Quick Take
“Token prices are crashing, but enterprise AI bills are skyrocketing. Welcome to the Jevons paradox of machine intelligence, where cheaper unit economics simply unlock a massive surge in autonomous compute.”
From what I’ve seen, this tension isn’t going away anytime soon. AI token prices have dropped sharply, with benchmarks like the Silicon Data LLM Token Expenditure Index settling near 97 cents amid intense competition.
What happened is straightforward enough: a wave of capable open-source models, low-cost options from China, and steady gains in inference tricks like quantization have cut the average cost per million tokens by as much as 75% over the past year.
Why it matters now comes down to a growing mismatch. Revenue per token keeps sliding while the capital outlays for massive data centers and GPU clusters stay sky-high, which puts real pressure on margins for clouds and providers alike.
The groups feeling it most are the frontier labs trying to hold pricing power, enterprise FinOps teams watching budgets stretch, and cloud operators dealing with fast-depreciating hardware.
The angle that often gets missed is what I call the token cost illusion. Lower unit prices lead companies to expect smaller bills, yet the cheaper access is spurring far more usage through reasoning models and agents, which drives total spending higher.
Deep Dive
AI tokens have become one of the most deflationary resources around. Much like coal or electricity reshaped earlier economies, machine intelligence is reshaping the digital side – except its price keeps falling hard over time. The Silicon Data index recently touched a fresh low, and data from platforms like Ramp shows the average cost per million tokens sliding from around $10 down to $2.50. At first glance, this looked like pure progress from better algorithms and custom chips such as Google’s TPUs or NVIDIA’s Blackwell line.
But the reality includes a sharp price war underneath. Frontier labs have trimmed rates to hold ground against solid open-source options like Llama 3 and the flood of inexpensive Chinese models. Analysts also point out that some of the deflation stems from a shift in usage, with teams routing simpler tasks to smaller models instead of defaulting to premium ones. The base layer of intelligence is turning into a commodity fast.
This drop collides with heavy infrastructure spending. Investors have noted how falling token prices run straight into record capital expenditures. Providers are putting $30,000 GPUs into service and locking up gigawatts of power, yet the revenue they pull per token is thinning out. That gap tests whether the cloud model holds up without relying on explosive volume growth.
The volume growth is already showing up, thanks to the Jevons paradox. Making something cheaper tends to increase how much of it gets used overall. In AI, lower token costs have not trimmed IT budgets; they’ve pushed aggregate LLM spending higher, often doubling it in places. The lower barrier has opened the door to complex agentic workflows that burn through far more compute.
Measurement itself is changing. CIOs and FinOps teams can’t rely on token list prices anymore. Task-level cost is what counts now. A basic chatbot might use a few hundred tokens, while an agent running chain-of-thought reasoning on code could consume tens of thousands in moments. As architectures lean more on reasoning, total consumption will likely outrun any unit-price savings, keeping infrastructure demand strong.
Stakeholders & Impact
- AI / LLM Providers — Impact: High. Facing severe margin pressure as token prices drop while upstream training and infrastructure costs soar.
- Enterprise FinOps & CIOs — Impact: High. Must shift focus from list prices to task-level costs; agentic token consumption requires strict budget governance and smart model routing.
- Infra & Hardware (Clouds, NVIDIA) — Impact: Medium–High. High utilization is guaranteed by elastic demand, but rapid GPU depreciation makes the ROI of AI data centers riskier as token revenue shrinks.
- Open Source & Alternate Models — Impact: Significant. Driving the commoditization of the bottom tier, effectively forcing frontier models to justify their premium pricing with advanced reasoning.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This independent, research-based analysis synthesizes enterprise spending data, commodity pricing theories, and the Silicon Data LLM Token Expenditure Index to unpack the realities of AI cost deflation. It is designed for enterprise CIOs, AI developers, and infrastructure investors who need to navigate the widening gap between token unit economics and total AI expenditure.
i10x Perspective
The drop in AI token prices is not a signal that the technology is losing value. It is the very thing pushing it toward everyday use. By driving unit costs close to zero, the market is moving away from selling raw intelligence and toward delivering concrete, task-level outcomes. Over the next five years, the real winners will likely be the infrastructure players and orchestration platforms that can handle the huge compute volumes created when large numbers of AI agents start interacting at scale.
