Why power, not chips, sets the value of AI compute, and where the real megawatts are being built
No power, no cheese.
INSIDE
1. The cheese shop, October 20. What SpaceX’s Q3 call tells us about the price of a megawatt
2. One megawatt, two chips. H100 vs. B300 on the same 1 MW footprint
3. The economics of one megawatt. Capital, revenue and payback, with sensitivities
4. What a megawatt is worth. Revenue per MWh and what it means for curtailment
5. Which scarcity are we in? Universal scarcity vs. forced hardware eviction
6. Where the megawatts actually are. Satellite-verified construction data from Vulcan
7. What to do next. Three steps and how Vulcan can help
WHAT YOU WILL TAKE AWAY
|
By the end of this paper you will be able to: 1. Convert any GPU rental rate into dollars per megawatt-hour, and see why a megawatt of Blackwell earns about twice a megawatt of Hopper. 2. Say which scarcity scenario we are in, and what happens to older chips when power, not silicon, is the limit. 3. Tell which announced datacenter megawatts are being built, using satellite evidence rather than press releases. |
|
$24.6M gross revenue per MW-year on B300 (H100: $12.6M) |
27% fewer GPUs per megawatt on B300 vs. H100 |
$2,812 revenue per MWh consumed on B300, vs. a $150 power cost |
39% of U.S. capacity listed for 2026–28 shows construction on satellite imagery (54 of 138 GW) |
The idea in one line: power is the scarce asset, so price every megawatt accordingly.
On October 20, SpaceX’s CEO told investors that the limit on deploying AI compute is now power, cooling and infrastructure rather than chips, and that SpaceX could rent out its compute several times over. He compared the hunt for spare GPUs to the Monty Python cheese-shop sketch, where every kind of cheese is requested and the answer is always no.
That is a price signal, and this paper turns it into a number you can underwrite. Hold the facility at one megawatt and a Blackwell B300 cluster hosts 27% fewer GPUs than a Hopper H100 cluster, yet it grosses about $24.6 million a year against $12.6 million. The market is paying for yield per megawatt, not for the chip’s sticker price.
What this paper is not: a GPU price forecast or a call on whether AI demand is a bubble. It prices a megawatt under the rates and costs shown, and shows you how to check the inputs yourself.
SpaceX’s third-quarter call is the freshest and most specific public statement on how tight compute is. Six signals matter for the price of a megawatt.
|
Signal |
What management said (paraphrased) |
So what for a megawatt |
|
The constraint |
The limit on deploying superintelligence is power, cooling and infrastructure more than chips. |
The scarce unit is the energized megawatt, not the GPU. |
|
The pace |
Nameplate compute rose from 1.4 GW in Q2 to 1.7 GW in Q3, with about 2.2 GW expected by end-October and 2.5–3 GW by year-end. Management is highly confident of roughly 10 GW more in 2027 (13 GW total) and has a stretch plan for 20 GW of added power and cooling. |
One buyer adding about a gigawatt of load in a quarter. Power equipment, not demand, sets the pace. |
|
The power clock |
Turbine suppliers reportedly deliver blades and vanes 12–18 months after the rest of the turbine. SpaceX says its rocket-engine casting capability lets it energize 12–18 months faster than rivals. |
Lead time on power equipment is the moat. Energization date, not chip delivery, decides who earns. |
|
The price |
SpaceX prefers short, spot-like rentals because it expects compute value per watt to rise. Short contracts are renewing “at higher dollars per watt, not lower dollars per watt,” and even old H100s reportedly rent for more than when new. |
Supports using spot indices (Ornn OCPI) over long-term rates, and a rising rather than falling rate path. |
|
The payback |
The CFO said new deployments pay back in under a year, at pricing slightly above the top of the $30–50 per watt range cited last quarter. The CEO said a six-month payback next year is not impossible. |
Our conservative B300 case pays back in 13.3 months on GPU cost alone. Management’s figures sit above it. |
|
The chip clock |
Terafab: a test chip targeted for mid-2027 and volume production as an aspiration for Q3 2028. |
Silicon relief is not a 2027 event. Power and chips both stay tight over this paper’s horizon. |
Source: machine-generated call transcript (Motley Fool, Oct 20, 2026). Verify figures against SpaceX’s investor materials before publication.
