Elon Musk wrote on X on 25 October 2026: “I think we’re going to see a lot of stranded compute in partially complete data centers.” The post, a reply to Yuchen Jin on X, gives no figures, sites or dates. The same week, our reporting found an Nvidia H100 renting for $2.88 per GPU-hour on Silicon Data’s index, against $2.00 on 9 December 2025: a rise of 44% by our arithmetic, as our H100 piece reported.
This piece sets Musk’s line beside both kinds of signal in our October reporting: the ones that point to scarce, dearer compute, and the ones that point to sites and spending under strain. The figures come from different sources and measures, so they are set side by side and not combined.
This Week In The AI Supercycle · VIDEO
The Business Engineer’s explainer for the week of 18 to 24 October 2026: the companies selling AI’s parts say they are booked for years, while the companies paying for the build found money dearer. It was made before Musk’s post and does not discuss it.
Business Pill · NEOCLOUDS
A short explainer of neoclouds: companies that rent AI chips by the hour. It teaches the general idea only and says nothing about Elon Musk or any company named here.
The key insight: As we read it, Musk’s line and the rent data need not contradict each other: the rents and sold-out remarks describe finished capacity, while Musk wrote of “partially complete data centers,” and our reporting holds campuses that need permits during a freeze and a withdrawn plan.
Musk was replying to a post by Yuchen Jin, who wrote that a “neolab friend” had said: “Even if we fail to train the best model, we locked in GPU capacity early. H100/B200/B300 rental prices have since gone up 2–3x. We can just resell the compute and print money.” Jin’s post names neither the lab nor a source for the 2–3x figure, and we have not checked it.
Musk’s reply is one sentence. It does not say which data centers are meant, why they would be partially complete, or when. It is a post on X, not a company statement.
Silicon Data H100 neo-cloud rental index, published readings from 9 December 2025 ($2.00) to 24 October 2026 ($2.88), in dollars per GPU-hour, as reported in our H100 piece. The 44% rise is our arithmetic.
What Was Posted on X
@Yuchenj_UW I think we’re going to see a lot of stranded compute in partially complete data centers
Elon Musk, replying to Yuchen Jin on X on 25 October 2026, writes: “I think we’re going to see a lot of stranded compute in partially complete data centers.” The post gives no figures, sites or dates.
$SPCX CEO Elon Musk says we’re going to see “a lot of stranded compute in partially complete data centers” as GPUs arrive faster than power infrastructure needed to run them.
That sounds bearish but could actually keep GPU rental prices higher for longer as companies with power already secured benefit from tight supply while idle chips lose value with every new generation.
Shay Boloor (@StockSavvyShay) highlighted on 25 Oct Elon Musk’s line that there will be “a lot of stranded compute in partially complete data centers” as GPUs arrive faster than the power to run them. Boloor’s view: that could keep GPU rental prices higher for longer, as companies with secured power benefit while idle chips lose value with each new generation.
AI data center tenants are now demanding penalties of more than $5 million per day if developers fail to deliver power and facilities on time, up from around $1 million earlier this year, according to The Information.
Wall St Engine (@wallstengine), citing The Information, says AI data center tenants are now demanding penalties of more than $5 million per day if developers fail to deliver power and facilities on time, up from around $1 million earlier this year.
Data center tenants now demand $5M+ a day in late-delivery fees, up from $1M this spring, spawning new “delay-in-delivery” coverage from brokers like Marsh (@anndavisvaughan / The Information)
(Visit Techmeme dot com for the link and full context!)
Techmeme (@Techmeme), summarising The Information’s Ann Davis Vaughan, says data centre tenants now demand $5M+ a day in late-delivery fees, up from $1M this spring, spawning new “delay-in-delivery” coverage from brokers like Marsh.
