LLM Models Will Be Worth Zero in the Future

The architecture was a gift from Google. The data was taken from a commons. The best free model on earth is now four months behind the best paid one, and closing. Models are not the asset. They were never going to be.

I. The Future

Large language models will be worth zero.

Not useless. Worth zero. The distinction carries the entire argument, so let me put it first.

A technology’s utility and its economic rent are unrelated quantities. Electricity is indispensable and generates almost no rent for the firms that produce it. TCP/IP carries every dollar of digital commerce and earns nothing for anyone. Linux runs the world’s servers and is free. The transistor built the modern economy, and the company that invented it licensed the patent for $25,000 and left the business.

Language models are heading to the same destination, faster, because they possess every property that destroys pricing power and almost none that preserves it. The architecture was invented at Google and published for free. The training corpus was scraped from a commons open to all comers. The compute is bought at list price from a vendor whose commercial interest is to sell to as many buyers as possible. The researchers are free agents who change employers every eighteen months. The finished artifact can be copied, distilled, or approximated within a quarter, and the best free version on earth currently sits fourth on the industry’s headline intelligence leaderboard, ahead of every model you have to pay for.

Meanwhile the firms that productionized this gift are marked at a trillion dollars.

That is the position. Models are the least defensible layer in the entire stack. They are being priced as though they were the most defensible. The correction will be violent.

II. Google invented it, published it, and lost every author

On 12 June 2017, eight Google researchers posted an eight-page paper called “Attention Is All You Need.” They insisted on equal credit and randomized the author order. The title was a joke about a Beatles song.

By 2026 the paper had been cited more than 250,000 times, which places it among the ten most-cited scientific papers of the century for a work under a decade old. GPT, Claude, Gemini, Llama, DeepSeek, Qwen, Kimi, GLM, Mistral: every one is a transformer derivative. The paper is to this industry what the transistor paper was to electronics.

Google put it on arXiv. For free. With the reference implementation. Under no restriction whatsoever.

It was not an aberration either. It was institutional policy sustained across a decade. Google published word2vec in 2013 with open code, seq2seq in 2014, and open-sourced TensorFlow in 2015. The transformer came in 2017, alongside the sparsely-gated mixture-of-experts paper. BERT followed in 2018 with released weights, T5 and the text-to-text framing in 2019, the Switch Transformer in 2021. DeepMind published the Chinchilla compute-optimal scaling laws in 2022.

Every architectural idea that matters inside a 2026 frontier model traces back to a paper somebody gave away. Mixture-of-experts routing, the technique that makes DeepSeek V4-Pro’s 1.6 trillion parameters economic by activating only 49 billion of them per token, is Google, 2017. The scaling laws that tell you how to allocate a training budget are DeepMind, 2022, published in full.

Then the second act. All eight authors left.

Aidan Gomez, a twenty-year-old intern when the paper was written, founded Cohere. Illia Polosukhin founded NEAR. Ashish Vaswani and Niki Parmar founded Adept and then Essential AI. Jakob Uszkoreit founded Inceptive. Llion Jones founded Sakana. Łukasz Kaiser went to OpenAI. Noam Shazeer founded Character.AI, was reacquired by Google in a 2024 licensing arrangement reported at roughly $2.7 billion, which is the price of buying back one author of a paper Google had given away seven years earlier, was installed as an engineering vice president on Gemini, and then left again, for OpenAI.

The people who wrote the paper that armed Google’s competitors walked out, one at a time, and became those competitors.

The instinct is to call this a blunder. It was not. Publishing was rational at the time. In 2017 Google could not have monetized the transformer internally at anything close to its eventual value; the compute did not exist at the necessary scale, the product surface was not obvious, and the norms of machine learning research made publication the price of keeping the researchers who did the work. Withholding it would have cost Google the people. Publishing it cost Google the invention.

That is not incompetence. It is a structural property of an industry whose core inputs are ideas, and ideas are non-rival, unenforceable in practice, and carried out of the building in the heads of employees who are free to resign. Every lab faces the same trap today. It is why there is no such thing as a proprietary architecture, and why there never will be.

The foundation was free by the deliberate choice of the firm that built it, and that firm kept none of it.