|
Why not take it at face value? SpaceX is both a seller of compute and a heavy buyer of power equipment, and both incentives favor describing scarcity. Others are saying similar things: Nvidia’s CEO said electricity will remain a major bottleneck for the next couple of years, and the President called for large datacenters to generate their own power. We treat the call as a signal to cross-check against independent price data (Ornn OCPI) and independent construction data (Vulcan). The rest of this paper does exactly that. |
The move from NVIDIA’s H100 (Hopper) to the B300 (Blackwell Ultra) is a large step in compute and memory, bought with a large step in power. The question for a power-limited site is whether the extra performance pays for the extra watts.
|
Specification |
H100 (Hopper) |
B300 (Blackwell Ultra) |
Change |
|
Release (announced / ships) |
Mar 2022 / late 2022 |
Mar 2025 / H2 2025 |
~3 years later |
|
Memory capacity |
80 GB HBM3 |
288 GB HBM3e |
3.6x |
|
Memory bandwidth |
3.35 TB/s |
8.0 TB/s |
2.4x |
|
FP16 compute |
989.5 TFLOPS |
2,250 TFLOPS |
2.3x |
|
FP4 compute (5th-gen Tensor Cores) |
Not supported |
14,000 TFLOPS |
New |
|
Board power |
700 W |
1,100 W+ |
~1.6x |
|
Launch price per unit |
$25,000–$40,000 |
$50,000–$60,000 (B200 base: $34,999) |
~1.7x at midpoints |
GPU board power understates the draw. We use the full 8-GPU HGX server (host CPUs, memory, NVSwitch, fans and networking) and then apply a facility Power Usage Effectiveness (PUE) of 1.15 for cooling and power distribution.
|
Per GPU, at the facility meter |
H100 |
B300 |
|
8-GPU server draw, sustained |
10.2 kW |
14.0 kW |
|
System power per GPU |
1.275 kW |
1.750 kW |
|
Total power per GPU (PUE 1.15) |
1.466 kW |
2.012 kW |
|
GPUs supported per 1 MW |
682 |
496 |
Figure 1. Same megawatt: fewer B300s, but about twice the gross revenue (Ornn OCPI rates, 75% utilization).
|
Near miss: “Fewer chips per megawatt must mean worse economics” It does not. Each B300 earns 2.7x the hourly rate of an H100 ($7.56 vs. $2.81) while drawing 1.37x the facility power. Revenue per megawatt rises 1.96x. The cost of the extra watts is small next to what the extra performance earns. |
We price both clusters off the Ornn Compute Price Index (OCPI), which reports cleared spot and rental prices. The H100 index is $2.81 per GPU-hour. To stay conservative we price the B300 at the OCPI B200 rate of $7.56, giving it no credit for its 288 GB of memory. Utilization is 75%, or 6,570 billable hours of 8,760.
|
1 MW facility, PUE 1.15 |
H100 cluster |
B300 cluster |
|
Capacity |
||
|
GPUs hosted |
682 |
496 |
|
Capital cost (GPU unit price only) |
||
|
Price at launch, per GPU |
$30,000 |
$55,000 |
|
Current price, per GPU |
$16,000 (used) |
$55,000 (new) |
|
Cluster cost at launch price |
$20.46M |
$27.28M |
|
Cluster cost at current price |
$10.91M |
$27.28M |
|
Revenue (Ornn OCPI, 75% utilization) |
||
|
Rental rate per GPU-hour |
$2.81 |
$7.56 (B200 proxy) |
|
Revenue per hour at full billing |
$1,916 |
$3,750 |
|
Revenue per month |
$1.05M |
$2.05M |
|
Revenue per year |
$12.59M |
$24.64M |
|
Revenue per MWh consumed (year average) |
$1,437 |
$2,812 |
|
Payback (GPU cost only) |
||
|
Months to payback, launch price |
19.5 |
13.3 |
|
Months to payback, current price |
10.4 |
13.3 |
|
First-year revenue ÷ cluster cost, current price |
115% |
90% |
Capital cost covers GPUs only. It excludes server chassis, CPUs, networking, facility build, power, staffing and financing, so full-system payback is longer than shown. Revenue per MWh is annual revenue divided by 8,760 hours at a constant 1 MW draw.
