Josef Chen

· 10 min read

The GPU Already Lost This Job Once

I made money flipping mining rigs as a teenager and lost some of it holding the last ones too long. The AI build-out is running on the mistake I made with the last ones.

J. M. W. Turner, The Fighting Temeraire, 1839. Oil on canvas. National Gallery, London.
J. M. W. Turner, The Fighting Temeraire, 1839. Oil on canvas. National Gallery, London.

CoreWeave raised $8.5 billion of debt in March to buy more GPUs. And if CoreWeave can't rent them all out? No worries. Under an order form it filed with the SEC last September, "NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032".

It reminded me of another wave of infrastructure investment and depreciation: bitcoin mining. At fourteen I started buying bitcoin miners with my own money, the USB kind with a massive heatsink on top, and I made a lot of money flipping cloud mining rigs before I burnt myself holding onto the last ones too long. The machines never stopped working. A number on a spreadsheet did, and the H100s are sitting on that same line today.

I soon realised the machine on my desk was just one way to own what I actually wanted: hashrate. There are really five seats at the hashrate table:

  • Rig owners, who buy and operate ASICs
  • Slice owners, who buy TH/s generated on someone else's ASICs
  • Host owners, who buy shares in the companies that run the facilities
  • Chip owners, who buy shares in the chip makers
  • And the ASIC designer, which I was fairly sure took most of the money. That was Bitmain, and unfortunately for would-be shareholders, it stayed private.

For about two years, I traded between the first four. I learned how to separate wheat from chaff by spending many hours on forums like Bitcointalk.

All five seats exist again. The rig is an H100 server. The slice is a GPU rented by the hour from a neocloud. The host is a neocloud like CoreWeave, which you can buy shares in. The chip company, currently king, is NVIDIA. The fifth seat is more complicated: Google, Amazon and Meta build TPU, Trainium and MTIA chips for their own use, captive to their own clouds. But Broadcom builds many of them, and Broadcom is listed.

The five seats, 2014 and 2026. A three-column table: the seat, 2014, 2026. The rig: Antminer U2; H100 server. The slice: hashlets and Hashnest; a GPU rented from a neocloud. The hosting company: the cloud miners; CoreWeave. The chip company: NVIDIA by analogy; NVIDIA. The ASIC designer: Bitmain, private; TPU, Trainium and MTIA, and Broadcom, which builds them and is listed.

2013#

In 2012, gamers' Radeon 7970 graphics cards could be used for bitcoin mining, generating $1 to $2 a day. But in January 2013 the first ASICs came out, purpose-built chips that could only do bitcoin hashes. During 2013 the network's total computing power increased about 450 times, and the price of bitcoin increased 85 times, from $13 to over $1,100.

So this was the best year bitcoin has ever had, and the 7970's wage still collapsed. It briefly made $7 a day in the April run, had its last day above its own power costs by the start of September, and made 45 cents on December 4th, the day bitcoin peaked at $1,135.

The GPU lost this job once already. A Radeon HD 7970's bitcoin earnings in dollars a day, January 2012 to June 2014, on a log scale: about $1.50 a day through 2012, a peak of $7.26 on 7 April 2013, then a fall below its 60-cent power line for good on 1 September 2013 as the first ASICs shipped and the network's hashrate rose 450-fold in the year.

My machines#

The Radeon's 250 watts generated 6 cents a day by March 2014. A tiny, two-watt Antminer U2 USB miner generated 15 cents.

The U2 generated 5 cents a day by June and 1 cent by December, and bitcoin was still trading at $379, so the price wasn't the main thing. In March, bitcoin's price had hovered at $567. Even if the price had stayed there all year, the U2 still ended it at a cent and a half.

My U2 sat in a laptop in a school library and made fifteen cents a day. Nothing about it changed. It hummed away at the same hashing speed, but by December it made a measly cent a day.

The U2 did not lose money, electricity costs considered, but it wasn't making me rich.

What happened? The U2 had competition. The bitcoin network adjusts the difficulty of mining every 2,016 blocks, about every two weeks, to compensate for increases or decreases in total computing power applied to the network. In 2014, computing power hit a new all-time high eleven months in a row. My little U2 was insignificant.

But I'd moved on to GAW Miners. This fraudulent operation sold $20 million of "hashlets", a promise of future earnings, to more than 10,000 people in five months. Hashlets were touted as "always profitable and never obsolete". They were transferable between people. They promised to pay out every single day. People on Bitcointalk kept asking which pool the hashlets were even mining in. Nobody could find one. The SEC later concluded that GAW Miners "directed little or no computing power toward any mining activity".

