This article explores the complex (and somewhat terrifying) nuances of the tipping point where a tool stops being a massive efficiency boost and starts becoming an expensive infrastructure dependency.
From the mechanical rhythm of the typewriter to the ubiquitous screen of the smartphone, the evolution of modern technology has been driven by a singular, relentless pursuit: eliminating friction from routine tasks. Each leap forward shifted the boundaries of human productivity. The typewriter transformed messy, handwritten manuscripts into standardized, legible documents; the telephone bridged vast physical distances, turning days of mail transit into instantaneous voice communication. When computers arrived, they digitized filing cabinets and automated complex calculations via spreadsheets, laying the groundwork for the smartphone to ultimately consolidate all of these tools into a single, pocket-sized device.
For both individuals and organisations, embracing these advancements has never just been about adopting new gadgets–it has been a continuous strategy to offload cognitive and administrative burdens, freeing up time to focus on higher-level creativity, strategy, and connection.
Since the advent of the computer, each of these gadgets has come with an operating system (OS). The OS is the main system software that manages a computer’s hardware, memory, and files, and allows other programs to run. Examples of operating systems include Microsoft Windows, macOS, and Linux.
In contrast, Microsoft Office is a software suite (a collection of applications) which includes Word, Excel, and PowerPoint, and runs on top of an operating system. These applications require an existing OS to function. Today, many operating systems are provided free of charge when you purchase a gadget, or they can be downloaded for free. However, much application software must be purchased, either via a one-time fee, a time-based subscription, or usage-based tokens–though many free alternatives also exist.
To complicate matters for the uninitiated, we have Google (the search engine/company) and Wikipedia (a website). These are not software programs that run on top of an OS in the traditional sense; they are web-based services accessed via a browser like Chrome or Edge.
AI Platforms
A modern evolution of application software includes generative AI platforms like Gemini, ChatGPT, and Claude. Like traditional suites, these are applications that run on top of an OS, though they operate primarily as cloud-based web applications accessed via an internet browser.
While they are technically applications, using the word “software” doesn’t quite capture their scale, because most of the “thinking” doesn’t happen on one’s actual gadget. It happens on massive external servers.
Because of this, the tech industry usually describes them using three specific terms:
Generative AI Applications: This is the most common consumer term. It means they are application software designed to generate new content–like text, code, or images–based on human prompts.
Cloud-Based Web Applications: Here, individuals and organisations do not need a powerful computer to run Gemini or ChatGPT because the software lives in the “cloud”, i.e. on Google or OpenAI’s cloud servers. One’s gadget is just a window showing the results.
AI Platforms / Ecosystems: These are increasingly called “platforms” because they are not just single tools anymore. They are ecosystems where Individuals and organisations can build their own custom chatbots, analyse massive data files, or connect to other software (like Google Workspace or Microsoft 365).
AI Pricing Structures
The AI platforms listed above perfectly mirror modern software pricing structures discussed earlier: they offer (1) free basic access,(2) time-based monthly subscriptions for advanced capabilities, and (3) usage-based “token” pricing for developers integrating AI into their own systems. All three pricing structures use a process called tokenization where raw input data is chopped into fundamental building blocks that a machine learning model can mathematically digest. Tokens are basically the thought units of AI – every word it reads, every word it writes, every decision it makes relies on a token.
As a rule of thumb for English text, one token averages out to about four characters or three-quarters of a word, meaning common phrases remain whole while rare or complex terms are split into multiple fragments. For multimodal inputs, calculation methods depend entirely on the data type: images are typically mapped to a set number of tokens based on their pixel dimensions and resolution grids, whereas audio and video files are calculated at fixed token rates per second of playback. [See Appendix One for more details on Tokenization].
Note that daily or time-based usage limits for individual and organizations on monthly subscriptions are calculated fundamentally differently from standard pay-as-you-go API keys. While the underlying AI models always process information using tokens, subscription-based tiers wrap this technical reality into user-facing constraints to manage server demand and prevent platform abuse. Thus, Individuals and organizations using basic AI access and time-based monthly subscriptions have a daily limit, and once reached, the user is shut out from further access for a specified time period.
A serious Claude Code user runs through roughly 10 billion tokens a year. If such a person paid for that usage through a standard API, those 10 billion tokens would cost him or her around $15,000 a year. That is the real, unsubsidised price. No discounts, no incentives, just the raw compute costs.
