two stops, stand independently of the analogy. Be equally clear about what the template cannot answer: how long the top tier's pricing power lasts (past the end of the compute shortage? past the open-weight catch-up?) — that is the three-pillar question the pricing-power column left open, and it is not retried here. One more thing surfaces on a close re-listen to the original: Ghodsi's actual phrasing was that even with Samsung and Intel in the market, TSMC remains very valuable — a sentence with the oligopoly already hiding inside it, papered over by the word "interchangeable." The most defensible version of this prophecy may have been the tiered version all along, compressed into "it's all a commodity" somewhere in transmission.

Two places where the template genuinely doesn't fit deserve honest listing. First, generation lifespans differ by more than an order of magnitude: a TSMC node lead holds for years; a model's leaderboard crown lasts a short stretch before changing heads. If capability leads really are that short-lived, the model layer's top-tier "monopoly" ought to be far more fragile than the foundry's. Right now it is not acting fragile — share is still concentrating — and the likelier explanation is the one from stop one: buyer stickiness makes the effective lead run far longer than the leaderboard lead. Second, old nodes meet completely different fates: TSMC's depreciated old lines run at close to pure profit (Stratechery), while the model layer's "old nodes" are taken outright by open-weight models: a 96% discount is not an old model cutting its price; it is someone else giving the equivalent away. Which means real foundry economics is actually kinder than the model layer's: it lets the incumbent keep earning the commodity base; model labs cede that base outright. For the "endgame of the labs," this aggravates rather than mitigates: they depend more than TSMC does on holding the top.

Where the analogy breaks: TSMC doesn't build its own chips, and model labs do

There is one more break point, and it matters more than the two above, because it lands directly on the "money flows to the application layer" leg of the prophecy.

The core of the TSMC model's value is a self-imposed constraint: pure foundry, never design your own chips (Stratechery). That "we will not compete with our customers" promise is what let NVIDIA, Qualcomm, and Apple hand over their crown jewels, and what grew the entire fabless semiconductor industry. Samsung's foundry share sits in the single digits partly for this exact reason: it builds its own phones and its own chips, and customers don't trust it. Model labs are running in the opposite direction: the three most complete agent products on the market — Claude Code, Codex, and Antigravity (Google's agent development tool) — are all built in-house by model makers. This piece unpacks only Claude Code: it is the one with a public revenue figure and enterprise-share anchors to discuss; the other two have no reliable decomposition this round. Claude Code at $2.5 billion annualized is, structurally, foundry house-brand revenue taken directly out of the mouths of application-layer coding-tool customers. And Anthropic has already once blocked users from wiring their subscriptions into third-party tools — users who, within days, switched to a cheaper open-weight setup and carried on (Lex Fridman Podcast). The vendor-versus-customer conflict of interest is no longer a deduction; it has already happened.

In semiconductor vocabulary: the frontier lab's structural position is not TSMC's pure-play foundry but Samsung's integrated device manufacturer — selling the process (the API) while also selling first-party agent products, competing in the same market as its own application-layer customers. That is a structural correction to the prophecy's second leg. "Money flows to the application layer" fails for every application a lab can reach: the fattest one — coding — has already been eaten by the labs themselves, and by the same logic the next sufficiently fat adjacent application will be eaten too. Draw the boundary of this judgment fairly, though: coding is the one use case that grows right next to a lab's own engineers — they are its first users. Move to enterprise applications that need industry knowledge and distribution, and the labs have no guarantee of success there — Ghodsi's own favorite statistic is that ninety-five percent of enterprise AI pilots fail. So the ceiling covers only the adjacency zone, not the whole application layer; the application layer's real safe zone is wherever the labs' hands don't reach: industry depth, proprietary data, offline workflows. And this in turn explains why the neutral distribution layer — model supermarkets in the cloud, governance gateways like Databricks — has a business at all: in an industry where the foundry will poach its customers' business, neutrality itself becomes the scarce good. Ghodsi's company is selling exactly the TSMC-style promise; he just never puts it that way.

Did the money actually move? Reading the cash registers at the three destinations

The last audit is the prophecy's landing spot: is the money actually flowing to data, applications, and governance? The most direct measurement is to line up the cash registers — with the definitions declared first: everything below is unaudited annualized revenue, the industry's customary run-rate; private-company figures are mostly relayed, so compare magnitudes and slopes only, not precise values.

The mid-2026 readings. Model layer: Anthropic, official figure, $47 billion annualized. It was $9 billion at the end of 2025. Data layer: Databricks, official figure $4.8 billion annualized, relayed figure roughly $6.9 billion as of June 2026, growing about eighty percent year over year. Application layer: Glean, relayed, roughly $300 million, up markedly over the past year and change. Governance and security: no pure LLM-governance company discloses revenue at any magnitude; the nearest proxy, data-security firm Cyera, runs about $150 million annualized (Sacra and other roundups). Now add the slopes: third-party research group Epoch AI measures Anthropic growing at roughly 10x per year since crossing $1 billion annualized (recently moderating to about 7x), against roughly 3.4x for OpenAI (Epoch AI). Next to that line, the data layer's eighty percent and the application layer's doubling-or-tripling are gentle grades. Nor is this the small-base illusion: "fast growth" can usually be dismissed with "small base," but Anthropic's base is now roughly seven times Databricks' — the bigger one is growing faster.

