AnswersPublished21 sources
AI Gets Cheaper While the Infrastructure Gets More Expensive
AI is becoming cheaper to use but more expensive to supply: comparable inference costs have fallen by roughly 95% in some estimates, while Samsung is raising selected advanced chip prices by up to 15% and SK Hynix exp... OpenAI’s latest cuts—80% for its Luna model, 20% for Terra and more than 20% for frontier Sol ac...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- AI is becoming cheaper to use but more expensive to supply: comparable inference costs have fallen by roughly 95% in some estimates, while Samsung is raising selected advanced chip prices by up to 15% and SK Hynix exp...
- OpenAI’s latest cuts—80% for its Luna model, 20% for Terra and more than 20% for frontier Sol access—show how quickly model pricing is compressing.
- Nvidia’s $96.2 billion fiscal Q2 2027 revenue, including $89 billion from data centers, shows that infrastructure demand remains exceptionally strong even as the economics of AI services deteriorate.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: How is the AI industry’s emerging economic squeeze—marked by a roughly 95% collapse in token prices since early 2024, the appearance of free. Article summary: The industry is becoming a barbell: AI capability is rapidly commoditizing for users, while the physical inputs needed to supply it—advanced fabrication, high-bandwidth memory, servers, power, and data-center capacity—re. Topic tags: general, news, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts w
The AI economy is splitting in two. The cost of using intelligence is falling rapidly, but the cost of building the infrastructure that delivers it remains constrained by advanced chipmaking, high-bandwidth memory, servers, networking, electricity and data-center capacity.
That divergence is more than a pricing story. It is changing which companies can defend margins, how investors should evaluate AI spending and where credit risk may accumulate. The likely winners are not necessarily the companies selling the most tokens, but those controlling scarce infrastructure, distribution, proprietary data or software tied to a verifiable business result.
The AI price war is moving value away from tokens
Several estimates put the decline in equivalent AI inference costs at roughly 95% over a period of about two years, although the comparison depends on the models, workloads and quality levels being measured. 39 The direction is clearer than the exact percentage: capabilities that once commanded premium API prices are increasingly available through cheaper models and open-weight alternatives.
OpenAI’s recent price reductions illustrate the pressure. The company cut its smaller GPT-5.6 Luna model by 80% and its mid-tier Terra model by 20% in July 2026. 34 In August, it announced a further reduction of more than 20% for its frontier GPT-5.6 Sol model for developers.
33
For users, this is good news. Lower inference costs make it more economical to run agents, automate support, generate content, write code and process larger volumes of data. For model providers, however, every price cut reduces revenue per unit unless usage expands enough to compensate. That makes a strategy based solely on selling premium access to undifferentiated model capability increasingly fragile.
The key distinction is between cheaper intelligence and cheaper AI systems. Token prices may fall while the total cost of operating a reliable product—including retrieval, monitoring, security, latency, integration and human review—remains substantial.
The physical supply chain is moving in the opposite direction
The infrastructure required to serve more AI workloads is not expanding as quickly as demand. Samsung reportedly raised prices for some advanced contract-chipmaking services by as much as 15% on new orders as AI demand tightened capacity. 1
Samsung has also said that chip shortages could become more acute and continue into 2028, while pursuing multiyear supply arrangements with major data-center customers. 2 Its memory business has described a framework under which roughly 60% to 70% of capacity could be covered by longer-term agreements if negotiations are completed.
5
6
Those contracts change the nature of the market. Capacity is no longer simply available to whoever pays the spot price. Large customers are reserving supply years in advance, potentially improving their own availability while leaving smaller buyers more exposed to shortages and price swings.
The memory outlook is especially severe. SK Hynix CEO Kwak Noh-jung said the industry could face its worst supply year in 2027 and that demand could remain above production capacity beyond 2030. That is a company forecast, not a guaranteed outcome, but it underscores how long the supply response may take. 17
Nvidia shows why infrastructure still has pricing power
Nvidia’s latest results demonstrate that falling model prices have not yet weakened demand for the hardware layer. The company reported fiscal second-quarter 2027 revenue of $96.2 billion, up 106% year over year, with data-center revenue of $89.0 billion, up 117%. 49
50
This is the central tension in the AI market. More efficient models can reduce the amount of compute needed for a particular task, but lower prices can also stimulate much greater usage. If demand expands faster than efficiency improves, total infrastructure consumption can continue rising even as the cost per inference falls.
The risk is that bottlenecks migrate. Accelerators may be available, but memory, advanced packaging, networking, electricity or grid connections may not be. Nvidia’s reported gross margin was 75% in the quarter, while other reporting has highlighted the possibility that rising memory costs could put pressure on future margins. 49
53 The lesson is not that infrastructure economics have broken, but that strong demand does not make every part of the supply chain equally profitable.
Hyperscalers can absorb the squeeze—up to a point
The largest cloud companies are better positioned than smaller AI vendors to manage the mismatch between falling service prices and rising infrastructure costs. They can spread investment across cloud computing, advertising, enterprise software and consumer platforms. They can also bundle AI into broader contracts, use their distribution to drive utilization and negotiate long-term supply commitments.
