AI token index falls 56% as cloud sales grow

Silicon Data’s AI token price index stood at 92 cents per million tokens on October 27, about 56% below a May 28 peak the company reported. AI providers charge developers for tokens, small pieces of text processed by models, when software uses services such as chatbots and coding assistants. Current index Reported peak Token definition

Microsoft and Google’s latest earnings show cloud sales continuing to grow. Microsoft reported Azure revenue of $33.2 billion for July through September, up 48% from a year earlier. Google Cloud revenue rose 103% to $30.7 billion. Both companies released results on October 28. Microsoft results Alphabet results

Alphabet reported that its own models processed nearly 32 billion tokens a minute through developer interfaces, more than four times its rate a year earlier. That covers Google’s first-party model activity; Silicon Data measures prices across its sample. Alphabet’s usage disclosure

Those figures leave investors with a more specific question than whether cheaper AI access undermines a data center boom: how much additional service customers buy, what providers spend delivering it, and which businesses earn the revenue. Silicon Data’s average price supplies only part of that calculation.

Customers can lower the average by choosing cheaper models

Silicon Data combines provider prices with consumption observed through services that connect customers to different models. Its benchmark weights prices by usage. The company does not disclose its proprietary raw usage inputs or private commercial terms. Customers choosing inexpensive models can lower the average while every supplier leaves its price unchanged. How the index works

In its own published explanation, Silicon Data said its sample had moved toward cheaper, faster models. It also warned that a token’s usefulness changes across model releases, limiting what a price comparison says about how much customers will pay for capability. Its historical peak was $2.07 per million tokens; comparing that with October’s 92-cent reading gives a decline of 55.6%. Silicon Data’s explanation

There is a competitive argument for lower prices. In September 1 reporting, CNBC quoted Silicon Data research chief Steve Hou suggesting that cheaper and premium models together might already offer enough capability for most tasks. That was an interpretation of market conditions, rather than a measurement of every application’s requirements. CNBC’s reporting

More tokens can outweigh a lower price

A simple illustration shows why a falling unit price does not decide revenue. Assume a provider sells one billion tokens at $2 per million, earning $2,000. Cut the price by 55%, to 90 cents, and unchanged volume generates $900. Double volume and revenue reaches $1,800. Triple volume and revenue reaches $2,700. These are hypothetical quantities and prices, chosen to explain the calculation.

With that 55% cut, the provider needs roughly 2.22 times its original token volume to keep revenue unchanged, an increase of about 122%. This arithmetic holds the product mix constant. An average drawn from several companies and models cannot reveal how any single provider’s volume or revenue changed. These figures describe sales before expenses; delivering a larger volume still has a cost.

Cloud sales include several kinds of products. Alphabet says Google Cloud earns fees from computing services, subscriptions to Workspace communication and collaboration software, and sales of AI computing systems. It said those hardware sales contributed meaningfully to revenue growth. Investors therefore need to separate purchases of computing equipment from recurring payments for services when assessing how AI demand produces sales. Alphabet’s segment explanation

The $165 billion figure covers June-quarter property and equipment purchases

The four-company figure can be reproduced from cash-flow statements for the quarter ended June 30. Amazon paid $54.208 billion for property and equipment, Alphabet $44.924 billion, Microsoft $35.802 billion and Meta $30.116 billion. Together, that is $165.050 billion. These are company-wide purchases, covering each business’s investment needs. Amazon Alphabet Microsoft Meta

That total measures cash paid for assets during one quarter. Treating all of it as spending to serve AI tokens would assign uses the statements do not establish. It also puts a June-quarter cash total beside a price index that changes daily. Comparing the two can frame an investment question, but cannot produce a return on investment.

More recent statements show the spending continuing. Microsoft’s September-quarter cash additions to property and equipment were $39.367 billion. Alphabet reported purchases of $51.355 billion for the same three months. Meta’s September-quarter release, published October 21, recorded $33.942 billion. Microsoft cash flow Alphabet cash flow Meta cash flow

Companies’ headline capital-spending figures can use different definitions. Meta called its September-quarter capital expenditures $34.91 billion, including $964 million of payments that repay financing for leased assets. Its separate property-and-equipment purchases were $33.942 billion. A comparison needs to choose the same treatment of cash purchases and lease payments across companies. Meta’s reconciliation

Temporary discounts complicate longer budgets

Published prices also contain deadlines. Google currently lists standard Gemini 3.8 Flash input at 75 cents per million tokens and output at $3.75 through December 31. Starting January 1, 2027, those rates become $1.50 and $7.50. Developers budgeting around today’s offer therefore face a scheduled doubling for the same token categories. Google’s price table

Google announced that introductory price schedule with Gemini 3.7 Flash on August 13. The dated announcement and current table show why a low present price should be checked against the period covered by a customer’s budget. August announcement

Independent benchmarking company Artificial Analysis calculates what buyers pay for an API benchmark task, using posted token prices and measured token consumption. An API connects a customer’s software to a model provider. Longer answers or more internal reasoning can raise that charge even when two models have identical token rates. Different models also divide text into tokens differently. Benchmark methods

For a team choosing a model, the practical comparison includes whether the software completes its intended job, how many attempts it needs and how much staff review it requires. A cheaper unit of generated text helps most when it reduces the cost of delivering an acceptable answer or completed action.

Amazon’s next results provide a near-term update to the investment figures. Its third-quarter earnings call is scheduled for October 29 at 5 p.m. Eastern. That release will allow a four-company comparison for the September quarter. As of October 28, June remains the latest quarter with results from all four companies. Amazon’s announcement