I got lucky with GAW Miners though: I got out of hashlets a few months in with a profit.

GAW Miners' hashlets promised daily payments forever. So did the legitimate contracts from Bitmain's Hashnest. That was the one I held too long.

On 1 September 2014 you could buy 1 TH/s for 1.6 bitcoin on Hashnest, about $759. The contract also cost $3.24 a day in US dollars. It made more than its cost for 134 days and failed to do so on day 135. It failed to do so on 352 of the 365 days of 2015. At its best it had returned $272, less than half of its original $759 cost. It still never made a profit. Bitmain made better chips every few months, chipping away at the value of those it'd already sold.

The only reason that contract ever made any money was that people believed bitcoin would keep going up forever. Sounds like CoreWeave and NVIDIA to me. That's the part of the AI demand story people don't seem to want to believe.

Two things I bought in 2014. Left, the Antminer U2 in cents a day from March 2014: 14 cents on the March average, 5 in June, 0.9 by December, with a dotted line holding bitcoin at its March price that still falls to 1.5 cents. Right, one terahash of Hashnest Round 1 in dollars a day net of its fixed fee: above water for 134 days from 1 September 2014, then below it from 13 January 2015 and on 352 days of 2015, having returned $272 of the $759 paid.

The one I didn't buy#

Next came Ethereum. Ethereum launched in July 2015 using a memory-hard algorithm which made ASICs unfeasible, breathing new life into GPUs that'd been collecting dust since bitcoin squeezed them out in 2013.

I didn't jump on that bandwagon. Within six weeks, Ethereum's developers put a difficulty bomb into the codebase, a feature of the code designed to make mining harder over time and encourage Ethereum to leave proof of work behind. The people I trusted on Bitcointalk took the bomb seriously, and so did I. I still don't get how people financed GPU purchases with The Merge on the calendar.

On the morning of Thursday 15 September 2022, at block 15,537,394, earnings went to $0 for all 826 TH/s of GPUs on the Ethereum network, with ETH at $1,472. The day before, a stock RTX 3080 had made $2.77; at the May 2021 peak it made a cool $9.86 a day. To be clear, the GPUs all worked just fine. The difficulty bomb had been kicked down the road six times in seven years, and then the job ended in one block.

The job left on a Thursday. The dollars a day earned by a stock RTX 3080 mining Ethereum from January 2021 to December 2022: $5.41 a day on the January 2021 average, a peak of $9.86 on 11 May 2021, $2.77 on Wednesday 14 September 2022, and zero from Thursday 15 September, when the network switched to proof of stake and 826 terahashes per second of graphics cards lost their only job.

Three clocks#

Every machine has three clocks: its book life for depreciation, the length of its customer contracts, and how long it actually earns, which is never published. Depreciation is really about the expected period of earnings, not wear and tear: how long until a superior alternative comes along.

Bitcoin miners changed their depreciation schedules once their actual period of earnings became public. MARA and CleanSpark cut the useful lives of their rigs from five years to three; Riot went from two to three.

In AI the books went the other way. Amazon cut the book life of some servers to five years and took a $1.4 billion charge, "due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning". CoreWeave uses a six-year useful life for its GPUs. Will they still be earning in six years? That's the whole question.

Three clocks on one machine, twice. Mining, as it turned out: book life cut from five years to three, no contract clock because the wage repriced publicly every day, and earning clocks of 134 days for a Hashnest terahash and about 3.75 years above five-cent power for an Antminer S9. AI, as underwritten: a six-year book, contracts averaging three years under a five-year loan, and an earning clock nobody discloses, drawn as an empty track.

The same seats, with more money#

In January 2026 NVIDIA agreed to buy another $2 billion of CoreWeave equity. CoreWeave's $8.5bn of investment-grade debt in March, priced at SOFR+2.25% with final maturity in 2032, was underwritten against customer contracts that run as long as the loan. In August, the company priced a five-year loan for $2.6bn that's below investment grade and is priced SOFR+5.50%. CoreWeave's customer contracts there match an average duration of three years.

This circularity reminds me of Hashnest at a smaller scale. Under a $6.3 billion order form filed with the SEC back in September 2025, "NVIDIA is obligated to purchase the residual unsold capacity through April 13, 2032". As the August press release notes:

By financing these shorter-date commitments, lenders are signalling confidence in the long-term value of NVIDIA GPUs running on CoreWeave's cloud platform and a willingness to underwrite renewal risk.

