Most founders I know are quietly running their entire businesses on AI subscriptions that, on paper, look like the bargain of the century.

A $20 ChatGPT Plus account. A $200 Claude Max plan. A few hundred dollars a month in Cursor and Codex and Copilot. Maybe an Anthropic API spend that's growing faster than they'd like but still feels manageable. The whole stack costs less than a junior hire, and it's doing what would have taken three of them eighteen months ago.

It feels too good to be true. It is.

I've been watching the last few weeks of news in the AI economy and there's a pattern forming that almost no founder I've spoken to has fully connected yet. A handful of unrelated-looking stories - a class action lawsuit, a Tim Cook interview, big enterprises pulling back on AI spend - are actually all the same story.

The story is that the bill is about to land. And when it does, the founders who built their economics around tokenmaxxing are going to feel it first.

Let me walk through what's actually happening, because I think this changes a few things on your roadmap.

See below to continue…

Join the waitlist for SoPhy - the networking intelligence engine!

Soon, you'll be able to quantify your business relationships - but for now, we're starting with the essentials:

Quick update: the app is almost ready - just a few finishing touches, and we should be able to open it up in the next 2 weeks. Here is a short demo video to check out ahead of the release.

Join the waitlist to be one of the first to try it once it's out.

Sign up here.

The founders who close rounds fastest aren't just the ones with the best pitch; they're the ones who show up prepared.

Carta Launch gives early-stage founders a clean, investor-ready cap table for free, so when the conversation gets serious, you're already serious.

Manage equity, issue shares, and track dilution, then instantly share it with a built-in data room that gives investors exactly what they need, exactly when they ask.

The best time to get your cap table right is before you need to.

Learn more about Carta Launch and get started for free.

Continue…

Here's what nobody at the big AI labs likes to talk about openly.

The price you pay for AI right now is not the price it costs to deliver. Not even close. For the last two years, OpenAI, Anthropic and Google have been quietly losing money on most paid plans, because venture capital has been making up the difference. The $20-a-month ChatGPT subscription has never covered the compute it consumes. It was a customer acquisition cost dressed up as a product.

That worked when most people were just chatting. It does not work when those same people are running agentic workloads, Cursor sessions, Claude Code projects, and long-running automations that consume tokens around the clock.

The receipts are now everywhere. Anthropic is being sued in a proposed class action over its Claude Max plans. The lead plaintiff, a Washington DC user named Karl Kahn paying $200 a month for the Max 20x plan, says a single five-hour coding session burned through 15% of his weekly quota - and that the "20x" plan in practice delivers six to eight times the Pro plan, not twenty. Anthropic's own internal economics are reportedly so stretched that the $200 Claude Code plan lets power users consume between $600 and $1,500 worth of API-priced compute every month.

That gap isn't a pricing mistake. It's the entire industry's business model.

The most telling stories of the last six weeks aren't the lawsuits. They're what's happening inside the companies that were supposed to be the poster children for AI adoption.

I'll be honest here: I came across a widely-shared video claim that Microsoft told around 100,000 of its own engineers to stop using a specific AI coding tool by month-end because the per-engineer cost exceeded the engineer's salary, and that Uber burned through its entire 2026 AI budget by April. I haven't been able to independently verify either claim through mainstream reporting at the time of writing, so I want to flag them as unconfirmed - though both are circulating widely enough to be worth keeping an eye on.

What I can verify is the direction of travel from people speaking on the record. Engadget reports that "right now, venture capitalists are offsetting those costs, and if anything, the problem is likely to become more acute when Anthropic and OpenAI go public." Gizmodo notes that "open-source models like DeepSeek have meanwhile been growing in popularity as businesses and individuals get priced out of using proprietary models", and that "models have also been getting more token-hungry as they become more agentic."

The pattern is real, even if some of the specific anecdotes circulating right now are still rumour. People who actually deploy AI at scale are quietly stepping back. Not because it doesn't work - it does. But because the math doesn't.

While all this is playing out on the software side, the hardware side just got significantly worse - and most founders haven't registered it yet.

