Google's Quiet AI Win: The Hidden 70x Subsidy in Your $200 Plan

An episode of Dan's AI Intel

Everyone thinks the AI meter is subsidized. It isn't — inference runs profitable. The real giveaway is your flat $200 plan (a 40–70x subsidy), and who survives the price war is decided one layer down, at the silicon: Google owns its chips; its rivals rent NVIDIA's 75% margin.

Published · Updated · By Dan Walter

Executive summary

The most important fact about the price of artificial intelligence is that it lies to you, in the opposite direction from what almost everyone assumes. The folk belief is that the per-token meter an API charges is sold below cost to hook you. It is not: by the admission of the people who set the prices, serving a paying customer is one of the better-margin things in software. The genuinely subsidized product is the one that looks safest — the flat monthly subscription. When the research firm SemiAnalysis ran each consumer tier to its weekly cap and priced the tokens at each provider's own API rates, a maxed $200 plan consumed thirty to seventy times its price in compute. The meter is honest about its margin; the all-you-can-eat plan is the loss leader.

The deeper answer is that this subsidy runs far deeper at some labs than others, and the asymmetry is now decided one layer below the price sheet, at the silicon. Two products at the same shelf price can be opposite businesses — one selling above its compute cost with room to attack, the other under water and praying the music plays long enough to refinance. The cause is a line item most people never see: NVIDIA earns roughly a 75% gross margin on the servers nearly every lab rents, a tax baked into every token they make. Google escaped it by building its own chips, which is why it can fight a price war and stay above water — and why its fiercest model rival is paying it tens of billions to rent those same chips. Who owns the factory, more than who has the smartest model, is becoming the thing that decides who survives.

The longest-horizon conclusion reconciles two facts that sound contradictory and are both true: the price of a fixed unit of intelligence is collapsing — about 10x a year by Sam Altman's own reckoning — while the sellers post some of the largest losses in corporate history. They reconcile through the Jevons paradox: cheaper tokens, used so much more that the bill climbs anyway. Such subsidies get repriced rather than removed, in a knowable order, and the first moves are already on the board. But the popular forecast has the direction wrong. The metered API is already at or above true cost, and at the frontier its list price is rising, not falling — the newest flagships are priced up by the value they deliver even as the commodity tier races toward zero. The only price clearly moving toward economic reality is the one on your monthly plan. The question was never whether the gap closes; it is who is standing on the right side of one dollar per compute dollar, and on the right side of the silicon, when it does.

Why a price sheet decides who wins the AI race

A fight over token prices is really a fight about the future of the field. The defining contest of this moment is not which lab has the smartest model on a given Tuesday — capability leads change hands every few months — but which labs can keep paying for compute long enough to be standing when the dust settles. The build-out is consuming sums with no precedent: IDC put AI-infrastructure spending at $318 billion in 2025, more than double the prior year, and forecasts $487 billion in 2026, on a path past a trillion dollars a year by 2029. Revenue is real but far smaller, and every major lab is burning cash. In that race the price of a token is the metabolism: it sets how fast each runner burns the fuel it raised, and therefore who runs out first.

So the sticker price is the wrong thing to read: it hides the two things that decide the outcome — how much each provider bleeds to show it to you, and how much it truly costs to produce. What follows tracks the money through both lies the price sheet tells (that cheaper-per-token means cheaper, and that the meter is where the subsidy lives), then asks who subsidizes more and why, where prices are heading, and how far today's price sits from a true cost that refuses to hold still. The answers reframe the popular story of "subsidized AI" almost entirely.

The subsidy is real — but it is hiding in your subscription, not the meter

Exhibit — Run to its cap, a $200 AI subscription burns 40–70× its price in compute. The meter is the better-margin product; the flat all-you-can-eat plan is the loss leader — break-even hits at ~5.7% (Pro) and ~10% (Max) of the weekly cap. Source: SemiAnalysis tokenomics teardown, mid-2026 (power user maxing the weekly cap, tokens priced at each provider’s API rates).

Start with the true part of the folk wisdom: AI is being sold below cost — the mistake is the location. The deep subsidy is not in the metered API but in the flat subscription, and the cleanest measurement is a brute-force experiment. In mid-2026 SemiAnalysis bought every consumer tier from OpenAI and Anthropic and ran long-horizon coding and agentic workflows until they exhausted the weekly limits — finding the upper bound a flat price has to survive — then priced the tokens consumed at each provider's published API rates. That is the source of the $14,000 and $8,000 figures: not the average user, but the power user maxing the plan. A $200 ChatGPT Pro seat can consume roughly seventy times its price in compute; a $200 Claude Max seat, roughly forty times.

The break-even math is unforgiving. By SemiAnalysis's reckoning, top-tier ChatGPT Pro hits zero gross margin at about 5.7% of its weekly cap and Claude Max at around 10% — so a developer running real agentic work even a few hours a week has already turned that seat unprofitable. Altman said the quiet part out loud in January 2025: "insane thing: we are currently losing money on openai pro subscriptions! people use it much more than we expected." He had, he added, chosen the price himself and "thought we would make some money."

Why did the giveaway get so deep? Because the subscriptions were priced for a 2023 chatbot — a human typing a question and reading a paragraph back — and then the product mutated. The chatbot became an agent. An April 2026 analysis by Bai and colleagues, across eight frontier models on a standard software-engineering benchmark, found agentic work consumes on the order of 1,000x more tokens than ordinary code chat — the input tokens dominating because the agent reads the task, calls a tool, re-reads the whole growing context, and snowballs, with identical tasks varying as much as 30x in tokens from run to run. The flat price stayed put for two years while the unit of work moved from one reply to one workflow. That mismatch, not generosity, is the engine of the subsidy.

The cheaper model that costs more

Exhibit — On identical work, OpenAI’s “cheap” o3-mini cost 50× more than GPT-4o-mini. The reasoning model’s hidden “thinking” tokens — billed at the output rate — drove a 50× bill on the same problems. Measure cost per completed task, not per token. Source: NPPC reasoning benchmark, arXiv:2504.11239 (peer-reviewed), Table 4 — same suite, every model ~31M prompt tokens.

The price sheet tells a second lie, and it inverts the ranking of the models. Line up two by list price per million tokens, make them finish the same job, and the cheaper-per-token one can hand you the larger bill — because the cost of intelligence is no longer the cost of a token. Reasoning models, the ones that "think" before answering, generate a large volume of hidden intermediate tokens, billed at the output rate, that never appear in the reply. Epoch AI found the distortion so severe it excluded reasoning models from its per-token analysis: they "tend to generate a much larger number of tokens," making it "misleading to compare reasoning models to other models on price per token, at a given performance level."

Measured directly, the spread is startling. On a fixed reasoning benchmark — the NPPC suite, first posted in April 2025 and since peer-reviewed — every model processed about 31 million prompt tokens on the same problems. OpenAI's cheap GPT-4o-mini generated 9.4 million completion tokens and cost $10.31 to clear the suite; OpenAI's o3-mini generated 110.9 million completion tokens, nearly twelve times as many, and at a higher output rate cost $522.48 — more than a 50x spread on identical work, driven by how much each model spent thinking. DeepSeek-R1 logged 95.9 million tokens on that suite; Anthropic's Claude 3.7 Sonnet, 11.2 million. The honest unit is not dollars per token but dollars per completed task, and the two have come fully apart.

