The Token Tax:
The Price War We Have Seen Before

First, Understand the Meter That Is Running

Before any of the geopolitics or the equities or the rate decisions matter, you need to understand one thing: how AI actually gets charged for. Because the entire argument in this piece rests on a unit of measurement that most people have never heard of, and that the financial markets are betting hundreds of billions of dollars on. That unit is a token.

Think of a token as a small chunk of language — roughly three to four characters, or about three-quarters of a word. When you type a message to an AI and it writes back, the entire exchange is broken down into tokens: your question, every word of the answer, punctuation, even spaces. Every single token processed by an AI model requires computation — electricity fed through chips to run mathematical operations millions of times per second. The longer the question, the longer the answer, the more tokens. The more tokens, the more computation. The more computation, the more electricity. The more electricity, the higher the bill.

This is the engine behind every AI product you use. When a company embeds an AI assistant into their software, or when a researcher runs thousands of queries through an API, they are paying per million tokens. In early 2025, using OpenAI's frontier model ran roughly $15 to $30 per million output tokens for the most capable versions.¹ That is the revenue model. That is also the vulnerability. Because what China just demonstrated is that the same answer — delivered at the same quality — does not have to cost anywhere near that much.

Tokens are the AI equivalent of kilowatt-hours. You do not think about them when you flip a light switch, but the entire economics of the power grid is built around them. The question that markets have not yet fully priced is: what happens when someone else's grid becomes dramatically cheaper?

The Efficiency Race and Where It Leads

The AI industry has spent two years obsessing over model efficiency — finding ways to answer your question using fewer tokens, fewer operations, less compute. This is not altruism. It is competitive survival. When the cost of inference (the technical term for running a model to generate a response) is your primary variable cost, cutting that cost by 30% is the equivalent of a manufacturer cutting their labor bill by 30%. It flows straight to the bottom line.

The approach that has driven most of the efficiency gains is called Mixture-of-Experts, or MoE. Instead of a single massive neural network processing every token with every one of its parameters — think of it like routing every customer complaint to every employee in the building simultaneously — an MoE model activates only a small subset of specialized sub-networks for each piece of text. DeepSeek V4, for example, has 671 billion total parameters but only activates 37 billion for any given token.² The other 634 billion stay dormant. The computation required drops by roughly ten times compared to a comparable dense model.

This efficiency trajectory matters, but it has a ceiling. Models are already operating near the theoretical floor of what is computationally necessary to perform the tasks users demand. The low-hanging fruit of architectural optimization has largely been picked. What remains is incremental. And once the efficiency gains plateau — which the trajectory of the last twelve months strongly suggests is approaching — what you are left with is not a software race. You are left with a commodity race. And commodity races are decided by one thing: the cost of inputs.

China Has Played This Game Before. America Always Acts Surprised.

If the AI compute story sounds familiar, that is because it has happened before — just with different products. The mechanism is consistent enough that it should have a name: the Chinese price compression cycle. It works like this: American companies build a technology, invest heavily in domestic manufacturing and infrastructure, assume the premium will hold, and price accordingly. Chinese competitors study the product, streamline the production process, benefit from subsidized energy and government-backed capital, automate aggressively, and then enter the market at a price that does not make sense — until it does, because the volume and the margins they are willing to accept are not comparable to what American companies require.

The solar panel industry is the most instructive case. In 2010, the United States supplied roughly 30% of the world's polysilicon — the raw input for solar panels.³ China was a fraction of that. By 2015, China was the largest solar panel producer on the planet. By 2025, China controlled approximately 80% of global solar panel production. American companies filed trade petitions, the Department of Commerce levied duties ranging from 31% to 250%, China retaliated with tariffs on American polysilicon — and U.S. polysilicon's share of global production fell from 30% to roughly 10% over seven years. The infrastructure the United States had built became economically uncompetitive not because it stopped working, but because the price China was willing to charge rendered the American alternative unjustifiable to any rational buyer.

Consumer electronics. Flat-panel displays. Electric vehicles. The pattern repeats itself with remarkable consistency. The American industry builds. The Chinese industry scales. The Chinese industry undercuts. The American industry petitions for protection. The market, despite the tariffs, drifts toward the cheaper product. The American infrastructure — the factories, the supply chains, the capital that went into building them — suffers impairment that the investment case never fully modeled.

Now apply that template to AI compute. The United States has spent hundreds of billions of dollars building out data centers, sourcing NVIDIA chips at $27,000 to $40,000 each for the H100,⁴ and wiring those centers to power grids that charge some of the highest industrial electricity rates in the developed world. The entire investment thesis rests on the assumption that this is the infrastructure the world must rent time on. DeepSeek's architecture challenged that assumption directly — not with a superior product, but with an equivalent product built at a fraction of the cost, running at a fraction of the inference price.

The Energy Shock Makes the American Position Worse, Not Better

Here is where the story stops being hypothetical and starts being about your energy bill, your portfolio, and the Federal Reserve's next move.