III. What a real moat looks like, and the test the labs fail

Someone will point out that free foundations are the normal case for enormous companies. TCP/IP was DARPA-funded and given away. CERN released HTTP and HTML into the public domain in 1993. Google itself is worth roughly three trillion dollars, built entirely on protocols it neither invented nor owns.

True, and it sets exactly the right test.

Google’s moat was never HTTP. It was four things that came after the free protocol, none of which appeared in any paper.

A crawl and index at capital-prohibitive scale, whose cost grew with the web itself, so the gap widened over time rather than narrowing. A ranking loop with proprietary feedback, where every query and click improved the next result and no rival could observe the data. Scale in usage produced scale in quality, which produced scale in usage. A two-sided auction with genuine network effects: advertisers went where the users were, users went where the results were, and the auction grew more efficient with every additional participant. And distribution defaults, purchased with the profits from the first three.

Compounding, proprietary, unobservable, unbuyable. Google’s lead over a well-funded rival in 2005 was not four months. It was structural, and it was widening.

Now run the same test on a frontier lab. What has been built on top of the free transformer that a competitor cannot replicate?

The honest inventory of genuine invention is short, and all of it is product. Instruction-tuning at scale. Reinforcement learning from human feedback in production. The chat interface itself, which was a first-rate product insight. Later, agentic coding harnesses. Real work, and the people who did it deserve the credit.

Every one of those became standard practice across the industry within about a year.

RLHF was the crown jewel of 2022. By late 2025, reasoning-capable post-training had become a routine stage in open-weight releases rather than a differentiator, partly because the distillation pipelines were themselves published. There is no ranking loop here that rivals cannot see. There is no two-sided market. There is no index whose cost grows faster than a challenger’s budget. There is a model, and four months later there is a free one that does the same thing.

Google built something unbuyable on top of a free protocol. The labs built something replicable on top of a free architecture. That is the difference, and it is the whole difference.

IV. Every input is free, purchasable, or mobile

Decompose the production function and ask of each input whether it is scarce, whether it is excludable, and whether it can simply be bought.

The architecture is free. Google, 2017. There is no differentiation available, because everyone is building the same thing.

The training algorithms are published. Scaling laws, mixture-of-experts, RLHF, reinforcement learning on verifiable rewards, all in the literature within months of deployment. Diffusion is measured in quarters.

The training data came from a common pool. Every lab drank from the same river, which means no differential access and therefore no differential rent. More on this in section VIII, because it is the piece most often argued backwards.

The compute is purchasable at list price from a supplier whose commercial interest is to have many buyers and no monopsonist. Nvidia’s explicit strategy is to proliferate model-makers. Anyone with capital can buy the same chips.

The talent is mobile. All eight transformer authors left Google. Shazeer left, was rebought for $2.7 billion, and left again. Nobody owns these people. Compensation escalates and the knowledge walks out the door.

And the weights themselves are replicable within roughly four months, or distillable at a fraction of a percent of the original training cost. That is a wasting asset with a one-quarter exclusivity window.

Not one input to a frontier model is simultaneously scarce, excludable, and non-purchasable. Not one.

Compare TSMC, whose yield learning is scarce, excludable, unbuyable, and accumulated across thirty years of defect analysis that cannot be extracted by querying an API. Or ASML, whose EUV supply chain took two decades to assemble and cannot be reconstituted with money alone. Those are moats. A model trained on public data using a published architecture on rented silicon by researchers who will be working somewhere else next year is not a moat. It is a manufactured good with an unusually short shelf life.

An industry in which every factor of production is free, buyable, or free to leave will not sustain rents at the layer where those factors combine. It will sustain them upstream, where physics binds, and downstream, where switching costs and proprietary outcomes accumulate. The middle gets crushed. This is not a forecast. It is the standard result, and it has already started.

V. The proof is already shipping, and it is free

The commoditization thesis stopped being a prediction somewhere in 2026. Look at what you can download today, for nothing, under an MIT licence.

Z.ai shipped GLM-5.2 on 13 June 2026. Three days later Artificial Analysis scored it fourth overall on the industry’s headline intelligence index, ahead of every model you cannot download. Open weights, MIT licence, one-million-token context. GLM-5.3 followed on 14 August, improving across the board.