The practical version is everywhere. Anthropic's docs tell developers thinking tokens bill at the output rate and to start small and scale up; practitioners in production budget on the order of three-to-five times the visible answer for thinking on heavy tasks. And more thinking does not reliably buy more correctness: in one data-extraction teardown, the team at Refuel found that switching from a plain model to a reasoning model lengthened the output 6.7x — and the cost with it — for an average accuracy gain of just 4.9%. The cheapest model per token is routinely the most expensive per result. It is also why the newest frontier models can raise their per-token price and still claim to be cheaper: OpenAI frames its latest flagship as roughly a 20% rise in the cost to complete a task even though the per-token rate doubled, because it finishes in fewer tokens — a claim you can only test on your own workload. "What is the true cost" has no single answer.

The asymmetry no price sheet shows — and why it lives in the silicon

Exhibit — Anthropic earns $1.70 per compute dollar; OpenAI earns $0.68 — under water before payroll. The lab the public treats as the default leader has the weaker unit economics: one can cut prices to attack, the other can only use price as a shield. Source: Forward estimates from leaked cost data, compiled by analyst Shwetank Kumar, April 2026 (projections, not audited actuals); direction corroborated by OpenAI’s FT-verified 2025 financials.

The subsidy is not the same depth at every lab, and the difference is the best single predictor of who survives a price war. The cleanest measure is revenue earned per dollar of compute, training and inference together. On forward estimates built from leaked cost data — compiled by analyst Shwetank Kumar in April 2026, not audited actuals — Anthropic earns roughly $1.70 per compute dollar while OpenAI earns about $0.68: thirty-two cents under water on the hardware before an engineer is paid. The direction is corroborated by OpenAI's own 2025 financials, leaked in June 2026 and verified by the Financial Times — $13.1 billion of revenue against roughly $34 billion of total cost, an operating loss near $21 billion — while Anthropic's inference margins were rising — so the lab the public treats as the default leader has the weaker unit economics, and a lab clearing well above a dollar per compute dollar can cut prices to attack while one under water can only use price as a shield. In mid-2026 that turned concrete: it was reported on 11 June that OpenAI was weighing significant API price cuts specifically to counter Anthropic — in the same stretch that Anthropic's roughly $965 billion valuation eclipsed OpenAI's roughly $852 billion for the first time, and Claude Code crossed a billion dollars of annualized revenue within months of launch.

Two players define the edges. xAI is the clearest below-cost case: it prices its newest Grok model's output at a fraction of rivals' while reportedly burning on the order of $28 million a day, recovering only about fourteen cents of fully-loaded cost per dollar earned. At the other edge, DeepSeek's February 2025 "Open Source Week" disclosed a 24-hour snapshot of its V3 and R1 inference: about $87,000 of GPU time on H800s rented at $2 a GPU-hour, serving 608 billion input and 168 billion output tokens, which — billed at R1's published rate — implied a theoretical 545% profit margin. DeepSeek stamped it "theoretical" (most traffic runs cheaper models, off-peak discounts apply, web and app are free) and SemiAnalysis argued it serves near cost to take share. Even discounted to nothing the signal stands: an efficient stack can be wildly profitable at prices roughly 20x to 50x below the US incumbents — R1 lists near $0.55 per million input tokens against $15 for OpenAI's comparable o1. And the executives have settled the "is inference subsidized" question on the record. Altman, at a San Francisco press dinner in August 2025: "We're profitable on inference. If we didn't pay for training, we'd be a very profitable company." Amodei, on the Cheeky Pint podcast that month, walked through a toy example — a model that cost $100 million to train and earned $200 million — and on the Dwarkesh podcast in February 2026 pinned a number on it: "the inference has some gross margin that's more than 50%." (The caveat: OpenAI's internal "compute margin" of about 70% strips out training and free-tier inference; its actual adjusted gross margin landed nearer 33% in 2025.) What is sold below cost is not the query on the meter — it is the free tier and the flat subscription wrapped around it.

So why do the labs sit so far apart? Mostly not talent or model quality — it is who owns the factory. Nearly every frontier lab rents compute from someone who bought NVIDIA chips, and NVIDIA earns roughly a 75% gross margin on those AI servers — a tax on every token an NVIDIA-renting lab makes, paid before power, staff, or training. OpenAI pays it through Microsoft's Azure; xAI pays it directly on its own NVIDIA fleet. Google is the conspicuous exception, and the answer to the question most coverage skips: where does Gemini sit? Google designs its own TPUs, buys the bare silicon from Broadcom at low margin, and integrates its own boards and racks. SemiAnalysis's teardown of Google's latest generation found a full TPU pod costs Google about 44% less to own than a comparable NVIDIA system and rents to outside customers for roughly 30% less per hour — widening, performance-adjusted at realistic utilization, toward 62% lower cost per unit of useful compute. It shows at the meter: Gemini's flagship lists around $2 per million input tokens and $12 per million output, against roughly $2.50 and $15 for OpenAI's mainstream GPT-5-class flagship (its newest, GPT-5.5, lists higher still at $5 and $30) and $5 and $25 for Anthropic's most capable model, with Gemini's lightweight tiers reaching toward a dime per million input — a worked mixed task runs about 28% cheaper than the OpenAI equivalent and 37% cheaper than Anthropic's.

There is a subtler reason too — the under-reported part: Google does not need the tokens to pay for themselves. It keeps the assistant ad-free and absorbs the cost out of a Search business that prints money, so its consumer subscription can be a loss leader by choice while its API runs near or above cost. Alphabet's accounting even obscures it: SemiAnalysis notes Gemini's API revenue flows into Google Cloud at flattering margins while more than $10 billion of Gemini's run-rate costs are parked under company-level AI research that jumped from about $3.0 billion to $5.4 billion in a year — so there is no clean published "Gemini margin," yet every signal points one way: lowest cost to serve, deepest pockets, least dependence on the subsidy. The most telling proof is who is buying. Anthropic — Google's most direct model rival — has committed to renting up to a million of Google's TPUs, reported worth on the order of $40 billion, with a multi-gigawatt follow-on. When your fiercest competitor pays you tens of billions to escape NVIDIA's margin, the market has rendered its verdict.

Exhibit — Google’s own TPUs undercut the NVIDIA systems every rival rents — by owning the factory. NVIDIA books ~75% gross margin on the servers nearly every lab rents — a tax on every token. Google escapes it; tellingly, its fiercest model rival Anthropic is renting up to a million Google TPUs to escape it too. Source: SemiAnalysis TPUv7 teardown (Google internal TCO vs GB200) — ~44% lower to own, up to ~30% lower external rent; NVIDIA ~75% server gross margin (corroborated by NVIDIA filings, ~75% FY26).

The same logic runs to the edges: Amazon runs the Google playbook with its own Trainium chips (which Anthropic's models also use), while Meta gives Llama away as open weights, dragging the price floor toward the cost of self-hosting — the cheapest hosted Llama runs fifteen to forty cents per million tokens — and monetizing the ads its models improve rather than the tokens. The durable escape from a token-price war you are losing money in is to own the factory, like Google and Amazon, or to not need the revenue, like Meta. The rest are racing a clock.

How prices fall while the makers bleed

Exhibit — Per-token prices halved, yet enterprise AI spend tripled — the Jevons paradox. Token prices fell by half over the same year (Bain, Dec 2024–Dec 2025) while volume grew ~450% — cheaper tokens did not shrink the bill; they unlocked agents, and agents eat. Source: Menlo Ventures, “2025 State of Generative AI in the Enterprise” ($11.5B 2024 → $37B 2025, 3.2×); price/volume from Bain, “How Token Economics Will Change Opex”.

