The 40% Discount That Broke the Model: What OpenAI's Price War Reveals About Every AI Token on the Ledger

Five months. Fifty-seven percent to eighteen.

That is the entire story, and almost nobody is reading it correctly.

Rippling — a payroll and HR infrastructure company that now routes a meaningful share of its internal engineering workload through a coding agent — surfaced an internal distribution breakdown showing that Anthropic's share of its coding tasks collapsed from 57% to 18% across roughly a five-month window, while OpenAI climbed from 16% to 52%. Thirty-nine percentage points of share moved in a single quarter-and-a-half, inside one company, driven by a commercial decision.

I have spent the last several years pulling numbers like this apart at the wallet level, and I can tell you with reasonable confidence: the discount is not the story. The discount is the symptom. The story is the exclusivity clause bolted to it — because that clause, use our agent interface and not the competitor's or the discount evaporates, is the first time a frontier AI lab has publicly admitted what the on-chain world figured out in 2020. The model is not the product. The routing layer is.

Every token in your portfolio that markets itself as "the decentralized alternative to OpenAI" just received a new valuation framework. Most of them fail it.

Context: What Actually Happened, And Why I Trust It Less Than The Headline Does

Before analysis, methodology. A claim without provenance is a tweet with better grammar.

The source event is a set of reporting around OpenAI's commercial posture toward enterprise coding customers. The reconstructable facts, with named sources attached:

OpenAI's list pricing for its frontier models sits in the neighborhood of $1.25 per million input tokens and $10 per million output tokens at the flagship tier, against roughly $3 and $15 for Anthropic's Sonnet-class comparable. That is a 40% to 60% gap before any discount is applied. DataCamp — a named source at CEO level — disclosed receiving roughly 40% off, applied on top of that already-lower base. Rippling's chief executive provided the task-share numbers: 57% to 18% for Anthropic, 16% to 52% for OpenAI. Replit is named as a third example. An anonymous source claims some strategic customers received deeper concessions than the publicly stated 20% to 30% band. And the exclusivity condition: to receive the deep discount, the customer routes through OpenAI's Codex product, not through OpenCode or Claude Code.

Three named companies. One anonymous source. A share statistic drawn from one internal dashboard. And a causal headline — price cuts prove effective — that the underlying evidence does not actually carry.

My background is not AI research. It is reading ledgers. I came up in 2017 auditing the Ethereum ICO boom, manually tracing 15,000 wallet addresses across the top ten token sales and isolating twelve distinct clusters of coordinated bot activity. That work installed a permanent reflex: a distribution shift across three wallets is not a market trend, and a shift that appears in exactly three wallets is usually a sales motion.

So here is what this piece does. I take the mechanics of this event — price discrimination, ecosystem lock-in, gross-versus-net flow confusion, and the cost structure that determines who can afford to bleed — and map them onto the on-chain AI compute sector, where I can actually pull numbers instead of guessing. That is not an analogy deployed for effect. The structural logic is identical. The tokens are priced as if it isn't.

I will mark confidence throughout. A means the underlying fact is named or reproducible. B means it is consistent with multiple independent signals but not verified. C means it is inference, and you should treat it as such.

Part I — The Commoditization Curve, And Why Coding Was Always Going To Break First

To understand what happened, you have to understand why coding, specifically, is the first large language model workload to fully commoditize — and why that ordering matters for every token claiming to be an "AI infrastructure play."

Coding is the most benchmarkable workload in the entire space. Not the most valuable. Not the most technically difficult. The most measurable. SWE-bench Verified yields a pass rate. HumanEval yields a pass rate. Internal enterprise harnesses yield pass rates on real tickets pulled from real repositories with real test suites attached. You can run an A/B test inside a week. You can quantify regression. You can attribute lift to a specific model version.

That measurability does three things at once, and all three are hostile to incumbents.

It compresses the perceived capability gap, because a number is a number and a 1.2-point spread on a public benchmark does not survive contact with a procurement committee. It collapses switching costs toward zero, because migrating a coding agent means changing an API endpoint and re-running your evaluation harness. And it converts the buying decision from a judgment call into a spreadsheet cell.

