AI is getting more expensive, and it will keep getting more expensive for the foreseeable future. If you control for capabilities, the cost per token is actually falling. But the capabilities you need to stay competitive keep increasing, which means your AI bill is going up regardless.
At TJ Digital, we run AI through every workflow across roughly 40 to 50 active client campaigns. Our token costs have increased meaningfully over the past year, and we expect them to keep climbing.
The math still works for us because we’ve built systems that produce about four times the workload at the same rates as a traditional agency. But that ratio only holds if you’re deliberate about how you’re using AI.
Here’s the thing. The question that matters for most businesses is whether the output justifies the spend. Reducing your AI costs is the wrong goal if it means falling behind.
Why Is AI Getting More Expensive?
You might assume AI would follow the same trajectory as most technology, where prices drop over time. Hardware costs have done exactly that. Computing and software prices are roughly 74% lower today than they were in 1997.
AI inference is different. The latest frontier models (GPT-5.5, Claude Opus, Gemini) launch at higher per-token prices than their predecessors. OpenAI’s GPT-5.5 debuted in 2026 at roughly double the rates of the previous version.
The reason is a compute capacity wall. AI demand is growing faster than providers can build data centers. Experts estimate that if usage grows 10x, providers need approximately 100x more compute to keep up. That gap between demand and supply is what pushes prices up.
You can absolutely use cheaper models. Google’s Gemma 4, for example, is incredibly capable and runs at a fraction of the cost of frontier models.
In independent benchmarks, Gemma 4 ran complex business simulations at about $0.20 per run, compared to $3 to $8 for top proprietary models. It even outperformed several frontier models on return-on-investment despite costing 15 to 40x less.
But here’s the problem. If you lock yourself into last year’s model capabilities while your competitors are using the latest models, you fall behind.
The cheaper open-source models give you “yesterday’s frontier” quality. That’s good enough for many tasks, but not if your competitors are using the newest tools to move faster and produce better work.
How AI Token Costs Compare to Payroll
The companies at the forefront of AI adoption are now targeting token costs at about 50% of their overall payroll costs. Nvidia’s CEO Jensen Huang has suggested that engineers should be spending roughly half their salary on AI tokens. Nvidia’s Bryan Catanzaro has said the cost of compute on his team already exceeds the cost of the employees.
One to two years from now, I don’t think it will be unusual for token costs to exceed payroll costs entirely. Some companies are already burning through AI budgets in weeks. CFOs increasingly treat token expenditure as a major line item on the same level as payroll.
| Cost Category | Current State | 1-2 Year Outlook |
| AI token costs | ~50% of payroll at leading companies | May exceed payroll entirely |
| Frontier model pricing | Rising with each new release | Continued increases as demand outpaces supply |
| Open-source model pricing | Extremely cheap (15-40x less than frontier) | Will stay cheap, but capabilities lag behind frontier |
| Traditional payroll | Stable | May shrink as AI handles more tasks |
This raises an obvious question. If payroll costs are effectively doubling when you add AI, what has to be true for that math to work out?
How Much More Output Do You Actually Need?
At first, you might think if costs are doubling then output has to double. But why would anyone go through the effort of implementing AI if the end result is just twice the payroll for twice the output? For this to make sense, the tokens have to be more cost-efficient than payroll.
So maybe output has to be something like three times what it is now. But there’s another factor. If all companies are becoming more efficient (or at least those that are staying competitive), then what you can charge for the same output is going to go down.
This is already happening across digital services. AI is compressing pricing across software categories. Automated workflows mean customers can get the same content, code, or analysis at far lower marginal cost.
As one analyst put it, if your token cost rises 40%, did your execution velocity rise 40%? If not, you’re just automating noise.
In reality, your output probably needs to go up by four to eight times in the next one to two years with the same payroll. To be clear, I’m referring to work that can be done on a computer. Plumbers don’t need to increase their output by eight times.
But for anyone in a knowledge-work industry, that’s the math.
Where Does the Extra Demand Come From?
That’s the supply side. But where does all that demand come from? How are you going to find four to eight times as many customers?
That part is actually simpler than it sounds. Most of your competitors aren’t going to do this. They’re not going to stay competitive.
Their customers become your customers. A smaller and smaller group of companies will eat up more and more of the market share.
This is the consolidation effect that market analysts are already flagging. Growth-stage companies need to decide now how to use AI, because these choices will determine whether they scale into category leadership or face consolidation pressure. The companies with the highest intelligence density per dollar will take the most market share.
We’re already seeing this in digital marketing. Traditional agencies running the same playbook from 2020 are losing clients to AI-native agencies that can deliver more work at lower cost. The pattern will repeat across every knowledge-work industry.
What Should Businesses Do About Rising AI Costs?
There are three things I’d recommend to any business owner thinking about this.
Build your AI systems around documentation, not individual tools. The specific models will keep changing. What won’t change is your need for organized, structured information about your business that AI can use.
At TJ Digital, we call this the Brand Ambassador. It’s a curated set of documents that teaches AI everything about a client’s brand. The better your documentation, the more value you extract from every token you spend.
Use the right model for the right task. You don’t need frontier models for everything. Use cheap, capable open-source models for routine tasks. Save the expensive frontier models for work that requires their capabilities.
This is how you control costs without falling behind.
Measure output, not spend. Track what your AI investment actually produces. If you’re spending more on tokens but not seeing a proportional increase in output quality or volume, something is wrong with your process, not your budget.
Will AI Costs Eventually Come Down?
If you’re okay sticking with the capabilities of today’s frontier models, then yes, you can expect those same capabilities to get cheaper over time. That’s how technology works. What was top-of-the-line last year becomes commodity pricing this year.
The problem is that six months from now, today’s capabilities won’t be enough. If your competitors are using the latest models and you’re not, you’re not going to keep up. So the effective cost for staying competitive will keep rising even as the absolute cost of any given capability falls.
The best analogy is smartphones. A phone from 2020 is incredibly cheap now. It’s also not good enough if your business depends on having the latest tools and apps.
How an AI-Native Agency Manages Token Costs
We use AI in every workflow, with humans reviewing everything. That combination lets us deliver more work at better quality for less money than a traditional agency would charge.
Our approach is built on three principles. First, we invest heavily in documentation and knowledge systems so every token produces maximum value. Second, we’re transparent about using AI. We don’t hide it or pretend our work is done entirely by hand.
Third, we stay on the frontier. We test every major AI tool the moment it releases, and we pay for all the major platforms because that’s the only way to know what’s actually best for each type of task.
The economics of AI are changing fast. But for businesses that build the right systems and keep up, the rising cost of AI is still far cheaper than the cost of falling behind.
Talk to an AI-Native Marketing Agency
If you want to see how an AI-native agency approaches your marketing, we’d like to hear from you. We start every engagement with a two-week strategic assessment where we build your Brand Ambassador and put together a data-driven plan.