THE MW DESK / BLOG
The Landlords of Compute: Does a Data Center Freeze Hand Hyperscalers the Pricing Power?
The thesis: stall new data centers and whoever already owns powered capacity sets the price of AI. We tested it against rental indexes, earnings calls, contract terms and the best counterarguments. It mostly holds, with two big corrections: the squeeze hits renewals rather than the spot screen, and the winner is whoever holds energized megawatts, not just whoever has a cloud logo.
Here is the thesis, in plain English. If new data centers keep getting stalled by moratoriums, grid queues, angry county meetings and transformer backlogs, the companies that already own powered capacity end up holding the scarcest input in AI. Scarce inputs get expensive. So the hyperscalers win, everyone renting from them pays more, and the price of compute could go vertical.
It’s a good thesis. Chamath Palihapitiya made a version of it on the November 7 episode of All-In (episode list). We checked what he actually said against a third-party transcript before quoting him. A slowdown, he said, “benefits the hyperscalers because they have existing capacity,” is “very bad for the neoclouds because they’re renters,” and leaves the frontier labs exposed “because they are so short this capacity.” Later: “The dollar cost per megawatt per month is going to go through the roof. The dollar cost of a GPU hour is going to go through the roof.” David Sacks added the timing: fully permitted projects probably get built in 2027, but by 2028 he expects the pipeline to be “either closed or significantly reduced.”
Smart people saying something on a podcast doesn’t make it true. So we went looking for evidence that would break it.
Exhibit A: the price of a GPU-hour already turned
For two years, the safest trade in AI infrastructure was waiting. H100 rental prices fell almost every quarter from the 2023 peak, and in June 2025 AWS cut on-demand H100 instance prices by 44% (Compute Census). That era is over.
- Contract rates. SemiAnalysis’s survey of one-year H100 contracts bottomed around $1.70 per GPU-hour in October 2025 and reached about $2.35 by March 2026, roughly 38% higher in five months (Compute Census, summarizing SemiAnalysis).
- Executed trades. Ornn’s daily H100 index settled at $3.37 on November 8, up 27% in a month. September ended at $2.69 and October at $3.00 (Ornn).
- List prices. Nebius raised rates across its fleet on October 1. Lambda announced increases of about 15% effective November 16 (H100 from $3.99 to $4.59, B200 from $6.69 to $7.69). Verda raised H100 SXM prices ten times in fifteen days (GPUJournal).
The tell is what is getting repriced: old chips. Nobody launched a new GPU in October. IREN told investors on November 5 that it had recontracted older H100s with a frontier lab “at double the original pricing,” and IREN’s Kent Draper stressed on the same call that it was a multiyear term, not a one-month spike (IREN Q1 FY27 call). A three-year-old chip that gets more expensive is not a chip story. It’s a story about the building it’s plugged into.
The per-megawatt math says the same thing. CoreWeave said in September it was signing three-to-six-month contracts at about $40 million of annualized revenue per megawatt (CoreWeave). IREN quoted a five-year deal at $22 million per megawatt and a shorter one above $40 million. In other words, buyers who need capacity now are paying nearly double what long-term buyers pay. That premium is the market putting a price on time to power.
Exhibit B: the biggest buyers say they’re supply-constrained
If hyperscalers had spare capacity, they wouldn’t be talking like this:
- Microsoft trimmed its calendar-2026 capex to about $170 billion from $175 billion, but not because of weak demand. CFO Amy Hood said third-party GPU capacity deliveries slipped into the first half of 2027 (CNBC). Her summary: “We still have demand that exceeds supply” (The Deep Feed, citing Fortune).
- Alphabet held $195–205 billion for 2026 and said outside computing power “came online only at the very end of Q3” (CNBC). Google is renting capacity.
- Andy Jassy, on the a16z podcast: “‘26 is sold out… ‘27 is largely sold out. A good chunk of ‘28 is sold out” (DCD).
Then there’s Wyoming. On November 6, Google acquired the southeast Wyoming site where Crusoe had handled site selection, government approvals and a power strategy with Tallgrass Energy (Crusoe). TechCrunch reported that Google, which Crusoe had tried and failed to land as a tenant, took the project over after raising concerns about cost and timeline (TechCrunch). Chamath cited the same deal on air. When the tenant buys the building, you know where the leverage sits.
Exhibit C: the supply pipe really is narrowing
Heatmap Pro counts at least $260 billion of U.S. data center investment canceled in 2026 after sustained local opposition. About half of that, roughly $130 billion, came in the third quarter alone, and about $1 trillion of projects now face meaningful opposition (Heatmap). Those figures were published before the November 3 election. ConstructConnect recorded $9.4 billion of new data center construction starts in September, less than half of August’s record $25.8 billion. That’s one month, but it’s the earliest signal we get (Yahoo Finance).