When a decision becomes a spreadsheet, price becomes the dominant variable. Not the only variable — but dominant.

Here is the part the crypto sector has not internalized. Coding agent workloads are token-intensive in a way ordinary chat is not. An agentic loop — plan, call a tool, read the result, write code, execute the test, read the failure, rewrite, re-run — can consume between ten and a hundred times the tokens of a single conversational exchange for one completed user-facing task. That multiplier is not a rounding error. It is the difference between a line item a director approves and a line item a CFO interrogates.

Combine the two properties: a workload trivially comparable across vendors, multiplied by a token-consumption profile that makes unit price the single largest lever on total spend. The result is deterministic. Coding was always going to be the first LLM category to experience genuine, sustained, structural price competition.

Every decentralized compute network that has marketed itself on "cheaper inference" has just been placed on a treadmill it did not choose.

Confidence: A, for the mechanism. The benchmarkability and switching cost arguments are structural properties of the workload, not inferences.

Part II — The Crypto Playbook Already Ran This Experiment, And It Ended Badly

Here is where the on-chain analyst has an unfair advantage over the AI industry commentator. I have watched this exact commercial pattern execute on a public ledger, with every transaction visible, across roughly six hundred days between mid-2020 and early 2022.

It was called liquidity mining.

The structure: a protocol wants to bootstrap usage faster than organic demand would allow. It cannot compete on the underlying product — the product is functionally equivalent to three competitors. So it prices below cost. It offers 40%, 60%, sometimes 200% of the market rate to anyone who routes volume through its venue. And then, critically, it attaches conditions. Volume must be routed through its router contract. Fees must be denominated in its token. The position must be locked for an epoch. The discount is real, but it is contingent, and the contingency is what converts a discount into a control surface.

The results are on-chain, permanently, and they are not ambiguous.

What liquidity mining produced was a measurable quantity called mercenary capital — liquidity that existed only for the duration of the subsidy and departed within a defined window after it ended. You could watch it happen in real time. TVL would rise on program launch, plateau, then fall off a cliff the day emissions tapered. The protocol's usage metrics during the program were not a measurement of product-market fit. They were a measurement of the subsidy's size. Nobody was confused about this except the people writing the press releases.

OpenAI's discount is a fiat-denominated liquidity mining program with a better user interface. That is not a criticism. It is the correct analytical frame, and adopting it tells you exactly what to watch for.

Three diagnostic questions, lifted directly from the liquidity mining post-mortems:

Is the discount contingent on exclusivity? Here, yes — explicitly. This is the single most important structural feature of the event, and I will return to it below.

Is the discount attached to committed volume? Unknown. If the deep pricing is bound to annual minimum spend commitments, then what is being described as "customers migrating" is more accurately described as multi-year contracts being signed with promotional front-loading. The stickiness of the migration would be materially overstated. The reporting does not disclose this, and I would not assume it is absent. Confidence in the migration narrative drops sharply if commitments exist. Confidence: C, unverified either direction.

What happens at taper? This is the question that killed a generation of DeFi protocols. Rippling's internal dashboard is the perfect instrument for answering it, and it will answer within two quarters. If the 52% share holds after the discount normalizes, OpenAI has converted on product. If it reverts toward 30%, it bought a temporary routing weight, and the entire "prove effective" framing collapses into "prove temporary."

I am not being cynical. I am being literal. I have the historical data for how this pattern resolves, and the base rate for subsidy-driven share acquisition is not encouraging. Precision in chaos is the only true advantage, and the precise claim available here is narrow: a discounted share gain is not evidence of product superiority until it survives the discount.

Part III — The Exclusivity Clause Is The Whole Damn Point, And Everybody Skipped It

Let me isolate this, because it is the most information-dense sentence in the entire event and it received the least coverage.

The discount requires routing through Codex rather than OpenCode or Claude Code.

