Dark GPUs

Laying dark fiber
Stealth Communications CC-BY-SADark Fibre article starts:

Dark fibre originally referred to the potential network capacity of telecommunication infrastructure. Because the marginal cost of installing additional fibre optic cables is very low once a trench has been dug or conduit laid, a great excess of fibre was installed in the US during the telecom boom of the late 1990s and early 2000s. This excess capacity was later referred to as dark fibre following the dot-com crash of the early 2000s that briefly reduced demand for high-speed data transmission.

These unused fibre optic cables later created a new market for unique private services that could not be accommodated on lit fibre cables (i.e., cables used in traditional long-distance communication).

Many are analogizing the current AI bubble to the dot-com era, pointing out that although the dark fiber eventually became very valuable infrastructure, the companies that invested in laying it suffered. Wikipedia notes that the buildout:

was based on the assumption that telecoms traffic, particularly data traffic, would continue to grow exponentially for the foreseeable future. The advent of wavelength-division multiplexing reduced the demand for fibre by increasing the capacity of a single fibre by a factor of as much as 100.

According to Gerry Butters, the former head of Lucent Technologies's Optical Networking Group at Bell Labs, the amount of data that could be carried by an optical fibre was doubling every nine months at the time. This progress in the ability to carry data over fibre reduced the need for more fibres. As a result, the wholesale price for data communications collapsed and a number of these companies filed for bankruptcy protection. Global Crossing and Worldcom are two high-profile examples in the United States.

The analogy being suggested is that the AI data centers will eventually be valuable infrastructure even if delays in construction and lack of real demand mean that the companies building them will suffer.

Below the fold I investigate this analogy.

Flawed Analogy

There are some significant problems with the analogy between dark fiber and AI data centers:

• Depreciation: Decades later the fiber in the ground was still fully functional. The capacity could be sold for very little additional investment. Estimates are that about 40% of the cost of an AI data center is the physical infrastructure (buildings, power connections, cooling, etc.) which might still be useful a decade later. But 60% of the cost is the (mostly Nvidia) racks full of state-of-the-art chips.
  
  SourceMind The GAAP Again and its predecessors, this is unrealistic. Because:
  
  • There is intense price competition in AI.
  • Power is a major cost of both inference and training. Power effectively sets the price floor for inference.
  • Each technology generation improves power efficiency by a significant factor.
  • Thus the profit from running a rack doesn't decay linearly, but exponentially.
  • Thus racks more than a couple of generations old are uneconomic.
  • Nvidia has announced that they plan to move from a two-year product cadence to a one-year product cadence.
• Overproduction: Wikipedia correctly notes what drove the massive overproduction of dark fiber:
  

  Because the marginal cost of installing additional fibre optic cables is very low once a trench has been dug or conduit laid, a great excess of fibre was installed

  Building data centers does not have the same low marginal cost. While it is true that there are economies of scale in construction driving up the size of AI data centers, these are both much smaller than for laying fiber and apply only to the smaller part of the overall cost.
• Carrying Costs: The fiber in the ground required very little maintenance. The physical infrastructure of unused AI data centers could survive until eventual re-use only with significant maintenance costs.
• Technology evolution: Developments in optical technology increased the capacity of the unused fiber with no additional investment. The evolution of AI racks has meant that data centers built a few years ago cannot host current racks, and that the design of data centers for Nvidia's future racks is a research problem.

Dark fiber thus had two important attributes that AI data centers lack; the surplus initial capacity was both long-lived and cheap.

Construction Delays

Building and activating an AI data center is a much bigger, more expensive and long drawn-out task than digging a trench and laying a cable in it. This problem grabbed the headlines when, as Sridhar Natarajan et al reported, Oracle Cites ‘Force Majeure’ to Shield Itself on Controversial Data Center. Almost two years into the project:

the campus designed to handle 2.45 gigawatts — enough electricity to power roughly 1.8 million homes at any given moment — has hit serious setbacks including the denial of a permit key to its plans for energy resources. It’s becoming a symbol of the growing backlash against the broader AI data center boom and a political flashpoint ahead of the US midterm elections that threatens to turn races into referendums over the build-out.

The amount of "construction in progress" for the hyperscales, representing data centers they are struggling to bring on-line, is enormous. Ed Zitron writes:

Considering Oracle’s CIP number from its latest quarterly earnings was at $48.5 billion and Google’s Q2 2026 “assets not yet in service” were at an astonishing $122.8 billion, up from $108.5 billion in Q1 2026 and $78.5 billion at the end of 2025, it’s reasonable to believe that Microsoft has at least $50 billion of construction in progress. I also think it’s fair to assume that Amazon has, since the beginning of 2026, added at least $25 billion to CIP, though we’ll find out at the end of the year.