DeepSeek shipped V4 in April 2026. V4-Pro runs 1.6 trillion total parameters with 49 billion active, a one-million-token window, MIT weights. V4-Flash launched at $0.14 input and $0.28 output per million tokens, roughly 36 times cheaper than GPT-5.5 on input and more than a hundred times cheaper on output. DeepSeek’s own technical report positions V4-Pro-Max as trailing the state of the art by three to six months, which is unusually candid for a vendor and roughly matches the independent measurements.

Moonshot released Kimi K3 on 27 July 2026, a trillion-parameter agentic model with published weights, currently the ceiling of the open tier.

All three beat GPT-5.5 on at least one flagship benchmark. All three ship weights you can host yourself. All three run at five to thirty times less per token than the closed frontier.

And then Qwen, which is the number that should end the argument.

Alibaba’s Qwen family crossed three billion downloads in six months. Google’s models managed 418 million across all of 2026. Meta’s managed 227 million. Qwen has open-sourced more than 460 models, and the ecosystem has produced over 300,000 derivatives. Hugging Face’s state-of-open-models report, published 14 August 2026, called Qwen one of the largest foundations of the open AI ecosystem and observed that it has become part of the default workflow for developers deciding what to fine-tune and deploy. Chinese open models overtook American ones as the most-downloaded on Hugging Face in the summer of 2025 and have not looked back since.

Read that again in economic terms. The world’s most widely adopted foundation model has a price of zero and a market share measured in billions of downloads. Not a cheap competitor. Not a budget alternative. The default.

The independent measurements agree on what remains. Epoch AI’s Capabilities Index puts the average lag between the best open-weight model and the closed frontier at four months over January to May 2026, about 8 ECI points, roughly the distance from GPT-5 to GPT-5.5. The UK AI Safety Institute, evaluating cyber capability separately, found leading open models matching closed models released four to seven months earlier, down from six to ten months through most of 2025.

Four months. That is the entire asset. That is what a trillion dollars is buying: a rolling four-month head start on something the Chinese labs give away, and a head start that has been shrinking rather than growing.

One delusion deserves naming here. Western labs spent 2023 and 2024 arguing that capital intensity would protect them, that training runs would become so expensive only three or four firms could afford to play. It was a coherent argument and it was wrong. DeepSeek V4 was trained on Huawei Ascend silicon at a reported fraction of the cost of the Western flagships it competes with. Capital intensity did not build a wall. It built a very expensive way of arriving at the same place a few months early.

VI. The price of intelligence is collapsing toward its marginal cost

This is what commoditization looks like on a chart.

When GPT-3 first shipped it was the only model on earth scoring around 42 on MMLU, at roughly $60 per million tokens. By March 2026 several models cleared that same bar at six cents per million or less. A thousandfold decline in four years, the phenomenon a16z clocked at roughly ten times per year and named LLMflation.

At the frontier the same thing happened on a shorter clock. DeepSeek’s arrival at $0.14 per million input tokens in December 2024 was a frontier-class model at one hundredth of GPT-4’s launch price. R1 followed at $0.55 and $2.19 against o1-preview’s $15 and $60 from four months earlier: a 97 percent discount on comparable capability, delivered in a single quarter.

This is precisely what theory predicts when a homogeneous good is produced by several firms carrying large sunk development costs and facing no capacity constraint at the relevant margin. Price collapses toward marginal cost and the fixed cost is never recovered. It is the standard outcome in airlines, DRAM, container shipping, and long-haul fibre. It is happening here on a compressed timescale because the product can be copied outright rather than merely competed with.

Note where the money went in each of those industries. Not to the operators. The fibre laid in 1999 was genuinely useful, the companies that laid it went bankrupt, and the value accrued to whoever bought the strands out of receivership at ten cents on the dollar.

VII. The spot market has been built

A good does not become a commodity because it is homogeneous. It becomes a commodity when an institution exists that prices homogeneous units against one another at low transaction cost. Wheat became a commodity when the Chicago Board of Trade standardized grades and contracts, not when farmers started growing similar wheat.