The verdict is blunt: as of mid-2026, "value leaving the model layer" cannot be read on any cash register — the model layer's own register is ringing loudest and accelerating hardest. One structural observation goes on the record beside it: of the prophecy's three destinations, the one delivering most concretely is the data layer — exactly the layer its author stands on; the one delivering least is governance — the layer farthest from him, and the one that most resembles narrative filler. That does not prove the prophecy wrong, but it matches the classic shape of a position-holder's narrative, and the discount applied at the top of this piece gets cashed here.

Of course, revenue is not value. This audit leaves one legitimate retreat: labs are capital-heavy cash furnaces, the upper layers are capital-light businesses, and on profit or free cash flow the scales could swing back: Databricks says its free cash flow is positive, while the labs are still raising at enormous scale. But the retreat is narrowing: if Anthropic's first profitable quarter is confirmed in its coming financial disclosures, the "revenue is vanity, only profit counts" defense falls too. That is the first item on the watch list below.

Where we land

The "LLMs are a commodity" prophecy supplied its own three gauges, and as of July 2026 all three point the opposite way: buyers stack models but don't switch vendors (only 11% switched in the past year); model-maker margins and flagship prices are both rising — on unaudited figures; and enterprise spend share is concentrating toward a single lab, Anthropic (40%, top three at 88%). Meanwhile the prophecy's own skeleton — the TSMC analogy — bites back once checked against real foundry economics: the foundry business's true shape is a monopoly top that raises prices and a commodity base that bleeds, which maps onto the model layer as exactly the two lines this series measured in its previous two columns — hard tasks holding a premium, flagship price tiers moving up: one structure, different cross-sections. Commoditization is real, but it is a line climbing up from the base, not a switch that flips the whole layer. This piece's new judgment is the analogy's break point: TSMC grew the fabless industry by never competing with its customers; model labs run the opposite way — Claude Code at $2.5 billion annualized is foundry house-brand revenue — so "money flows to the apps" has a ceiling: labs will eat the adjacent applications they can reach, and the application layer's safe zone lies only in industry depth and proprietary data. What would prove this wrong: (1) enterprise spend share starts flattening (the next Menlo enterprise report — the mid-year update to the roughly December 2025 edition; an observation window this publication set for itself) — the counter-indicator goes quiet and the commodity thesis recovers; (2) Anthropic's listing documents show audited margins far below the current figures (carried over from the pricing-power column's checks); (3) an actual price cut in the flagship tier, or the August 31 Sonnet promotional price failing to return to list (carried over from the pricing-power column; the date is Anthropic's own announced timing); (4) within a year the labs fail to internalize a second adjacent application beyond coding, while third-party applications hold their ground on the labs' own turf (an observation window this publication set for itself) — the integrated-device-manufacturer reading gets downgraded; (5) once the compute shortage clears, data- and application-layer revenue slopes overtake the labs' (the checkpoint tech analyst Benedict Evans himself supplies from the commodity camp: the shortage is the background variable behind all of today's pricing power).

What this means for you

• For anyone allocating AI exposure: "models commoditize, buy the application layer" is the most popular lazy conclusion of the moment. This audit says it is at minimum early — and that its error is treating a moving line as a switch. The operable version: price the top — the flagship agentic tier — as a pricing-power asset and the base as a commodity, and watch the speed at which the line climbs; the gauges are the five prove-this-wrong conditions above.
• If you buy or negotiate for an enterprise: the data hands you two facts: 81% of your peers run multiple vendors, but only 11% actually switched their primary one — and your vendor knows both numbers, which is why it dares to raise prices. Your leverage is not the threat of "we'll switch" (a 58% migration-failure rate makes that threat incredible); it is tiering workloads by difficulty. Base-tier workloads genuinely can move anytime — open-weight discounts run to 96%. Make that mobility real first, then use it to negotiate the price of the top tier.
• If you build applications or agents: before choosing a vendor, answer one question — does what you're building sit on a lab's adjacency zone? The lesson from coding is that the foundry will launch a house brand, and it will be better than yours, because the model and its orchestration layer are trained together — the previous column's conclusion. The safe zone is industry depth, proprietary data, and the compliance-heavy work labs won't touch. And architect for "the vendor blocks third-party access" as a real, already-observed platform risk.
• If you watch model-lab strategy: the most counterintuitive line in this piece is that real TSMC economics is the gentler version of the labs' bad news — TSMC still earns from its depreciated old lines, while the labs watch open weights take their old tiers for free. That means the labs have less of a fallback than the foundry and a harder obligation to hold the top — which goes a long way toward explaining a capex arms race that otherwise looks irrational: for them, falling off the frontier is not a demotion; it is an exit (the semiconductor-history version, per Stratechery: no company that fell off the leading edge of Moore's Law ever came back). The signal to watch: whether the labs internalize a second adjacent application beyond coding within the year (falsifier 4 above).

Written from the same research and judgments as the Traditional Chinese edition; every claim links to a primary document.

EN English edition|繁 中文版 Traditional Chinese →