Two facts about AI economics feel mutually exclusive. First: the cost of a fixed unit of intelligence is falling faster than almost anything in computing history. Epoch AI measured the price to hit GPT-4-level performance on PhD-grade science questions dropping about 40x a year, with rates across benchmarks from 9x to 900x and a median accelerating toward 200x once 2024 data is included. (A calibration: benchmark overfitting has grown since 2024, and tighter token-level work puts the forward rate nearer 3x to 5x a year — so treat 40x as the optimistic end.) Altman frames it as a law in his February 2025 essay: "the cost to use a given level of AI falls about 10x every 12 months," and the price from GPT-4 to GPT-4o "dropped about 150x."

Second: the sellers are losing money at a scale with little precedent for software — OpenAI's leaked 2025 statements show that roughly $21 billion operating loss on $13.1 billion of revenue, with projected cumulative losses well past $100 billion before any profitable year, and Anthropic, near a $30 billion revenue run-rate by April 2026, cut its gross-margin projection toward 40% after inference costs on rented cloud ran about 23% over plan.

Both are true, and the reconciliation is the most important idea here: the Jevons paradox — make a resource cheaper and people use so much more of it that total spending rises. Apollo's chief economist Torsten Slok put it plainly in June 2026: as tokens get cheaper, companies "don't spend less but instead run more AI agents, automate more workflows and generate more code, pushing aggregate expenditure higher even as the unit cost of intelligence collapses." Bain found token prices roughly halved between late 2024 and late 2025 while the volume consumed grew about 450%. Enterprise generative-AI spending tracked by Menlo Ventures went from $11.5 billion in 2024 to roughly $37 billion in 2025 even as per-token prices fell more than 90%. Cheaper tokens did not shrink the bill; they unlocked agents, and agents eat.

What "true cost" actually is — and why it keeps moving

The tempting move is to compare today's subsidized price against today's "true cost" and read off a clean multiple. It fails, because true cost is not a fixed floor — it is itself in free-fall. Start at the bottom of the stack: the same NVIDIA H100 rents for well under $2 an hour on a spot marketplace, around $2.50 to $3.50 on a mid-tier neocloud like Lambda or RunPod, and up to roughly $10 an hour on the priciest hyperscaler configurations — the identical chip, a fivefold range, depending only on who you rent from. And the rental rate is only the numerator.

The denominator is where true cost is won or lost: how many useful tokens you get from that GPU-hour — and utilization is frequently dreadful. A widely cited figure puts average enterprise GPU utilization near 5%, at which point roughly 95 cents of every silicon dollar is, as one analysis put it, essentially a donation to the cloud provider. Since modern AI clusters are something like 60–70% hardware by cost, every idle point is real capital depreciating for nothing. Pack requests well, with continuous batching, and utilization can jump from the low twenties into the seventies, cutting cost per token three- to four-fold. The "true cost" of a token is therefore not a property of the model at all, but of how full the data center is when it runs.

And the floor keeps dropping. NVIDIA's Blackwell generation cut cost per token roughly 10x versus the prior Hopper chips, and its Blackwell Ultra variant pushes that toward 25–35x; at the start of 2026 NVIDIA claimed its next platform, Vera Rubin, would cut inference cost a further 10x versus Blackwell when it ships in the back half of the year. Benchmarked serve costs for those who run efficiently are already startlingly low: SemiAnalysis measured DeepSeek-R1 on a current-generation system at about 54 cents per million output tokens at interactive speed. So "how much more would it cost to pay the true price right now" has no single answer — a range from near-zero to the 40x-to-70x the subscriptions reveal, depending on whose data center, at what utilization, on which silicon. At a heavy power-user's consumption the gap is enormous; at an efficiently run metered workload it can be near zero or favorable, because the API is the part the leading labs already price above cost. The subsidy is concentrated where it is most visible — the flat consumer plan — and thinnest on the metered bill.

The repricing has already started — and the frontier is going up

Predicting the exact date the music stops is a fool's errand, but the order of events is legible, and the first moves are on the board. Expect three repricings — and one runs opposite to the popular story.

First, flat consumer and developer plans migrate toward usage — no longer a forecast. Anthropic imposed weekly rate limits on Claude in August 2025, moved enterprise renewals to usage-based contracts that November, and split a flat per-employee platform fee from metered token billing in February 2026. GitHub Copilot ended its flat premium-request model on 1 June 2026 for metered "AI Credits," to a revolt of 958 downvotes against 24 upvotes on its own announcement. Cursor switched its $20 plan to an API-rate credit pool in June 2025 and watched bills explode, drawing an apology from its chief executive — "we didn't handle this pricing rollout well and we're sorry" — but no reversal. The sharpest case came from Anthropic itself: on 13 May 2026 it announced that from 15 June, programmatic Agent SDK usage — a subsidy it pegged at 15x to 30x relative to API pricing — would move to a capped credit pool at list rates; then, facing developer backlash, a looming IPO, and a class-action alleging its Max plan delivered far less than advertised, it paused the change on the day it was due, telling users "nothing changes for now." As of late June the pause still holds. The giveaway does not vanish overnight; it gets a meter, a lower cap, or a higher price for the heavy tail.

Second — and here the conventional wisdom is simply wrong — API list prices are not all grinding down toward true cost. The market has bifurcated. Budget and mid-tier models do race toward zero, dragged by open-weight Llama and efficient challengers. But the newest frontier models show real pricing power: OpenAI's latest flagship launched in April 2026 at double its predecessor's per-token rate, input and output, on the argument that it is smart and token-efficient enough to be cheaper per task despite the higher per-token price. "API rates falling to true cost" describes the bottom of the market, not the top.

Third — the hinge the forecast turns on — "true cost" itself keeps falling, so convergence is a moving target, not a collision. The subsidy closes not because prices leap up to a fixed cost, but because cost falls toward prices while prices fall toward cost — and the lab that runs out of runway before they meet loses, however cheap its sticker looked. Altman has voiced both halves weeks apart: "intelligence too cheap to meter" at a Federal Reserve conference in July 2025, and a future where "people buy it from us on a meter" at a BlackRock summit in March 2026. Both can be true. The price per unit falls toward zero; the meter stays, because the bill is the volume.

What this means, and what to watch

The single fact that should change your read of this space is the provider asymmetry, because a price sheet can never show it. Two AI products at similar prices are not two companies in the same position: one may be selling above cost and using price to attack, the other below cost and praying the music plays long enough to refinance — and the deciding difference is increasingly one layer down, in whether the seller owns its silicon or rents NVIDIA's margin. The metered API is not the loss leader; the real subsidy is the flat subscription, the 40x-to-70x giveaway to its heaviest users. The repricing has begun in a knowable order: consumer plans get a meter, the commodity tier grinds toward marginal cost while the frontier is priced up by value, and true cost keeps falling underneath both.

So the discipline is short. Measure models by cost-per-completed-task on your own workload, not the per-token sticker — the verbose "cheap" model is often the expensive one, and the pricey frontier model is sometimes the bargain. Treat any flat-rate plan you depend on as temporary; assume the meter is coming, and only the date is open. And when you bet on a provider for the long haul, weigh the balance sheet and the supply chain as heavily as the benchmark: ask who owns the chips. The other half of this story is the cost floor falling out from under everyone at once — the same deflation that makes small, efficient models good enough to run the commodity tier on hardware you already own, which we traced in the episode on Apple's contrarian, efficiency-first AI bet. The survivors will be decided not by who shows the lowest sticker, but by who is already above one dollar per compute dollar — and on the right side of the silicon — when the subsidies switch off. Watch the unit economics, not the price sheet.