Now, that same user on a flat rate Max subscription pays around US$100 per month or US$1,200 for an entire year for the same workload. From $15,000 down to $1,200 – a 92% discount.
How is this possible?
Enter the murky world of ‘Tokenization’. This murky world leads us to the AI pricing trap that is the subject of this article.
The Cheap Pricing Illusion
Individuals and organisations with AI subscriptions ranging from US$22 to US$100 appear to have the “deal of the century”, because, providing even a low-level user using tools like Claude Code is estimated to actually costs about US$15,000 a year to run for Anthropic, the company providing the platform. However, the subscriber is only paying a fraction of that because Venture Capitalists (VCs) are footing the bill.
Imagine walking into a dealership, picking out a car priced at $15,000, and being told you only owe $1,200 because someone somewhere else covered the rest. It does not make sense, and that is the illusion created by this pricing model.
This introductory price is often referred to as a “honeymoon period” or “bait-and-switch” depending on how deceptive it is. These pricing traps of selling a service below cost are done to attracts a massive customer base and hook one into an ecosystem. The costs are usually subsidized by future price hikes. However, in the case of AI platform, most likely that the usage-based “token” pricing developers will be subsidizing the ‘free basic access’ and ‘time-based monthly subscription’ users.
For today’s developers, this is a temporary illusion built to get them hooked before the price tags change. But the money is running out. There will be temporary fixes like the IPO of Chat GPT that will rake in trillions, but when this house of cards eventually collapses, the tools developers rely on every day will either vanish or cost them10 times more.
How long will this illusion last?
The answer lies in OpenAI’s own financial projections that show that the company is on track to lose $14 billion in 2026 ( Muppidi and Palazzolo, 2025). This means that a $22 monthly subscription covers about 1.7% of what an active power user actually costs to serve. Clearly, such users are not customers, they are bait. Every prompt typed, every line of code generated, every late-night chat session is being paid for by investors, and they are betting that nobody will be able to live without this product when the real bill finally lands.
Whole industries are being signed up at a loss. Law firms running document review at 5 cents on the dollar. Marketing agencies turning out campaigns at prices that would have been impossible 18 months ago. Hospitals trialling diagnostic tools at sticker prices that no model provider could actually sustain at scale. Every single deal is being propped up by patient venture capital that expects 10 times returns. If companies are losing money on every user they sign up, why are they racing to sign up more? Because we have seen this exact playbook before, and we know how it ends.
Freemium pricing
Companies commonly use “freemium” models, conventionally called penetration pricing, of charging artificially low prices or nothing at all to capture a massive audience. Once customers rely on the product, the company raises prices. There are notable examples of companies that successfully used this strategy. A good example is Uber, that subsidized rides with massive discounts to kill local taxi competition. Prices then surged dramatically to maximize profits. Spotify grew its user base by offering an ad-free trial and cheaper family plans and has since steadily raised the standard monthly fees across global markets. Adobe allowed users to buy perpetual software licenses once. They then shifted to a subscription, eventually raising the overall cost of ownership for long-term users.
For example, Uber’s take rate, i.e. the slice of every fare the company keeps, was, in 2022, around 32 cents of every dollar a rider paid. By 2024, that figure had climbed to roughly 42 cents. Drivers got a smaller share. Riders paid more. After years of spending billions of investors’ cash to expand, the company eventually turned $1.8bn loss into $1.1bn profit in 2023 (Jolly and Wearden, 2024).
Now it is happening in AI. It is the same investors, the same playbook, and the same pricing memo. Industry analysts expect consumer subscription tiers to roughly double in price over the next two years. Anthropic has rolled out new rate limits that gently push power users toward higher priced plans. Google is testing premium-only Gemini features that used to be free. A 100% price hike is not a rumour. It is already pencilled in on the calendar.
Enterprise contracts are following the same curve. Custom deals signed in 2024 are being quoted much higher at 2026 renewals. It is the same product, just costing multiple times the price. But there is a key difference between ride sharing and AI. Ride sharing only had to do one thing: move a car from point A to point B. The cost of doing that does not explode as usage rises. More drivers, more density, better routing – if anything it gets more efficient. AI works differently. The underlying math of “thinking” does not get cheaper in the same way.