That institution now exists for intelligence, and it was assembled in about thirty months.

OpenRouter aggregates hundreds of models from dozens of providers behind a single normalized endpoint, with live telemetry on price, latency, throughput and availability, routing each request to whichever supplier wins on the buyer’s criteria. LiteLLM does the same thing self-hosted. Portkey, Cloudflare AI Gateway, Bifrost and a dozen others occupy the adjacent ground.

In August 2026, Stripe agreed to buy OpenRouter. The company that built the metering and settlement layer for internet commerce has bought the metering and settlement layer for model inference, having spent the year building token-based billing. That is not a neutral corporate event. That is the exchange opening for business.

The demand side reorganized around it some time ago. F5’s 2026 State of Application Strategy report, surveying more than 1,100 IT decision-makers, found 52 percent of organizations chaining or orchestrating multiple models, using an average of seven each.

Seven. When the median sophisticated buyer holds a portfolio of seven interchangeable suppliers behind an abstraction layer that reprices on every request, no individual supplier has pricing power. It has a listing.

The strategic advice circulating openly in venture circles makes the point without embarrassment: a team locked to one frontier model pays frontier prices forever, while a team that can swap the model underneath its product inherits every price cut the moment it lands. Buyers are being coached to treat model providers as fungible, and they are complying.

VIII. The data was a commons, which is why nobody has a data moat

The popular argument runs like this. Models are valuable because they absorbed a vast quantity of human knowledge, therefore data is the asset, therefore whoever holds data holds the value.

This gets the economics backwards, and correcting it strengthens the case considerably.

Rent requires differential access, not importance. Ricardo settled the question in 1817. Land generates rent not because agriculture requires land but because parcels differ in fertility and the marginal parcel is worse than the inframarginal one. A universally accessible input, however essential, generates no rent at all. Air is necessary for combustion and priced at zero.

The pre-2023 web was a common-pool resource. Every serious lab ingested substantially the same corpus: Common Crawl, Wikipedia, GitHub, arXiv, Stack Overflow, the digitized book collections. This is not a side observation. It is the mechanical explanation for why all these models resemble each other. Fit broadly comparable function approximators to a nearly identical corpus and you get broadly comparable functions. The convergence everyone now treats as a startling empirical finding was overdetermined from the moment the scraping began.

So the correct formulation is not that the data was free and therefore data is the moat. It is that the data was free and undifferentiated, therefore it produced no moat at all, and the convergence we observe today is its direct consequence.

The great data land-grab of 2020 to 2023 was an enclosure race over a commons that could not be enclosed, run by firms who failed to notice that the value of an appropriated commons is competed away the instant the second firm appropriates it too.

The commons is closing now and the prices are public: Anthropic’s $1.5 billion author settlement, News Corp’s roughly $250 million five-year deal with OpenAI, Google paying Reddit around $60 million a year, Shutterstock booking $138 million in a single year. But this is a tax, not a fence. It raises everyone’s forward costs, incumbents most of all, grants no forward licence, and does not bind at all on labs training in jurisdictions where American copyright judgments are advisory. The models that closed to four months were not trained on a pristine corpus.

The genuinely scarce data now is what never existed on the open web at all: real-world video, professional audio, business process records, and above all workflow exhaust, meaning the record of what a specific user did, in what context, whether it worked, and what happened next.

All of which is generated at the application layer. The model provider sees a prompt and returns a completion. The application sees the prompt, the completion, the edit the user made to it, whether the code compiled, whether the contract closed, whether the ticket was resolved, and what happened three weeks later.

The lab gets a sample. The application gets the label. That is not strategy or positioning. It is a fact about who is standing where the outcome becomes observable, and no amount of capital reverses it.

IX. A wasting asset that appears on no balance sheet

The economic life of a trained model is not set by physical wear or by its owner’s choices. It is set by rivals’ release schedules.

A 787’s useful life is a function of metal fatigue. A model’s useful life is a function of when somebody else ships something better and your capability tier reprices to the new floor. Call it endogenous obsolescence. It cannot be hedged, insured, or extended by maintenance spending. With several well-capitalized rivals racing, each firm’s asset life is the minimum over everyone else’s release dates, which shortens as the field grows. Every new entrant shortens everyone’s asset life, including its own. There is no equilibrium in which this stops.