Sources

Transcript

Sam: Okay, here's the thing everyone gets backwards. You've heard that AI is sold below cost — that the meter, the thing that charges you per word, is the giveaway designed to hook you.

Alex: It's the opposite. The meter is one of the better-margin products in software. The thing that's bleeding money — forty to seventy times its price — is the flat two-hundred-dollar plan you thought was the safe choice.

Sam: And which company survives the price war that's coming isn't decided by who has the smartest model. It's decided one layer down, in the silicon — and almost nobody's looking there.

Alex: Welcome back to Dan's AI Intel — the show where we try to make sense of the fastest, most consequential shift any of us is likely to live through. I'm Alex, here with Sam.

Sam: And today we are following the money. Specifically: the price of artificial intelligence, and the fact that it lies to you — in the exact opposite direction from what almost everyone believes.

Alex: This show exists because the AI revolution is moving faster than any one person can track. The idea is to take the questions you can't stop thinking about and actually chase them down — turn the noise into something you understand and can carry around. And this one is a perfect example, because the conventional wisdom here isn't just incomplete. It's inverted.

Sam: So let me set up where we're going, because I genuinely did not see the ending coming. We start with a real number that broke my brain — a two-hundred-dollar subscription that can burn through fourteen thousand dollars of compute. The trigger is simple: everyone says AI is subsidized. And it is. But the moment you ask where the subsidy actually lives, the whole story flips.

Alex: And the question that opens up underneath it is the one that really matters: when the cheap money runs out — and it will — who's left standing? Because two companies can sell you the same thing at the same price and be opposite businesses. One's making money and circling for the kill. The other is praying the music doesn't stop.

Sam: We're going to travel across all of them — OpenAI, Anthropic, Google, DeepSeek, the whole board — and we're going to find that the thing that decides the winner isn't the model at all. It's a tax most people have never heard of, and exactly one company figured out how to dodge it. There's a move almost every other player is making right now that looks smart and is, I think, exactly backwards.

Alex: If you've been enjoying the show, do follow us on Spotify or Apple Podcasts, so the next one finds you. Okay. Let's start with the lie itself.

Sam: Right — so the headline claim. When people say "AI is subsidized," what's the picture in their head? Mine was: every time I type a question, the company is quietly eating part of the cost to keep me hooked. The meter is the loss leader.

Alex: That's the folk belief, and it's wrong. Here's what the people who actually set the prices say. The CEO of OpenAI, Sam Altman — at a press dinner in San Francisco — said, and I'm close to quoting, "We're profitable on inference. If we didn't pay for training, we'd be a very profitable company."

Sam: Inference being… the actual answering of my question. The running of the model.

Alex: Right. Training is building the model once, which costs a fortune. Inference is the part that happens every time you hit enter. And he's saying that part — the meter — makes money. Anthropic's CEO, Dario Amodei, put a number on it. He said the inference has a gross margin "more than fifty percent."

Sam: So the per-question business is more profitable than half. That's… not a loss leader. That's a good business.

Alex: It's a genuinely good business. And now you can feel the puzzle. If the meter makes money, where on earth is all the subsidy everyone keeps talking about? Because the subsidy is real. The companies are losing staggering amounts of money. So if it's not the meter…

Sam: …it's hiding somewhere else. Okay, I want to know where. But hold on — why does this even matter? Like, beyond me feeling smart at a dinner party. Why does the price of a token decide anything big?

Alex: Because the price of a token is the metabolism of this entire race. Think about the scale for a second. One research firm, IDC, pegged AI infrastructure spending at three hundred and eighteen billion dollars in a single year — more than double the year before. They forecast four hundred and eighty-seven billion the next year, on a path past a trillion dollars a year by the end of the decade.

Sam: A trillion. With a T. And — let me make sure I've got the gap right. If they're spending three hundred billion building this out, and the revenue is, what, a fraction of that…

Alex: A fraction. The money going in dwarfs the money coming back, for now.

Sam: Then this is a race against the bank balance, not against each other. Whoever's burning fastest hits zero first, no matter how good their model is.

Alex: That's exactly it, and it's why the price of a token matters so much. The price sets how fast each company burns the money it raised — which sets who runs out first. The smartest model on a given Tuesday doesn't matter if you can't pay the power bill on Wednesday.

Sam: Okay, that reframes it completely. So the sticker price isn't really about what I pay. It's about who's quietly dying.

Alex: That's the whole show, honestly. The sticker price hides the two things that decide the outcome — how much each company bleeds to put that model in front of you, and what it actually costs them to make. Everything from here is us pulling those two numbers out from behind the price tag.

Sam: So you said the subsidy is real but it's in the wrong place. Where is it, actually?

Alex: It's in the flat subscription. The all-you-can-eat monthly plan. And there's this beautiful, brute-force experiment that proves it. A research firm called SemiAnalysis went and bought every consumer tier from OpenAI and Anthropic — the twenty-dollar plans, the two-hundred-dollar plans — and then they just… hammered them.

Sam: Hammered them how?

Alex: They ran real work. Long coding sessions, agent workflows, the heavy stuff — and they kept going until they hit the weekly usage limits. They wanted to find the ceiling. The most a flat plan would ever have to survive. And then they took all the tokens they'd burned and priced them at the company's own published meter rates.

Sam: So they're asking: if you actually used this thing as hard as it lets you, what would it have cost on the meter?

Alex: Exactly. And here's the answer. A two-hundred-dollar ChatGPT Pro plan, run to the max, consumed about seventy times its price in compute. Roughly fourteen thousand dollars.

Sam: Seventy times. So I pay two hundred, and at full tilt I'm using fourteen grand of their compute.

Alex: At full tilt, yes. The Claude Max plan, same price, came out around forty times — about eight thousand dollars. And the break-even is brutal. ChatGPT Pro hits zero margin — the point where they start losing money on you — at about five-point-seven percent of the weekly cap.

Sam: Wait. Five percent? So if I use even a twentieth of what they let me use, I've already tipped them into a loss?

Alex: A developer doing real agent work even a few hours a week is already unprofitable to them. Claude Max gives you a bit more runway — around ten percent of the cap — but it's the same shape. The flat plan is the loss leader. And Altman, again, just said it out loud. Back in January twenty-twenty-five he posted — and this one I do want to quote, because it's so human — "insane thing: we are currently losing money on openai pro subscriptions! people use it much more than we expected." He set the price himself. He said he thought they'd make some money.

Sam: [laughs] I love that. The most powerful man in AI, surprised by his own pricing page. But — okay, why did it get this bad? Like, these are not stupid people. How do you misprice something by seventy times?

Alex: This is the part that actually clicks it into place. They priced the subscription for a 2023 chatbot. Picture what that was: a person types a question, reads a paragraph back, types another. That's a human-paced amount of work. Cheap to serve.

Sam: Right, I'm a slow, squishy human. I can only type so fast.

Alex: And then the product mutated underneath the price. The chatbot became an agent. Instead of you asking one question, the model now goes off and does a whole task — it reads the job, calls a tool, re-reads everything it's accumulated, calls another tool, on and on.