Zitron quotes a Frebruary 2026 estimate from Sightline Climate:

We’re tracking 190GW across 777 large data centers and AI factories (>50MW) announced since 2024. At least 16GW of capacity is slated to come online in 2026 across roughly 140 projects. Yet only about 5GW is currently under construction. Around 11GW remains in the announced stage with no visible construction progress, despite typical build timelines of 12–18 months.

Anyone who has built themselves a house will understand that construction takes longer than it takes.

Overproduction

Of course, only 40-50% of this "construction in progress" is actual construction, the rest is the racks. There are two ways to estimate what proportion of the racks Nvidia has shipped are actually powered on in a data center as opposed to sitting in a warehouse waiting for the construction to stop being "in progress".

Ed Zitron bases his estimate on the financial picture. He adds up the "construction in progress" and assumes 50% is the racks:

if we assume the number is roughly $374 billion, plus (at minimum) $50 billion from Microsoft, plus another (at minimum) $20 billion from Amazon, that puts us around $444 billion, with Google’s share — $120.8 billion — being 60% GPUs and related hardware, for a total of $72.48 billion, putting us at ... an estimated $234 billion in GPUs and TPUs sitting in warehouses.

He goes on to compare this $234B number with the value of the racks sold since the start of 2023, which he estimates at $561.5B. Clearly, the systems that shipped early in this time-span are likely to have been powered up, so he estimates that around 50% of recent shipments are idle. He looks in detail at Microsoft's financials and and estimates:

Nadella mentioned in November 2025 that he had “...a bunch of chips sitting in inventory that [he couldn’t plug in],” but didn’t make any mention of how many there were.

In other words, Microsoft is warehousing anywhere from $50 billion to $100 billion in GPUs, and has barely gotten $50 billion worth installed in the last four years.

If Microsoft is struggling, everybody’s struggling, and we may have a very inconvenient truth: that NVIDIA has potentially sold hundreds of billions of GPUs years in advance.

SourceOK, so where are all these data centres?, Bryce Elder wrote:

A theme we’ve touched on a few times is the mismatch between consensus forecasts for chip sales and US data centre completions.

Morgan Stanley measured the gap earlier this week by estimating the shortfall in available power, concluding that more than half of the GPU servers sold between 2026 and 2028 might not have anywhere to be plugged in. New research from Jefferies analysts reaches a similar conclusion using a very top-down method: space.

Tracking US data centre construction sites by satellite imagery shows between 16 and 18 gigawatts of gross capacity that can be energised this year, the broker says, citing SynMax Intelligence. (Gross capacity means the site’s all-in power requirement, not just the draw for IT equipment.)
...
In the longer term, power availability is still the bottleneck — along with labour. And transformers. And cooling equipment. And backup generation. As Jefferies says: “The gap between planned capacity and physical execution remains the central issue.”

Thus there is an emerging consensus that something like a quarter of the AI investment boom is going into rapidly depreciating equipment quite unlike the dark fiber.

Demand Overestimation

The frontier labs and the hyperscalers claim that there is overwhelming demand for AI compute, which justifies the planned AI data center buildout. But the frontier labs are private and don't publish audited numbers to back up their claims. The hyperscalers are public but their accounts don't include explicit numbers for their AI compute. Nvidia is public and its audited accounts reveal the number of AI systems it shipped each quarter. So the easy way to estimate the demand for AI compute is to base it on Nvidia's shipments.

There are three problems with such estimates. Assume this estimate is D and the real demand from end-users is D':

• As we have seen, it appears that about half of Nvidia's shipments go to a warehouse not a data center. Thus D' = D/2.
• Ed Zitron estimates that 80% of D comes from OpnAI and Anthropic. So D/10 comes from other sources, which we will generously assume is all end-user demand.
• In Thoughts on The Curve and the Future of AI Thomas Wright reported an estimate:
  

  that the labs spend 58% of their compute on R & D and 32% on servicing customers.

  Thus the labs end-user demand is D*0.8*0.32, or about D/4.

Adding these, we have D' = D*0.35. The real end-user demand is likely only around 1/3 of the estimates.

SourceWhat’s really going on with AI token price deflation?, Bryce Elder points out that they don't:

The data [Jay Goldberg] collects shows a similar pattern to the Silicon Data index, with token prices plunging for frontier LLMs from $50 per million tokens in late 2023 to about $6 now. More interestingly, he finds the median token price has held fairly steady at just below $1 per million since mid-2024 as the LLM marketplace became more competitive. A floor price of about $0.11 (basically just the cost of electricity) has barely budged.

@arakharazian tweets data from Ramp::

AI Spend FALLS in the latest Ramp AI Index.

