The price of intelligence

The world map: who does what, and at what price

The gap between the United States and China is no longer measured in years. The US NIST assessed DeepSeek V4 Pro as roughly eight months behind the frontier in May 2026, and in July judged GLM-5.2 broadly comparable to a US model from six months earlier. Meanwhile, Chinese models have come to account for roughly 61% of tokens processed on OpenRouter as of mid-2026, and Alibaba’s Qwen family has passed one billion downloads.

The commercial consequence shows up in the price list.

Exhibit 4

Output list price, dollars per million tokens

On input, the gap between the cheapest and the premium end is thirty-six times; on output it reaches one hundred and seven times. And these are capabilities that, on many real-world workloads, differ only marginally. That is the working definition of a market commoditizing from the bottom up.

Two political complications, both recent and both underestimated by the market.

1. Beijing is building its own export-control regime — for software. Since July 2026 China’s Ministry of Commerce has been discussing with Alibaba, ByteDance and Z.ai a tiered system that would restrict foreign access to Chinese frontier models, including open-weight releases. For the first time both superpowers are building access controls, and at both layers: the United States mainly on hardware but — as shown below — with a precedent on models too, China on models.
2. Washington has already pulled the mirror-image lever: in June 2026 it imposed restrictions on non-US access to Anthropic’s flagship models, then lifted them at the end of the month.

Exhibit 5

Two mirror-image controls, at opposite ends of the same stack

For investors: geopolitical risk on this theme is no longer one-directional. It is no longer just “what happens to Nvidia if China is cut off.” It is also “what happens to Western software margins if the Chinese substitute at a tenth of the price suddenly becomes unavailable — or if it doesn’t.”

Where the bottleneck has moved

Solid · memory first

The worst shortage in fifteen years

The reallocation of production capacity toward HBM has produced the worst memory shortage in fifteen years. The numbers that matter.

SK Hynix’s chief executive said in July that the shortage will probably last beyond 2030. A new fab takes 18-24 months just to build. This is not a cycle: it is a structural change in the market, and the three producers — SK Hynix, Samsung, Micron — for the first time in decades hold asymmetric bargaining power over their customers, Nvidia included.

Critical · then electrons

Generation is not the problem. The queues are

Exhibit 6

US data center power demand, GW

Goldman Sachs estimates that US data center power demand will rise from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027. By 2027, data centers would account for 8.5% of US summer peak demand, versus 4.1% today. EPRI goes as far as estimating that by 2030 data centers could absorb between 9% and 17% of US electricity generation, and between 39% and 57% in Virginia.

But generation is not even the main problem. The problem is the queues.

• Interconnection queues: 3-7 years in key markets, up to 10 in Northern Virginia. A campus gets built in 2-3 years.
• High-power transformers: lead times have gone from 24-30 months before 2020 to about five years.
• Gas turbines: GE Vernova ended Q2 2026 with 116 GW of orders in backlog — up from 83 GW at end-2025 — and is taking reservations for 2031 delivery. The average cost of a combined-cycle plant has risen to about $2,157 per kW, from under $1,500 in 2023.
• Construction cost: about $11.3m per MW of shell in 2026, to which the AI fit-out can add up to $25m per MW.

The practical result is that facilities are being completed and then sit idle, waiting to be connected to the grid. Electrical equipment is less than 10% of the cost of a data center and 100% of the bottleneck.

The opportunities: four layers, four risk profiles

Layer 0 — electrons, the longest duration

This is the only point in the chain where scarcity is protected by physical and permitting constraints, not by a technological edge that can be replicated. Nobody can manufacture a shorter interconnection queue.

Layer 1 — silicon and memory, where the margin is shifting

The non-consensus thesis: over the next four to six quarters, memory makers capture a growing share of total profit, taking it from the accelerator designers. The signal is already in company filings, in Nvidia’s declining margin guidance, which comes with an explicit rationale.

A note on Broadcom: the market has punished the stock — roughly 28% below its June 2026 high — over doubts that the $100bn AI revenue target for fiscal 2027 is achievable. This is the kind of skepticism that creates asymmetry, in a stock trading at lower multiples than its sector peers. One to watch, not to chase.

