Cheaper AI: follow the earnings path

The prices announced on July 30

The following standard API prices describe the July announcement, as reviewed on July 31 in Japan. They are historical prices rather than a statement of today's tariff. Amounts are US dollars per million tokens.

ModelInputCached inputOutputJuly 30 change
GPT-5.6 Sol$5.00$0.50$30.00Standard pricing unchanged
GPT-5.6 Terra$2.00$0.20$12.0020% reduction
GPT-5.6 Luna$0.20$0.02$1.2080% reduction

OpenAI also introduced Fast mode for Sol, advertising up to 2.5 times the speed of standard processing at twice the price. This gives customers a separate choice about paying for faster completion. It does not imply that every task runs 2.5 times faster.

The July 30 announcement describes the reductions and speed offering. Model documentation can change as later pricing updates take effect, so it should not be used to replace a historical tariff without checking the announcement date.

A completed task is the right unit of cost

The invoice for tokens is only part of an application's economics. Retries, tool calls, human review, latency and errors can all increase the cost of obtaining a usable result.

A cheaper model can improve margins if it completes the required work reliably. It may offer little saving if a business must repeat requests or employ people to repair its output. Companies therefore need to measure cost per successful task alongside their quality standard.

Price cuts can also make a previously uneconomic workflow worth deploying at scale. Document classification, routine support and repetitive operational work are possible examples. A pilot becomes commercially meaningful when it turns into sustained use and reduces total delivery cost.

For a listed application provider, the evidence would include retention, paid usage, revenue per customer and gross profit after all processing and review costs.

More use does not automatically mean more hardware

An analytical way to frame compute demand is:

Number of uses × tokens processed per use × compute required per token.

Lower prices can increase the first term. More efficient models, better caching or shorter workflows can reduce the other terms. Total hardware demand depends on how those changes combine.

Existing servers may absorb additional traffic through better utilization. New capacity becomes necessary only when sustained demand exceeds that capacity at the required service level. The size of an API price cut alone cannot determine future GPU purchases.

This is why usage growth and efficiency gains can coexist without producing an immediate increase in semiconductor capital spending.

The transmission path differs by business

Japanese AI application and implementation providers are closest to the change in input costs. They may improve profitability or attract customers with lower prices. Competitors receive the same opportunity, however, so some savings may be passed through rather than retained.

For Sakura Internet (3778) and other cloud providers, GPU utilization, contracted demand, power availability and the cost of new capacity are more informative than API pricing headlines. A provider must sell sufficient computing services to justify the infrastructure it builds.

Fujikura (5803) and Furukawa Electric (5801) could benefit if growing traffic supports additional optical connections and data-center construction. Orders, product mix and margins would confirm whether that demand reaches their businesses.

Advantest (6857), Tokyo Electron (8035), DISCO (6146), SCREEN Holdings (7735) and KOKUSAI ELECTRIC (6525) are further along the chain. Inference growth would first need to support GPU, memory or advanced-packaging demand, then semiconductor manufacturers' investment and finally equipment orders.

Lasertec (6920), with its exposure to EUV mask-related equipment, has an even more specific transmission path. Its orders cannot be inferred directly from the number of everyday AI requests.

Cheaper access also increases competition

Lower model costs reduce the barrier to launching an AI service. Existing providers may face more competitors and customer pressure to share the savings.

A business with strong distribution, workflow integration or valuable proprietary information may retain more of the benefit. A service whose only advantage is access to a model has less protection. Rising usage can coexist with falling revenue per task and weaker gross profit.

Infrastructure suppliers face a different risk: capacity added for optimistic demand assumptions may remain underused if efficiency improves faster than workload volume. The result would be weaker returns on capital despite a growing AI market.

What would confirm an earnings benefit?

For applications, look for recurring paid use and lower cost per completed task without deterioration in quality or gross profit. For cloud providers, look for sustained utilization and orders that support the investment in power and servers. For optical suppliers, check actual bookings and margins. For semiconductor equipment, follow customer capital expenditure, orders and shipments.

The July price cuts widen the range of economically feasible AI applications. The stock-market opportunity lies in identifying which companies turn that wider use into earnings and cash, rather than assuming that every AI-related business benefits equally.

Related Pages

Sources