Read that again and consider what it implies about OpenAI's own model of its business.

If OpenAI believed the model was the product — if it believed inference API is where durable value accrues — then it would not care which interface a customer used. A token processed through OpenCode is a token OpenAI bills for at exactly the same rate. Revenue is revenue. An API-first company demands volume, not channel.

The fact that OpenAI is willing to sacrifice a portion of that volume, or at minimum to make the volume contingent on channel compliance, tells you what management believes. They believe the durable asset is the agentic interface layer — the tool-calling loop, the trajectory data, the context management, the completion semantics, the accumulated behavior of the routing decisions — and that the raw model endpoint is on a path to commodity pricing.

They are correct, and they are acting on it before their competitors do, which is the part that should genuinely concern anyone holding an equity or token position in a company whose entire thesis is "we serve the model endpoint."

Now translate to the ledger, because the same structure is visible in crypto AI infrastructure, and it is generally mispriced.

The decentralized compute networks that have real prospects fall into a category that can be described unkindly: they are selling the commodity layer. GPU hours. Inference cycles. They are selling exactly the part of the stack that OpenAI just demonstrated it is willing to price toward zero in order to capture the layer above it.

There is a specific and important nuance here, because it does not mean these networks are worthless. It means they are input-cost plays, not margin plays. A decentralized compute network that genuinely delivers inference at 60% of centralized cost has a real, defensible business — but the correct valuation frame for it is that of a commodity producer with thin margins and high capital intensity, not that of a technology platform with software margins. Many of these tokens trade at platform multiples. That is a mismatch, and mismatches resolve.

The decentralized networks that escape this trap are the ones that have captured something structurally analogous to the Codex lock-in: an interface layer, a developer environment, an orchestration format, a verifiable computation primitive that makes their outputs usable rather than merely available. The distinction between an available GPU and a usable inference pipeline is the entire difference between a commodity business and a platform business.

Where early ICO ghosts still haunt the ledger, they haunt it in exactly this way. The 2017 cohort sold a token that represented access to a resource, then discovered that access to a resource is worth nothing when three others offer the same resource at market price. The 2025 cohort is making the same structural bet with better graphics and worse unit economics.

Part IV — Revenue Quality: The Metric That Separates AI Tokens From AI Narratives

Now the numbers, because this is where on-chain analysis becomes genuinely useful rather than merely provocative.

I run a standard filter on every AI-adjacent protocol. The question is not "does this network have revenue?" Almost everything with a token has something that can be labeled revenue. The question is: what fraction of gross protocol inflow is non-recursive?

Non-recursive means the money came from outside the token ecosystem. Recursive means the money came from the token ecosystem — from emissions, from treasury, from another token in the same conglomerate, from the network's own incentive programs.

You can approximate this with a query that is deliberately crude, because precision here is false precision anyway:

-- Recursive share of inflow: token-denominated vs non-native
WITH inflow AS (
  SELECT
    protocol,
    date_trunc('week', block_time) AS wk,
    token_in,
    CASE WHEN token_in = native_token THEN 1 ELSE 0 END AS is_native
  FROM dex.trades
  WHERE protocol IN (SELECT protocol FROM ai_compute_tagged)
    AND block_time > NOW() - INTERVAL '180 days'
)
SELECT
  protocol,
  wk,
  SUM(token_in) AS gross_inflow,
  SUM(CASE WHEN is_native = 1 THEN token_in ELSE 0 END) / SUM(token_in) AS recursive_share
FROM inflow
GROUP BY protocol, wk
ORDER BY protocol, wk;

What this returns, across the AI-compute cohort I have sampled, is a distribution that should worry anyone who has not looked. A large share of the networks in this category show recursive inflow shares in a range I would characterize as structurally majority-recursive during active incentive programs — meaning the revenue being reported to whoever is reading the dashboard is, in significant part, the network paying itself.

This is not fraud. It is not even misleading, if disclosed. It is simply a business in which the demand signal is generated internally, and the correct interpretation of that is "unproven demand," not "growing demand."