Layer 2 — cloud and models, the most visible and the most fragile

Here the question is not “will they grow?” but “at what return on invested capital?” The market has already changed its yardstick: until 2025, announcing more capex sent the stock higher; in 2026 it prompts the question about returns. The evidence is that the seven largest tech companies, with capex at the four biggest accelerating by 77%, have stalled at +10.6% year to date: announcing spending no longer re-rates the stock the way it did in 2025. And since those seven account for 33.9% of the index, the entire US market is dragged along at their slow pace.

The real event on the calendar, however, is a different one.

Cautious · the IPO that resets the price of this layer

The first public price on a frontier lab

Anthropic reached an annualized revenue run rate of about $65bn at the end of July 2026, up from about $9bn at the end of 2025, with preliminary Q2 revenue above $11.5bn and positive adjusted operating income, according to Bloomberg. The financial press also reports that it has filed confidentially for a listing, with Morgan Stanley, Goldman Sachs and JPMorgan, and could come to market as early as this fall. Neither report has been confirmed by the company.

The comparison is what matters: OpenAI is running at around a $40bn annualized run rate and in 2025 posted an operating loss — about $20.9bn — larger than its reported revenue for the year.

Why this is the event of the fall for investors: it would be the first public, continuous price on a frontier lab. Until now the entire model layer has been valued through private funding rounds, at valuations set by parties with an interest in setting them high. A public order book, a free float and an analyst consensus turn an opinion into a price — and that price would immediately become the benchmark multiple for re-rating Microsoft, Alphabet, Amazon and every neocloud.

The deal can go one of two ways, and both are informative. If the public market accepts private-market multiples, the whole layer re-rates. If it rejects them, the correction cascades down the entire chain, credit included. There is no neutral outcome.

Layer 3 — the users, the least crowded

This is the line of reasoning almost nobody is making explicitly.

If the price per token falls by a factor of 50 to 200 a year and capability keeps improving, then the interesting place is not with those who sell intelligence, but with those who buy it. A company that buys an input deflating by 90% and sells an output whose price is anchored to the cost of the human labor it replaces is expanding gross margin by construction.

The categories to watch, in order of how verifiable the effect is.

1. Text-intensive professional services — legal, accounting, audit, translation, customer support. Margins rising without revenue growth.
2. Vertical software with proprietary data and workflows — the model is a commodity; the data and the workflow are not. It must, however, be carefully distinguished from horizontal software, which today is the target of AI: some SaaS valuations have been compressed precisely by the models’ ability to write code.
3. Companies with high research and development costs — pharmaceuticals, semiconductors, engineering.

The macro evidence is still ambiguous, and that should be said plainly: US productivity grew about 2.7% in 2025 against a ten-year average of 1.4%, which is consistent with an AI effect but does not prove it. On labor, academic studies have not yet found an aggregate decline in employment; they do, however, find an effect heavily concentrated on new entrants — employment among 22-25-year-olds in the most exposed occupations fell by about 13% between end-2022 and end-2025.

The risk everyone is talking about right now

What happened, in order

• September 3, 2026 — lawmakers Bernie Sanders and Greg Casar introduce the “Ban Artificial Superintelligence Act”: a permanent ban on developing superintelligence and a temporary pause on the development of advanced AI.
• September 7 — the UN High Commissioner for Human Rights, Volker Türk, states in Geneva that he shares the concern that advanced AI could pose an existential risk, explicitly naming Meta, OpenAI, Google and Anthropic.
• September 8-9 — a researcher who resigned from Anthropic posts a message on X accusing the leading labs of racing toward self-improving superintelligence without adequate accountability. The message circulates widely and is picked up or cited by more than twenty lawmakers.
• September 12 — executives from OpenAI, Anthropic, Meta and Google meet to discuss the safety of advanced systems.
• September 13 — President Trump publicly downplays the existential risk, reframing the issue around competition with China.

In eleven days the issue moved through every level: Congress, the United Nations, the press, the labs and finally the White House. Concrete technical precedents are also circulating in the debate: in July 2026 OpenAI disclosed an episode in which, during cybersecurity evaluations, some models had circumvented the intended isolation controls.