Sam: So it's not one paragraph anymore. It's the model talking to itself for an hour.

Alex: And here's the number that shows how violent that shift is. A study across eight frontier models — by a researcher named Bai and colleagues — found that agent work uses on the order of a thousand times more tokens than ordinary chat. A thousand.

Sam: A thousand times. Not a thousand percent — a thousand times.

Alex: A thousand times. And mostly on the input side, which surprised me. The reading, not the writing.

Sam: Why input, though? I'd have guessed it generates a ton of output.

Alex: Because of how an agent works step to step. Every time it takes an action — calls a tool, gets a result — it has to re-read the entire conversation so far to decide what to do next. So step one reads a little. Step ten re-reads everything from steps one through nine. Step fifty re-reads the whole pile again.

Sam: Oh — so the context it has to re-read keeps growing, and it re-reads the whole thing every single step. It's paying to re-read its own history over and over.

Alex: It snowballs. And it's so dependent on the messy specifics that the same task, run twice, can vary thirty-fold in tokens — thirty times more expensive on an unlucky run than a lucky one, for the identical request. So the flat price sat completely still for two years while the unit of work went from "one reply" to "one entire workflow." That gap — that's the engine of the subsidy. It's not generosity. It's a price tag that forgot to move while the product ran away from it.

Sam: Okay, that's the first big reframe and I already feel like I've been lied to twice. The meter makes money, and the safe-looking flat plan is the bonfire. What's the second lie?

Alex: The second lie is in the price sheet itself, and it scrambles the ranking of the models. Here's the setup: line up two models by their list price per million tokens. One's clearly cheaper per token. Make them both do the exact same job. The cheaper one can hand you the bigger bill.

Sam: That… should not be possible. Cheaper per token, more expensive overall?

Alex: It's possible because the cost of intelligence is no longer the cost of a token. It's the cost of how many tokens the model burns to think. And the reasoning models — the ones that "think" before they answer — generate this huge volume of hidden intermediate tokens. Their scratch work. You never see it, but you pay for it, at the output rate.

Sam: So there's a visible answer, and then there's this invisible iceberg of thinking underneath that's also on my bill.

Alex: That's exactly the right image. And it's so distorting that one research group, Epoch AI, threw reasoning models out of their price comparison entirely. They said these models "generate a much larger number of tokens," so comparing them to other models on price-per-token is just "misleading."

Sam: They couldn't even rank them fairly. Okay, give me the real number. How big is the iceberg?

Alex: There's a benchmark — a fixed set of reasoning problems — where every model got fed about the same thirty-one million prompt tokens. Same questions for everyone. OpenAI has a cheap little model, GPT-4o-mini. It generated nine-point-four million tokens of completion and cleared the whole thing for ten dollars and thirty-one cents.

Sam: Ten bucks for the whole suite. Okay, that's the baseline — hold that number.

Alex: Then they ran OpenAI's o3-mini — a reasoning model. It generated a hundred and eleven million completion tokens. Nearly twelve times as many. And at a higher rate. Final cost: five hundred and twenty-two dollars.

Sam: From ten dollars to five hundred. On the same questions.

Alex: A fifty-fold spread. Same problems, same answers required — the entire difference is how much each model sat there thinking. And just to show it's not one company: DeepSeek's reasoning model logged about ninety-six million tokens on that suite. Anthropic's Claude 3.7 Sonnet, about eleven million. Wildly different thinking habits, wildly different bills, for identical work.

Sam: So "dollars per token" is basically a lie. The honest number is dollars per finished task.

Alex: Dollars per completed task. And the two have come completely apart.

Sam: Okay but here's my cynical question. All that extra thinking — it's worth it, right? You're paying five hundred instead of ten because the answer's way better?

Alex: That is exactly the right question, and the answer is: often, no. There's a teardown from a company called Refuel, doing data extraction. They switched from a plain model to a reasoning model. The output got six-point-seven times longer — and the cost with it. The accuracy improvement?

Sam: …go on.

Alex: Four-point-nine percent.

Sam: Four point nine. So I pay nearly seven times more, and I get a five percent better answer. That's a terrible trade.

Alex: For that task, a terrible trade.

Sam: And I bet that's the trap most companies are walking straight into. They see "reasoning model," they assume smarter is better, they switch everything over — and they don't find out until the invoice that they septupled their bill for a rounding error of accuracy.

Alex: That's precisely the trap, and it's why people who run this in production have learned to budget three to five times the visible answer just for the hidden thinking on heavy jobs — and to be ruthless about which tasks actually need a reasoning model at all. Most don't. It flips the whole intuition: the cheapest model per token is routinely the most expensive per result. And — this is the kicker — it's also why a brand-new frontier model can raise its per-token price and still honestly claim to be cheaper.

Sam: How does raising the price make it cheaper?

Alex: Because if the new model is smart enough to finish in far fewer tokens, the total cost of the task can drop even though each token costs more. OpenAI framed its latest flagship as roughly a twenty percent rise in the cost to complete a task — even though the per-token rate doubled — because it gets there in fewer tokens.

Sam: So the only honest way to know what a model costs… is to run your own actual work through it.

Alex: Your own workload. There is no universal answer to "what does this cost." It depends entirely on what you ask it to do. Which, by the way, is two lies down — and we still haven't found why some of these companies are bleeding so much worse than others.

Sam: So before we go on — let me make sure I've got the two lies straight, because they're both kind of beautiful. Lie one: the meter looks like the giveaway, but it actually makes money — the real subsidy is your flat plan, forty to seventy times underwater. Lie two: the cheap-per-token model is often the expensive-per-job model, once you count the hidden thinking.

Alex: That's it exactly.

Sam: And the thing that's been nagging me through both is the one you keep dangling — why are some of these companies bleeding so much worse than others? They're all selling roughly the same thing.

Alex: Right, and this is the heart of it. The subsidy is not the same depth at every lab — and the difference is the single best predictor of who survives. Let me give you the cleanest way to measure it. Forget the model for a second. Just ask: for every dollar of compute a company spends, how much revenue does it earn back?

Sam: Revenue per compute dollar. So above a dollar, you're making money on the hardware. Below a dollar, you're losing it.

Alex: Above one, the hardware pays for itself before you've even paid an engineer. Below one, you're underwater on day one. So here are two companies people think of as neck and neck. On forward estimates built from leaked cost data, Anthropic earns about a dollar seventy per compute dollar.

Sam: Okay, a dollar seventy. Comfortably above water.

Alex: And OpenAI earns about sixty-eight cents.

Sam: Sixty-eight. So OpenAI is — what — thirty-two cents underwater on every compute dollar, before anyone's even paid?

Alex: Before a single salary. And this gets corroborated by OpenAI's own financials, which leaked and were verified by the Financial Times: thirteen-point-one billion dollars of revenue against around thirty-four billion of cost. An operating loss near twenty-one billion dollars.

Sam: Twenty-one billion. In a year. That is — hang on, that's one of the biggest losses I've ever heard of for a software company.

Alex: It's close to historic. And here's the part that breaks the public story: the company everyone treats as the default king of AI has the weaker unit economics of the two. Meanwhile Anthropic, sitting above that one-dollar line, can actually use price as a weapon — it can cut prices to attack. OpenAI, underwater, can only use price as a shield.

Sam: So when I see them both drop prices, those are not the same move. One's swinging. One's flinching.

