Report: The Speed Gap: AI’s Fastest Layers Are Growing at Triple Digit Rates, the World Economy at 3 Percent, and Trillions Are Being Repriced in Between
Report page: summary and downloads →
The artificial intelligence economy is growing far faster than the world it is being built to change, and in the week of 14 September 2026 the markets began to price the distance between the two. The Edge has published a new report examining how the fastest layers of the AI economy, growing at triple digit rates, are now governed by the pace of the slowest, and why that gap repriced trillions of dollars of value in a single week.
One economy, many speeds
The report sorts the AI economy into five layers by the clock each one runs on. Frontier models improve in months, chip and memory suppliers respond in quarters, the hyperscalers and specialist clouds commit in years, power and grids and metals move at the pace of permitting, and the enterprises whose spending has to repay all of it change slowest of all. The layers are not comparable by design, because capability, sales, capital expenditure, electricity and output are different quantities. What the ladder measures is the order of magnitude at which each party moves. At the fast end, global semiconductor sales rose 135.1 percent in the year to July 2026 and Nvidia’s revenue rose 106 percent in its latest quarter; at the slow end, the IMF expects the world economy to grow 3.0 percent in 2026, electricity demand 3.6 percent, and United States labour productivity 2.2 percent. When the fast layers are financed by the slow ones, on our reading the slow ones set the price.
The suppliers are running fastest
The chip industry sits at the top of the ladder. Global semiconductor sales reached 146.8 billion dollars in July 2026, 135.1 percent above a year earlier and the seventeenth consecutive monthly increase. Nvidia’s latest quarter is the cleanest expression of the pace: revenue of 96.2 billion dollars, up 106 percent, data centre revenue up 117 percent, and guidance that the company still describes as supply constrained. The rest of the chain reports the same shape, with Broadcom’s AI semiconductor revenue up 221 percent and SK hynix’s revenue up 257 percent. The market has begun to demand more than record results: SK hynix shares fell 9.6 percent on the day it reported those figures, because the profit, though a record, missed estimates. On our reading, growth of 200 percent is now priced for growth of 200 percent, and a supplier can miss by growing merely very fast.
The buyers spend faster than they earn
The four largest hyperscalers and Oracle are the AI economy’s demand, and their growth is strong by any standard except the one above them. Google Cloud revenue rose 82 percent to 24.8 billion dollars, Microsoft’s Azure 43 percent past 100 billion dollars of annual revenue, Amazon Web Services 37 percent to 42.2 billion, and Oracle’s cloud infrastructure 121 percent to 7.4 billion with a contracted backlog of 664 billion dollars. The spending is growing faster than the revenue. Amazon’s free cash flow for the year to June was an outflow of 7.6 billion dollars, Alphabet’s second quarter figure was minus 5.9 billion and Oracle’s first quarter minus 5 billion, and two of the five turned to the markets in the quarter, Alphabet raising 49.6 billion dollars of stock and 20.3 billion of notes and Oracle completing a 20 billion dollar share sale. Sell side estimates cited by Reuters put the 2026 capital outlay of the five at around 795 billion dollars, and part of the buildout has become a credit exposure that answers to the yield curve.
Where the billions went
The repricing did not begin with the economy but with the developers. After the head of one AI laboratory called publicly for a slower pace of capability development and the heads of two others endorsed it, the Philadelphia semiconductor index fell 5.9 percent on 14 September, SoftBank lost 10.72 percent in Tokyo and the Kospi 3.26 percent. The rate leg followed: the 10 year Treasury yield touched 5 percent for the first time since 2023 and then rose to 5.039 percent, its highest since 2007, and on 16 September the Federal Reserve raised interest rates for the first time in more than three years, to a range of 3.75 to 4.00 percent. On our calculation from CNBC quote data at the 16 September close, thirteen listed parties to the AI economy stand about 5.1 trillion dollars below their respective 52 week peaks, most of those peaks set in May and June, with Oracle 56.6 percent below its high, Broadcom 31.4 percent and Nvidia 9.6 percent.
Why it matters: The parties in this cycle now depend on the same variable, the pace of the buildout, so their shares move together even when the companies operate in different industries. Internal cash has given way to equity, notes and leases, which makes a 10 year Treasury yield at 5 percent the same input for a data centre lease, a utility bond and a chipmaker’s multiple; grid connections are set by queues measured in years; and monetisation is set by enterprises whose share reporting any profit impact from AI has not moved from 37 percent in a year. A statement about model capabilities, made by developers who sell no chips and own no grids, moved ASML, SK hynix, SoftBank and the cybersecurity vendors, each in the direction its position on the ladder dictated.
Outlook: The report sets out three paths to 2027. In convergence, enterprise conversion broadens, free cash flow at the buyers turns positive despite record spending, and the chip cycle slows from triple digit growth without contracting. In the managed gap, its central case, demand keeps exceeding supply but the buildout is funded further into debt and leases at a 5 percent risk free rate, and episodes like July and September recur as repricings of pace rather than demand. In reversal, a higher cost of capital and a further year of flat enterprise conversion reduce orders, and the equity drawdowns become credit events for the most leveraged builders. Six observable series, from the ratio of cloud growth to capital expenditure at the five buyers to the Federal Reserve’s path, will decide which one is taken.
Sources: The Edge.
Download the full report
⬇ Download the report in English
⬇ تحميل التقرير بالعربية