AI executives continue to say that computing is getting cheaper every year, and the unit economics will work out over time. It is not exactly a lie – it is more like a half-truth. It is a fact that the price of running a query through a model has dropped year-over-year. Chips are more efficient. Models are leaner. Each individual word an AI generates is genuinely cheaper to produce than 18 months ago. And that is the part AI executives want people to hear.
Now here is the part they do not tell you.
Token Tax
Modern agentic workflows – the kind that power Claude Code and ChatGPT’s deep research tools – burn through anywhere from 5 to 30 times more tokens than simple chat sessions of two years ago. When you ask a code assistant to fix a bug, it doesn’t write 50 words of response. It quietly spawns subtasks. Then it rereads your files. It checks its own work. It writes draft after draft, throws most of them away, and quietly runs tests in the background. A single user request can chew through hundreds of thousands of tokens before any answer shows up.
A model might be slightly cheaper per word than before, but it is also producing far more words per request. The total bill is shooting upward. This is known as the token tax. It is bankrupting scrappy AI startups burning through their seed rounds, and it is threatening to wipe out one of the most profitable business models in the history of the internet – Google’s search model.
Killing the Search Goose
For 25 years, Google has printed money through a brutally simple formula. A user types in a query, Google returns 10 blue links, and the total cost to Google is a fraction of a cent per search. The ads next to those results generate much more than that – the margin has been one of the great financial miracles of modern times. Now Google is rebuilding that entire system on top of generative AI.
A single AI-powered search response – the kind that writes a paragraph-long answer instead of showing links – costs significantly more to produce than a traditional keyword search. Multiply that across billions of queries a day, and the most reliable profit machine of the 21st century begins to strain. If Google fully replaces traditional search with AI overviews, the margins that have funded YouTube, Android, Waymo, and Gmail begin to dry up. It does not take much scenario modelling for financial analysts and management accountants to map out the worst-case scenarios, and the results are catastrophic for Alphabet (the parent company of Google).
It does not take much to envisage that Google’s advertising model may soon become redundant too. When AI just gives you an answer, nobody clicks on the links, so why would advertisers pay? Google is staring at a future where it serves more queries than ever before, costs more to run than ever before, and earns less revenue per query than at any point in its modern history. Tech giants are willingly cannibalising their most profitable businesses. They have decided the only thing more dangerous than killing a cash cow is letting a competitor kill it first. Business school has a name for this: the innovator’s dilemma (Christensen,1997).
This dilemma is based on the theory that the adoption rate of innovations is non-linear; it is slow at first, then rapidly rises before flattening out again as it reaches market saturation. Such trajectories of growth are commonly known as the S- curve. There are many examples of great companies that refused to kill their cash-cow and their demise was fast. Kodak is a prime example of the S-Curve and the innovator’s dilemma, where it refused to cannibalize its highly profitable legacy business model (analogue film) itself in favour of a disruptive technology (digital cameras) at that time (Lucas and Goh, 2009).
The Cost vs. Quality Compromise
What happens when a product or service becomes too expensive to provide at a competitive price? Why, you just reduce the quality.
An AI model used to one-shot your code. Now it forgets your project halfway through. A chatbot used to write five paragraphs at a stretch. Now it cuts off at three. An image generator that used to render a flawless portrait in 30 seconds now spits out something with seven fingers and asks for an upgrade to the next tier. Nobody is imagining these things. The product is getting worse.
When the numbers stop working, the easiest lever a provider can pull is to quietly water the service down. Message caps that used to refresh every 5 hours suddenly refresh every 8. The default model in an app gets quietly swapped from a flagship to a smaller, cheaper version. Memory features get rolled back. Advanced reasoning gets locked behind a higher price tier. Reddit threads about AI tools are full of users who swear their assistant has gotten lazier. Engineers are posting side-by-side screenshots showing the same product producing visibly worse output than 6 months earlier. Companies almost always deny it, sometimes releasing selected benchmarks – clean prompts, controlled conditions, optimised scenarios designed to demonstrate performance at its best. It buys them some time, but it does not fix the bigger problem.
The Financing Roundtrip Scam
If unit economics are this bad, such that quality and service delivery are being compromised, how are these same AI companies posting record AI revenues on Wall Street every single quarter? That is where the financing sleight of hand gets interesting. It is called ‘Round-tripping’, and it no different to methods used by money launderers and bitcoin investors.
