The AI Data-Centre Bust Will Look Like a Boom

NVIDIA is arranging half a trillion dollars to finance GPUs. Neocloud stocks are ripping. Etched just hit a ten billion dollar valuation for an ASIC built to take GPUs' steadiest work, and OLIX raised $312 million at $3.3 billion eleven days behind it.

Start with the bullish case, because I hold it. Demand is enormous and still growing: about half of US adults now use AI chatbots and a quarter use them daily. Nobody halves the token budget, and a better model can create its own market. That is why I am bullish in a way I never was on mining, and why custom inference silicon looks like one of the better hardware bets of the decade: the volumes are finally large enough to pay for a chip that does one thing. Even so, that is a quarter of one rich country, and the demand story runs well ahead of the paying base, which is precisely the condition in which build-outs overshoot.

None of that saves the machine.

I started mining Bitcoin at fourteen. On 1 September 2014, Bitmain posted an announcement on the Bitcointalk forum: computing power from its Antminer S2 miners, sold through Hashnest at 0.0016 BTC per GH/s, maintenance fee deducted daily. I was on the other side of that page, buying not the coin but the machinery under it: Antminers, steel boxes built to perform one calculation, and Hashnest claims on somebody else's machines. Later I bought GPU rigs for Ethereum. Demand never killed any of them. Newer chips killed them on a spreadsheet while Bitcoin was booming: a computer can work perfectly and still be a terrible asset.

That was my first data-centre asset class. The data centre was smaller.

Twelve years later, NVIDIA started talking like a bank. On 10 August 2026 it announced plans with six of the largest names in capital markets, Apollo and BlackRock and Blackstone among them, to mobilise more than $500 billion of outside capital. The agreements were not final. Jensen Huang then said NVIDIA might, project by project, support the value left in the machines for up to 25% of one financing opportunity. The heroic number matters less than the shape of the deal: the firm selling the chips is helping to organise the money, and residual value has become part of the sales pitch before the current build-out is finished.

The bust will not begin when people stop using AI. It will begin when the predictable inference jobs that were supposed to pay for a GPU fleet leave before the debt does.

Memory is the bottleneck, which is the whole argument#

Serving a token is mostly a memory problem. The machine reads the model's weights and the running conversation out of memory for every token it emits, and much of the accelerator waits. It is the most important fact about this hardware, and it cuts both ways.

It is the bull case. A chip built for one fixed model can lay those weights out on the die and stop paying for the trip. The saving is not arithmetic but a journey no longer made, and nothing about a general-purpose GPU beats it on work that repeats.

It is also the bear case, for the same chip. Holding the model on the die means being welded to the model, and there is no version of the trade that gives the win without the welding. So the question was never whether ASICs beat GPUs. It is which work sits still long enough to be welded to.

So far, the boring work. Meta says its MTIA chips serve recommendation models while unsupported models remain on GPUs. Amazon reported in April that it had landed more than 2.1 million AI chips in twelve months, more than half Trainium, alongside more than one million announced NVIDIA GPUs from 2026. It buys both, and the simplest explanation is that the jobs are different.

My bet through 2030 is narrower than "custom chips win". The largest platforms move stable, repeated, high-volume inference onto chips they control; GPUs keep the models that change, the software those chips do not support, demand spikes and customers who need broad compatibility. Move enough predictable queues and the GPU survives. The spreadsheet that paid for it does not.

The obvious objection is Google, and it is the best one anybody has put to me. Google has served its own steady inference on chips it designed for about a decade, through several generations of them, and NVIDIA's data-centre business compounded through the whole period with Google buying GPUs the entire way. If migration alone were enough, my bust would have arrived years ago. So the call is not that predictable work starts to move, because it has been moving for ten years. The call is that the fleet losing that work is now bought with borrowed money against a six-year book.

Three things prove me wrong by 2030: custom chips stay too painful to program, model changes eat the savings, or old GPUs keep named workloads and steady cash yields through a major hardware transition. Any one of the three kills the call. More AI demand settles none of them.

The middle one is heaviest, and it is the memory fact turning against me. The calculation my Antminers raced to perform never changed, which is the only reason a chip welded to it could win. Transformers are not fixed. Mixtures of experts, longer context and models that think for longer before answering all make the work less regular. Weld too early and the chip arrives fitted to last year's assumption. None of this needs a view on whether AGI is coming: if the models keep compounding, the predictable work migrates sooner, and if they stall, the debt still outlives the fleet.

The machine dies on a spreadsheet#

Bitcoin mining is cruel in one useful way: its economics are hard to hide.