🚨 @ElonMusk ON WHY THERE’S NO AI DEMAND PROBLEM:
“We’ve actually tried to offload some of our inference compute to other cloud providers so that we can use more compute for training, and we get the Cheese Shop Sketch”
He says no one has any chips left to rent:
“We’ll ask them, ‘What about some GB300s?’ ‘Nope, don’t have any of those.’ ‘What about some TPUs?’ ‘Nope.’ ‘How about some H100s?’ And they’re like, ‘Well, we might have some overseas somewhere.’ And then they’ll come back and say, ‘Nope, we don’t have that either'”
A post by @wolf_terminal at 04:00 UTC on 8 November resurfaces a clip first posted by @CEOinterview on 25 October, in which Musk says: “We’ve actually tried to offload some of our inference compute to other cloud providers so that we can use more compute for training,” and describes being told there were no GB300s, TPUs or H100s to rent, adding: “I don’t think we’re gonna have a demand issue here.” It is one executive’s account of tight GPU supply, set beside the rental-price data in this piece.
The Squeeze Signals
Silicon Data’s published H100 readings went from $2.00 on 9 December 2025 to $2.88 on 24 October 2026, as our H100 piece reported. A second index points the same way: a chart published by a16z, using Ornn Data as of 19 October and indexed to 1 January 2026 = 100, shows the H200 ending near 260 and the H100 SXM near 135, as our H100 piece reported in an update. The chart prints no values, so those readings are approximate.
Amazon chief executive Andy Jassy said of AWS computing capacity, Reuters reported on 23 October: “2027 is largely sold out.” Jassy also said a good chunk of AWS’s 2028 is sold out, as our sellout piece reported. On the a16z podcast, as quoted by The Transcript, Jassy said: “even spending $220B in CapEx this year, we’re leaving a meaningful amount of demand on the table.”
Signal
Figure
Source
H100 rent
$2.00 to $2.88 per GPU-hour (+44%, our arithmetic)
Silicon Data, via our H100 piece
H200 rent index
Near 260 (1 Jan 2026 = 100), approximate
a16z chart of Ornn Data, via our H100 piece
AWS capacity
2027 “largely sold out”
Jassy, via Reuters, via our sellout piece
Amazon capex
About $220B in 2026, demand left “on the table”
Jassy on the a16z podcast, via our sellout piece
Texas permits
Six campuses, 8.7GW, need air permits during the freeze
Sources: as listed in each row, as reported in our earlier pieces. The rows use different measures and are set side by side, not combined.
The Strain Signals
Some sites are stuck. The research firm SemiAnalysis said six Texas data center campuses planning their own on-site power, “8.7GW in all,” need air permits while Governor Greg Abbott’s permit freeze is in force, as our Texas piece reported. Abbott’s letter of 21 September tells the state environmental regulator “to pause the issuance of all permits related to data center projects until ERCOT completes its review.” ERCOT will present its findings on 10 December, the San Antonio Report said.
Some plans have been dropped. Amazon has withdrawn its plan to build four data center buildings on the former George Washington University campus in Ashburn, Virginia, which it bought for $427 million, according to an agreement announced by Loudoun County Board of Supervisors Chair Phyllis Randall and reported by Fortune, as our Amazon Ashburn piece reported.
The buyers are concentrated. Ramp’s 14 October letter says “80% of spend on OpenAI and Anthropic is driven by 1% of firms” and records “two straight months of virtually no growth in token spend,” as our Ramp piece reported. Apollo’s Torsten Slok wrote that “mainstream corporate America is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins,” as our Apollo piece reported.
And one scenario prices the downside. TS Lombard’s crash scenario of 22 October puts credit losses on AI-related debt at $125 billion to $190 billion, against about $20 trillion of losses for equity holders, and says the strain is “going to appear upstream,” as our TS Lombard piece reported. It is a scenario, not a forecast of losses.
How Both Can Be True
As we read it, the two sets of signals describe different capacity. The rent indexes and Jassy’s remarks describe compute that is built and running. The Texas freeze and the withdrawn Ashburn plan describe capacity that is not finished: SemiAnalysis said the six Texas campuses need air permits for their on-site power during the pause, and Amazon has withdrawn the four Ashburn buildings. Musk’s phrase, “partially complete data centers,” points at the second group, though Musk’s post does not say which sites are meant or why.
On that reading, scarcity in finished capacity and stranded half-built sites are not opposites. Our sources do not measure how much capacity is partially complete, so this is our interpretation of the signals, not a finding.