The accounting has not caught up. Michael Burry put the hardware version on the tape in 2025: hyperscalers depreciating AI equipment over five to six years against a two-to-three-year product cycle, understating depreciation by roughly $176 billion across 2026 to 2028, and inflating reported operating income at firms like Oracle and Meta by more than 20 percent against economic reality. Satya Nadella conceded the substance in passing when he said he did not want to be stuck with four or five years of depreciation on one generation.

The hardware is the second-order problem. The first-order problem is that the intangible those chips produce has a shorter life than the chips, a terminal value of exactly zero, and appears nowhere in the accounts. There is no line item reading “GPT-5.4, net of accumulated amortization.” Tens of billions of dollars of sunk training spend, producing an asset with a four-month competitive half-life, no salvage value, no secondary market, and no alternative use. The mismatch is invisible by construction, which is the only reason it has not yet been priced.

X. The arithmetic of a trillion dollars

OpenAI is marked around $852 billion against roughly $25 billion of annualized revenue, about 34 times run-rate, having filed a confidential S-1 in June 2026 while projecting a 2026 loss near $14 billion, with outside analysts modelling cash burn as high as $27 billion. Anthropic’s most recent mark is $350 billion post-money, following an investment from Google of up to $40 billion. Neither company is cumulatively profitable.

Work backwards from the number.

At a generous terminal multiple of 25 times free cash flow, which is what the market pays for high-quality compounders with durable moats, a $1 trillion equity value requires $40 billion of steady-state annual free cash flow.

These businesses are converging on infrastructure margins rather than software margins. Frontier lab gross margins currently run between 39 and 60 percent, with inference sitting permanently in cost of goods sold. Assume a mature free cash flow margin of 20 to 25 percent, which is optimistic. That implies $160 billion to $200 billion of annual revenue.

Two hundred billion dollars approaches the entire annual revenue of Microsoft. It is 13 to 17 percent of all global enterprise software spend, captured by a single vendor, in a market with five well-capitalized rivals and a free MIT-licensed substitute four months behind at a thirtieth of the price.

Then the decomposition nobody performs. Revenue is price times quantity, and price is in freefall. Growing revenue eightfold from $25 billion to $200 billion while price per unit of capability falls even three times a year for four years, an 81-fold cumulative decline, requires volume to rise by a factor of roughly 650.

Now zoom out to the aggregate. Hyperscaler capital expenditure for 2026 runs around $725 billion, up 77 percent from roughly $410 billion in 2025, with analysts projecting past $1 trillion in 2027. Much of it is deployed on behalf of pure-play AI vendors whose combined revenues remain a fraction of that figure. When infrastructure spending exceeds the revenue of the firms it serves by an order of magnitude, either the revenue is about to grow by an order of magnitude, or the capital is being destroyed.

The fibre was useful. The fibre companies went bankrupt.

XI. Google is winning, and that proves it

Gemini has reached somewhere between 750 and 900 million users. Sensor Tower data has ChatGPT falling below 50 percent of the global assistant app market for the first time in March 2026, landing near 46 percent by May, with Gemini around 28 percent. Apple confirmed Gemini powers Siri across 1.4 billion iPhones. Google Cloud has been compounding at 48 to 63 percent with a backlog in the hundreds of billions.

Doesn’t that show the inventor captured its invention after all?

No. The way Google is winning is the cleanest confirmation of this entire thesis.

Google is not winning because it invented the transformer. It captured precisely zero from that, by its own choice, permanently. Google is winning on the complements.

Silicon first. Sundar Pichai has positioned TPUs as a decade-long moat, citing seventh-generation Ironwood alongside a 78 percent reduction in Gemini serving unit costs across 2025 and a 33-fold reduction in energy per query. More than 75 percent of Gemini runs on TPUs, and Gemini 3 was trained primarily on them. That is a vertically integrated hardware position built over a decade: scarce, excludable, unbuyable.

Distribution second. Search, Android, Chrome, Workspace, and now Siri. Structural, unpurchasable, compounding.