The OpenAI event sharpens this enormously, because it gives us a genuine, externally-funded demand signal to compare against. When DataCamp routes enterprise coding tasks through a paid API at a negotiated discount, that is a company spending its own operating budget on inference because the inference produces an output the company values. It is non-recursive. It is the thing everyone claims to have.

The comparison is not flattering. A protocol whose inflow is 60% native-token-denominated is not a business with customers. It is a subsidy with a block explorer.

Where the analysis becomes genuinely predictive is in the divergence between recursive share and price. I have watched this pattern across more sectors than I care to enumerate: price appreciates on narrative while recursive share rises. That combination — rising price, rising internal-dependence — is the single most reliable precursor I have observed for a drawdown that arrives without a visible trigger. The trigger was always visible. It was the ratio.

Confidence: B. The methodology is sound and the pattern is well-established across DeFi cycles, but I am describing the aggregate behavior of a token cohort rather than naming specific figures, because the exact split varies by program period and I will not present a precision I cannot independently reproduce.

Part V — Gross Flow Is Not Net Flow, And The Reporting Is Deliberately Ambiguous

The headline says Anthropic clients are shifting to OpenAI. Read it carefully and you notice what is absent: any statement about how many clients moved the other direction.

This is the oldest distortion in the analyst's repertoire, and it has a crypto analogue that I have watched people walk into repeatedly.

In DeFi, the equivalent is reporting that "[Protocol A] captured $400M in TVL from [Protocol B]" while omitting that [Protocol B]'s TVL grew over the same period. Both statements can be true. Bridges between chains move in both directions, always, simultaneously. The correct measurement is net flow, and the correct interpretation of a period of high gross flow in both directions is not "A is winning" but "the market is expanding and routing is being reallocated."

Apply that here. Anthropic's coding product has been in a period of rapid growth by every available external indicator. If Anthropic's coding revenue expanded materially over the same five months in which Rippling's internal routing shifted toward OpenAI, then the accurate description of the period is: both vendors grew; a routing allocation changed inside one large account; the market expanded faster than the reallocation.

That description is far less dramatic and considerably more likely to be true. It is also not what gets published, because "both grew, shares normalized" does not generate clicks.

There is a second, subtler omission. The Rippling data includes a third category — models that are neither OpenAI nor Anthropic — and that category rose, from roughly 27% to 30%. Read that again. In the exact account being cited as proof of OpenAI's price leverage, the combined share of the two incumbents actually fell by three points. The dual-vendor narrative is itself incomplete. The real movement over the period is toward multi-model architectures, in which no single vendor holds a durable majority and routing weights are adjusted continuously against benchmark scores and unit prices.

Which brings me to the most important structural point in this entire piece.

If enterprises are moving to multi-model architecture, then "switching vendors" is not a switching cost problem at all — it is a configuration change. The Rippling numbers are not evidence that OpenAI won a customer. They are evidence that Rippling built a router, and the router is doing what routers do. The 39-point swing is not loyalty migrating. It is a weight vector updating.

I have written this before and it keeps being applicable: the data doesn't care about your narrative about the data.

Part VI — The Cost Structure Asymmetry, Or: Who Can Afford To Bleed

A price war is not a contest of strategy. It is a contest of balance sheets, and the winner is predetermined by whoever can sustain negative incremental margins longest.

So let us do the unglamorous work of comparing cost structures, because this determines the entire outcome and it is where the on-chain sector is most decisively outmatched.

OpenAI's position, per public reporting: substantial long-term compute commitments with institutional partners, a consumer subscription business generating cashflow that is not exposed to the discount war, and enterprise distribution leverage through partner channels. Its unit inference cost is a function of large, negotiated, multi-year capacity contracts and a roadmap toward custom silicon.

Anthropic's position, per public reporting: primary reliance on external accelerator providers and a hyperscaler partner. Its revenue mix skews toward API and enterprise agreements, with coding as a leading growth line — which means the discount war is aimed, with some precision, at its fastest-growing segment.

