Rafomac News, Tech & Trends That MatterAI was supposed to slash costs and boost productivity. Instead, companies like Uber and Meta are burning through AI budgets faster than expected, turning AI into a costly challenge. What’s driving this, and how are businesses coping?
When Using AI Becomes a Spending Spree
The surge of AI adoption in recent years was hailed as a transformative force that would revolutionise work and cut expenses. But for many companies, it’s become a money pit. By May 2026, the US alone had seen nearly 97,000 job losses due to AI disruption. Yet ironically, those same AI tools—meant to save wages and time—are now adding to corporate budgets in unexpected ways.
Tech firms encouraged employees to use AI heavily, even embedding usage metrics into their performance reviews. This frenzy, dubbed “token maxing,” encouraged workers to maximise how much AI they used to appear ambitious, sparking a costly token consumption race.
Uber exhausted its entire yearly AI budget within four months. Meta’s token-related expenses reportedly topped $1 million per employee. Even smaller companies face mounting bills; one report noted monthly AI costs at around $50,000 for a mid-sized firm. If AI was once seen as a cost saver, it now borders on an economic headache.
Token maxing essentially means unrestrained use of AI model tokens—the tiny bits of data that power AI interactions. Each prompt or response, whether a simple greeting or a hefty document analysis, consumes tokens that translate directly into dollar signs. Costs that once were negligible have skyrocketed. For instance, in 2023, the cost of processing one million tokens on top AI models like GPT-3.5 hovered between $0.5 and $1.5. Today, spending that same amount on GPT-5.5 Pro’s outputs can cost an eye-watering $180.
Why Are AI Token Costs Spiking?
A big part of the problem is simple: most workplace tasks just don’t need the highest-powered model. Yet employees often default to the most advanced—and expensive—AI version because it promises better results. In reality, the quality difference for routine requests, like summarising meeting notes or correcting emails, is minimal.
Jensen Huang, Nvidia’s CEO, pointed out that if a software engineer’s AI token use doesn’t cost half their salary, it suggests underutilization. This highlights the ugly truth: AI giants like Nvidia profit when workers demand excessive token consumption. But productivity gains only justify the cost if the extra spending doubles an engineer’s output, not when it barely improves efficiency.
Practical Ways Companies Are Cutting AI Costs
The answer isn’t reducing AI use—it’s smarter usage. Companies are adopting multiple strategies to tame the token beast.
- Cheaper default models: Like setting a printer to black-and-white by default, organisations are deploying cost-effective AI models for mundane tasks, reserving premium ones for complex needs.
- Model routing: This system directs simple jobs to inexpensive models and escalates only challenging tasks to powerful AI, without changing the user experience.
- Caching repeated responses: AI systems reuse answers to common questions instead of regenerating them each time, saving loads of tokens.
- Keeping context lean: Since AI fees depend on conversation length, trimming old chat history and restarting sessions for new topics drastically cuts costs.
For example, instead of sending a hefty legal contract to the AI multiple times, caching means the system remembers and reuses the processed data. Similarly, starting fresh conversations limits needless expensive context processing. Even converting long chats into concise project instructions preserves consistency without hefty token bills.
Hardware Could Flip the AI Cost Equation
Typically, AI models run on cloud servers, charging for every token. But owning your AI hardware—called inference rigs—lets companies run AI locally, with costs mainly for electricity, not tokens. While building rigs with high-performance GPUs can be pricey upfront, it slashes recurring AI bills.
Even individuals can start small. A laptop GPU coupled with tools like LM Studio can run basic models like Google DeepMind’s Gemma offline, while more serious users grab devices like the Mac Studio, which has soared in popularity for this reason. On the large scale, corporate interest is exploding; inference hardware already accounts for 60 to 70% of AI compute demand, a steep rise from 40% just two years ago.
New Frontiers: Turning the AI Crisis into Opportunity
The AI boom isn’t just about smart models but building infrastructure to make AI usable and affordable. This wave is creating fresh business avenues. Experts root for companies that build token-optimising software, rent inference rigs, or offer services customizing open-source AI models for enterprise use.
Employees stand to benefit, too. Imagine a video editor whose AI usage halves company spending on tokens. If that saving is shared, they could earn an additional bonus without extra effort. Efficiency in AI use will soon become as valued as raw output.
This shift mirrors earlier tech surges, where not only the inventors but the enablers flourished—think cloud platforms and payment gateways during the internet boom. Today, companies making AI scalable and affordable will lead the pack.
The question now is: will you be part of this new wave or get drowned in its rising costs?
