Something crossed a line this month that the industry should have noticed, and mostly didn't. Epoch AI โ the non-profit research outfit that maintains the most-cited public database of AI training compute โ had its estimates picked up by the New York Times graphics desk. That is the entire event. No model shipped. No wafer was taped out. No round was raised. A set of inferred numbers describing how much computation went into training frontier models was rendered into a visual, and the visual walked into the mainstream.
I have spent twenty-nine years watching markets fall in love with charts, and I will tell you the most dangerous sentence in finance is not "the fundamentals are weak." It is "look at the curve."
The curve here is training compute โ the total number of floating-point operations consumed to train a model. Epoch's signature finding, first published by Sevilla and co-authors in 2022 and updated repeatedly since, is that training compute for notable models grew by roughly a factor of ten every two years through the deep-learning era, with frontier models recently accelerating to a four-to-five-fold increase per year. That number is the load-bearing beam of the entire AI capital cycle. Every GPU order, every gigawatt power contract, every hyperscaler capex guide traces its justification back to a curve that says this line goes up.
And now that curve has a New York Times byline. We do not build in the dark; we audit the light. So let us audit what just happened.
Context: Who Epoch Is, and Why the Chart Is the Argument
Epoch AI is not a vendor. It does not sell models, APIs, or subscriptions. It maintains a database called Notable AI Models โ a registry of thousands of systems, each annotated with an estimate of training compute, parameter count, and dataset size. The estimates come from several methods, all of them inference rather than disclosure: reconstructing FLOPs from published hardware configurations multiplied by reported training duration; computing them from model architecture times dataset scale; and, where loss curves or learning-rate schedules are public, back-solving the compute from the optimization trajectory.
This is genuinely valuable work, and I want to be precise about why. Epoch did not invent a new model architecture; it invented a measurement standard. That is a "meta-technical" contribution โ it changes nothing about how models are built, but it changes how the industry measures whether they are improving. It is the closest thing AI has to a GDP statistics bureau: imperfect, contested, and indispensable precisely because everyone cites the same number.
The lineage matters. The original public attempt to chart this relationship was OpenAI's 2018 "AI and Compute" analysis, which first popularized the idea that compute was growing at an extraordinary rate. Epoch picked up that thread, formalized it, and built the dataset that researchers now reach for by default. When a topic acquires a default dataset, it has acquired a center of gravity.
A short technical detour, because the whole edifice rests on it. The scaling hypothesis, in its modern form, holds that model loss falls as a predictable power law in compute, parameters, and data. Kaplan and co-authors formalized the compute relationship in 2020; the Chinchilla result in 2022 refined it, showing that many large models of that era were badly under-trained relative to their parameter count and that the correct recipe was more data per parameter. Epoch's database is the empirical spine that lets anyone test whether these power laws still hold at the frontier โ or whether, as some labs have quietly suggested, the exponent has started to decay.
The reason the NYT visualization matters is not the data. The data has been public for years. What changed is the audience. When the graphics desk of a paper of record selects a dataset for visual treatment, that dataset stops being a technical artifact and becomes a cognitive frame. Millions of readers who will never open a terminal now hold a picture in their heads of a line that goes up and to the right, labeled "AI compute." Policy staffers, corporate boards, and retail investors inherit the same image. That is the news: not a number, but the mainstreaming of a number.
There is a mechanical reason this matters, and it is not conspiratorial. Data journalism of the NYT variety is agenda-setting by design: the selection of a dataset for visual treatment is itself an editorial judgment about what the public should care about. The graphics desk does not visualize everything; it visualizes what it has decided is consequential. So when training compute is chosen, the choice signals that the subject has cleared a bar โ that it is now considered foundational knowledge rather than specialist curiosity. That bar-clearing is the real event, and it is invisible in the chart itself.
It also lands at a very specific moment. The past two years have hosted a genuine, unresolved argument over whether scaling laws have hit a wall โ whether the returns to ever-larger training runs are flattening. Epoch's curve is the most powerful piece of quantitative ammunition in the "scaling is not dead" camp. Its adoption by the Times is therefore not a neutral act of science communication. It is one side of a live dispute receiving the most credible possible platform, while the other side gets a paragraph.
I have seen this movie before, in a different theater. In 2021, when I built a probability model of Bored Ape Yacht Club rarity distributions and published the resulting analysis under the title "The Mathematics of Hype," the point was never that apes were worthless. The point was that a cultural object had been quietly converted into a financial one, and nobody had run the conversion rate. Codifying the intangible is the same problem whether the asset is a Bored Ape or a training run measured in 10^26 FLOPs. You take something nobody can price, you assign it a unit, and the unit starts driving capital.
Core: The Transmission Chain, Audited Line by Line
Here is the mechanism that the NYT visual sets in motion, and here is where I start pulling on threads.