Here is how it works. Microsoft commits very publicly to investing $13 billion into OpenAI. The press release is slick, the headlines dramatic, stock prices rise. But read the fine print and a different story shows up. A big chunk of that investment never actually hits OpenAI’s bank account. It arrives in the form of Azure cloud credits – essentially a gift card that can only be redeemed at Microsoft’s own data centres. OpenAI records that sum on its balance sheet as capital raised; i.e. debit Azure cloud credits (asset) and credit share capital (equity). In Microsoft’s books it shows a debit of share investment (asset) and the cloud usage is recorded as a credit (revenue). It is an investment and a sale at the same time. Wall street is happy!
OpenAI has separately committed to spending up to $250 billion on Azure services, locking the loop in for years to come. Then Nvidia announces tens of billions in commitments to OpenAI, OpenAI uses that capital to buy Nvidia GPUs, Nvidia’s quarterly revenue posts a record, and their stock price soars. The whole cycle takes a few months and almost no real money has actually changed hands – it has simply been given a different name at each stop. Add Oracle, CoreWeave, and AMD to the list. Each company invests and then sells services to the next, recording revenue as the same dollar flows through the cycle.
The technical name for this in money laundering jargon is ‘round-tripping’, i.e. a circular scheme where illicit funds are sent out, routed through multiple offshore accounts or shell companies, and returned to the perpetrator’s home country masked as “clean” wealth, such as legitimate foreign investments, consulting fees, or loans (Karhunen, et.al., 2021).
In Silicon Valley, it is called a strategic partnership.
The Hardware Debt Trap
In 2026, big tech is projected to spend roughly $700 billion to $1 trillion on AI infrastructure. Data centres, GPUs, cooling systems, power delivery – entire grids are being reinforced to handle it. Meanwhile, total global consumer spending on AI services is only about $12 billion. Hundreds of billions are flowing out while only $12 billion flows in. The gap is the size of an entire mid-sized country’s economy, and it is being filled not with revenue but with debt: corporate bonds, structured credit, and private lending (Lee and Greenbaum, 2026).
Meta alone raised $30 billion in bond markets in late 2025, with another roughly $30 billion through a Morgan Stanley-arranged joint venture set up to keep liabilities off Meta’s public balance sheet. Microsoft has signed a 20-year power purchase agreement to restart Three-Mile Island (the site of a nuclear accident in 1979). Google has partnered with NextEra Energy to reopen nuclear power plants. These promises don’t go away if AI revenue underperforms.
The hardware itself does not last. A high-end Nvidia GPU has a useful life of just one to three years before the next generation makes it outdated, losing most of its book value the moment a new generation hits the market – which now happens roughly every 18 months. Compare that to the dotcom bust. When that bubble popped in 2000, telecom companies left behind millions of miles of fibre optic cable buried in the ground. New companies bought it for pennies on the dollar and built YouTube, Netflix, and Spotify on top of it. The crash was brutal, but the wreckage was useful. This AI bubble will leave behind warehouses full of useless silicon, locked into 20-year power contracts and concrete shells in the middle of nowhere (De Vynck, 2024). Utilities will pass higher electricity rates on to households for decades, regardless of whether the AI revenues ever show up.
The Mass AI Extinction
Roughly 40% of AI startups launched in 2024 have already been shut down or “Acqui-hired” by bigger players (CB Insights, 2026). Acqui-hire is the polite term for a fire sale where a struggling company is sold for cents on the dollar to a rival. The buyer is not really buying a business – they are getting the engineers, shutting down the product, and absorbing whatever talent they can. These were not hobby projects. These were companies that closed Series A rounds with serious investors. They had revenue, paying customers, and glowing TechCrunch profiles. Then within 18 months, the lights went off.
The reason is almost always the same. Their cost of goods sold – the money they pay to model providers like OpenAI, Anthropic, and Google – is so high it wipes out any margin they could hope to charge. A startup that wrapped a polished interface around GPT-4 might charge $50 a month, but the API usage that same customer generates can cost the startup $80. Every active user is negative revenue. The more successful the marketing, the faster the company bleeds out. An impossible management accounting dilemma.