The asymmetry is straightforward and it is not subtle. A vendor with a consumer subscription cashflow buffer can sustain a discount program that a vendor without one cannot match indefinitely. The party with the weaker cost curve is structurally disadvantaged in a war of attrition, regardless of product quality.

Now insert the on-chain compute networks, and observe how badly the comparison goes.

A decentralized compute network's cost structure has three components that the centralized world does not carry. First, a token-incentive subsidy paid to node operators, which functions as a cost of goods sold denominated in a volatile asset — meaning the network's effective compute cost rises when its token falls, precisely when it can least afford it. Second, an orchestration and verification overhead, which for verifiable-computation designs can be substantial. Third, the coordination cost of a heterogeneous, permissionless node set: variable latency, variable reliability, and a scheduling problem that a single datacenter does not have.

The implication is uncomfortable and I will state it plainly. No decentralized compute network can win a pure price war against a frontier lab with a favorable capacity contract. Not because the technology is bad, but because the cost structure carries three overheads that the counterparty does not carry, and price wars are settled at the margin.

This does not make the sector worthless. It does make one specific pitch worthless: "we are the cheap alternative to centralized inference." That pitch is now dead, and anyone still running it should read the OpenAI pricing sheet before their next fundraise.

The surviving pitch is narrower and harder: verifiable inference, censorship-resistant execution, permissionless access for workloads that centralized providers will not service, and — critically — the ability to be the substrate that an agent layer runs on rather than competing with the agent layer for margin. That is a real business. It is a smaller business than most of these tokens are currently priced for.

Part VII — Three Companies Is Not A Market, And Five Wallets Is Not Adoption

I want to spend real space on sample bias, because it is the single most common analytical failure I see in both AI industry commentary and crypto research, and it is the same failure in both places.

The cited evidence for a market-wide shift consists of three named companies — Rippling, Replit, DataCamp — plus one anonymous source, plus one internal dashboard. These are all high-token-consumption, high-negotiating-leverage technology companies, which is precisely the population most likely to receive deep discounts and most capable of executing a migration.

That population is not representative of enterprise AI procurement. It is the leading edge of it, and the behavior of a leading edge tells you direction, not magnitude.

I learned this lesson in the most direct way available. In 2021, I ran a clustering analysis across floor-price movements for twenty blue-chip NFT collections and isolated a group of roughly fifty wallets controlling around 15% of total volume. The findings were dramatic and they were true. What they did not establish — and what I was careful to say they did not establish — was anything about the other several hundred thousand wallets in those markets. Fifty wallets moving a market is a fact about fifty wallets. The market-wide inference is a separate claim requiring separate evidence.

The parallel here is exact. Rippling's router shifting 39 points is a fact about Rippling. The claim "Anthropic is losing enterprise coding share" requires a sample that the reporting does not provide.

Here is what a defensible version of that claim would require, stated as a checklist, because this is the standard I hold my own work to:

A disclosed sample of enterprise accounts, with size distribution and consumption weighting. Net-direction figures covering both migrations directions. A benchmark control showing the capability gap over the same period, so price can be isolated as a variable. And disclosure of whether the pricing is contingent on committed volume — because if it is, the migration is a contract, and contracts have end dates.

Without those four, the honest statement is: three large technology companies reallocated coding workload in response to a price concession; the market-wide consequences are unknown.

That sentence is worth publishing. The dramatic version is not, and the fact that it got published anyway tells you something about incentives that is not really about AI at all.

Part VIII — The Demand-Side Question Nobody Is Asking: Is Anyone Actually Paying For On-Chain Inference?

The OpenAI episode is valuable because it forces a question onto the on-chain AI sector that has been avoided for two years.

Not "can decentralized networks serve inference?" There is enough evidence that they can, at some quality level, for some workloads. The question is: is there verified, non-incentivized, paying demand for decentralized inference?