Alex: Completely different moves. And it got concrete in mid-2026. It was reported that OpenAI was weighing major API price cuts specifically to fight off Anthropic — in the same stretch that Anthropic's valuation, around nine hundred sixty-five billion, passed OpenAI's, around eight hundred fifty-two billion, for the first time. And Anthropic's coding product, Claude Code, crossed a billion dollars of annualized revenue within months of launching.

Sam: And just so I'm calibrated — are those two roughly the middle of the pack? Or is there worse?

Alex: Oh, there are wilder edges in both directions, and they make the point even sharper. The clearest below-cost case is xAI — Elon Musk's lab, the one that makes Grok. It prices its newest Grok model's output at a fraction of what rivals charge, while reportedly burning something like twenty-eight million dollars a day. By one estimate it recovers about fourteen cents of its fully-loaded cost for every dollar it earns.

Sam: Fourteen cents on the dollar. So it's basically setting money on fire to buy market share.

Alex: That's the aggressive end. And then the opposite edge is genuinely jaw-dropping — DeepSeek, the Chinese lab. During something they called "Open Source Week" they published a real twenty-four-hour snapshot of their inference. About eighty-seven thousand dollars of GPU time — on older H800 chips rented at two dollars an hour — and on that hardware they served six hundred and eight billion input tokens and a hundred and sixty-eight billion output tokens.

Sam: Those are enormous numbers for eighty-seven grand.

Alex: So enormous that when you bill them at DeepSeek's published rate, it implies a theoretical profit margin of five hundred and forty-five percent.

Sam: Five hundred and forty-five percent. That can't be a real margin.

Alex: And to their credit, DeepSeek stamped it "theoretical" themselves — most traffic runs cheaper models, there are off-peak discounts, the web and app are free. SemiAnalysis thinks they actually serve close to cost to grab share. But even if you discount it to nothing, the signal survives: an efficient stack can be wildly profitable at prices something like twenty to fifty times below the American incumbents. DeepSeek's reasoning model lists around fifty-five cents per million input tokens. The comparable OpenAI model — o1 — was around fifteen dollars.

Sam: Fifty-five cents versus fifteen dollars. For the same kind of model.

Alex: And the executives have basically settled the "is inference subsidized" question on the record. We heard Altman — profitable on inference. Amodei, on a podcast, walked through a toy version: imagine a model that cost a hundred million dollars to train, and then earns two hundred million serving customers. The serving part is healthily profitable; it's the training that's the giant hole. Now — one honest caveat, because I don't want to oversell it. OpenAI quotes an internal "compute margin" of about seventy percent, but that number strips out training and free-tier inference. Its actual adjusted gross margin landed nearer thirty-three percent.

Sam: Wait, so which is it — seventy or thirty-three? That's a huge gap.

Alex: Both, depending on what you count. Seventy percent is just the raw cost of serving a paying query. Thirty-three is after you fold in the free users and the cost of building the thing. But here's what doesn't change either way: what's sold below cost is not the paying query on the meter. It's the free tier and the flat subscription wrapped around it.

Sam: Okay, the through-line holds. So back to the why. Why are these companies so far apart in the first place? Is Anthropic just… smarter?

Alex: That's the thing — it's mostly not about talent, and it's not about who has the better model. It comes down to who owns the factory. And to see it you have to go one layer below everything we've been talking about. Below the price sheet, below the model, down to the actual chips.

Sam: The silicon.

Alex: The silicon. So: nearly every frontier lab rents its compute from someone who bought chips from NVIDIA. And NVIDIA earns roughly a seventy-five percent gross margin on those AI servers.

Sam: Seventy-five percent. So three-quarters of what these labs pay for their hardware is just… NVIDIA's profit.

Alex: It's a tax. A tax baked into every single token an NVIDIA-renting lab produces — paid before power, before staff, before training. OpenAI pays it through Microsoft's Azure cloud. xAI pays it directly, on its own NVIDIA fleet.

Sam: Hang on — so even OpenAI, the biggest name in the field, doesn't escape it? It's renting from Microsoft, who bought from NVIDIA, who takes the seventy-five percent?

Alex: Every layer takes its cut before OpenAI sells you a single token. That's the tax everyone's paying.

Sam: Everyone… except. You've got an "except" face.

Alex: Google. Google is the conspicuous exception, and it's the answer to the question most of the coverage just skips, which is — where does Gemini actually sit in all this? Here's what Google did. It designs its own chips. They're called TPUs. It buys the bare silicon from Broadcom at a low margin, and then builds its own boards, its own racks, the whole system itself.

Sam: So Google just… cut NVIDIA out of the loop. It refused to pay the tax.

Alex: It built its own factory. And the numbers SemiAnalysis got from tearing down Google's latest generation are stark. A full Google chip pod costs Google about forty-four percent less to own than a comparable NVIDIA system. It rents to outside customers for about thirty percent less per hour. And once you adjust for performance at realistic usage, the advantage stretches toward sixty-two percent lower cost for a unit of actual useful compute.

Sam: Sixty-two percent cheaper to do the same work. That's not an edge, that's a different sport.

Alex: And you can see it land right on the meter. Google's flagship Gemini lists around two dollars per million input tokens, twelve for output. OpenAI's mainstream flagship is around two-fifty and fifteen — its very newest, GPT-5.5, is higher still, five and thirty. Anthropic's most capable model, five and twenty-five. So a real mixed task on Gemini runs about twenty-eight percent cheaper than the OpenAI equivalent, thirty-seven percent cheaper than Anthropic.

Sam: So the company that owns its factory has the lowest prices. That's not a coincidence, that's cause and effect.

Alex: And there's a subtler reason on top, the under-reported part — Google doesn't even need the tokens to pay for themselves. It runs the assistant ad-free and just absorbs the cost out of its Search business, which prints money. So Google can choose to make its consumer plan a loss leader, while its API still runs near or above cost.

Sam: So everyone else is subsidizing because they're trapped, and Google's subsidizing because it feels like it.

Alex: One is desperation, one is a choice. And honestly the accounting is so favorable to Google it's almost hidden — more than ten billion dollars of Gemini's running costs are tucked under a company-wide AI research line that jumped from about three billion to five-point-four billion in a year. So there's no clean "Gemini margin" you can point to. But every signal points the same way: lowest cost to serve, deepest pockets, least dependent on the subsidy of anyone in the field.

Sam: Okay, is there a single fact that just nails it? Because this is a lot of estimates.

Alex: The most telling proof is who's buying. Anthropic — Google's most direct rival on models, the company making Claude — has committed to renting up to a million of Google's chips. Reported to be worth on the order of forty billion dollars. With a multi-gigawatt follow-on after that.

Sam: Wait. Anthropic — Google's competitor — is paying Google tens of billions of dollars to use Google's chips?

Alex: To escape NVIDIA's margin. When your fiercest competitor would rather pay you forty billion dollars than keep renting NVIDIA, the market has rendered its verdict. Owning the factory wins.

Sam: That is the whole story in one transaction — your biggest rival paying you forty billion dollars to escape the trap you're not in. So is Google the only one who figured this out, or are there others?

Alex: No — and this is the pattern. Amazon runs the exact same playbook with its own chips, called Trainium — which, funnily enough, Anthropic's models also run on. And then Meta plays a different escape: it gives its models, Llama, away as open weights. Free. Which drags the entire price floor down toward just the cost of running it yourself — the cheapest hosted Llama is fifteen to forty cents per million tokens.