Tim Cook gave an interview to the Wall Street Journal last week describing the current memory and storage chip shortage as a "hundred-year flood" - a phrase a CEO of Apple does not use lightly. He told the Journal: "I've never seen anything like it in any area in over 40 years". Memory prices have risen more than sixfold over the last year, according to Morgan Stanley's Shawn Kim.

Hyperscalers like Google, Microsoft, Meta and Amazon - spending around $750 billion combined on AI infrastructure in 2026 - have hoovered up the supply with long-term contracts and prepayments, leaving everyone else fighting for what's left.

Apple, the largest memory and storage buyer on the planet for years, has just announced it can no longer absorb the costs. "Unfortunately, price increases are unavoidable," Cook told the Journal. Research firm TechInsights estimates the iPhone 18 Pro will need to go up by roughly $270 to maintain margins. Microsoft's new Surface devices are already roughly 50% more expensive than last year's. The component cost of memory and storage in last year's iPhone 17 Pro was about $50; in the iPhone 18 Pro, it's around $200.

This matters for founders for two reasons. First, every laptop, server, and device your team uses is about to get more expensive. Second - and this is the part nobody is saying clearly - the same memory squeeze that's making your next iPhone $270 dearer is also making it more expensive for OpenAI and Anthropic to add compute capacity. The infrastructure these labs need to keep your AI bills artificially low is now structurally more expensive than it was six months ago.

The frontier labs are trapped. If they keep prices low, they keep burning capital. If they raise prices to reflect real cost, users leave - because open source models like DeepSeek are now genuinely good enough for most workloads at a tiny fraction of the cost.

This trap won't last forever. Both Anthropic and OpenAI are targeting IPOs this year at trillion-dollar valuations. OpenAI is reportedly already considering significant cuts to its current token prices as a sales tactic, while Microsoft is positioning its own MAI-Thinking-1 model as a cheaper alternative, with Mustafa Suleyman saying that "many people are urgently looking for alternatives" to Anthropic's expensive models.

The moment Anthropic and OpenAI file for IPO, the real numbers stop being private. Public companies cannot indefinitely lose money on every paying customer.

Which means one of three things will happen, probably in sequence. Subscription limits will tighten further - this is already happening, see Google cutting its AI Plus plan from $7.99 to $4.99 and its top-tier from $250 to $200, and Anthropic moving to block third-party tools that let users run high-volume workloads at flat consumer rates. Prices will rise. And the moment they do, the cheaper open source models will start eating the bottom of the market.

For founders, the implications are uncomfortable but useful. If your unit economics quietly assume current AI pricing - if your product margin only works because you're using a $200 plan that gives you $600 to $1,500 of real compute - your margin is on borrowed time. You're not running a profitable business. You're a downstream beneficiary of a venture-funded subsidy that won't survive the IPO.

I'm not telling you to panic. I'm telling you to model the scenario nobody is modelling.

Pull up your AI cost stack. For every recurring subscription and every API line, ask: what does my P&L look like if this cost doubles in the next 12 months? Because somewhere between the lawsuits, the IPOs, the chip shortage, and the open source migration already underway, doubling is a perfectly reasonable base case.

The founders who win the next 24 months won't be the ones who used AI the most. They'll be the ones who built their economics on the assumption that AI would, eventually, cost what it actually costs. That means hybrid stacks with open source fallbacks. That means caching aggressively. That means picking the cheapest model that's "good enough" for each job, not the most capable one by default. That means architecting for substitution before you're forced into it.

The bill is coming. The only question is whether you've designed your business to survive it - or whether you're still tokenmaxxing on someone else's runway.

✅ Know a founder running their whole stack on Claude Max or Cursor and not thinking about what happens next? Forward this their way. The earlier they model what their unit economics look like at real AI prices, the better the decisions they'll make over the next 12 months.

(👉 Vote now — we’ll share the results in next week’s issue. All votes are anonymous.)

(Last week’s poll results)

All the UK’s best startup events in one place.

Access here and please share it with others.