The AI bubble bursting isn't the end of AI. It's the price cut.

Telecoms raised $1.6 trillion, went bust, and by 2004 bandwidth cost 90% less — the glut built the next web. The same arithmetic is running on AI right now, and the companies that map their judgment points before it clears will buy their advantage at the bottom.

Andrew Ramsden 7 July 2026

There is a line in your budget for AI, and if you are anything like me you have glanced at it, decided it was small enough to live with, and moved on. I did that for most of last year. What eventually stopped me was not my own invoice. It was a number from somewhere else entirely, and the question it left behind.

In the first three months of this year, investors put $300 billion into startups. Not the year. The quarter. Crunchbase counted it across 6,000 companies, and four cheques took $188 billion of that: OpenAI at $122 billion, Anthropic at $30 billion, xAI at $20 billion, Waymo at $16 billion.

Most of us read a number like that as confidence, and then read no further. The size is not the interesting part. The question I could not put down is where the next $300 billion comes from.

The arithmetic nobody says out loud

Money at that scale does not appear. It is allocated. Every dollar in those rounds was sitting somewhere else last quarter: in bonds, in property, in an index fund, in a pension member's balance in Melbourne. To keep funding a run rate like this, investors have to find genuinely new money or sell something else to buy this.

Annualise it and the discomfort is obvious. One quarter at $300 billion is more than a trillion dollars a year of allocation, aimed at a sector where the largest companies are not yet profitable. I have no idea what month it turns, and this is a supply question rather than a timing call. Everybody in that queue is asking for money at the same time, from the same pool, and the pool does not care how good the technology is.

So some of them will not get what they are asking for. That is a prediction about queues. And when a few of them miss, the pricing everybody downstream has been quietly enjoying stops making sense.

Your token price is somebody else's loss

This is the part that lands on your invoice.

OpenAI's 2025 figures, as reported, show $13.07 billion of revenue against a $20.92 billion operating loss. Round it and the company spent something close to $2.60 for every dollar it earned. The headline net loss of nearly $39 billion is mostly a one-time accounting adjustment from the restructure, so quoting that one would overstate the case. The operating loss is the honest number, and the honest number is bad enough.

That gap is currently your discount.

The per-token price you are budgeting against is the cost of serving you, minus what an investor is willing to absorb to buy the market. It is the cheapest AI will ever be relative to what it costs to produce, and every business case built on it inherits the assumption that somebody keeps absorbing the difference.

I think you already know this, in the way we all know things we have not said in a meeting. You have seen the training bills in the news, then looked at your own API line, and wondered how those two numbers live in the same universe. They don't. One of them is being paid by somebody else, for now.

Meanwhile, nobody can find the value

The other half of the squeeze is on your side of the ledger.

PwC asked 4,454 chief executives across 95 countries what twelve months of AI investment had returned. Fifty-six per cent said they had seen neither higher revenue nor lower costs. One in eight, 12%, reported both.

The easy read is that AI doesn't work. I think that read is incomplete, and expensively so. Where the money went explains it better. The dominant enterprise move of the last two years has been the seat: a licence on every desktop, a chat window beside the work, a dashboard counting how many people opened it. That move produces exactly what PwC measured, high adoption and no movement in the P&L, because a person with a faster way to draft an email is not a process that costs less to run.

The value is real. It sits one layer below where most companies are looking for it.

Which sets up the squeeze. Keep buying seats and you build a cost base on subsidised prices, for capability that isn't moving your numbers. Freeze everything until the market settles and you hand the compounding to whoever kept going. Both roads cost money. The wrong answer is expensive in both directions.

What the last burst actually did

The assumption underneath most bubble commentary is that a burst deletes the thing that burst. The last one didn't.

The Nasdaq peaked at 5,048 in March 2000 and bottomed at 1,114 in October 2002, a fall of nearly 80%. Everybody remembers that. Far fewer people remember what was happening underneath it.

Before the crash, telecom companies raised $1.6 trillion on Wall Street and floated another $600 billion in bonds to lay fibre across the United States: 80.2 million miles of it, 76% of all the digital wiring installed in the country to that point. Then the market repriced them. Companies died. Investors were wiped out. By any market measure it was a catastrophe.

The fibre stayed in the ground.