Balance sheet third. Between $175 billion and $205 billion of 2026 capital expenditure, funded from an advertising business that prints cash.

Chips, defaults, and capital. Not the architecture. Google gave the architecture away and is recapturing value from everything around it, which is the smiling curve exactly as drawn: rents accrue upstream at the physical bottleneck and downstream at distribution, and get competed to nothing in the middle.

And then the detail that should settle it. Anthropic, a company that exists because of a paper Google published for free, has contracted for up to a million Google TPUs and, as of April 2026, 3.5 gigawatts of next-generation TPU capacity, alongside a Google investment of up to $40 billion at a $350 billion post-money valuation.

Google gave away the invention, now rents the silicon to the company built on it, and owns a slice of that company. The architecture was worth nothing. The compute and the equity were worth tens of billions.

The margin structure tells the same story with no interpretation required. Nvidia earns a 75 percent non-GAAP gross margin on $81.6 billion of quarterly revenue, with data centre up 92 percent. TSMC earns 67.7 percent gross margin and better than 50 percent operating margin on roughly 72 percent of top-ten foundry revenue, with advanced packaging the binding physical constraint on the whole industry. The frontier labs earn 39 to 60 percent, converging downward toward infrastructure norms rather than upward toward software norms.

The rents are upstream, where lead times run in years and process knowledge does not distill.

XII. What is actually worth something

If the model is worth zero, what isn’t?

Upstream, the physical bottlenecks. Fabs, advanced packaging, high-bandwidth memory, power generation and interconnect. Multi-year lead times, decades of accumulated process knowledge, genuine capacity constraints. You cannot download a fab.

Downstream, the application layer, which owns four things the labs cannot reach. Workflow embedding: swapping a model is a configuration change, while replacing a legal team’s system of record is a project. Outcome data: the application observes whether the thing actually worked and the lab never will. Liability absorption: somebody must be accountable when the output is wrong, that somebody is contractually the application vendor, and the model providers explicitly disclaim it, which makes liability a scarce input that gets priced. And distribution, meaning control of the surface where demand arrives.

Application-layer gross margins are lower than classic SaaS. ICONIQ’s 2026 data puts pure application-layer AI products near 45 percent, with roughly $230,000 of every million dollars in revenue leaving as inference cost. That number is going one way. Application cost of goods falls automatically with every price cut, provided the product is built for model fungibility, while lab margins converge downward. These lines cross. And sunk capital at the application layer is a rounding error against a training run: 45 percent on $50 million invested is a superb business, while 55 percent on $30 billion sunk against a four-month asset life is not a business at all.

The growth already reflects it. Lovable reached $100 million in annual recurring revenue in eight months and $200 million four months after that, with no paid acquisition. Cursor passed $2 billion annualized by early 2026. Bessemer’s data has AI-differentiated companies growing 70 percent faster than SaaS peers in the $1 million to $5 million band.

And then the human complement. When the price of a complement collapses, the value of everything genuinely complementary to it rises. The data tells us who qualifies. Humlum and Vestergaard’s Danish registry evidence shows that within high-exposure occupations, workers in the top earnings quartile adopt generative AI at rates 40 to 50 percentage points higher than those in the bottom quartile. The tools cost twenty dollars a month, so this is not a story about access. It is a story about possessing something the tool does not supply. The BIS and EIB study of more than 12,000 European firms finds AI adoption raising labour productivity around 4 percent on average, but each additional percentage point spent on workforce training adds 5.9 percentage points to the gain. The tool is not what binds.

The mechanism is straightforward. As generation cost approaches zero, the scarce activity becomes verification, and verification cost is anchored to human attention and domain knowledge, so it falls far more slowly. Any activity whose cost falls more slowly than its complement’s becomes the bottleneck, and the bottleneck captures the rent. Add Kremer’s O-ring logic, where production is a multiplicative chain and automating one link raises the marginal return to quality in every remaining link. If a model handles 70 percent of a workflow well, the human 30 percent determines the quality of the whole and commands a disproportionate share of the surplus.

The result is superstar dynamics. Compressed routine capability, with an extended right tail for people who can decompose a problem, specify what success means, and detect when a fluent answer is quietly wrong. That skill is thin, largely tacit, learnable, and priced nowhere near its marginal product.