That makes hyperscalers plausible shock absorbers for the industry. They may accept lower margins on AI services to protect cloud relationships or gain market share. But the trade-off is capital efficiency: if prices fall faster than usage rises, the return on enormous data-center investments may take longer to materialize.
For smaller providers, the situation is less forgiving. A company that rents expensive capacity and resells generic model access has little protection if API prices fall faster than its customer base grows. Sustainable businesses will need differentiation through workflow ownership, proprietary data, security, reliability, distribution or pricing linked to an outcome rather than token volume.
Cheaper models do not guarantee faster productivity gains
Lower prices should encourage experimentation and adoption, but adoption is not the same as realized economic value. Companies still need to redesign processes, clean data, integrate systems, train employees and establish controls before AI can produce durable productivity gains.
That creates a timing risk for investors. Spending can accelerate ahead of measurable returns, particularly when companies purchase capacity or software because competitors are doing so. If the promised productivity gains arrive later than expected, infrastructure commitments may remain fixed while AI revenue assumptions are revised downward.
The upside scenario is a demand expansion: much cheaper inference leads to so much more usage that total spending, cloud consumption and productivity continue to rise. The downside scenario is a margin and utilization squeeze: model prices collapse, usage grows too slowly, and some data-center projects generate weaker returns than their financing plans assumed.
Credit risk will depend on contracted demand
The AI buildout is also becoming a credit-market question. Data centers, infrastructure acquisitions and component commitments require large amounts of capital, much of it long-dated. Debt is more defensible when it is supported by creditworthy customers with multiyear contracts and predictable usage.
It becomes more vulnerable when repayment depends on speculative future demand. Lower AI prices, slower utilization growth, higher power costs or component shortages could all weaken the cash flows supporting an infrastructure project.
A larger junior debt tranche or higher credit-insurance costs would indicate that lenders are demanding compensation for greater downside risk. Neither would prove that a credit event is imminent. They would, however, make the quality of the underlying contracts and the resilience of the borrower’s cash flow more important than headline AI demand.
The macroeconomic effect will be mixed
AI can be disinflationary in digital services. Cheaper inference may reduce the cost of coding, analytics, customer support, content production and some administrative work. Those savings could eventually reach businesses and consumers if competition passes them through.
The infrastructure buildout can have the opposite effect in specific markets. Scarce memory, advanced chips, servers, construction capacity, land, electricity and grid connections can become more expensive. The near-term pattern may therefore be lower prices for AI-enabled digital work alongside inflationary pressure in the capital goods and energy inputs needed to produce it.
This is unlikely to produce one uniform “AI inflation” or “AI deflation” effect. The outcome will depend on whether efficiency gains spread faster than bottlenecks and whether companies pass savings on to customers or retain them as margin.
Supply chains are becoming strategic assets
Multiyear capacity reservations also increase the geopolitical importance of semiconductors. Samsung and SK Hynix have greater leverage in memory, while advanced logic and packaging remain concentrated in a small number of suppliers and regions. When capacity is booked years ahead, export controls, subsidies and regional disruptions have larger consequences because supply cannot be redirected quickly.
Chinese memory producers may find opportunities in less restricted product segments as global demand rises. Closing the gap in leading-edge HBM and other AI-critical components is more difficult, however. Customer qualification, reliability standards, technical complexity, export controls and procurement rules can all limit how quickly alternative suppliers gain share.
Who ultimately pays for the gap?
The unresolved issue is who absorbs the difference between cheaper AI consumption and more expensive AI production. Several outcomes are possible:
- Hyperscalers can absorb it through lower margins while using AI to defend broader ecosystems.
- Chip and memory suppliers can retain scarcity-driven pricing power while capacity remains constrained.
- Investors can absorb it through lower returns on AI infrastructure spending.
- Lenders can absorb it through wider spreads or losses if projects are overleveraged.
- Customers may pay more if providers repackage infrastructure costs into higher-value, outcome-based products.
The most plausible near-term result is a combination of the first three. AI access will continue getting cheaper, but the infrastructure required to deliver it will remain expensive until supply catches up—or demand proves unable to justify the commitments already being made.
The strategic implication is straightforward: falling token prices are not the end of the AI investment story. They are a test of where durable value sits once intelligence becomes abundant and the machines, memory, power and distribution needed to use it remain scarce.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "AI Gets Cheaper While the Infrastructure Gets More Expensive"?
AI is becoming cheaper to use but more expensive to supply: comparable inference costs have fallen by roughly 95% in some estimates, while Samsung is raising selected advanced chip prices by up to 15% and SK Hynix exp...
What are the key points to validate first?
AI is becoming cheaper to use but more expensive to supply: comparable inference costs have fallen by roughly 95% in some estimates, while Samsung is raising selected advanced chip prices by up to 15% and SK Hynix exp... OpenAI’s latest cuts—80% for its Luna model, 20% for Terra and more than 20% for frontier Sol access—show how quickly model pricing is compressing.
What should I do next in practice?
Nvidia’s $96.2 billion fiscal Q2 2027 revenue, including $89 billion from data centers, shows that infrastructure demand remains exceptionally strong even as the economics of AI services deteriorate.