The global risks of the AI illusion

On June 1st, thousands of software developers began their work week by logging into GitHub Copilot, the AI coding tool that had become a mainstay of their trade over the past several years. They were met with an unwelcome surprise. 

GitHub Copilot provides developers with a one-stop interface to use AI in their workplace. For many developers, this is simply a case of using AI to help them troubleshoot, review, and propose fixes for their in-progress code, using plain English prompts: jobs that once required a developer to spend many hours trawling through documentation and forum posts.

Notoriously, however, a growing cohort of developers have taken to outright delegating the job of writing and structuring code to Copilot. A significant share of developers in the corporate and startup worlds now habitually “vibe code” at work — rather than write code themselves, they tell Copilot what to do, wait as it generates a codebase for a programme or a feature, and then play around with the result and give it feedback.

Regardless of the intensity of use, most of these developers and their employers have been able to enjoy Copilot at a flat subscription rate of around $39 per month. However, on logging in on June 1st, they found that this was no longer the case — instead, they’d be paying based on their intensity of usage. The result, across the board, has been developers finding that their bills have gone up by orders of magnitude.

The subreddit of GitHub Copilot is a near-universal bloodbath of outrage. The only developers unaffected seem to be those who used Copilot to help them with gentler orchestration and configuration tasks for their projects, rather than actual development work. For those who use Copilot to troubleshoot code, usage-based pricing has increased their projected bills by around ten-fold. Among those who use it to outright vibe code, roughly hundred-fold increases in bills appear extremely common.

Most developers are wise to the reason this is happening: Copilot had to have been making a considerable loss for its owner, Microsoft.

It’s long been known that computing costs for the large language models (LLMs) that underpin tools like Github Copilot are heavily subsidised by the big four players in the space: Microsoft, Amazon, Google, or Meta. On the surface of it, this may have initially made sense for the big four, since it resembled the playbook that built the modern software and cloud computing industry in the first place: build a compelling offering at a loss, establish a dominant position with an insurmountable competitive moat, and then pursue commercial viability once you have a captive userbase.

However, this game plan is running into the rocks when it comes to AI. The sheer cost of the data centres, GPUs, power, and engineering expertise needed for the technology is beyond any comparable investment in the industry’s history. To achieve this, the big four have had to issue significant debt: one of the highest-profile examples being the Hyperion data centre project by Meta, which demanded a $30bn bond at 6.6 per cent interest.

This means that to break even, the return on invested capital for these projects needs to at least be around 7 per cent. However, as reported by Andrew Orlowski in The Telegraph, forecasters believe that even this modest 7 per cent return will require the big four to make $7tn in revenue in the next three years — all to merely break even and avoid selling off assets to pay back their creditors.

Since the high market capitalizations of these firms are based on their perceived growth potential, an inability to deliver a significant return for the remainder of the decade would be catastrophic for their finances. And this is a scenario that already relies on the big four raising their revenue up from around $1.6tn in 2025 to an average of $2.3tn every year for the next three years. In short: to achieve this disappointing return, revenue must increase by $700bn per annum.

With all this in mind, usage-based billing for GitHub Copilot starts to look more like a move of desperation than a cynical attempt to squeeze profit out of a captured market.

Over the next three years AI is expected to raise productivity by just 1 per cent economy wide

The reason for the desperation can be explained by a damning paper published back in February by the US National Bureau of Economic Research: 90 per cent of firms have experienced zero productivity growth because of AI since 2022, and over the next three years AI is expected to raise productivity by just 1 per cent economy wide.

The consequences of this paper destroy the case for AI generating a return on investment capable of sustaining even a 7 per cent return on investment. That’s because, in the long run, the total amount of revenue that a vendor can make from a tool will be constrained by the productivity growth their tool enables: firms do not pay for tools that cost more than what they produce.

Focusing on the US market, which has a GDP of around $32tn, 1 per cent productivity growth means around $320bn in extra output per annum can be expected by 2029 due to AI. Of this $320bn per annum, the economic reward to the providers will be the “Schumpeterian profit” — the fraction of the surplus output captured by innovators who offer productivity-raising technology.

According to Nobel Prize winner William Nordhaus, the historical Schumpeterian profit ratio in the US is 2.2 per cent. This means that of the $320bn additional output we may expect AI to generate in the US, AI firms themselves will likely take home only $7bn per annum. This is 100x lower than the $700bn per annum revenue the big four need to cover their borrowing costs.

For the big four, this is extremely dangerous. If their balance sheets begin to increasingly reflect the above — AI’s modest productivity potential means that Schumpeterian profits are orders of magnitude below the costs of the sunk investments — then the market correction would be brutal. Thus, there’s a clear logic for moves like GitHub Copilot’s price increase: prove that there are enough customers willing to pay at such a higher level, indicating that productivity expectations for AI are still buoyant.

This gives the GitHub Copilot price hike a new dimension. It’s not about the product itself breaking even, but demonstrating that AI as a category is creating a market that will drive significant revenue.

If this gambit fails, then the market’s pricing of productivity expectations could sour. At which point, there’s good reason to expect a market correction for the big four — along with the valuations of their own vendors, such as Nvidia, Oracle, and any potential listing from SpaceX, OpenAI, and Anthropic.

This is worrying, because US policy has made us all bag holders for any such correction.

The US has posted a continuous trade deficit for decades, averaging around $300bn a year since 2008. Mathematically, a current account deficit demands a capital account surplus — if you can’t cover your imports with exports, you need to sell some assets. For a long time, the US has sustained this with a seemingly win-win solution: the stock market.

Foreign investors have risen from owning 5 per cent of US stocks in 1994 to around 18 per cent today, with the extra demand having fed the virtually continuous US bull market since 2010. Because of this surplus foreign demand, the US now makes up 72 per cent of the benchmark MSCI World Index. By contrast, the US made up just 30 per cent at the start of the 1990s — and around 60 per cent on the eve of the 1929 crash.

If the illusion lifts, the AI correction will bring a significant risk of financial contagion that pushes an already-strained global economy beyond its limits. Without them knowing it, a lot was riding on the responses to those software developers who logged into GitHub Copilot on June 1st.