Training compute does not exist in isolation. It is the head of a supply chain. Each order-of-magnitude increase in training compute implies a synchronized upgrade in GPU cluster size, interconnect bandwidth, high-bandwidth memory, advanced packaging, cooling, and โ the constraint that actually binds โ electrical power. When a mass audience is shown an exponential compute curve, the intuitive takeaway is not "interesting measurement." It is "buy the picks and shovels." The demand expectations for NVIDIA, TSMC, the hyperscaler capex lines at AWS, Azure and Google Cloud, the data-center REITs, and the power and cooling suppliers are all downstream of that single rendered line.
Trace the dollar. A doubling of training compute requires roughly a doubling of the accelerator fleet, which requires a corresponding increase in high-bandwidth memory stacks, in advanced packaging capacity at the foundries, and in the switch fabric that binds the cluster together. Each of those is its own bottleneck with its own lead time. When the market is shown a compute curve, it is being shown, indirectly, a demand forecast for every one of those links. That is why the visualization is not merely informational โ it is a capital-allocation prompt aimed at a public that will act on it.
Put a number on the stakes, if only to make the abstraction concrete. If frontier training compute is compounding at four to five times a year, then a cluster that costs a hundred million dollars to build today implies a multi-billion-dollar cluster within three years, with a power draw to match. That is not a rounding error in a spreadsheet; it is a land-acquisition and grid-interconnection problem. The moment the curve is accepted as a planning input rather than a descriptive statistic, it stops describing the past and starts dictating the future. A measurement that becomes a target ceases to be a measurement โ the oldest rule in the book, and the one most likely to be forgotten in a bull market.
That transmission is real, and it is why I take the event seriously. But an audit does not stop at the direction of the arrow. It examines the error bars, and this is where the visual fails the discipline.
Epoch's estimates for open models can be reasonably tight. For closed frontier models โ GPT-4, Gemini, and their successors โ training compute is inferred, not disclosed. The error on those estimates can span a factor of several, and a smooth exponential curve rendered in a national newspaper communicates a precision that the underlying data does not possess. A visualization is an argument about certainty. Plot a clean line and you have silently asserted that you know where the points are. In this case, for the systems that matter most, we do not.
There is a second distortion, deeper and more technical. The hardware-inference method assumes it can see the computation. It cannot see what vendors have chosen not to disclose. If a lab trained with mixture-of-experts routing, structured sparsity, or aggressive expert parallelism, the raw FLOP count that Epoch reconstructs diverges from the compute that actually did useful work. Training compute and effective compute โ the fraction actually utilized, what the field calls model FLOPs utilization โ are not the same quantity, and Epoch's estimates generally assume a high utilization that may overstate effective computation. The curve may be measuring the size of the engine, not the distance traveled.
Then there is the linear assumption smuggled in by the whole framing: compute equals hardware multiplied by time. That identity ignores the countervailing force of algorithmic efficiency. Better architectures, better data curation, and better training recipes reduce the compute required to reach a given capability. Some of that progress is folded into Epoch's "notable models" selection, but the visualization, as a genre, cannot show it. The viewer sees one variable rising. The reality is a ratio, and the denominator is moving too.
And note what the curve omits. Epoch's flagship series is training compute. But the compute that will dominate the industry's energy bill over the next decade is inference โ the compute spent serving a model every time someone uses it. A visualization of training compute alone tells you about the construction phase of the AI economy and nothing about its operating phase, which is where the recurring cost and the recurring revenue actually live. Readers will not know they are looking at only half the equation.
There is also a selection problem hiding in plain sight. "Notable models" is a curated set. The models that get included are, by construction, the ones that were announced, documented, and deemed significant. That is survivorship bias built into the sample โ the compute curve is drawn only over the systems we chose to notice. A dataset of the notable will always look more dramatic than a dataset of the typical.
I ran a version of this audit in 2017, when I built a forty-point checklist for ICO whitepapers and put fifty early Ethereum projects through it in Beijing. The lesson that stuck was not "most projects are frauds." It was that the structure of a claim tells you more than the confidence of its author. A whitepaper that hides its token distribution is making a structural confession. A chart that hides its error bars is making the same confession at a higher resolution.
And I have watched infrastructure narratives outrun their demand before, in my own backyard. The data-availability layer story is the cleanest recent example. The pitch was that every rollup would need dedicated data availability, that this was the inevitable bottleneck, and that the capacity had to be built now. The reality is that the overwhelming majority of rollups do not generate enough data to need a dedicated DA layer at all. Dedicated infrastructure built ahead of real demand is a familiar pattern, and it is worth asking whether the training-compute buildout carries the same whiff of capacity chasing a narrative rather than a need. I am not saying compute demand is fake. I am saying the discipline is to separate the demand that exists from the demand that is projected because the chart says it must.
Contrarian: This Was Never a Story About Compute
Now the part that most coverage will miss, because it requires reading the news event against its own grain.
The interesting question is not "how much compute did the frontier consume?" The interesting question is "who gets to define the unit?" That is the real competition, and it is not happening at the model layer. It is happening at the data-and-research-institution layer, where Epoch, Stanford's AI Index, and a scattering of investment-bank research desks are fighting over standard-setting authority. Epoch's differentiator is method transparency, downloadable data, and frequent updates. Being adopted by the NYT is the flywheel turning: adoption breeds citation, citation breeds dependence, dependence converts a dataset into a de facto fact standard.
