AI companies are spending heavily on chips, data centers and model usage while many businesses are still waiting for measurable returns from adoption.
Rising token costs can quickly increase AI bills, making model selection, context management, caching and usage monitoring important for controlling spending.
Safety requires additional investment in security, skilled teams, governance and testing, while stronger controls can reduce exposure to costly AI failures.
AI companies now face rising costs and safety challenges at the same time, and each problem makes the other harder to solve. Spending on chips, data centers and daily model use is growing fast, while clear proof of profit is still thin. Safety work adds more expense, since secure and well-tested systems need time, skilled staff and money. Business leaders now ask a simple question about every AI project, which is whether the return justifies the bill. This article explains the main cost drivers, the safety risks experts worry about most, and the steps that can help.
Also Read: Bill Gates on AI Regulation: Why he Says Governments Need to Monitor AI Companies
Every AI request is priced by the amount of text a model reads and writes, which is measured in tokens. One request costs very little, yet a company that runs millions of them sees a large monthly total. Chandan Bhattacharya, writing on Medium, names three main drivers. They are the choice of model, the amount of context sent with each request and the overall volume of use. A test that costs a few dollars a day can grow into thousands of dollars a month once a whole organization starts using the tool.
Large technology firms lead the spending race, and the numbers show how fast the gap between cost and proof is widening.
| Who | What the Data Shows |
|---|---|
| Five large US cloud firms | Planned capital spending of USD 660 billion to USD 690 billion in 2026, nearly double the 2025 level |
| Four biggest hyperscalers | Roughly USD 900 billion spent on AI capital projects over two years |
| Startup Swan AI | A USD 113,000 AI bill in one month for a team of four |
Smaller firms feel the same pressure on a smaller scale.
Higher spending has not yet produced matching profit for most firms. Research reported by Phys.org found that only about 5% of firms using AI report clear productivity gains so far. The same study argues that bold early investment can still pay off, since heavy investors with lower profits have about a 4% yearly chance of a large productivity jump, against 1.6% for a typical firm. Harvard Business School's Hise Gibson advises leaders to judge AI tools by business return and not by technical precision alone, so that projects grow beyond small pilots.
Rising token bills have become a boardroom topic. An EY survey found that 82% of senior leaders at firms investing in AI worry about token use and related costs, and 98% of leaders using token-based tools said the costs made them rethink their approach. Reports also say some companies used their full annual AI budget in three months, while others saw monthly bills double or triple. Forbes reported that Uber's chief technology officer spent the 2026 budget on AI costs by the start of the second quarter.
Safety problems add a second layer of cost, and IBM's research shows how wide the gap in protection is.
Only 24% of generative AI projects were secured, according to IBM.
The global average cost of a data breach was USD 4.88 million in 2024.
IBM's 2026 report puts that average at USD 4.99 million and shows AI-driven attacks up by 56%.
Harvard's Gibson adds that poorly protected AI can expose firms to data poisoning and cyberattacks, which is why he urges risk checks before any launch.
Also Read: From Chatbots to Robots: Asia’s Startups Shift Focus to Physical AI
The MIT AI Risk Initiative asked 272 experts to rank the dangers. They judged that 18 of 24 risk areas carry at least a 10% chance of catastrophic outcomes within five years if current practice continues. Catastrophic here means more than one million deaths or more than USD 100 billion in damage. The five most severe areas are dangerous capabilities, competitive dynamics, weapons and cyberattacks, power centralization and false information. Even with practical safeguards, all 24 areas kept a chance above 5%. In September, calls to slow frontier AI also unsettled markets and raised worries that spending may outrun revenue.
The MIT study found a responsibility gap. Users and the general public are most exposed to harm, while developers and governments hold most of the duty to prevent it. Experts say voluntary action alone is not enough, since any developer that slows down for safety pays a competitive price. They call for rules that can be enforced, such as liability, transparency duties and mandatory insurance. Information, finance and national security were rated the most vulnerable sectors. IBM adds that firms should keep audit trails and records of human decisions, so that someone can be held accountable when a system fails.
Companies can lower both cost and risk with a few clear habits.
Match the model to the task, since model routing can cut spending by 50% to 80% in many setups.
Send only the information a model needs, as better retrieval has reduced costs by 70% or more.
Reuse earlier answers through caching instead of paying for the same reply twice.
Train every employee in AI skills and risks, not only the technology team.
Write AI-specific incident response plans and verify every access request.
Judge each project by the return it delivers to the business.
A governance group with members from human resources, security and strategy can review these steps every quarter.
The MIT study found a responsibility gap. Users and the general public are most exposed to harm, while developers and governments hold most of the duty to prevent it. Experts say voluntary action alone is not enough, since any developer that slows down for safety pays a competitive price. They call for rules that can be enforced, such as liability, transparency duties and mandatory insurance. Information, finance and national security were rated the most vulnerable sectors. IBM adds that firms should keep audit trails and records of human decisions, so that someone can be held accountable when a system fails.
Companies that act early on these steps will gain a real edge. Lower token bills free up money for safer systems, and safer systems protect the trust that customers give. Firms that treat cost control and safety as one shared goal will be better placed to earn steady returns as AI use grows across every industry.
India Pushes US for Mythos AI Access as Cyber Threats to Power Grids, Banks Grow
Trump-Tech CEO’s Meeting: AI Renamed ‘Super Intelligence’, Safety Pact & More
Australia Calls OpenAI and Anthropic CEOs to Senate AI Inquiry
1.Why are AI costs increasing for businesses?
AI costs are rising as companies expand model usage, increase data processing and invest in computing infrastructure. Token consumption can also grow quickly when AI tools move from limited pilots to organization-wide deployments.
2.Why are companies concerned about AI token costs?
Token costs can become significant when employees and applications make millions of model requests. Recent enterprise discussions increasingly focus on monitoring usage, selecting suitable models and reducing unnecessary context to control monthly AI spending.
3.How can companies reduce AI costs?
Companies can reduce AI spending by matching models to specific tasks, limiting unnecessary context, improving information retrieval and using caching. Usage monitoring can also identify expensive workflows before they become large recurring expenses.
4.Why does AI safety increase business costs?
AI safety requires testing, monitoring, security controls, skilled employees and governance processes. Companies also need plans for incidents involving data exposure, cyberattacks or unreliable outputs, adding operational costs to AI deployment.
5.How should businesses balance AI spending and safety?
Businesses can evaluate AI projects using both financial returns and risk measures. Regular governance reviews, access controls, audit records, employee training and AI-specific incident plans can help companies scale adoption with stronger safeguards.