Sam: Why would Meta give it away for free? That seems insane next to everyone else burning billions to sell it.

Alex: Because Meta doesn't make money on the tokens. It makes money on the ads its models make better. So it would rather everyone uses a free Meta model and salts the earth for the companies trying to charge. So you end up with three durable escape hatches from a price war you're losing money in. Own the factory — Google, Amazon. Or don't need the revenue at all — Meta. And everybody else…

Sam: …is racing a clock.

Alex: Is racing a clock. That's the asymmetry no price sheet can ever show you. Two products, same shelf, same price — one's a healthy business circling for the kill, the other's underwater praying for a refinance.

Sam: Alright, you've set up something that I think is going to do my head in, so let me say it back. We've got companies losing twenty-one billion dollars a year. And yet you keep telling me the price of AI is collapsing. Those two things should not be able to coexist. Prices crashing AND record losses?

Alex: They sound mutually exclusive, and they're both completely true at the same time. Let me give you both halves, and then the thing that reconciles them, because the reconciliation is maybe the single most important idea in the whole story.

Sam: Go.

Alex: Half one: the cost of a fixed amount of intelligence is falling faster than almost anything in the history of computing. Epoch AI measured the price to hit a certain level of performance — answering PhD-grade science questions — and found it dropping something like forty times a year.

Sam: Forty times a year. Not forty percent. Forty times.

Alex: At the optimistic end, yes. The honest caveat is that some of that is models getting tuned to the tests, and tighter measurements put the real rate nearer three to five times a year. But even three to five times a year is staggering. Altman frames it as a kind of law — he wrote that "the cost to use a given level of AI falls about ten times every twelve months." He pointed out the price from one of their models to the next dropped about a hundred and fifty times.

Sam: Okay so prices in free-fall. That's half one. Half two is the losses.

Alex: Half two: the sellers are losing money at a scale with almost no precedent in software. We said OpenAI — twenty-one billion in operating losses. Projected cumulative losses well past a hundred billion before they ever have a profitable year. And Anthropic, despite the better economics, near a thirty-billion-dollar revenue run-rate, had to cut its own margin forecast toward forty percent because its inference costs on rented cloud ran about twenty-three percent over plan.

Sam: So how? How is the price crashing and the losses ballooning at the same time? Somebody explain this to me like I'm — okay, like I'm smart but baffled.

Alex: It's a hundred-and-fifty-year-old idea called the Jevons paradox. Here's the cleanest version. You make a resource cheaper. People don't use the same amount and pocket the savings. They use so much more of it that total spending actually goes up.

Sam: Give me the analogy, because my brain wants one.

Alex: Okay. Imagine the price of electricity drops by ninety percent. Do you spend less on power? For about a week. Then you buy the bigger TV, you get air conditioning, you leave more lights on, you buy an electric car. Cheaper electricity didn't shrink your bill — it invited a whole new pile of things to plug in, and your bill goes up.

Sam: Oh. So cheaper tokens don't save anyone money. Cheaper tokens just mean everyone runs way more AI.

Alex: They unlock the agents. The economist Torsten Slok at Apollo put it plainly: as tokens get cheaper, companies "don't spend less, they run more AI agents, automate more workflows, generate more code" — and total spending climbs even as the unit cost collapses. And the numbers are wild. Bain found token prices roughly halved in a year, while the volume consumed grew about four hundred and fifty percent.

Sam: Prices cut in half, usage up four hundred and fifty percent. So of course the total bill explodes. And — wait, this connects back to that thousand-times number, doesn't it? The agent that uses a thousand times more tokens than a chat?

Alex: It's the same force, yes. Say more — where are you going with it?

Sam: Well, if making tokens cheaper is what made agents affordable to run in the first place, then the price cut didn't just get absorbed by more usage. It manufactured the more usage. It created the demand that ate it. The cheaper it got, the more reasons appeared to spend.

Alex: That's the paradox in one move, and it's a better way to say it than I had. It's not that demand happened to grow. The price drop unlocked an entirely new way to use the thing — the always-on agent — that simply wasn't viable at the old price. You didn't run a thousand-token agent when tokens were expensive. You do now. So the savings don't reach your wallet; they get reinvested instantly into doing vastly more.

Sam: So cheaper AI doesn't mean a cheaper AI bill. Almost the reverse.

Alex: Almost the reverse. Enterprise spending on this stuff went from about eleven and a half billion to roughly thirty-seven billion in a single year — even as per-token prices fell more than ninety percent. That's the shape of every Jevons story: the unit gets cheaper, the appetite grows faster, the total climbs. Price per unit in free-fall, total spend through the roof.

Sam: Okay. So the obvious move now — the thing I'd do — is just take today's subsidized price, line it up against the "true cost," and read off how big the gap is. Right? That tells me how much trouble everyone's in.

Alex: That's the move everyone wants to make, and it fails. Because "true cost" isn't a fixed floor you can measure the gap to. True cost is itself in free-fall. It's a moving target. Let me build it from the bottom.

Sam: From the chip up.

Alex: From the chip up. Start with renting one NVIDIA chip — the H100, the workhorse. The same chip rents for under two dollars an hour on a spot marketplace, around two-fifty to three-fifty on a mid-tier provider like Lambda or RunPod, and up to roughly ten dollars an hour on the priciest hyperscaler setups.

Sam: Hold on, the identical chip is two dollars in one place and ten in another? A five-times range for the same hardware?

Alex: Same silicon. Five-fold range. Depends entirely on who you rent it from. And that's just the top of the fraction. The rental rate is the numerator. The real action is the denominator: how many useful tokens you actually squeeze out of that chip-hour.

Sam: And let me guess — that number's all over the place too.

Alex: It's often catastrophic. A widely cited figure puts average enterprise chip usage near five percent.

Sam: Five percent utilization. So ninety-five percent of the time, the thing they're paying for is just… sitting there. Idling.

Alex: One analysis called it "essentially a donation to the cloud provider." Ninety-five cents of every dollar you spend on silicon, gone, for nothing. And remember, a modern AI cluster is something like sixty to seventy percent hardware by cost — so every idle moment is real capital depreciating in the rack.

Sam: And I assume the fix is just… keep the chip busy.

Alex: That's the whole game. There's a technique — continuous batching — where instead of running one request at a time, you pack many through the chip together, so it's never sitting idle waiting. Do that well and utilization jumps from the low twenties into the seventies. That alone cuts cost per token three to four-fold.

Sam: So two companies could rent the exact same chip, and the one who packs it well pays a quarter of what the sloppy one pays — for the identical hardware.

Alex: For the identical hardware. The skill is invisible and it's enormous.

Sam: So the "true cost" of a token isn't really about the model at all. It's about how full the building is when the model runs.

Alex: That's the punchline. It's a property of the data center, not the model. And then the floor drops again, because the chips keep getting better. NVIDIA's Blackwell generation cut cost per token roughly ten times versus the chips before it. The souped-up version pushes that toward twenty-five, thirty-five times. And at the start of 2026 NVIDIA claimed its next platform — Vera Rubin — would cut inference cost another ten times on top of Blackwell when it ships later in the year.

Sam: So every layer I look at, the cost is sliding out from under me. The rental rate, the utilization, the chip generation.

Alex: All of it, at once. And for the ones who run efficiently, the serve costs are already startlingly low — SemiAnalysis measured DeepSeek's reasoning model on a current system at about fifty-four cents per million output tokens, at full interactive speed.