When a foundation model provider releases a new feature, it often kills 10 startups overnight. ChatGPT launches native voice mode – say goodbye to half a dozen voice agent startups that closed Series A rounds last quarter. Claude releases native PDF reading – a whole crop of document tools became useless in a single product update. An ecosystem of independent AI companies is falling apart under the weight of compute costs that nobody can profitably absorb. When startups die, cloud providers lose the round-trip revenue that made foundation model investments look like good business in the first place. And that is when the final phase begins.
The Finding New AI Drug Pushers
Venture capital firms are no longer willing to cover losses in the hope of future glory. They want to see a path to profit in writing, with quarterly milestones, and they want to see it now. For foundation model companies, that means one of two options for the AI addicts.
The first is a brutal sudden repricing. A $20 consumer plan becomes a $100 plan, or it quietly disappears and is replaced by a pro tier that costs 10 times more for the same features. A Claude Code user who paid $1,200 a year suddenly faces an invoice closer to the $15,000 the API actually costs. A freelance designer who relies on a $10 image generation subscription gets an email explaining that their plan is moving to a new structure. Small businesses that built workflows on cheap AI face a choice: pay 10 times more or go ‘cold turkey’ and go back to doing it the old way. Most will opt for their AI addiction somehow to be fed, at whatever price!
The second option is worse for the addict. The AI services they are so dependent on simply gets shut down. We have already seen the first signs. Smaller AI companies have folded with 30 days’ notice, leaving customers scrambling to move years of work to whatever competitor is still standing. They must feed their AI addiction. Specialised models for legal research, medical imaging, and customer support have been pulled because their economics never worked. An era of cheap AI ends with a thousand small invoices and a thousand small shutdown notices.
The deeper truth is uglier than a price hike. AI in 2026 is on track to become a luxury, not a basic product. The cheap versions trained an entire generation to need it. The expensive version is the only one that balance sheets now allow to exist. Big companies that can afford the new pricing tier will lock in their advantage. Freelancers, small businesses, and the people who powered early adoption – the ones who created the buzz – will be priced out first.
Summary
History says crashes do not take a year to play out. The dotcom bust took two years from peak to trough. The AI bubble has more leverage, more concentration, and more debt baked into its foundations. When it tips, it can move in months – maybe weeks. When the margins shrink, when the first big enterprise customer publicly walks away from a renewal, that confidence can vanish overnight.
The tools millions rely on every day were never as cheap as anyone thought. They were being held up by investor money that is finally starting to dry up. An AI age might still be coming. But a cheap AI age – one that fooled an entire generation into rebuilding their working lives on top of it – is already over. The bill simply has not arrived yet. And when it does, that price will never feel real again.
References:
CB Insights (2026), The top 9 reasons startups fail, CB Insights, March 5. https://www.cbinsights.com/research/report/startup-failure-reasons-top/
Christensen, Clayton M. (1997), The Innovator’s Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business School Press, Boston, MA, p. 658.
De Vynck, Gerrit (2024), “The AI boom may unleash a global surge in electronic waste”, The Washington Post, October 29. https://www.washingtonpost.com/technology/2024/10/29/ai-electronic-waste-recycling/
Jolly, Jasper and Wearden, Graeme (2024), “Landmark moment as Uber unveils first annual profit as limited company”, The Guardian, Feb 8. https://www.theguardian.com/technology/2024/feb/07/landmark-moment-as-uber-unveils-first-annual-profit-as-limited-company
Karhunen, Päivi; Ledyaeva, Svetlana and Brouthers, Keith D. (2021), “Capital Round-Tripping: Determinants of Emerging Market Firm Investments into Offshore Financial Centers and Their Ethical Implications”, Journal of Business Ethics, 181(3): 117–137.
Lee, George and Greenbaum, Lucas (2026), “Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out”, Insights, Goldman-Sachs. https://www.goldmansachs.com/insights/articles/tracking-trillions-the-assumptions-shaping-scale-of-the-ai-build-out
Lucas, Henry C. and Goh, Jie Mein (2009), “Disruptive technology: How Kodak missed the digital photography revolution”, The Journal of Strategic Information Systems, 41(1):46-55.
Muppidi, Sri and Palazzolo, Stephanie (2025), “OpenAI Boosts Revenue Forecasts, Predicts $111 Billion More Cash Burn Through 2030”, The Information, Dec 30. https://www.theinformation.com/articles/openai-boost-revenue-forecasts-predicts-112-billion-cash-burn-2030