Tokenomics: smart strategies to deal with the costs of AI and infrastructure

AI has been heralded as having the potential to change not only the way we do business but to reshape entire economies. The industry is ballooning and investment in infrastructure has sucked in trillions of dollars. This funding needs a return and pressure is building to charge for AI.

AI usage has come at a cost for several years but the average consumer is unlikely to have noticed. Entry-level chatbot use, the “what can I cook with x, y and z” type of query, remains free — but more intensive users can access many more tools.

Consumer-facing or small-enterprise off-the-shelf subscriptions for Claude and ChatGPT start at $8 per month. For “longer conversations” this is $100 or more. These “per seat” prices are the norm for AI assistants in non-technical roles.

More intensive users are having to pay by token. The first model to charge this way was ChatGPT in 2022 and others soon followed. As usage becomes more widespread and tasks more complex, for instance with the increased deployment of AI agents, the consumption of tokens and the size of bills for AI have both rocketed.

The economics of this technology continue to evolve as developers and providers seek a return. This report looks at the mechanisms behind tokens and their pricing, how companies can manage usage and how they might measure the value provided.

Tokens and costs

Tokens are pieces of information that an AI system uses to digest, understand and generate text. The unit size varies between providers and can consist of anything from one to several characters. Different models from each developer break down a query into a set number of tokens but it will cost more to process the same query using more advanced models because deeper reasoning requires more computing power, known as “compute”.

For instance Claude, the model family produced by Anthropic, includes Haiku (its least sophisticated), Sonnet, Opus and Fable. Each uses the same protocols but Haiku tokens are the cheapest. Fable has the greatest reasoning capability and so costs more to use.

Any given query and its response involves input tokens (those provided by the user in a prompt) and output tokens (those generated for an answer). Input tokens are cheaper because reading text needs little computation. Output tokens cost more because producing text requires repeated forward passes and more computation.

The eventual bill for a task is determined by the processing needed. The cost will be a factor of the number of tokens related to the query and the computing power required to read input and generate output. Even if a query and its answer are short or simple, frontier models come with a higher charge.

For all the sophistication, it is still open to debate whether AI applies logic and reasoning to come up with answers or simply uses sophisticated pattern-matching.

Pricing evolution

When AI models launched, providers focused on hooking customers onto their platform. Much like the early days of Uber cabs, AI companies gave their product free to increase adoption and model reliance. As model sizes and training costs surged, requiring more power and infrastructure, providers could no longer subsidise the load and investors pushed them to monetise.

At the same time, as capacity has grown and developers have found they cannot lock in customers, aggregate token prices have fallen. In early October, the LLM Token Expenditure Index compiled by Silicon Data fell to 95 US cents, the lowest in its nine-month history. Switching costs are low because AI models are infrastructure agnostic, which explains why Google is trying to create stickiness by bundling cloud storage with its Gemini per-seat subscriptions.

Tech for Growth Forum

The Tech for Growth Forum looks at how technology can be used to achieve the growth goals and objectives of organisations, consumers and society as a whole.

Through a series of reports, events and sharing of expertise, it seeks to inform leaders on how they can harness technology to make real change.

💾 View and download the entire report

🔎 Visit the Tech for Growth Forum hub for more from this series

Even as token prices fall, heavy users of AI are paying more, mainly due to the proliferation of agents that demand high computation loads.

“The entire dynamics change with the introduction of agentic workflows,” says Sara Hooker, a co-founder and chief executive of Adaption Labs, a frontier AI company that specialises in innovative solutions. Before the advent of agents, she says, a single query made to a single model produced a straightforward answer, but with agentic workflows everything is “long horizon”.

“You might ask a model to figure out how to get you efficiently from point A to point B, and it is calling a ton of tools and then it decides, whoa, I’m not quite sure about my answer. So it pulls a ton more tools.” Essentially, the agent looks up data until it is satisfied that its answer is complete, a process that agents do autonomously, which is hard to interrupt.

This has utterly changed the dynamics of pricing, increasing unpredictability. “Depending on the complexity of the task, it is going to be way more tokens.” As a result, Hooker says, the amounts being billed are much greater.