This can go badly without demand falling at all. The August loan has a five-year maturity, but the contracts average only three years. In 2029, if GPU rental rates don't cover the debt payments, whoever financed the GPUs has to sell into whatever the market is at the time; they can't sit tight and hope things work out. And if lots of leveraged GPU owners hold the same paper, it isn't a gradual GPU market decline but a steep one.

Two loans, five months apart. Horizontal bars in years on a shared axis from zero to seven. March 2026: an $8.5 billion facility rated A3 and A (low) at SOFR plus 2.25%, maturing March 2032, drawn as a six-year bar above a six-year bar for customer contracts that run through the maturity of the debt. August 2026: a $2.6 billion facility rated Ba2 and BB+ at SOFR plus 5.50%, drawn as a five-year bar above a three-year bar for the underlying contracts, with a dashed outline for the two unsigned years. The 325 basis points between the two facilities is the lenders' price for the years after the contract ends.

CoreWeave's investment-grade issuance in March and its below-investment-grade issuance in August have a spread of 325bps. Part of that pays for the worse credit profile, and part pays for a risk public markets weren't previously accounting for: the last two years of the August issuance have no contracted revenue.

Yes, NVIDIA has a stake in CoreWeave and has agreed to buy remaining capacity, but only unsold capacity under one order form. That is not a residual-value agreement on the fleet, and there's none in the credit agreement.

The whole thing is a bet, with NVIDIA circularity no less, that GPUs have more economically meaningful years than the market expects. CoreWeave's chief executive, on the second-quarter call:

The fact that we have been able to go ahead and sell a GPU whose architecture was from 2020, in a contract that was fully priced out to 2029, really provides some insight into what the future is going to look like as this infrastructure comes off contract.

I do find it remarkable that a GPU could stay relevant for nine years. We seem to be betting the farm on H100 rental prices staying elevated for a very long time. The H100 rental index is down 57% in under three years (one-year rates are 38% off the October lows, but supply is constrained, with availability booked out since February). Across the industry, the bet that the debt gets repaid is a bet that GPUs will still be generating meaningful rental revenue nine years down the track. Happy to be wrong on that, but it's not a certainty to my mind.

How durable an income-generating asset is the GPU? Predictable inference workloads are going to keep moving away from GPUs and onto custom chips. Broadcom, mainly custom chips, had $10.8B in AI chip revenue in one quarter this year, up 143% year over year. Amazon deployed 2.1 million AI chips last year, more than half of them its home-brewed Trainium. Meta is deploying MTIA. Google has used TPUs for inference for around a decade.

The steady work is moving. Three rows. Meta, MTIA: serves its recommendation models on its own chips, and models MTIA cannot run stay on GPUs. Amazon, Trainium: 2.1 million or more AI chips landed in twelve months, more than half of them Trainium, with one million or more NVIDIA GPUs announced for 2026, and OpenAI committed to about two gigawatts and Anthropic up to five gigawatts of Trainium capacity. Google, TPU: has served its own steady inference on its own chips for about a decade, and bought NVIDIA GPUs the whole time.

The counterargument is that last fact. Work has been leaving GPUs at Google for ten years and NVIDIA's data-centre revenue grew massively the whole time, with Google buying GPUs throughout. Workloads moving to ASICs doesn't break anything on its own, and Google can afford to watch GPUs fall out of its fleet because it earns about $100 billion a year.

What it does is expose whoever owns the fleet the work leaves, to the extent that they scaled up assuming demand would stay put and took on debt to buy the machines. The question is how long CoreWeave's workloads last before they migrate.

Caveats: CoreWeave is not MARA or any of the other bitcoin miners. I'm not trying to trash CoreWeave's stock or debt. I'm not even bearish. Just skeptical of their ability to roll this out without problems.

But I have watched a boom fail to save a machine twice. Bitcoin's price rose by a factor of 85 in 2013 and the graphics cards still became obsolete. Hashrate set eleven consecutive monthly records in 2014 while my U2's daily earnings slid from fifteen cents to one cent.

GPUs remind me of bitcoin ASICs. They too depreciate once progress renders them obsolete. Hardware doesn't have to be broken to be broken; the only question is when lenders notice.

In 1839 Turner painted the Temeraire: a warship, masts still up, being towed to the breaker's yard by a steam tug. Note what's doing the towing.

Show me a customer and a workload and I'll gladly consider buying more GPU racks, but unspecified "AI demand" won't sway me. I've owned this asset before, from four seats. It was a mining rig with better branding.

All braindumps