The glut it left behind meant that by 2004 the cost of bandwidth had fallen by more than 90%, even as internet usage kept doubling. The next wave of companies, the ones running video, cloud and social at prices that would have been impossible in 1999, were built on infrastructure somebody else had already paid for and lost money on.

The crash did not delete the internet. It transferred it, from the people who financed it to the people who were ready to use it cheaply.

The obvious move, and why it fails

The instinctive response to all of that is to wait. Let the correction happen, see what survives, buy in at the bottom with a clear head. I understand the instinct. I have argued for it myself in rooms where it was the responsible-sounding thing to say.

And yet the dust is where the discount is, and you cannot pick it up empty-handed.

What gets cheap in a correction is compute and talent. What does not get cheap, and takes the same six to eighteen months whatever the market is doing, is knowing precisely where in your own operation a machine judgement belongs. Companies that arrive at the bottom of the cycle already knowing that spend the cheap tokens immediately. Companies that arrive with a licence list start the mapping work then, and finish two years later, buying at whatever the price has become. The gap looks small now. In three years it is uncatchable.

There is a fair objection, and it deserves a straight answer. The price of compute was falling before any of this, and will keep falling without a crash to help it. Open-weight models have already commoditised much of what the frontier labs charged a premium for, and underneath them the chips, the power draw, the cooling and the hardware lifespan all sit on curves that have bent downward for decades. Richard Sutton named the dynamic in 2019, in an essay called The Bitter Lesson: general methods that leverage computation win in the end, because they ride the falling cost of computation. So if cheap intelligence is arriving regardless, why hurry?

Because falling prices are an argument for the map, not against it. Cheap compute is only worth something to a company that knows where to point it, and that knowledge is the one input in this story with no downward cost curve. It takes the same months of process work whatever a token costs. Meanwhile the bills are real today: at these subsidised prices, badly shaped usage, a model re-reading everything on every transaction, an agent looping because nothing told it to stop, already produces numbers that make CFOs flinch. The waste is in the shape, not the price, and a ninety per cent price cut on a wasteful shape is still waste.

Judgment points, not desktops

For almost every role in your company, the valuable form of AI is a discrete judgement point dropped into a process that was previously too hard to automate.

Think about the processes you have never automated. They were never blocked by the mechanical steps: the data movement, the form filling, the routing. Those have been automatable for twenty years. They were blocked at one or two moments in the middle where somebody had to read something and decide. Does this invoice match this contract? Is this complaint about the product or the delivery? Does this application need a human? Is this photo showing the damage the claim describes?

One judgement, mid-process, was enough to keep the whole thing manual.

That is what changed, and "AI can do knowledge work" is the framing that sold the seats. The specific, boring, enormous change is that a bounded judgement can now be made by a machine, cheaply, at any point in a workflow, with a confidence signal attached and a human escalation behind it.

The economics of that shape are nothing like the seat. A judgement point is a small, predictable, measurable number of tokens per transaction: one call, on one bounded question, with a defined input. A chat window is an unbounded conversation with no unit of output, priced per person per month whether or not anything changes. When token prices rise towards their real cost, and they will, because $2.60 spent per dollar earned is not a business model, a judgement point re-prices from trivial to slightly less trivial. A seat estate re-prices into a board conversation.

Which brings us to the people, because this is where most programmes lose their nerve or their soul. When you put a judgement point into a process, hours come out. If those hours have no named destination they become invisible slack, the dashboards stay green, and nothing improves. What works is to redeploy the freed people into the parts of the business that cannot be automated and are currently under-served: the difficult customer conversation, the bespoke work you turn away, the follow-up nobody has time for, the quality pass, the relationship. That is craft. It is the layer your competitors cannot buy from a vendor, and the only durable answer to the question of what your people are for.

Judgement points take the cost out. Craft puts the difference in. One without the other is either a redundancy programme or a hobby.

The Judgment-Point Map

Take one end-to-end process, not a department or a transformation: one you could describe to a stranger in a minute, such as claims triage, onboarding a client, month-end close or quoting a job.

Write down its steps in order, and label every step with exactly one of three words.