The rising expenditure on AI has exposed the amount of computing power that companies chug through daily: collective outlay has climbed exponentially.

Ramp, the AI vendor management platform, says spending on AI in the 18 months to June 2026 grew nearly sixfold as companies consumed more tokens with more complex tasks. Enterprises’ aggregate expenditure on AI computing has risen by 60 per cent in the past nine months.

Even mainstream businesses are spending more. Median expenditure per employee per month (PEPM), which covers most companies in the early adoption phase, has risen from $2.86 in January 2024 to $11.95 in July 2026. For the top 1 per cent of spenders, those that have embraced the technology, the median PEPM figure rose from $755 to $7,400. JP Morgan, citing data published in June by Ramp AI, says the top 1 per cent of AI users, generally those that perform complex coding tasks, consume 600 times as many tokens as a median user.

Hooker believes token usage will continue to grow and that prices will rise because of the imminent initial public offerings of Anthropic and OpenAI, which will make the companies accountable to investors.

“There’s a clear understanding in the industry that tokens are not priced properly right now,” Hooker says. “Once the IPOs happen, there is going to be immediate pressure on those margins to be more sensible.”

For those who argue that tokens will soon be commoditised because of the huge capacity build-out, Hooker says this will be easily absorbed by the growing requirements of agentic AI. “It is a completely different breed of workload. Agentic workflows are much more compute-intensive and much more unpredictable about the amount of compute they need.”

Providers of computing capacity “are in a market where they can set the price of how they sell compute to frontier labs”. This will have knock-on effects for pricing along the chain.

Value shift

The imperative for frontier models to make a return has changed the way that enterprises use the technology. With AI usage costing serious money, return on deployment becomes a greater consideration; gone are the days of free experimentation with pilot programs.

At the same time, the value delivered by increasingly complex and expensive AI is harder to identify. Agents may improve productivity but as their reasoning and tool-use chains become more elaborate, costs could outweigh benefits.

Measuring the value of AI is notoriously difficult, especially with regard to complex tasks such as coding, but some proxies can help. One way to do this is to look at the time saved on tasks previously carried out without AI. Reimagined workflows can save time across entire teams and mapping savings against expenditure will give an idea of AI’s worth to a company.

How to control the cost

The adoption of AI is no longer optional, so managing its cost is critical. Enterprises will need to control token purchasing even though pricing is uncertain and token requirements vary widely, making costs hard to estimate.

Companies can manage costs by shopping around. Dave Blair, the chief executive of Ridgeline, which builds an AI-native front-to-back platform for asset and wealth managers, says more models now offer quality options. “In the past six months, Claude had something better than the competition and people were willing to pay for it,” he said.

However, Claude’s lead is reducing and not every company can afford the $300mn that Salesforce expects to pay Anthropic in 2026. This sum might be an outlier but higher costs have nevertheless pushed companies to look elsewhere.

Blair says: “The other model providers continue to raise capability and then you can look at what models you use: people are looking to [open-weight models] to save money.”

While this comes with concerns over where you source, host and access these models – especially Chinese AI, given the perceived security risk and political restrictions – “at the end of the day they will use open source and cut their costs”.

This is backed by data. The Ramp AI Index showed that in July 2026 more than 6 per cent of AI-adopting businesses used open-weight AI, up from 4.5 per cent in January.

Open-weight models, such as DeepSeek or Meta’s Llama, allow users to host the technology on their own server and to fine-tune the system to specific requirements using proprietary data. Closed-weight models – OpenAI’s GPT series and Anthropic’s Claude, for instance – disallow such access, so the only way users can fine-tune responses is by refining prompts.

Another approach is to fit the tool to the task: basic functions do not need the most complex, costly frontier model. “There is a lot of overreliance on the best model … and it’s simply not needed for most things,” Blair says.

With token prices in the same model family covering a wide range – Claude models charge from $1 to $10 per million input tokens and from $4 to $50 per million output tokens – using simpler models can make a big difference. This can be facilitated by model routing, using software such as OpenRouter, which directs a prompt towards the most economical model for the complexity of the problem.