The network pays out a roughly bounded pot, miners split it by the computing power they contribute, and every machine that joins shrinks everybody else's slice. Then electricity sends the bill. Every generation was a race between revenue per unit of computing power, which miners call hashprice, and energy cost, and a newer application-specific chip, or ASIC, ran the same calculation far more efficiently. The old box did not break. It crossed a line on a spreadsheet and became irrational to run.

Energy per hash fell about 984-fold between two sparse endpoints, a 2013 Avalon at 9,351 joules per terahash and 9.5 for Bitmain's 2025 S23 Hyd. That is what the old machine was up against.

Using April 2026's average Bitcoin network revenue and electricity at five cents per kilowatt-hour, a 29.5 J/TH Antminer S19 Pro failed to cover electricity alone. A 13.5 J/TH S21 XP still cleared it. Both worked; only one paid its power bill, and cooling, labour, financing and the price of the machine were all still waiting outside the calculation.

Miners have a name for a machine on the wrong side of its line: a space heater. The name is unfair, and the unfairness is the best objection to everything I argue after it. A rig underwater at one power price clears at another. Operators undervolt them, trade speed for a smaller bill and move them to power nobody else wants, which is why rigs written off at five cents were still hashing years later. My chart is drawn at one price. Cheap power buys the old machine time, which is not the same as buying it a job.

The resale market moved before the machines wore out. Hashrate Index reported that three ASIC price indexes fell between 82% and 87% during 2022. Bitcoin and its miners carried on, the boxes now worth much less.

Ethereum's Merge then sent its machines looking for another chain. Across equal thirty-day windows, Ethereum Classic's estimated computing power rose 264% while revenue per unit fell 79%, and the revenue pool itself fell only 23%, so most of the per-machine collapse was dilution. Compatibility found the machines another calculation without manufacturing another revenue pool.¹The data are entirely within Ethereum Classic and do not identify which GPUs or ASICs moved, where they came from, or how much equipment shut down. The revenue estimate excludes unobserved miner income and every cost.

An old GPU has more possible lives than an old Antminer, which gives its owner a convenient answer whenever residual value comes up: we can redeploy it. I always want the next sentence. To what? Which model fits in this card's memory, which software supports it, and after networking, support, electricity and debt, does the work still pay? "It can run inference" describes compatibility, not residual value. The seller makes the spec-sheet case officially: NVIDIA's financing post calls the factory "flexible and fungible" and stretches the A100's economic life "toward a decade". An old GPU gets no residual value in my model until somebody names the workload.

Mining had hashprice; AI has the dollar-per-GPU-hour rate every rental contract is built on.

Hashnest sold me a claim on a machine somebody else operated; the neoclouds sell that product at institutional scale, financed with debt against the machines. Even the hiring rhymes: the scarce people are not model researchers but power engineers, site developers and whoever can get a grid connection energised.

Bitcoin at least published its executions: one difficulty number, one hashprice, everyone repriced in the open. AI reprices in private, contract by contract, queue by queue, so the owner of an undifferentiated GPU learns the price of obsolescence later, and alone. The absence of a halving is not the absence of a cliff; it is a cliff without a published schedule.

The chip wins, the vendor usually does not#

Being right about a chip is not the same as owning the company that sells it. Mining ran that experiment, and nobody in inference hardware brings up the result.

Demand for mining chips beat every forecast. Computing power on the Bitcoin network rose about 351 times between January 2017 and April 2026. Knowing that in 2013, you would have bought the vendors.

You would have lost almost all of it. Revenue per unit of that computing power fell 95% across the same months, out of the same file. Record capacity, earnings per machine down to almost nothing, and only the second line pays for anything.

The buyers went first, diluted by every machine that shipped after theirs. Then the vendors went. KnCMiner, a Swedish maker of Bitcoin ASICs, raised $32 million from investors including Accel and Creandum, shipped real machines to real customers, and declared bankruptcy in May 2016. Its chips worked. Its chief executive blamed the coming halving: "Effectively our cost of coin...will be over the market price".

The reason is market structure, not engineering. A merchant vendor sells a machine whose value is set by a wage it does not control and cannot see, commits to a wafer months before delivery, and has to win the next generation on the last one's cash. Selling the machine is the worst seat in a business where the machine is the thing that depreciates.

Bitmain won, and it did not only sell machines. The forum post at the top of this essay is a manufacturer renting out the output of its own Antminers, so a stale generation had somewhere to work: its own hashrate business, while the fresh one went to everybody else. It was the only seat where a short-lived chip never had to be sold to a stranger.