Conveyor. The mechanical steps: moving data, filling fields, routing, chasing, formatting, reconciling against a rule. No judgement required. No AI belongs in this column. These are rules, scripts and integrations, the deterministic automation software has done reliably for twenty years, at a fraction of a token's cost and with the same answer every time. If you have not automated these, AI was never your blocker, and a process that is mostly conveyor has a cheaper answer than tokens.

Judgement point. The steps where somebody reads something and decides, and where the decision is bounded: a classification, a match, an exception flag, a summary that feeds the next step. The test is that you could write down what a right answer looks like, and check it after the fact.

This column is the only place AI goes, and that restriction is the whole economic argument. One bounded call, on one question, at one step, instead of a model sitting across the whole process, reading everything and charging you for the privilege. A judgement point costs a predictable fraction of a cent per transaction because you have told it exactly what to decide. That is the difference between a token bill you can forecast per unit of work and a token bill that scales with how chatty your staff are. And each judgement point carries an escalation: when the model's confidence is low, the case goes to a person. The machine takes the ordinary volume, the human keeps the hard cases, and you pay full human attention only where it changes the answer.

Craft. The steps that carry relationship, negotiation, accountability, taste, or a judgement nobody can write down. These are not "not yet automated". They are the reason customers choose you.

Craft is more than the escalation queue for the column beside it. Catching the cases the model wasn't sure about is the floor of what your people do here, not the job. The job is the under-served work named above, and this is where the hours freed by the judgement points get redeployed, and if you cannot name that destination, the map is not finished.

Then answer one question underneath it. When this process gets faster, where do those hours go, and who moves them? A name and a step, not an aspiration.

A mapped process is also a process you can now build, and building it is where your tokens should be concentrated rather than spread across a licence estate. Agentic tooling has made the construction work cheap enough to do properly: reading the existing system, writing the integration, wiring the exception paths. That spend is large, deliberate and once. What you are left with afterwards runs the way processes have always run: deterministic steps at effectively no cost, calling out to a model only when it reaches a judgement point. Front-loaded build, cheap run, for years.

The map ends in one of three words.

Positioned. You have named judgement points and a craft destination for the freed hours. You know exactly what you buy when compute gets cheap, and exactly where the people go. Protect this map. It is worth more than your licence budget.
Half-mapped. You have found the judgement points but not named where the hours land. You will automate successfully and capture nothing, because saved time with no destination is the most reliably wasted asset in business.
Unmapped. No judgement points identified. Either the process is pure conveyor and wants ordinary automation at a fraction of the cost, or it is pure craft and should be left alone and resourced properly. Both are useful answers, and both save you a year of pointing AI at the wrong work.

One boundary, honestly. This map cannot tell you when the correction comes, and it cannot price your tokens. Nobody can do the first, and no vendor will do the second. What it does is make sure that whenever the repricing happens, up or down, you already know which parts of your operation want compute and which parts want people. That knowledge is the thing that stays expensive when everything else gets cheap.

The winter is the opportunity

When it comes, the correction will be read as a verdict on the technology. It won't be. It will be a verdict on the capital structure that funded it: a queue where more people asked for money than could be paid, at prices that assumed everyone got theirs.

Underneath that, the fibre stays in the ground. The models don't get worse when a share price falls. The data centres do not switch off. The chips, the weights, the tooling, the people, all of it survives the correction, and most of it gets cheaper, exactly as bandwidth did for every company that walked into 2003 knowing what it wanted to build.

The companies that come out of this owning their markets will not be the ones that predicted the timing. They will be the ones that spent the noisy months mapping judgement points, moving people into craft, and building an operating model whose token spend is a per-transaction line they can defend rather than a per-person line they cannot.

Start in the middle of one process. Find the step where somebody reads something and decides. That step is the reason the whole thing is still manual, and it is the cheapest thing you will automate this decade. Thirty minutes from now you will have every step labelled, a named destination for the hours, and something useful to say in the meeting where somebody finally says "bubble" out loud.

Attribution

Written by Andrew Ramsden. AI tools assisted research and drafting; all outputs verified.

Accountable

Andrew Ramsden.

Limitations

OpenAI's revenue and loss figures are as reported by TechSpot, not from a filing. The dot-com fibre figures rest on one 2018 essay.

References

Six sources, retrieved and hashed on 26 and 27 August 2026, each figure quoted with a locator in the SOURCED sidecar.