Silicon Data Says Declining Token Unit Prices Mainly Reflect Model Substitution, While GPU Rental Prices Continue to Rise

Article

Opinion: Halving of LLM Token prices does not imply weakening AI demand; the GPU compute rental market continues to show strong demand

BlockBeats news, September 24: Professional data firm Silicon Data wrote that the market has recently frequently cited its LLM Token Index as grounds for bearish AI trades, but a decline in the indicator does not mean AI demand or model-layer profits are weakening.

Silicon Data stated that its LLM Token Index measures the price API users in the sample actually pay per million Token, namely the “usage-weighted price,” rather than Token usage or total spending. The index fell from USD 2.07 on May 28 to USD 1 on September 21, a decline of 52%, but it had risen 67% during the period from January to the May peak this year. Therefore, price changes alone cannot directly serve as an indicator of AI industry conditions.

The firm noted that the recent decline in the index mainly reflects more tasks being allocated to inexpensive, fast models, while AI labs continue to introduce more mid-tier models. Because Tokens do not fully represent AI capabilities, the indicator alone also cannot distinguish between model substitution and more efficient Agent routing.

In contrast, the GPU compute rental market continues to show strong demand. Silicon Data data show that H200 rental prices among non-hyperscale cloud providers rose from an average of USD 2.87 in June to USD 3.29 currently, briefly reaching USD 3.32 on September 19; B200 currently stands at USD 5.76, up 7% over the same period and 31% year to date; B300 has risen 44% since its launch at the end of April.

Silicon Data believes that as Agents are deployed at scale, the vast majority of future Tokens may come from inexpensive, fast models. Token prices may therefore continue to decline while total Token usage still grows substantially. The firm said that truly high-value tasks may remain concentrated in frontier models, while broad AI adoption could mean demand for compute is “more, not less.”

Original link https://m.theblockbeats.info/flash/368754

Key points

Silicon Data said its LLM Token Index measures the price API users in the sample actually pay per million Token.
The index fell from USD 2.07 on May 28 to USD 1 on September 21, a decline of 52%.
H200 rental prices among non-hyperscale cloud providers rose from an average of USD 2.87 in June to USD 3.29 currently.
B200 currently stands at USD 5.76, up 7% over the same period and 31% year to date; B300 has risen 44% since its launch at the end of April.
Silicon Data believes that as Agents are deployed at scale, Token prices may continue to decline while total Token usage still grows substantially.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

Token prices reflect the amount paid per million Tokens, not Token usage or total spending. When more tasks shift to inexpensive models, unit prices may decline; meanwhile, rising GPU rental prices may indicate that some compute demand remains strong.

Why it matters to readers

This suggests that Token unit prices alone should not be used to assess AI demand or model service provider profits; total Token usage, total spending, and changes in the GPU rental market should also be considered.

For beginners, a lower unit price does not necessarily mean the market is deteriorating: if usage grows quickly enough, total demand may still increase.

Risks and unknowns

The above data and assessments come from Silicon Data’s publication, and the provided material does not include independent verification.
The provided material does not specify the sample size, coverage, or total Token usage of the LLM Token Index.
This index alone cannot determine the specific effects of model substitution, more efficient Agent routing, total revenue, or total profit.

Silicon Data Says Declining Token Unit Prices Mainly Reflect Model Substitution, While GPU Rental Prices Continue to Rise

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center