The platforms designing their own accelerators sit in that seat. They own the queue, the model, the software and the customer, so a chip welded to their workload is welded to something they control, and when the model shape moves they are the ones moving it. OpenAI has signed for ten gigawatts of accelerators it designs itself, racks targeted from late 2026 through 2029, inside every six-year book now being written. Anthropic, which still buys NVIDIA systems at scale, signed for multiple gigawatts of next-generation TPUs expected from 2027 and runs Claude across Trainium, TPUs and NVIDIA GPUs to match workloads "to the chips best suited for them".

The merchant startups sit in KnCMiner's. Etched, Fractile and OLIX raised more than $800 million this year, and between them they have first silicon, chips still described as future, and first deliveries targeted for the second half of 2027. Samsung SDS says it launched a subscribable service on FuriosaAI hardware, the furthest public step in the group. I found no proof that any of them has a sustained paid workload. That is not a verdict on their engineering, which may well be better than anyone else's. It is that they must guess which model shape somebody else will still be serving when the wafer starts. Bitmain never had to guess. The test is the first invoice attached to a named workload that keeps paying.

Three clocks on one machine#

Alphabet's disclosed token throughput went from seven billion a minute in October 2025 to about twenty-two billion by July, though what the company says that number counts changed along the way, and its undefined Gemini "serving unit costs" fell 78% during 2025. Neither can value a fleet: price, profit, model mix and utilisation are all missing. Both show why the boom stays loud, because more use can arrive beside a lower cost to serve it.

Microsoft reported $41 billion of company-wide capital spending in its fourth financial quarter of 2026, roughly two thirds of it in what it called short-lived assets, primarily CPUs and GPUs. Yet major owners generally spread broad server and network costs across five or six accounting years. The books move slowly while chips and workloads move quickly. Mining companies converged on roughly three years from opposite directions while the selected AI equipment categories mostly lengthened. None of those policies isolates a GPU, but the mismatch is hard to miss.

Amazon extended server lives from five years to six in 2024, then shortened a subset to five a year later because AI and machine-learning technology was changing faster. The reversal added $1.4 billion to 2025 depreciation and amortisation expense. CoreWeave spreads technology-equipment cost across six years while its weighted-average committed customer-contract term was about five years at the end of 2025.

The filings do not show a cemetery of formally impaired GPUs, and I am not claiming one exists. Only that the accounting life, contract life and earning life of one machine already run on three different clocks. There is no line item for space heaters.

The pot is the one place the analogy genuinely breaks, and I would rather break it myself than have it broken for me. Bitcoin pays its machines from a pot designed to shrink: issuance halves on a schedule, the two step-downs in the first chart, and every machine that joins dilutes the rest within two weeks. AI has no such ceiling. But an open-ended pot secures the aggregate, not the machine, because what killed my Antminers was never the size of the pot; it was a better machine taking the same job, in rising markets as often as falling ones. Underneath the pot sit three different machines.

	Bitcoin ASIC	AI GPU	inference ASIC
computes	one fixed calculation	any parallel job	one model family
the wage	hashprice, public	$ per GPU-hour, private	$ per token, contracted
book life	about three years	five to six years	too new for a book
second life	space heater	we can redeploy it	none, welded to the job
dies of	a better ASIC	custom chips take the steady work	the model moves

Who gets paid when the chip changes#

One rule sorts this industry: prefer the business that gets paid when the chip changes, and distrust the business that needs the chip not to change before the debt is repaid. Bitmain passed that test; its customers did not. Three positions are left, the merchant vendor having been sorted above.

The chip seller is the safest seat, because it is also selling the system around the chip. NVIDIA reported $75.246 billion of Data Center revenue in its latest quarter and roughly $14.8 billion of networking revenue, and it is working to make that network useful around accelerators other companies design. Microsoft already says Maia 200 uses its own transport over standard Ethernet, which reads to me like a polite way of routing around it, and NVIDIA's own filing warns that customer-designed chips may not need every feature in its systems. Nobody discloses how much a vendor keeps when the accelerator belongs to the customer. The purest version of the seat is Broadcom's: it builds custom accelerators for the platforms, booked $10.8 billion of AI semiconductor revenue in its latest quarter and guided the next quarter to $16 billion. It gets paid when the chip changes, which is also why the clean short on this thesis does not exist.

