Written by Ralph Sun
Open-weight models and more efficient inference are making AI cheaper. That can damage the scarcity premium built into parts of the AI trade, while making the technology useful to far more customers. The decisive question is whether adoption and monetization expand faster than the capital required to serve them.
The most durable effect of open models may not be that they replace every frontier model. It may be that they reset the price customers expect to pay for a large portion of AI work.
The premise already has evidence behind it. Stanford’s 2025 AI Index found that the inference cost of a system performing at roughly GPT-3.5 level fell by more than 280-fold between November 2022 and October 2024. The same report found annual hardware-cost declines of 30%, annual energy-efficiency gains of 40%, and a selected-benchmark gap between open-weight and closed models that narrowed to 1.7%.
The current market data show that token prices are already falling. Ornn’s Anthropic Token Price Index settled at $0.85 per million tokens on October 17, down 50.6% over 30 days from $1.72, with a $0.84 low during the period. The supplied chart, which plotted the series at $0.84 on October 15, is consistent with the daily settled values published by Ornn.
The index is meaningful but should be read carefully. It is a daily, volume-weighted blend of posted prices across Anthropic’s paid production models. It excludes free tiers and open-weight self-hosting, and it can be affected by changes in the mix of models customers use, caching and volume distribution—not only by headline list-price cuts. That makes it a practical measure of the cost of the average paid token, not a pure measure of any individual model’s list price or vendor margin.
This is an extraordinary rate of unit-cost deflation. It makes a large decline in the effective cost per useful AI task plausible over the next several years. It does not, however, guarantee that every provider’s token price will fall on a fixed timetable.
Token prices are not the same thing as AI value
A token is a unit of model input or output. The price a customer pays per token is only one component of AI economics. Reliability, response time, context length, tool use, security, support, integration, model quality and distribution can all carry a price. A frontier provider may still command a premium for the most demanding workloads even if general-purpose inference becomes extremely cheap.
Open weights matter because they create a credible alternative. Companies can download, adapt and host a model rather than train from zero or pay a frontier API price for every task. Microsoft’s 2026 open-weight statement describes this as the ability to reserve frontier models for genuinely frontier problems and use specialized, lower-cost models elsewhere. It is an advocacy document, but the economic mechanism is straightforward: more usable substitutes weaken pricing power at the routine end of the market.
Google’s Gemma releases show what that substitution can look like in practice. Gemma 3 introduced lightweight, open models with official quantized variants that reduce model size and computational requirements; Google said its 27B model was designed to operate on a single GPU or TPU. Gemma 4 extended the approach with edge-oriented models and a 26B mixture-of-experts model that activates only 3.8B parameters during inference.
These examples do not prove that every model can be run cheaply on a CPU, a phone or a modest GPU. Nor do they prove that all open models are equivalent to the leading closed systems. They show why the cost curve is moving down: smaller, quantized and sparsely activated models can handle an expanding set of useful tasks on less expensive hardware.
The correct investor statement is therefore more precise than “open source guarantees a token-price collapse.” Open weights make sharp cost competition likely in standardized use cases. They make premium pricing harder to sustain where switching is easy. But they do not erase the possibility of premium pricing for differentiated capability, trusted distribution or operational reliability.
Efficiency changes the denominator, not necessarily total demand
The bear case for AI infrastructure is intuitive. If a task can be performed with fewer accelerators, fewer watts and less memory, then fewer chips, servers and data centers may be needed for a given volume of work. Per-task capital intensity falls. For any supplier priced on a permanently rising compute requirement, that is a real risk.
But per-task efficiency is not the same as aggregate demand. Lower cost can make a task economical in more products, more countries and more customer segments. It can turn a pilot into a default feature. It can move inference from a central cloud environment to laptops, phones, retail locations and industrial sites. If the number of tasks grows faster than the compute required for each task falls, total accelerator, server and power demand can still increase.
The International Energy Agency’s scenario work makes this point indirectly. It projects global data-center electricity consumption rising from roughly 415 TWh in 2024 to around 945 TWh in 2030 in its Base Case. It projects electricity use by accelerated servers, driven mainly by AI adoption, growing about 30% annually. Yet the IEA also shows a High Efficiency case with a much lower footprint and a Headwinds case in which demand plateaus around 700 TWh in 2035.
Efficiency, in other words, does not settle the capex question. It makes the outcome more elastic. Hardware spending can remain high if adoption accelerates, or decelerate abruptly if adoption, utilization and pricing do not keep up with supply.