In March 2026, Brent crude surpassed $100 per barrel for the first time in four years, driven by the Strait of Hormuz crisis — a disruption to the passage through which roughly 20% of the world's seaborne oil trade flows.⁵ At its peak, Brent reached $126 per barrel. The International Energy Agency characterized the disruption as the largest supply shock in the history of the global oil market. Gasoline prices in the United States rose 18.9% year-over-year. Fuel oil climbed 44.2%. US CPI hit 3.8% in April 2026 — the highest reading in three years, with energy accounting for more than 40% of the increase.⁶

This matters directly to AI. Data centers are among the most energy-intensive facilities in the American economy. A large hyperscale data center can consume 100 to 200 megawatts of power — enough to power a mid-sized city. Every dollar that oil adds to industrial energy costs compounds the gap between what it costs to run AI compute in the United States and what it costs to run it somewhere the grid is cheaper and the government is subsidizing the bill. The energy shock is not a sideshow to the AI story. It is a direct amplifier of every competitive disadvantage American infrastructure already had.

China, meanwhile, runs its industrial base on energy sources — coal, nuclear, domestically supplied natural gas — that are structurally insulated from Hormuz-linked pricing. The disruption that is inflating American operating costs barely registers on the Chinese cost structure for compute.

The Fed's Difficult Hand — and What Rate Hikes Do to AI Equities

A persistent energy-driven inflation reading does not stay in the gasoline aisle. It travels. It enters wage negotiations. It shows up in Producer Price Index readings. It accumulates in core services components of CPI. And it puts the Federal Reserve in one of its least comfortable positions: facing an inflation problem that originated from a supply shock it cannot fix by raising rates, while equity valuations — particularly in AI — are still priced for a world where rates stay accommodative.

As of June 2026, markets are pricing approximately a 43% probability of a rate hike before year-end, up from 26% just one month ago.⁷ That shift in expectations alone was enough to produce the worst single-day sell-off of the year in AI stocks on June 5, 2026, as the Nasdaq absorbed the repricing. The math here is not complicated, but it is important to understand clearly: when interest rates rise, the present value of future earnings falls. Every dollar of profit an AI company is expected to earn in 2030 or 2035 is worth less today when the discount rate increases. AI companies — particularly the infrastructure plays like NVIDIA, which trades at multiples far exceeding industry norms — are among the most interest-rate-sensitive equities in the market precisely because they are being valued on earnings that have not yet materialized.

NVIDIA's data center revenue reached $115 billion in fiscal 2025, a 142% increase year-over-year.⁸ The growth is real. The question the rate environment forces investors to ask is how much of that growth is already in the price, and whether the multiple can hold if the cost of capital rises and the underlying unit economics of AI compute are simultaneously being undercut by a competitor whose input costs are structurally lower and whose pricing is already 20 to 50 times cheaper per equivalent output.

A Mispriced Asset Looks Like a Great Investment — Until It Doesn't

The broadest version of this argument is not just about AI. It is about the relationship between a great business story and a correctly priced stock. NVIDIA is a remarkable company. The hyperscalers are remarkable businesses. The build-out of AI infrastructure is a genuine secular shift in how computation is organized and monetized. None of that is in dispute.

What is in dispute is whether the equities tied to that build-out are priced at levels that already assume the best possible outcome — maximum compute adoption, American infrastructure dominance, stable energy costs, and favorable interest rates — while the actual operating environment is moving in the opposite direction on at least three of those four dimensions simultaneously.

Custom silicon now represents 20.9% of the AI chip market in 2025 and is expected to expand to 27.8% by 2026, as hyperscalers build their own chips to reduce dependence on NVIDIA.⁸ DeepSeek's architectural innovations mean that future model generations may require less GPU density per unit of output. Oil prices are structurally elevated. The Fed's ability to cut rates is constrained. And the global cloud market — projected to grow at a 46.4% compound annual growth rate in China through 2035⁹ — is expanding fastest in the geography where the cost structure is most favorable to the competitor.

There is a term in fixed income for an asset whose stated value does not reflect its actual risk: a mispriced security. In equities, the same concept applies. When you look at a stock and ask whether the current price is fair given all available information, the honest answer requires you to include the information that is inconvenient — not just the 142% revenue growth story, but the cost pressure from below, the energy shock from above, and the rate environment squeezing the discount rate from the side.

At the End of the Day, Would You Still Overpay?

Here is the simplest version of this argument, stripped of all the macroeconomics and the footnotes and the historical parallels. Imagine two products sitting on the same shelf. They do the same thing. They deliver the same output. One costs twenty to fifty times more than the other. The more expensive one is backed by hundreds of billions of dollars in American infrastructure, American energy costs, and American chip prices. The cheaper one was built by a team in China that figured out how to get the same answer with ten times less computation.

Now imagine that the expensive product's energy bill is going up — because the oil market just had its largest supply disruption in history and the power grid that feeds its data centers is repricing in real time. And imagine that the central bank is considering making it more expensive to borrow the money that funded the entire infrastructure buildout. And imagine that the historical record — solar, electronics, EVs — shows you exactly what happens when China decides to compete on price in a market America thought it owned.

The question is not whether American AI is good. It is. The question is whether it is priced correctly for a world in which a cheaper, equally capable alternative exists, energy costs are rising, rates are potentially heading up, and the playbook being run against it has been run before — successfully — multiple times.

If you are given two exact products and asked whether you would still choose to overpay for one of them, the rational answer is no. Markets, eventually, arrive at rational answers. The gap between where AI equity prices are today and where that rational answer points is the opportunity — and the risk — that this moment is quietly pricing in.