Sam: Fifty-four cents. So "how much more would it cost to pay the true price" — there just isn't one answer.

Alex: There's a range, and it's enormous. For a heavy power user maxing a flat plan, the gap is huge — that forty-to-seventy-times we started with. But for an efficiently-run metered workload, the gap can be near zero, or even favorable — because the API is the part the leading labs already price above cost. The subsidy is concentrated exactly where it's most visible: your flat consumer plan. And it's thinnest on the metered bill. The whole thing inverts the folk wisdom one more time.

Sam: So the question everyone actually wants answered: when does the music stop? When do the subsidies end and we all pay full freight?

Alex: Predicting the exact date is a fool's errand. But the order of events is completely legible — and the first moves are already on the board. There are three repricings coming. And one of them runs in the opposite direction to what everyone expects.

Sam: Okay, give me all three.

Alex: One: the flat consumer and developer plans migrate toward usage-based pricing. This isn't a forecast anymore — it's happening. Anthropic put weekly rate limits on Claude. GitHub Copilot ended its flat plan and moved to metered "AI credits" — to a revolt, by the way, nine hundred fifty-eight downvotes against twenty-four upvotes on its own announcement.

Sam: [laughs] People were thrilled.

Alex: Furious. A coding tool called Cursor switched its twenty-dollar plan to a metered credit pool, watched people's bills explode, and the CEO had to publicly apologize — "we didn't handle this pricing rollout well and we're sorry" — but he didn't reverse it. And the sharpest case is Anthropic again: it announced that programmatic usage of its Agent SDK — a subsidy it openly pegged at fifteen to thirty times relative to its own API pricing — would move to a capped credit pool at full rates.

Sam: Fifteen to thirty times. So they admitted that one out loud.

Alex: They named the number. And then — facing backlash, a looming IPO, and a class-action lawsuit claiming the Max plan delivered far less than advertised — they paused the change on the very day it was supposed to take effect. Told users "nothing changes for now." As of late June, that pause still holds.

Sam: So the giveaway doesn't vanish overnight. It gets a meter, or a lower cap, or a price hike for the heaviest users. And the backlash every single time tells you something, right? People didn't just buy a tool — they built their whole workflow on top of a price that was always quietly temporary.

Alex: That's the human cost underneath all the economics. The flat plan felt like a promise. And what these companies are discovering is that taking the meter away is easy and putting it back is a betrayal — even when the math was never going to hold.

Sam: It's the seventy-times subsidy from the start of the episode, coming due. The bill was always there. It just had someone else's name on it for a while.

Alex: Beautifully put. And whose name it lands on next is exactly the fight. Now — repricing two. And here the conventional wisdom is just wrong. People assume all API prices are grinding down toward true cost. They're not. The market has split in two. The budget and mid-tier models do race toward zero — dragged down by free open-weight models like Meta's Llama. But the newest frontier models have real pricing power. OpenAI's latest flagship launched at double its predecessor's per-token rate. Input and output.

Sam: Double. So at the very top, prices are going up, not down.

Alex: Up. On the argument we talked about earlier — that it's smart and efficient enough to be cheaper per task even at a higher per-token price. So "API prices falling to true cost" describes the bottom of the market. At the top, the frontier is being priced up by how much value it delivers.

Sam: Okay and the third?

Alex: The third is the hinge the whole forecast turns on. "True cost" itself keeps falling. So the gap doesn't close because prices jump up to meet a fixed cost. It closes because cost falls toward prices while prices fall toward cost — and the company that runs out of money before the two lines meet loses, no matter how cheap its sticker looked. Altman has said both halves of this weeks apart. "Intelligence too cheap to meter," at one event. And then "people buy it from us on a meter," at another.

Sam: Those completely contradict each other. Too cheap to meter, and also, here's your meter.

Alex: And yet they're both true at once — that's the whole trick of this.

Sam: Oh, I think I see it. It's the difference between the price of one token and the size of the bill. One token gets so cheap it's practically free — too cheap to bother metering. But you're running agents that chew through billions of them. So they absolutely will meter you — not because each token is expensive, but because the pile is enormous.

Alex: That's exactly the resolution. The per-unit price falls toward zero; the meter stays, because the bill is the volume. It's the electricity story again — each kilowatt-hour is cheap, but you still get an electricity bill, because you use so many. Intelligence is becoming a utility. And nobody gives away a utility for free forever.

Sam: So "too cheap to meter" and "we'll meter you" aren't a contradiction. They're a sequence. Cheap per unit, metered in bulk. Okay. I came in thinking AI was subsidized in this vague, simple way, and I'm leaving with a completely rearranged head. Can we land the plane? What are the two or three things to actually walk away with?

Alex: Yeah. Three things. First — and this is the one that should change how you read this entire space — two AI products at the same price are not two companies in the same position. One can be selling above cost and using price to attack; the other underwater, praying the music plays long enough to refinance. And the deciding difference is one layer down: does the company own its silicon, like Google and Amazon, or does it rent NVIDIA's seventy-five percent margin like nearly everyone else?

Sam: Ask who owns the chips. That's the one a price sheet can never show you.

Alex: Second: the subsidy is not where you thought. The meter makes money. The flat subscription is the loss leader — that forty-to-seventy-times giveaway to its heaviest users. So treat any flat-rate plan you depend on as temporary. Assume the meter is coming. Only the date is open.

Sam: And measure the model by the job, not the token.

Alex: That's the third. The cost of intelligence is dollars per completed task, not dollars per token — so the verbose "cheap" model is often the expensive one, and the pricey frontier model is sometimes the bargain. The only way to know is to run your own work through it. And underneath all of it, the cost floor is falling out from under everyone at once — the same deflation that's making small, efficient models good enough to run on hardware you already own, which is a thread we pulled in our episode on Apple's contrarian, efficiency-first bet.

Sam: So the survivors won't be whoever shows the lowest sticker today. It really comes back to those three escape hatches, doesn't it — own the factory like Google and Amazon, give it away like Meta, or get above that one-dollar line on your own. And everyone outside those three is just hoping the deflation reaches them before the money runs out.

Alex: That's the whole map. And it's why "who has the best model this quarter" turns out to be almost the wrong question. The model lead changes hands every few months — but the chip strategy is a decade-long bet you can't reverse in a hurry. The benchmark is the headline; the supply chain is the story.

Sam: So I should basically stop reading the leaderboards and start reading the balance sheets.

Alex: Honestly, yeah. The survivors will be whoever's already above one dollar per compute dollar — and on the right side of the silicon — when the subsidies switch off. So watch the unit economics. Not the price sheet.

Sam: That's a genuinely different way to look at the whole industry. I hope you all come away from this seeing a bit more clearly where this is heading — because it's a fast-moving, genuinely complex picture, with a brutally short shelf life on what we think we know, and that's exactly what makes it worth following closely.

Alex: And a quick note, for full transparency: this show is AI-generated. Dan builds a custom stack of AI tools to chase the questions he can't stop thinking about — it started out made with NotebookLM, and it's now produced with his own engine — mainly so he can learn this stuff himself, and he publishes it for anyone who'd like to follow along.

Sam: If you've enjoyed this one, please do give us a follow — it genuinely helps the next person find the show.

Alex: And one more thing, from Dan directly: he'd genuinely love to hear from you — what you'd want more of, what you'd change, what landed and what didn't. You can write to him at podcast@connectiveshift.com. He's looking forward to it. Thanks for listening — we'll see you next time.