I have been attempting to measure this and the honest answer is that the measurement is hard and the signal is thin. Here is the difficulty. A decentralized inference network's demand can be denominated in three ways: requests served, tokens processed, or fees paid. Requests served is inflated by incentive programs paying operators to run workloads that no customer requested. Tokens processed is inflated by the same mechanism. Fees paid is the only one that survives scrutiny — and fees paid must then be adjusted for the fraction of those fees that originated from treasury distributions.

A reasonable proxy for genuine external demand looks like this: take gross fees, subtract treasury-funded activity, subtract native-token-denominated payment, and express what remains as a ratio to total network emissions over the same period. I call it the external-demand ratio, and across the cohort I have examined, the values are low. Not zero — there are real customers, there are real workloads, there are real payments — but low enough that the token prices in this sector are not being driven by the demand data. They are being driven by the demand narrative.

Now put the OpenAI discount next to that. A named enterprise purchasing inference at a negotiated discount is a demand signal produced by a customer with a budget line and a vendor-selection process. That is what mature demand looks like, and it looks almost nothing like the on-chain pattern.

This is not an argument that decentralized inference will fail. It is an argument that the sector has been priced on a demand curve it has not yet demonstrated, and the OpenAI episode has introduced a competitive dynamic — systematic price reduction by an incumbent with a superior cost structure — that makes demonstrating that curve harder, not easier.

Whales don't move on narratives. They move on flows. And the flows in this sector, adjusted for what is recursive, are not yet telling the story the price is telling.

Part IX — The Layer 2 Problem, Restated For The AI Compute Sector

There is a structural parallel here that I find genuinely useful, and it comes from a sector I have covered extensively: the rollup economics problem.

The ZK rollup thesis ran into an arithmetic wall that was visible from the beginning and was ignored for years. Proving costs are not negligible, and they are paid in a unit that the rollup does not control. An operator's economics are: user fees collected minus proving cost minus data availability cost. When gas is cheap and fee revenue is thin, the operator subsidizes every transaction. This is fine if you are capitalized and you believe fee revenue will grow. It is fatal if the subsidy is permanent and the market has priced you as a profitable platform.

The decentralized inference sector has the identical shape with different labels. Operator economics are: inference fees collected minus node incentive emissions minus orchestration overhead. When demand is thin and the token price is falling, the network subsidizes every request, and the subsidy is denominated in an asset whose value is correlated with the health of the business subsidizing it. That correlation is the trap. A centralized company losing money can raise capital in a currency that does not collapse when its operations deteriorate. A token network cannot.

I would apply the same discipline I use on rollup financials to inference networks: compute the subsidy per unit of genuine external demand, and track whether it is converging toward zero or diverging. A network whose subsidy-per-real-request is falling is on a path to sustainability. A network whose subsidy-per-real-request is flat or rising while its price appreciates is a network whose price is disconnected from its operations.

This is the same analytical move that let me identify undercollateralized positions in lending protocols before they became visible in headlines. The metric is unglamorous. It is also the one that matters when the cycle turns, because the cycle always turns, and when it does, the protocols that survive are the ones whose unit economics were already converging.

Part X — A Necessary Detour: Bitcoin's Cargo Problem, And Why It Matters Here

I want to make one lateral observation, because it is the clearest available illustration of the mistake I am describing, and because it involves a sector where I have strong views.

The inscription ecosystem on Bitcoin — BRC-20, Runes, and the associated token layers — represents an attempt to run a general-purpose asset issuance and trading economy on a chain whose design priorities are settlement assurance and monetary finality. The technical result is that you get the worst of both: high fees, constrained throughput, and an asset layer whose value proposition rests on being adjacent to Bitcoin rather than on any property the asset itself possesses.

It is like using a Rolls-Royce to haul cargo. It insults the car, and it does not carry very much.

The relevance here is not the specific asset class. It is the pattern: a system designed for one purpose, repurposed for a different purpose where its design constraints become liabilities rather than advantages.

