AI token prices are hitting new record lows

The LLM Token Expenditure Index fell to 97 cents on Monday, the lowest since its late-year launch and more than halving from earlier this summer, signaling ongoing AI token price deflation. Open-source lower-cost Chinese models (e.g., Moonshot’s Kimi K3), OpenAI’s late-July GPT-5.6 price cuts, and “dynamic pricing” from frontier labs are adding downward pressure to token rates and could compress revenues for model providers like Anthropic and OpenAI. The article notes the Nasdaq fell ~1% and the S&P 500 declined 0.4% alongside broader tech weakness, raising profit-pressure risks for AI leaders despite continued compute spending by firms such as Nvidia and Microsoft.

Analysis

Token deflation is first and foremost a margin squeeze for the model layer, not a universal AI bear signal. The economic rent is migrating from raw model access toward products that can bundle AI into existing distribution and monetize workflow stickiness; that is a relative positive for GOOGL and, to a lesser extent, MSFT, because they can dilute per-token pricing pressure inside broader cloud, ads, and productivity stacks. The losers are the pure-play frontier lab economics and any public-market proxy where investors have been paying for scarce model pricing power.

The second-order risk is multiple compression in the AI infrastructure trade if the market starts to question incremental capex ROI. NVDA does not need token prices to rise for unit demand to stay strong, but the stock does need the narrative that each new dollar of compute creates expanding monetizable value; falling token ARPU weakens that link and can cap upside over the next 1-3 months if hyperscaler guidance turns more selective. The more immediate catalyst to watch is whether model providers keep cutting prices faster than usage grows, because that would pressure enterprise AI budgets and delay any public-market re-rating for model monetization.

Contrarian view: the consensus is likely underestimating elasticity. If cheaper tokens dramatically expand inference volume, gross profit pools can still grow even as price-per-token falls, which would favor GOOGL/MSFT over a simplistic "AI demand is broken" trade. The cleanest expression is therefore a relative-value short on the most narrative-sensitive AI capex beneficiary versus a long in the platforms with distribution and pricing power, with the thesis falsified if hyperscaler AI capex and cloud consumption re-accelerate despite lower token prices.

Key Decisions for Investors

• Pair trade for the next 1-3 months: long GOOGL / short NVDA on any AI-led bounce, targeting a 10-15% relative move if hyperscaler capex guidance softens; exit if NVDA data-center growth reaccelerates or GOOGL cloud margins compress unexpectedly.
• Add selectively to GOOGL on weakness over 1-3 months; lower token costs should expand query volume and reinforce Search/Cloud bundle economics. Falsify if AI monetization remains purely promotional and cloud attach rates stall.
• Reduce MSFT exposure tactically into strength until Copilot and Azure AI usage data prove that lower token pricing is driving volume faster than revenue dilution; consider a 2-3 month call-spread sale rather than outright shorting.
• Underweight or hedge NVDA for 1-3 months via call spreads or a short against a semicap basket; the risk/reward improves if the market starts pricing in slower AI capex growth rather than weaker near-term chip shipments.