XIII. What to do about it

If you are valuing a lab, model the trained model as a wasting intangible with a twelve to twenty-four month life and zero terminal value. Put the enterprise value in contracted compute, distribution agreements, consumer default position, and enterprise switching costs, because that is where it actually sits. If your discounted cash flow requires the current capability lead to persist for five years, it is not a valuation, it is a wish.

If you are building an application, architect for model fungibility on day one and treat every price cut as margin you have already earned. Instrument outcomes obsessively, because the outcome label is the only proprietary data you will ever own and no lab can see it. Sell against the cost of headcount rather than the cost of software; the market for work is two orders of magnitude larger than the market for tools.

If you are buying, you are the beneficiary of the largest deflation in the history of computing. Dual-source. Route on cost per completed task rather than cost per token. Refuse multi-year single-provider commitments, because you are being asked to underwrite somebody else’s fixed costs. Then spend the savings on the thing that actually binds. Per the EIB data, that is 5.9 points of productivity for every point spent training your own people.

If you are practising, verification is the bottleneck, the bottleneck captures the rent, and the bottleneck is you.

XIV. Coda

Eight researchers at Google wrote an eight-page paper in the spring of 2017, gave it away, and left.

The paper became the substrate of an industry now carrying more than a trillion dollars of private valuation. Google captured nothing from the invention itself. No licence, no royalty, not even the authors. It is now recapturing value the only way this industry permits: selling the silicon the models run on, owning the surfaces they are delivered through, and taking equity in the companies built on its gift.

Meanwhile the most-downloaded foundation model on earth is Chinese, open-weight, and free. The fourth-best model on the industry’s headline intelligence index is MIT-licensed and free. The cheapest serious coding model in the world costs a hundredth of the frontier and ships its weights. The gap between free and paid is four months, and it has been closing rather than widening.

There is no version of this where the weights are the asset. The thing that was given away for free stayed free, and everything scarce was somewhere else the whole time: in the fabs, in the defaults, in the proprietary exhaust of real workflows, and in the small number of people who know what to ask for and can tell when the answer is wrong.

Models will be enormously useful. Models will be everywhere. Models will be free.

Those three sentences describe every important technology in the history of computing, and only the third one shows up in the accounts.

Sources

The transformer and Google’s contribution: Vaswani et al., “Attention Is All You Need,” NeurIPS 2017 (arxiv.org/abs/1706.03762); Wikipedia’s entry on the paper for equal-contribution status and author departures (en.wikipedia.org/wiki/Attention_Is_All_You_Need); Steven Levy, “8 Google Employees Invented Modern AI,” WIRED, March 2024; 36Kr, “Where Are the Eight Fathers of Transformer Now?”, June 2026 (eu.36kr.com/en/p/3872257909068806).

Open-weight models and the Chinese labs: Hugging Face, State of Open Models report, 14 August 2026, via Fortune (fortune.com/2026/08/15/alibaba-qwen-open-ai-models-3-billion-downloads-meta-google/); South China Morning Post on Qwen’s share of global open-source downloads, April 2026 (scmp.com/tech/big-tech/article/3349552/alibabas-qwen-family-captures-over-50-global-open-source-downloads-report-finds); Tech Insider, “GLM-5.2 vs DeepSeek V4 vs Kimi K2.6,” July 2026, for the Artificial Analysis ranking (tech-insider.org/glm-5–2-vs-deepseek-v4-vs-kimi-k2–2026/); Codersera on comparative pricing, August 2026 (codersera.com/blog/glm-5–2-vs-deepseek-v4-coding-2026/); TechJack Solutions on DeepSeek’s three-to-six-month self-assessment, August 2026 (techjacksolutions.com/ai-tools/deepseek/deepseek-v4-vs-frontier-models/); Epoch AI, “Open models lag state-of-the-art closed models by 4 months,” May 2026 (epoch.ai/data-insights/open-closed-eci-gap); UK AI Safety Institute, “How Far Behind the Frontier are Leading Open Weight Models on Cyber?”, July 2026 (aisi.gov.uk/blog/how-far-behind-the-frontier-are-leading-open-weight-models-on-cyber).