The day after the election, Microsoft and Amazon each rose more than 6% while CoreWeave fell about 10%. JPMorgan’s James Sullivan told CNBC that existing capacity becomes more valuable when new supply is constrained (CNBC). The market voted for the thesis within 24 hours.
Now let’s try to break it
Break #1: hyperscalers are renters too
The cleanest version of the thesis says hyperscalers own the bottleneck. The earnings calls say otherwise. Microsoft’s capex miss came from third-party GPU-as-a-service contracts arriving late, and CNBC noted it relies on CoreWeave, IREN and Nebius, among others. Microsoft’s Q3 capex also included $7.4 billion of finance leases. Alphabet started renting outside compute this quarter.
So the correct version is narrower: the winners are whoever controls energized megawatts and has the balance sheet to wait. That’s mostly the hyperscalers. But it also includes neoclouds and former bitcoin miners sitting on live power, which is how IREN gets to double its H100 price. The losers are anyone who has to buy future megawatts with borrowed money.
Break #2: “skyrocket” isn’t what the data shows
Look closer at the price data and you see three markets moving at different speeds:
- The floor didn’t move. The cheapest verified H100 SXM offer on GPUJournal’s board was $2.25 an hour on November 8, unchanged since October 19. Community-tier H100, H200 and B200 prices were flat too (GPUJournal).
- Hyperscaler sticker prices barely move. They were already at a premium. Compute Census lists on-demand H100 at about $11 on Google Cloud and $12.29 on Azure, against a $3.90 market median, and notes that a hyperscaler price index held almost flat through the spring (Compute Census).
- The middle is where it’s happening: neocloud list prices, short-dated contracts and renewals.
So hyperscalers aren’t exercising pricing power by putting up a bigger number on the rate card. They’re doing it through allocation and terms: who gets capacity, for how long, and with how much cash up front. IREN’s five-year deal came with a customer prepayment covering roughly 90% of the GPU capex. Sold-out years, prepayments and take-or-pay commitments are pricing power. They just don’t show up on a price chart.
Break #3: the cost of a token is still falling
This is the strongest counterargument. Citadel Securities estimates consumption-weighted token prices fell about 50% in six months while token volumes rose roughly 5x, driven by cheaper open-weight models and better inference efficiency (Citadel Securities). Same note: H100 rentals up about 14% over six months, B200 rentals up 27.5%.
Put those together and you get a squeeze, not a spike. On Citadel’s illustrative math, a lab facing a 27.5% rise in compute costs and a 50% drop in token prices needs about 2.55x more throughput per GPU just to hold gross margin. For an enterprise buying tokens, that’s good news, for now. For labs and neoclouds stuck in the middle, it’s a vise. Citadel’s analogy fits: labs are the airlines, open-weight models are the low-cost carriers, and hyperscalers are the airports.
Break #4: custom silicon is a quiet price war
Google is selling and renting TPUs to outside customers. SemiAnalysis found TPUv7 Ironwood delivered up to 50% better performance per dollar than B200 and B300 in its September tests. On one agentic benchmark published November 5, Ironwood delivered 1.2x to 2.7x the tokens per dollar of B200, B300 and MI355X (SemiAnalysis). More useful work per megawatt is real deflation.
But notice who gets to sell that efficiency: the company that owns the TPUs and the halls they sit in. Custom silicon drives down the cost of intelligence. It doesn’t create new megawatts, and it widens the gap between hyperscalers and everyone renting Nvidia.
Break #5: the overbuild and finance case
Our own finance slowdown scenario book treats a severe freeze as a tail risk, not the base case, and flags a second danger: stranded silicon. The bear case deserves a fair hearing. Apollo’s Torsten Slok argues demand for compute is highly price-sensitive and hyperscaler returns on invested capital are falling. Larry Swedroe points to CoreWeave’s 2031 bonds yielding above 16%, CDS above 1,100 basis points, and a maturity mismatch: long leases funding shorter customer contracts (WealthManagement.com). Citadel flagged Ramp data showing enterprise AI spend flat in August and September.
Timing cuts both ways, too. Nomura says most capacity due in 2027 is already under construction, so the freeze bites hardest in 2028 (CNBC). Prices can’t “go through the roof” while a big 2027 delivery wave is still landing. And as Jason Calacanis argued on the same episode, training can move to Canada, Mexico or Malaysia.
