Tech Chat – July 2026
Just a token cost
Many moons ago in my first week at University I called home to say everything was okay. As I chatted to the parents on a payphone, I casually mentioned I’d had a RAMC medical and they’d discovered something wrong with my heart. The pips sounded – I had no more change – and the call ended.
I didn’t call home again for another 3 or 4 weeks. Despite messages and letters to ask me to call, I had plenty to do (and certainly not lectures) and so just never quite got round to it. Mum and Dad were distraught. As a result, I was one of the first to get a mobile – forced upon me. It was Christmas 1990. Yes – what a xxxxxx.
It was a novelty and a luxury. Call charges were over 40p a minute to use a bricklike device that looked like a military satphone handset. It was analogue, and just for voice calls – SMS was yet to come. The choice was between two networks, Vodafone (then part of Racal Telecom still), and BT’s Cellnet.
This brings to mind how voice and data charging developed – on a consumption basis, satisfying the initial capex of the deployment of the original mobile networks. Yet as voice and data commoditised, all-you-can-eat bundles evolved. Most of us will now be used to not even considering data caps, merely the routine monthly cost – while voice caps surely don’t concern anyone anymore, and competing landline services migrated to the same model.
Echoing these previous technologies, the chatter around AI should by now have evolved from hysteria, through apocalypse – and ultimately to cost. Every new technology follows a similar path: experimentation, proliferation and then optimisation. AI is entering that third phase. With the return to a consumption-based world for AI, among opex bases used to “all-you-can-eat” per seat charging, companies driving AI usage seem to have overlooked this: having set KPIs based on encouraging AI among their engineers, urging experimentation and take-up and ranking engineers based on token consumption. That worked well while costs were obscured by enterprise licences and innovation budgets. The challenge emerged when experimentation became habitual and token consumption became measurable.
Lesson learned: from 6th July, Tesla has capped AI spend on third-party AI tools (now $200 per week) compared with the spend of thousands of dollars per week the most “experimental” were consuming. Grok and Composer, platforms on xAI, have no budget counter, but even Musk employed engineers prefer Claude.
Uber had a similar experience: restriction-free consumption of AI burnt through the annual budget by April. Employees now have a $1500 monthly cap. www.thestreet.com reports all of this, (and cites similar restrictions at Meta, Amazon, Walmart and Coinbase). Microsoft has withdrawn some Claude licences and steered staff towards lower cost Copilot workflows; Atlassian has withdrawn unlimited access.
The pattern is familiar. First comes encouragement and adoption. Next, AI tools become embedded in daily workflows. Finally, the organisation discovers who the power users are and how much those users cost. Governance inevitably follows: individual spending caps, team budgets, approvals for premium models and a search for lower-cost alternatives, en route to commoditisation.
As “bill checkers” once made money auditing telecoms invoices before evolving into cloud-cost optimisation specialists, a similar ecosystem is beginning to emerge around AI token usage. ActiveOps recently highlighted AI cost control as an increasingly popular function of its ControliQ platform.
Threshold billing is the logical end game that most have, or will, arrive at: Dotdigital has long been a proponent, where revenue in their case is based on client payment for a set number of communications per month, at a fixed monthly cost, on a use-it-or-lose-it basis. Burst through the limit and pay premium rates, or ideally increase your recurring quota and increase DOTD’s monthly revenue. The more effective the DOTD platform is, with deployment of new AI agents, the more messages will be sent, even if fewer marketers are needed. AI deployment follows the same pattern: do what you do better, creating efficiency and delivering time.
This is the future for token consumption, returning to the mobile phone analogy: a fixed number of tokens, where eventually a provider will break ranks and return to an all-you-can-eat basis, with “reasonable use” restrictions.
As with the SaaS-pocalypse, initial hysteria needs to calm down. SaaS businesses with customer trust, domain expertise and budgets for development will stay ahead of their game and use AI to its best use (with caveats for content production). Current token-use concerns will calm down. Who knows, an LLM may one day write the code to create a cheaper LLM, instantly accelerating the commoditisation process. Initial excitement reverts to calm.
My son regularly laments that he burns through his monthly Claude tokens too quickly during his A Level project research, which has sharpened his approach to writing more precise prompts. This is the type of efficiency that will come from cost awareness, saving money for both him, and the AI providers when reasonable use allowances arrive.
Mum and Dad have forgiven but never forgotten the heart-scare phone call (it came to nothing beyond an overzealous doctor). It’s striking the extent that popular technology has changed beyond recognition since Christmas 1990. The economic behaviour behind it has not. Whether voice minutes, cloud compute or AI tokens, the pattern is remarkably consistent: enthusiasm, adoption, cost awareness and eventually commoditisation.
Happy Friday