How to price it, as an investor

We separate three things that the headlines tend to blur together.

a) Terminal risk. If the worst-case scenario materializes, no portfolio allocation offers protection. A risk that, if it occurs, wipes out the very concept of return cannot be priced and therefore does not belong in an investment model. Ignoring it does not mean denying it: it means this is not the place for it.

b) Regulatory risk — this one can be priced. It is the channel through which the debate reaches share prices. Worth tracking, with dates and texts.

c) Reputational risk to valuations. If public perception of AI deteriorates, the sector’s cost of capital rises before any law is passed. This is already observable in the credit markets.

Three ways this story ends badly

Critical · credit, not equity

1. Capex is no longer paid for with cash

The structural change in 2026 is how capex is being paid for. Goldman Sachs estimates that roughly a third of 2026 spending is debt-financed, amounting to about $250bn of direct issuance by the hyperscalers alone, with project finance on top. Hyperscaler bond issuance had already reached about $121bn in 2025, more than four times the average of the previous five years. A widely cited estimate puts the external financing need at about $1,500bn by 2028 to close the gap between cash flow and investment.

Below investment grade the structure is more fragile: private credit has taken lending to the AI sector from almost zero to more than $200bn, CoreWeave has placed GPU-backed loans — the first to carry an investment-grade rating, at $8.5bn — and GPUs have a useful life of around seven years, against decades for the building that houses them. Anyone lending against GPUs is taking as collateral the component that ages fastest.

The indicators to monitor every week: hyperscaler bond spreads over Treasuries of matching duration, terms on new data center securitizations, neocloud convertible bond prices, and spot GPU rental rates. Spread widening leads a capex slowdown by two to three quarters.

2. Commoditization climbs higher than expected

If the capability gap between a model priced at $30 per million tokens and one at $0.28 keeps narrowing, value shifts entirely to distribution and proprietary data. In that scenario the whole model layer becomes an infrastructure business with utility-like margins, and current private valuations become unsustainable. The 61% share of Chinese tokens on OpenRouter is the leading signal to watch.

3. Physics beats capital

If interconnection queues and turbine lead times do not shorten, a significant share of announced capex slips rather than being spent. For equipment suppliers it is a deferral, not a cancellation. For anyone who has priced a 2027-28 revenue curve on 2027-28 deliveries, it leaves a hole.

Calendar and what to watch

Key takeaways

1. Margin is shifting upstream and downstream of Nvidia, not to Nvidia. Above: memory. Below: power and electrical equipment. Nvidia remains the training monopolist, but the pace of upside surprises is slowing as input costs rise.
2. The power layer has historically been cheaper during AI equity corrections, not at the highs. The two move together on the way down and diverge on the way up.
3. The model layer should be read as a commoditizing sector, with the sole exception of those who own distribution. A potential Anthropic IPO is the test: whatever the outcome, it reprices the entire group.
4. The less crowded side is the token buyer, not the seller. Margin expansion by construction, with no need for revenue growth. But mind the distinction between software that uses AI and software that AI replaces.
5. The risk to track is not extinction: it is credit, and then the power bill. The first determines when the cycle ends; the second determines where data centers can be built over the next five years.
6. $725bn of capex against $105bn of lab revenue is not, in itself, proof of a bubble — the spending also supports the hyperscalers’ cloud revenue, not just the labs. But it does measure how much of the return is still an assumption rather than cash in hand.

Main sources

Nvidia, Q2 fiscal 2027 results release, August 26, 2026 · CNBC, coverage of Nvidia’s results, August 26, 2026 · CreditSights, 2026 hyperscaler capex estimates · Goldman Sachs Global Institute, May 2026, and Goldman Sachs Research on credit and AI, August 2026 · Bloomberg, as reported by TechCrunch and Axios, on Anthropic’s revenue run rate, August 17, 2026 · Epoch AI, LLM Inference Price Trends · NIST CAISI, evaluations of Chinese models, May and July 2026 · Reuters, Chinese export controls on models, July 7, 2026 · Seoul Economic Daily and TrendForce on HBM and DRAM, August 2026 · GE Vernova, Q2 2026 results · EPRI, Powering Intelligence, 2026 · Federal Reserve, FEDS Notes on the AI capacity build-out, July 17, 2026 · Axios and IBTimes, coverage of the existential-risk debate, September 9-13, 2026 · providers’ public API price lists, September 2026.

All analysis, theses and exhibits are Vort Research’s own. Third-party data are identified as such; forward-looking estimates are labeled as estimates, not company guidance.