Enterprises can also curb inefficiency by relying on their institutional knowledge. Daniel Fallmann, the CEO of Mindbreeze, an enterprise AI and knowledge management engine, says the tendency to give an AI query more context and have it search for all variants of an answer ignores whether this is an efficient approach, contributing to high AI bills.

“Right now we all celebrate if there’s a model that is capable of accepting more context, but we are feeding it with everything, letting it select what it finds useful and then come back with the correct information.” The fear of losing competitive advantage encourages users to query numerous models for the same topic, which is similarly wasteful.

Mindbreeze operates in the “insight engine” sector where providers help enterprise employees to access corporate knowledge. It creates and manages institutional repositories, built using proprietary data, which answer internal queries based on the host’s own context and history. In addition the system can automatically add information retrieved from external searches, preventing duplicate processing and cutting the computing power needed for new queries on similar topics.

Without this layer, a similar query asked by different employees would require the same context to be processed each time. It could result in an agent calling up not only the latest contract but also all previous versions. “That’s somewhat strange, because it’s not only the amount of tokens that you consume and pay for … but it’s also the computation power, the energy usage, the data-centre growth that’s needed to handle all that context again,” Fallmann says.

Caching such queries and leaning more on the knowledge available internally allows companies to become more efficient, serving answers tailored to their needs while minimising token wastage.

Monitoring employee usage is key to preventing runaway spend, which is where the measure “per employee per month” comes into its own: companies can see when individuals run up large token bills versus their peer averages and decide whether they have achieved something or simply used it incorrectly.

Blair of Ridgeline says: “There’s a whole new field now to harness the context engineering, because you can use Claude really stupidly and burn a lot of tokens. Now you have to be a lot more intelligent in your approach to using it and to using it as efficiently as possible.”

The incentive to build

Hooker, whose company helps businesses to create more customised and adaptable models, says open-source variants do not necessarily solve the conundrum of computing-power requirements.

“Even an open-weights model can be highly inefficient if it’s super general. We already can see [this] with the latest models like Kimi [from Chinese AI company Moonshot].These are 1tn parameter models, so they are also incredibly compute hungry.”

For many enterprises that carry out defined tasks, the computation burden can be reduced dramatically by creating a bespoke AI model. It can also be wiser, especially for niche operators, to avoid passing data to an external developer, some of which have used acquired knowledge to operate directly in the sectors they once worked for.

“Companies are realising they want to [build and] own their own models not just for the token dynamics but because the medium-term advantage is that they own the data. Why provide it to another company,” Hooker asks.

Building a model is not hard anymore. The ability to teach models to self-train, automate data and automatically self-improve was not present even a year ago, Hooker says, and these capabilities have de-risked development.

Building your own model also lowers the risk of exposure to pricing for major-name frontier models, which could change at any moment. “People don’t want to depend on a handful of companies,” Hooker says. They are now more interested in in-house development as a hedge against rising token costs than they are in sovereign AI, the previous justification for going in-house.

Large enterprises that adopted AI early and now face high bills have a clear incentive to build but there is also an opportunity for mid-tier companies to skip the teething pains of using frontier models. The technology is moving so rapidly that slower movers could end up with the advantage of going straight to building their own models.

Where next?

The future of AI token pricing is at a critical point. On the one hand, expanding infrastructure capacity could presage the commoditisation of tokens, which may drive down prices. On the other hand, higher demand for computing power, coinciding with IPOs by frontier-model developers, will see providers impose higher prices to give funders a return.

Regardless of which force prevails, the trend for enterprises committed to AI deployment remains the same: AI expenditure is growing rapidly, demanding a rethink of capital deployment.

Forward-thinking organisations are moving from viewing AI as an IT overhead expense to a primary driver of operational capacity. Executive teams are beginning to substitute traditional labour costs with token consumption – a shift accelerating rapidly in specialised fields such as software development. Highlighting the scale of this migration, Bain, the US consultancy, expects that enterprises will eventually reorientate 20 to 30 per cent of the cost of headcount directly into token spend.

However, substituting headcount budgets for token infrastructure introduces operational risk. The capital reallocation must translate into value so companies will need strict oversight, monitoring AI expenditure and creating safeguards against spurious token consumption.

The winners of this transition will be the enterprises that understand why they deploy AI and have a way to measure gain.