Update (25 Oct): late fees on unfinished sites. The Information reported that AI data-centre tenants are now demanding penalties of more than $5 million per day if developers fail to deliver power and facilities on time, up from around $1 million earlier this year, according to a relay of the report by Wall St Engine. The Information’s headline, by Ann Davis Vaughan, says the late fees are “spawning new insurance.” We have read only the headline and the relay, not the full article, and neither names the tenants or the sites. That puts a price on the gap Musk described: a partially complete data centre is now a daily liability for the developer, not only idle capital. Wall St Engine’s post.
Connecting the Coverage
Our H100 piece of 24 October traced the rent from $2.00 to $2.88, and its update added the a16z chart of Ornn Data. Our sellout piece of the same day ranked how far ahead suppliers and AWS say they are booked. Our compute squeeze piece set rising rents against falling token prices.
Our Texas piece of 24 October reported the permit freeze and the 8.7GW count. Our Amazon Ashburn piece of 23 October reported the withdrawn campus plan. Our Ramp piece and our Apollo piece reported how narrow the paying customer base is, and our TS Lombard piece reported the crash scenario. What this piece adds is Musk’s line, set against both sets of signals.
The Structural Read
As we read it, the squeeze evidence is about built capacity: Silicon Data’s H100 index read $2.88 per GPU-hour on 24 October against $2.00 in December, and Jassy said a good chunk of AWS’s 2028 is sold out.
As we read it, the strain evidence is about unfinished sites and the paying base: SemiAnalysis says six Texas campuses, 8.7GW in all, need air permits during the pause, Amazon withdrew four buildings in Ashburn, and Ramp says 1% of firms drive 80% of OpenAI and Anthropic spend.
As we read it, the missing number is partially complete capacity: none of our sources counts it, and Musk’s post names no sites.
“I think we’re going to see a lot of stranded compute in partially complete data centers”
Three Implications
FOR RENTERS OF COMPUTE Silicon Data’s H100 index read $2.88 per GPU-hour on 24 October, up from $2.00 on 9 December 2025; 44% is our arithmetic.
FOR DEVELOPERS SemiAnalysis says six Texas campuses, 8.7GW in all, need air permits during Abbott’s freeze, and ERCOT presents its findings on 10 December, the San Antonio Report said.
FOR LENDERS TS Lombard’s crash scenario puts credit losses on AI-related debt at $125 billion to $190 billion and says the strain is “going to appear upstream.”
The framework puts it this way: “The structural bottleneck isn’t computing, it’s electrons. Companies that solve the power equation will dominate the AI era; those that don’t will watch their chips sit idle.”
As we read it, this week’s signals sit on both sides of that line: finished capacity is renting for more, while the six Texas campuses SemiAnalysis counts need air permits for their on-site power during the pause. The framework was published in December 2025 and does not discuss Musk’s post or these figures.
What Is Not Established
Musk’s post offers no data, and we do not know which sites, if any, it refers to. The 2–3x figure in Yuchen Jin’s post comes from an unnamed source and has not been checked. Silicon Data’s index covers neo-cloud rates, not long-term contract prices, and the a16z chart prints no values. None of our sources counts how much data center capacity is partially complete. We did not contact the people or companies named.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
Musk wrote “I think we’re going to see a lot of stranded compute in partially complete data centers” in the same week that Silicon Data’s index showed an H100 renting for 44% more than in December, by our arithmetic. Our reporting holds signals of both kinds: sold-out finished capacity, and campuses that need permits during a freeze or plans that were withdrawn.
A note on sourcing. This piece quotes Elon Musk’s and Yuchen Jin’s posts on X and is otherwise built only from our own earlier pieces of 23 to 25 October 2026 and the sources they cite: Silicon Data, an a16z chart of Ornn Data, Reuters, The Transcript, SemiAnalysis, the Texas Governor’s letter, the San Antonio Report, Fortune, Ramp, Apollo and TS Lombard. Figures from different measures are set side by side, not combined. We did not contact the people or companies named.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