Pricing: TokenCost, “AI Price Index: LLM Costs Dropped 300x, 2023 to 2026,” March 2026 (tokencost.app/blog/ai-price-index); AI Superior, “LLM Inference Cost 2026,” March 2026 (aisuperior.com/llm-token-cost/); Andreessen Horowitz, “LLMflation,” November 2024.

Valuations, capital expenditure, margins: Value Add VC on OpenAI’s revenue and losses, August 2026 (valueaddvc.com/blog/openai-revenue-2026–25b-arr-14b-loss-and-why-profitability-waits-until-2030) and its AI spending tracker (valueaddvc.com/ai-spending); Futurum, “AI Capex 2026: The $690B Infrastructure Sprint,” February 2026 (futurumgroup.com/insights/ai-capex-2026-the-690b-infrastructure-sprint/); NVIDIA Q1 FY2027 CFO commentary, filed with the SEC (sec.gov/Archives/edgar/data/0001045810/000104581026000051/q1fy27cfocommentary.htm); TSMC Q2 2026 results via QuantFlow Lab (quantflowlab.com/ai-semiconductor-spending/); Paulo Carvao, “OpenAI And Anthropic Are Testing Two Very Different AI Business Models,” Forbes, May 2026 (forbes.com/sites/paulocarvao/2026/05/21/anthropic-openai-enterprise-ai-profitability/).

Google’s current position: 24/7 Wall St. on Ironwood TPUs and the 78 percent serving cost reduction, April 2026 (247wallst.com/investing/2026/04/27/googles-latest-tpu-advances-make-the-ai-hardware-race-feel-less-one-sided/); AI Business Weekly, “AI Market Share 2026,” for the Sensor Tower and Apple figures (aibusinessweekly.net/p/ai-market-share-2026); Pasquale Pillitteri on Anthropic’s TPU commitments and Google’s investment, June 2026 (pasqualepillitteri.it/en/news/1441/nvidia-vs-google-tpu-anthropic-ai-chip-2026).

Depreciation: Deep Quarry, “Depreciation of GPUs: between useful lives and useful myths,” December 2025 (deepquarry.substack.com/p/depreciation-of-gpus-between-useful); Dave Friedman, “The $176 Billion Accounting Question at the Heart of the AI Boom,” January 2026 (davefriedman.substack.com/p/the-176-billion-accounting-question).

Routing and multi-model adoption: Artificial Intelligence News on the Stripe and OpenRouter acquisition, August 2026 (artificialintelligence-news.com/news/stripe-openrouter-acquisition-ai-model-routing/); F5, 2026 State of Application Strategy Report.

Application layer and labour: ICONIQ Growth 2026 State of AI snapshot via Jeff Brokaw, July 2026 (jeffbrokaw.com/blog/ai-gross-margins/); SaaS Magazine, “The Rise of AI-Native SaaS,” April 2026 (saasmag.com/ai-native-saas-rise-born-ai-companies-scaling-faster/); Anders Humlum and Emilie Vestergaard, 2025, on Danish registry evidence of generative AI adoption; Aldasoro, Gambacorta, Pal, Revoltella, Weiss and Wolski, “AI adoption, productivity and employment,” BIS Working Paper №1325, 2026 (bis.org/publ/work1325.pdf); Michael Kremer, “The O-Ring Theory of Economic Development,” Quarterly Journal of Economics, 1993.

About the Author

Brian Curry is a Kansas City–based AI and Machine Learning researcher, data scientist, and the founder of Vector1 Research, where he works on the intersection of cognitive AI architecture, causal inference, and economic systems. His research introduced Memory-Node Encapsulation (MNE), an original data structure for artificial episodic memory, and the NeoCortex-M architecture for memory-driven AI. He is the creator of Papilon, an open-source framework for complex system optimization with agentic orchestration, PyCausalSim, a Python framework for causal discovery through simulation, and Daedalus, a platform for causal and economic analysis.

His practitioner work focuses on building production intelligence systems — causal marketing analytics, agentic systems, and the serving infrastructure that makes them deployable inside enterprise environments. He writes about that work at vector1.ai and on Medium.