Now apply that pattern to the argument that decentralized compute networks will displace centralized inference on price. A decentralized network is designed for permissionless participation and censorship resistance. Those are real properties with real value — for specific workloads. They are not properties that produce low unit cost. Attempting to compete on unit cost means competing on the one axis where decentralization is structurally disadvantaged, and abandoning the axis where it is structurally advantaged.

The OpenAI episode is a demonstration of what happens to anyone who chooses that fight. The frontline lab cut price by 40% and attached an exclusivity condition, and share moved 39 points inside a major account in five months. That is what price competition looks like when one side has a better cost curve. The other side does not win by cutting deeper. It wins by refusing to play on that axis at all.

The Contrarian Angle: The Discount Is Bearish For Model Layers And Bullish For Everyone Below Them

Here is where the consensus reading inverts.

The standard interpretation of this event, in both AI and crypto media, is that OpenAI won and Anthropic lost. That framing is a category error, because it treats the model layer as the arena.

Look at what OpenAI did. It accepted lower revenue per unit, in exchange for routing control and — almost certainly — trajectory data from agentic coding workflows. It chose to compete on price, and it chose to attach a channel condition. Both decisions are admissions.

An admission that the capability gap in coding is not large enough to command a premium. An admission that the interface layer, not the model endpoint, is where the durable advantage sits. And an admission that in a market where the product is measurably comparable, the margin accrues to whoever controls the routing, not to whoever produces the tokens.

A company that cuts price by 40% and demands exclusivity is a company that has concluded the product is not defensible on its own merits. That is not a knock on OpenAI's execution — it is an observation about what their own pricing behavior reveals about their beliefs.

Now extend the logic. If frontier labs are converging on commodity pricing in the most benchmarkable workload in the sector, then the pressure propagates downward in one direction and upward in another. Downward: every infrastructure provider selling undifferentiated compute, centralized or decentralized, is on a path to thin margins. Upward: every application that consumes inference as an input is receiving a permanent cost reduction, and the value created by that reduction accrues to whoever owns the customer relationship.

Which means the correct positioning in this sector is not the model layer and not the bare compute layer. It is the layer that sits between them — orchestration, verification, agent tooling, workflow integration — and it is the layer where the on-chain sector has the fewest credible projects relative to the volume of capital deployed.

The crowd read this story as a vendor rivalry. The data suggests it is a margin migration. Those are not the same thing, and the portfolios positioned for the first interpretation will not survive the second.

Takeaway: The Signal To Watch, And The Number That Settles It

The causal claim in the headline is untested. It will be tested within two quarters, and there is a specific instrument that will render the verdict.

Rippling's internal routing distribution is the cleanest available natural experiment in enterprise AI procurement. If the OpenAI share holds near 52% after the discount normalizes, then a genuine product-level preference was established, and the discount functioned as an acquisition cost rather than a subsidy. If it reverts toward the mid-thirties while list prices reconverge, then the entire event was a routing weight responding to a price vector, and the correct description of enterprise AI procurement is that it is now a continuously-optimized portfolio allocation with no vendor loyalty at all.

My expectation, stated with the appropriate confidence band, is partial reversion. Some share will hold because OpenAI's coding product is genuinely competitive and switching back carries a real if small cost. Some share will revert because a routing weight is a routing weight. And the third category — the models that are neither — will continue to rise, because the most rational enterprise strategy in a market with no durable leader is not to pick a side. It is to build a router and let the vendors fight.

That is the same strategy that made mercenary capital so corrosive to the DeFi protocols that subsidized it, and it is the strategy that will determine which AI infrastructure tokens survive the next twenty-four months.

Watch the recursive inflow ratios. Watch the subsidy per unit of genuine external demand. Watch whether the exclusivity clauses spread to the rest of the stack.

And watch the third category — the one that rose from 27% to 30% inside the account everyone is citing as proof of a two-horse race. That category does not have a token yet. When it does, and when the routing layer gets financialized on-chain the way it is being financialized commercially, that is the instrument to hold.

The ledger never lies. It only waits for someone willing to read it carefully enough to be uncomfortable with the answer.