Price cuts at the frontier, in addition to cheaper and more efficient models at standard and lite levels, are pushing down the cost of using AI.
Open source models remain <5% of business spend. This decline in spend is driven almost exclusively by competition between OpenAI + Anthropic

The Ramp data does show increased token volume, but flat to decreasing spending. The increased colume is likely due to successive models burning more tokens per query.

Market Reaction

We now have some evidence showing how the big banks rate GPUs as collateral for loans, i.e. how useful they are likely to be down the road.

On 5th October Paula Seligson and Gowri Gurumurthy reported that JPMorgan Dangles 11% Yield on Volta’s $5 Billion AI Loan Sale:

JPMorgan Chase & Co. is pitching a yield of about 11% on a $5 billion leveraged-loan sale on behalf of Volta Infrastructure Holdings Ltd., one of the highest borrowing costs seen in the market for risky debt that’s become popular for financing the AI boom.
...
A $3.7 billion chip loan will be used to purchase about 36,000 graphics processing units, or GPUs, made by Nvidia Corp., according to the people. A $1.3 billion loan will cash collateralize a letter of credit for a data center lease, meaning the money will be held in an account and used to make lease payments to Bitdeer to use the data center, the people said.

11% is a pretty high interest rate, indicating a signinificant risk of default. A week later Claire Ruckin reported that JPMorgan, Goldman Curb Exposure on $5 Billion Volta AI Loan:

JPMorgan Chase & Co. and Goldman Sachs Group Inc. have taken a rare move to limit their exposure on a $5 billion loan sale for AI startup Volta Infrastructure Holdings Ltd. as investors demand higher yields for risky debt tied to the artificial intelligence boom.

The two banks are leading the loan offering for Volta, a new artificial intelligence cloud company. The banks have underwritten a $1.3 billion term loan for Volta, while a $3.7 billion chip loan is not underwritten by the banks — a move to protect themselves from potential risks, according to people familiar with the deal.

In other words, these banks are happy to lend at 11% against cash collateral, but not at 11% against GPU collateral. The loan is said to be oversubscribed, but the part with GPU collateral is for the private credit suckers.

SourceCoreWeave’s Bonds Turn Distressed, Drawing ‘Yikes’ From Musk:

The AI cloud-computing provider’s unsecured debt due in 2032 traded below 80 cents on the dollar on Wednesday, according to Trace pricing data. That’s widely considered the threshold for when a borrower is financially strained.

At that price, the bonds yield more than 15% — a figure noted by Palihapitiya that led Musk to respond “yikes” on his social media platform.

Shares of the company, which rents out access to leading AI chips, fell as much as 17% on Wednesday. Even though its bonds are trading at distressed levels, CoreWeave’s stock is up more than 60% since its initial public offering in March 2025 and its market value is more than $36.5 billion.

Ruckin noted that:

CoreWeave’s credit-default swaps are back at a record high, and at current levels are pricing in a more than 50% chance of default within the next five years.

CoreWeave is the original "neocloud". As I understand it, its assets against its debts are (a) a set of commitments from companies like Microsoft and Oracle to pay for future access to compute, and (b) a set of data centers full of GPUs. The market is betting that these assets won't be "money good" in the near future.

Coda: A Sun Microsystems Story

Nvidia is just the most successful of the vast number of companies spawned by Sun Microsystems. In the run-up of the dot-com bubble Sun's Starfire systems powered the Web the way Nvidia's GPUs power AI today. But this market dominance was a double-edged sword. Once the dot-com bubble burst, the bankrupt Web companies flooded the market with barely-used Starfire systems. So not merely did much of the customer base vanish, but the remaining customers bought from the second-hand market at huge discounts. Demand for new systems collapsed, and Sun never really recovered. That's not the only reason Sun died. Scott McNealy admitted in a fascinating podcast that during the boom he lost control of Sun's spending. It is hard to maintain discipline when you are being flooded with cash.

Once the bubble bursts, Nvidia will face a similar problem. First, their customers will stop ordering new racks, because:

• Companies like CoreWeave that exist to get debt off the AI companies' balance sheets will be bankrupt because their customers can't afford their commitments.
• The assets of these bankrupt companies will be sold off at a loss, flooding the market with barely-used data centers and stopping construction of new ones.
• It has proven harder than expected to build data centers quickly because of public opposition, lack of skilled tradespeople, difficulties in securing power and so on. Thus it is likely that there will be warehouses full of Nvidia racks that were waiting for a data center to be built when the bubble burst. These unused racks will also be sold off, crashing the price of the barely-used ones further.

Second, in some ways the problem will be worse for Nvidia than Sun. Nvidia will have invested heavily in developing and securing wafer slots for their next-generation GPUs. Although these will be faster and more power-efficient than those in the written-down data centers, they won't be able to compete on price-performance. The pipeline between committing to a new design and market entry is now much longer than it was in the '90s. Nvidia will have relatively more capital tied up in the product pipeline than Sun did.