The blessing: AI-to-consumer penetration can widen fast
The biggest winners from cost deflation may be downstream businesses. For this article, AI-to-consumer, or AI2C, means consumer-facing products that use AI to improve search, creation, shopping, support, education, entertainment and personal productivity.
Cheaper intelligence lets a product offer more usage before a customer hits a paywall. It supports localized languages, real-time multimodal features and more frequent agentic interactions. It also lowers the cost of serving low-revenue users, where a high-priced frontier API would make a product uneconomic.
The adoption data suggest this is already more than a theoretical story. Stanford’s 2026 AI Index reports generative-AI adoption at 53% of the population within three years, organizational adoption at 88%, and an estimated $172 billion in annual value to U.S. consumers by early 2026. Sensor Tower projected time spent in generative-AI apps would more than double from 17.2 billion hours in the first half of 2025 to 36 billion hours in the first half of 2026. It also projected more than $4 billion in global AI-app in-app-purchase revenue for the first half of 2026, up 36% from the preceding six months.
Those figures do not establish the profit pool for every application. They do show that cheaper and more accessible models are arriving in a market where usage is still rising rapidly.
That is the blessing of commoditization. A customer can receive more capability for the same spend, or similar capability for much less. The addressable market expands. The application layer can test more use cases and reach users who were previously too expensive to serve.
The curse: value can shift away from the investors who funded the build-out
The same process can be painful for capital providers. When model quality becomes easier to obtain, returns may shift from the owner of the model or the owner of raw compute toward the company that owns the customer relationship, proprietary workflow, distribution channel or data feedback loop.
This does not mean that every chip supplier, cloud provider or frontier lab loses. Some have scale, integrated platforms, power contracts and large external cash flows that can extend their investment horizon. It does mean that investors should separate demand for AI services from returns on AI infrastructure.
Gartner’s October 2026 scenario illustrates the risk. It estimates that more than $3.5 trillion of AI-infrastructure spending could be exposed to impairment by the end of 2030 if investment continues at its current pace. Gartner forecasts nearly $1 trillion of spending in 2026 and another $6 trillion between 2027 and 2030, while projecting that attributable enterprise and consumer AI revenue would not support profitable returns on all planned capex.
That is a forecast, not an observed loss or an inevitable crash. But it identifies the core mismatch the market must test: depreciation on rapidly evolving hardware is real, while end-user willingness to pay is still being discovered. The risk is not merely that demand slows. It is that demand grows, but at prices or margins insufficient to earn a return on the infrastructure built for it.
The AI trade is now a race between two curves
The first curve is adoption: active users, paid conversion, retention, revenue per customer, workload frequency and the share of workflows that move from pilot to habitual use. The second is supply: capex, available compute, power capacity, accelerator efficiency, depreciation and the decline in inference costs.
If adoption and monetization rise faster than unit costs fall, lower prices can enlarge the total market. In that outcome, applications gain users, cloud utilization remains strong and infrastructure spending can keep growing even while the cost of each task declines.
If the supply curve outruns real demand, the same efficiency becomes destructive to returns. Model providers compete away pricing. Compute utilization weakens. Customers gain bargaining power. Equipment may need to be depreciated faster than investors expected. Hardware capex then decelerates—not because AI is unimportant, but because the industry built capacity ahead of cash-generating demand.
The base case may be more uneven than either extreme. Routine text, image and small-agent tasks can commoditize quickly. Frontier reasoning, high-reliability enterprise workflows and tightly integrated platforms can retain a premium for longer. Geography, regulation, power availability and distribution can produce different winners across layers of the stack.
What the market should watch
The adoption side should be assessed through recurring usage and paid behavior, not download headlines alone. Consumer and enterprise AI revenue, retention, conversion, workload frequency and real productivity-linked budgets matter more than model benchmark scores. Sensor Tower’s engagement and in-app-purchase data are one useful measure of consumer momentum, while company disclosures will be needed to judge durable enterprise monetization.
The capex side requires equal attention. Investors should watch hyperscaler capital-spending plans, accelerator lead times, power contracts, data-center utilization, cloud pricing, equipment depreciation policies and any write-downs. The IEA’s Base Case, High Efficiency and Headwinds paths are a useful reminder that hardware efficiency and AI uptake are both variables, not fixed assumptions.
The curse and blessing are therefore inseparable. Open models are likely to make intelligence cheaper and more available. That is good for customers and can expand AI2C adoption dramatically. It is less comfortable for any investor whose thesis requires a permanent scarcity premium on tokens, compute or hardware capacity.
The outcome will not be decided by whether AI is “real.” It will be decided by whether commercial adoption becomes deep enough, fast enough and profitable enough to absorb the capacity built in anticipation of it.