The borrower who owns the fleet is where the analogy bites hardest. CoreWeave is the cleanest public expression: at the end of June it disclosed $35.6 billion of debt principal and $16.3 billion of recognised operating-lease liabilities, and it spent $14.1 billion on equipment in six months. The defence is serious: $103.7 billion of work still owed under contracts, with 98% of quarterly revenue tied to customer commitments. The other side is that three customers supplied 72% of Q2 revenue and quarterly interest expense was $640 million. A backlog does not pay interest by itself. Contract cash has to cover interest, leases, replacement chips and the cost of finding the next job, and that arithmetic belongs to every leveraged fleet in the industry, listed or not.

In March CoreWeave closed an $8.5 billion facility at investment grade, SOFR plus 2.25, against GPUs wrapped in a customer contract. In August it closed a $2.6 billion facility at SOFR plus 5.50, below investment grade, a roughly five-year loan against contracts averaging about three years, with no residual-value guarantee anywhere in the filed agreement. The 325 basis points between those two loans is a first public price for the years a GPU must earn after its contract ends.

Nebius has a direct answer: large contracts bring material prepayments, expected payback is about twenty-two months, and pricing for older GPUs improved more than 30% from the prior quarter. I want that to be true. It has not met a chip refresh yet, and that is where such claims are settled.

The landlord is a different animal, and the market keeps pricing it as the same one. Applied Digital's core leases sell the powered building, cooling and electrical system rather than the accelerator inside. A mine can leave behind a power connection; it does not leave behind an inference customer, which is why this is the position I would rather hold through any hardware transition, with one condition: distance from the GPU only helps when the product survives the next GPU. Only 100 of 1,410 contracted megawatts was operating and earning revenue at year-end. Contracts still have to become buildings.

What happens next#

Predictions, so this essay can be graded. The footnotes arrive quarterly, and I keep score on a ledger that logs the prints cutting against me too.

The six-year GPU book will not survive the decade. Amazon's cut from six back to five was the first crack; before 2030 a hyperscaler or major neocloud discloses four or fewer for accelerators, and the sell side will call it prudence when it is a confession. Miners tried five and two and met at three.

AI will get its hashprice. The H100 rental series charted above fell 57% in under three years, and the sceptic's answer rides in the same series: one-year contract rates up 38% from their October low, on-demand capacity sold out since February. Hashprice rallied before every collapse too. None of it is yet a number a lender marks a fleet against, which is what made the difficulty adjustment lethal rather than merely interesting. One of these series hardens into that number before 2030: quoted continuously, for a named generation of hardware, cited in a credit agreement. The moment it exists the private cliff becomes a public one, and last-generation GPU-hours grind toward power cost the way the electricity line ground down my Antminers.

Custom silicon takes the payroll and the GPU keeps the frontier. By 2030 the majority of the largest platforms' own stable, high-volume inference, the feeds, the recommendations, the search-shaped work, runs on chips they control. The aggregate GPU market can grow while this happens, which is why almost nobody will notice.

The first liquid secondary market in accelerators prints the number everyone is avoiding, the way Hashrate Index printed ASIC prices in 2022, and the first large GPU impairment lands in a quarter of record demand, blamed on a customer event rather than the curve. That is when the accounting life and the earning life meet in public.

Most neoclouds will not die as clouds. The survivors will quietly become power and shell companies, the way the surviving miners did, because the durable asset underneath was always the grid connection and the building. The rest get absorbed for their contracts and their megawatts, and their GPUs go to the secondary market.

The bill arrives before the boom ends#

At fourteen I thought I was buying computing power. What I had bought was a temporary spread between mining revenue and electricity, wrapped in a machine. Turner painted the asset class in 1839: a working ship, nothing wrong with her, under tow behind the machine that took her job.

Today's version comes with hyperscale customers, long contracts, financing syndicates and scarce grid connections, every one of which can be real while an individual GPU investment fails. The bust arrives loud: record electricity demand, full pipelines, more tokens every quarter, and rental rates falling anyway. The phrase "AI demand" tells me none of that.

The AI data-centre boom is a repricing event dressed as a building programme: hundreds of billions of dollars a year underwritten on the belief that a general-purpose chip keeps its best job for five or six accounting years, when everything mining taught me says the best job leaves first, quietly, while the buildings are still full. Somebody ends up holding the one asset this industry has never learned to price: a space heater with a six-year book.

Before I underwrite another GPU rack, I want one answer:

Which named customer and workload will still pay for this machine if custom chips take more of the predictable work?

If the answer is "AI demand", I have seen this asset before. It was a mining rig with better branding.

Research cutoff: 16 August 2026. This is a research view, not investment advice.

Braindump by Josef Chen

---

1. The data are entirely within Ethereum Classic and do not identify which GPUs or ASICs moved, where they came from, or how much equipment shut down. The revenue estimate excludes unobserved miner income and every cost.