The verdict
The thesis is right on direction and right on who wins. “Skyrocket” goes too far, and the mechanism isn’t what most people picture.
- Power is the bottleneck, and the price of powered capacity is rising. Old H100s recontracting at double, $40 million-per-megawatt short-term deals, sold-out 2027 order books and a hyperscaler buying a developer’s stalled campus all point the same way. This was already happening before the election, and a stalled 2028 pipeline extends it.
- The winners are the landlords: owners of energized megawatts with balance sheets. Hyperscalers are the biggest of them. Their pricing power shows up in allocation, prepayments and take-or-pay terms, not in rate-card spikes. Neoclouds and former miners with live power win too, if their lenders let them wait.
- The losers are leveraged renters in the middle. That means labs and debt-funded neoclouds paying more for compute while selling cheaper tokens. A broad “compute shock” to the economy isn’t happening yet, because per-token prices keep falling. Watch 2028: that’s when a closed permit pipeline meets demand that hasn’t slowed. If both are true at once, Chamath’s “through the roof” stops being a podcast line.
What it means for you
Builders and developers. Your entitlements and your power position are now the product. Google buying Crusoe’s site, rather than leasing from it, is a warning. If you can’t deliver on time and on budget, a hyperscaler will buy the project and finish it without you. Sites with power in hand, clean entitlement records and community deals already signed will command a premium. List them where buyers can find them.
Buyers, meaning enterprises renting compute. Don’t benchmark against the cheapest spot listing. Benchmark against your renewal. If your workload is steady, lock multi-year terms now. Ask any provider selling future capacity whether it’s energized today, whose power contract backs it, and what happens if the phase slips. Microsoft absorbed a delayed delivery this quarter. You probably can’t.
Investors. Separate owned-and-energized from contracted-but-unbuilt. The spread between hyperscaler and neocloud cost of capital is the trade the market made on November 4. The catch for the landlords: renting more outside capacity and paying up for scarce power costs real cash: Alphabet reported negative free cash flow of more than $9 billion for the quarter (CNBC). $MSFT, $AMZN, $GOOGL, $META, $CRWV, $IREN and $NBIS are labels for companies in sourced reporting, not recommendations.
What to watch
- Renewal prints, not spot prints. More disclosures like IREN’s “double the original pricing” would confirm the thesis. Renewals at flat rates would weaken it.
- Weekly H100/B200 indexes, and whether the cheapest offers start moving. When the floor moves, scarcity has reached every buyer.
- Hyperscaler Q4 calls in late January: 2027 capex numbers, third-party capacity language and any mention of prepayments or capacity reservation fees.
- Token price vs. GPU-hour divergence. If token prices stop falling while rentals keep rising, the squeeze hits end customers.
- 2028 permits. Texas, Georgia and other new-governor states: whether moratoriums become statute, and which projects get grandfathered. Our legislation tracker follows them.
- Neocloud credit spreads and the next stalled campus a hyperscaler picks up.
Not investment advice. $cashtags are labels for publicly traded companies in sourced reporting, not recommendations to buy or sell.
Sources
- All-In, episode list (Nov 7, 2026): allin.com ↗
- WhipScribe transcript of the All-In episode: podcasts.whipscribe.com ↗
- Ornn, OCPI-H100 rental price index: data.ornn.com ↗
- Compute Census, “Are GPU Rental Prices Going Up?”: computecensus.com ↗
- GPUJournal, “Power is the new GPU shortage”: gpujournal.ai ↗
- CoreWeave, Sept 17, 2026 release: investors.coreweave.com ↗
- IREN Q1 FY27 earnings call transcript (Motley Fool): fool.com ↗
- CNBC, Microsoft and Google capex (Oct 29, 2026): cnbc.com ↗
- The Deep Feed, Q3 hyperscaler capex: thedeepfeed.ai ↗
- DCD, AWS capacity “largely” sold out in 2027: datacenterdynamics.com ↗
- Crusoe, Wyoming campus update: crusoe.ai ↗
- TechCrunch, Crusoe sells Wyoming project to Google: techcrunch.com ↗
- Heatmap, data center cancellations top $260 billion: heatmap.news ↗
- CNBC, midterms and the AI data center divide (Nov 5, 2026): cnbc.com ↗
- Yahoo Finance, UBS CEO warns of AI buildout slowdown: finance.yahoo.com ↗
- Citadel Securities, “Tokenomics 2.0”: citadelsecurities.com ↗
- SemiAnalysis, “The Iron(wood) Throne”: newsletter.semianalysis.com ↗
- WealthManagement.com, Larry Swedroe on AI capex risk: wealthmanagement.com ↗