And once a measurement standard exists, it becomes a governance instrument. This is the piece the visualization will never show you. Training compute is not merely an economic variable; it is the threshold variable of AI regulation. The 2023 US executive order used a 10^26-FLOP training-compute threshold to trigger reporting requirements for dual-use foundation models. The EU AI Act layers compute-based tiers into its obligations. When a regulator writes a number into law, the dataset that produces that number acquires legal gravity. Epoch's estimates are not academic footnotes anymore. They are the factual substrate of compliance regimes.
Which means the NYT visual was not a science story. It was, whether anyone intended it, a story about the machinery of AI governance reaching public consciousness โ and about a non-profit research group, funded in part by the effective-altruism and longtermism funding circles, quietly becoming the arbiter of a regulatory line. That is a concentration of definitional power worth naming, even when the methodology is honest. The ledger remembers what the narrative forgets: behind every neutral curve there is an author, and behind every author there is a set of choices about what to count.
It is worth being blunt about the funding structure, because it shapes what gets measured. Epoch's support has come from the effective-altruism and longtermism philanthropic network, the same milieu that produced the AI-safety research agenda. That does not make the data wrong โ the methodology is public and reproducible, which is the real safeguard. But it does shape which questions get asked first. A research program funded to worry about catastrophic risk will, naturally, prioritize measurement that supports governance. The measurement is honest; the frame is chosen.
And there is a tension the visualization cannot resolve. The same compute data that enables safety governance also enables a form of control: if compute is the regulated variable, then whoever measures compute holds a lever. The open-source and decentralization camps have argued for years that compute thresholds become de facto licensing regimes โ that a number written into law is a gate, and gates have keepers. The curve that looks like neutral science is, from a different angle, the blueprint for a permissioning system. Both readings are true, which is exactly why the framing fight matters more than the data.
There is also a reflexive trap that the article, and the visualization, both step around. Media coverage of compute growth raises the expected return on compute investment, which funds more compute, which produces more data points that appear to confirm compute growth. The prediction helps generate the evidence for itself. This is not fraud; it is a feedback loop, and it is precisely the kind of circularity I spent 2022 warning clients about when algorithmic stablecoins promised a yield that depended on new deposits to pay the old ones. The mechanism is different; the epistemology rhymes. Distinguishing narrative-driven demand from utility-driven demand is the whole job, and a mainstream visualization blurs exactly that line.
One more signal, easy to miss. The story surfaced in Crypto Briefing โ a crypto outlet, not an AI or hardware publication. That placement is not an accident. It tells you that compute and hardware have become a generic, label-ized theme, portable enough to be borrowed for traffic by any adjacent vertical. Cross-market theme coupling follows: speculative capital that once chased tokens now watches GPU supply chains, and the two narratives start trading on the same sentiment. When a crypto desk starts running AI-compute explainers, you are not looking at diversification. You are looking at a single risk factor wearing two costumes.
There is a final blind spot in how these visualizations are read. They tend to flatten the geography and the ownership of compute into a single global curve. But compute is not distributed; it is concentrated โ in a handful of labs, a handful of countries, a handful of power grids. A single smooth line hides the fact that the interesting dynamics are about concentration: which jurisdictions can host the training runs, which firms can afford them, and which communities end up hosting the data centers that feed them. The energy footprint that NYT-style graphics often fold in is not a footnote to the compute story. It is the physical bill for it, and it is being paid in specific places by specific people.
Takeaway: Track the Footnotes, Not the Curve
I do not expect the compute curve to bend tomorrow. I expect the opposite: the mainstreaming of this narrative will reinforce the capex cycle it describes, at least until a physical constraint โ power, land, or advanced-node packaging โ forces a reconciliation between the chart and the megawatts.
But the discipline is the same one I applied to token sales in 2017 and to algorithmic stablecoins in 2022. Do not audit the headline. Audit the footnote. Watch Epoch's methodology updates for whether they publish error intervals and whether they extend coverage to inference compute, not just training. Watch the NYT original to see whether it rendered compute, energy, or model scale โ because those are three different stories with three different investment implications. Watch the hyperscaler capex guides and the power purchase agreements, because those are where the narrative has to survive contact with reality. And watch the regulatory thresholds, because the moment a number becomes law, the argument over the number becomes an argument over the industry.
In 2026 I spent a year designing a framework to verify AI-generated content on-chain using zero-knowledge proofs, working with three major AI labs on proof-of-humanity protocols. The lesson from that work is the one that applies here. The value of a standard is not that it captures truth perfectly. It is that it makes claims auditable โ and auditable claims can be corrected. Epoch's curve is a step toward auditability. The NYT visualization is a step toward belief. Those are not the same direction, and the distance between them is where the next correction will come from.
The curve is not the truth. It is a claim about the truth, dressed for a mass audience. The question worth carrying forward is simple: when the next exponent gets rendered in the paper of record, will anyone in the frame be holding the error bars โ or will we keep mistaking a beautiful line for a settled fact?