Used H100s pay back faster (10.4 months) because their price has fallen by nearly half. But payback measures speed, not size. On the same megawatt the B300 cluster grosses $12.0 million a year more, and it keeps earning long after both have paid back. The megawatt is the scarce thing, so the right question is which chip earns the most from it.
|
The eviction math In a power-capped site, swapping an H100 megawatt for a B300 megawatt costs $16.4 million net ($27.28M for the new cluster less $10.91M from selling the used H100s) and adds $12.0 million of revenue a year. Payback on the swap is about 16 months. An operator who cannot add power has a strong reason to do this. One who can add power does not need to. |
The base case uses a single rate. The table below moves the B300 rental rate across the range implied by third-party forward curves and list prices.
|
Case (B300, 496 GPUs, 75%) |
Rate / GPU-hr |
Revenue / MW-yr |
Revenue / MWh |
Payback (months) |
|
Forward-curve low (B200, Mar 2027, Kalshi-implied) |
$5.30 |
$17.3M |
$1,972 |
19.0 |
|
Spot index (B200, Sep 2026, Kalshi-implied) |
$5.83 |
$19.0M |
$2,169 |
17.2 |
|
Base case (OCPI B200 proxy) |
$7.56 |
$24.6M |
$2,812 |
13.3 |
|
List-price high (B300 median on-demand, Oct 7) |
$9.50 |
$31.0M |
$3,534 |
10.6 |
Even the lowest case earns 37% more per megawatt than the H100 base case ($17.3M vs. $12.6M). The $5.30, $5.83 and $9.50 rates come from third-party reports and are not verified against the Ornn index directly.
|
Stress test: rack-scale reality The base case uses HGX server power. In an early-2026 podcast, SpaceX’s CEO put rack-scale GB300 deployments at roughly 300 MW per 110,000 GPUs, and about 1 GW at the generation level per 330,000 GPUs, once networking, storage, peak cooling and a service margin are counted. That is 2.7–3.0 kW per GPU against 2.0 kW here. At those figures a megawatt hosts 330–367 B300-class GPUs, and revenue falls to about $16–18 million a year at the same rental rate. Payback per GPU is unchanged. Applying the same overheads to the H100 cuts its revenue to roughly $8–9 million, so the ratio stays near 2x. The megawatt is a smaller business than the base case suggests, but the ranking holds. |
A chip swap reuses a megawatt that is already built. A new datacenter has to pay for the building, power and cooling as well. We assume $50 million per MW all-in for a new B300 datacenter. Subtracting the $27.28M of GPUs leaves an implied $22.72M per MW for facility, power, cooling and networking, which we hold constant for the H100 build.
|
Per 1 MW |
New datacenter, B300 |
New datacenter, H100 |
Chip swap, H100 to B300 |
|
Capital cost |
|||
|
GPUs |
$27.28M |
$20.46M |
$27.28M |
|
Less: resale of used H100s |
— |
— |
−$10.91M |
|
Facility, power, cooling, networking (implied) |
$22.72M |
$22.72M |
$0 (already built) |
|
Total cost per MW |
$50.00M |
$43.18M |
$16.37M |
|
Revenue (75% utilization) |
|||
|
Gross revenue per MW-year |
$24.64M |
$12.59M |
+$12.05M (incremental) |
|
Payback |
|||
|
Months to payback, gross revenue |
24.4 |
41.2 |
16.3 |
|
Months to payback, net of $150/MWh power |
25.7 |
45.9 |
16.3 |
|
First-year revenue ÷ total cost |
49% |
29% |
74% |
H100 build priced at launch cost ($30,000 per GPU). The chip swap adds no power cost because the megawatt is unchanged. Payback excludes staffing, maintenance, financing and depreciation.
The same new B300 megawatt at different all-in costs and rental rates:
|
Months to payback (gross revenue) |
B300 at $5.30/hr |
B300 at $7.56/hr (base) |
B300 at $9.50/hr |
|
$35M per MW |
24.3 |
17.0 |
13.6 |
|
$50M per MW (assumption) |
34.7 |
24.4 |
19.4 |
|
$65M per MW |
45.2 |
31.7 |
25.2 |
|
The read: the building is the slow part A new megawatt takes about 24 months to pay back, against about 16 months for a chip swap and 13 months for new chips in a shell that already exists. The gap is the cost of the building, power and cooling. That is why an energized megawatt is worth more than the chips on it, and why a site that can already power a better chip has a head start over one that has to build first. |
Revenue per MWh is the value of power to the business running the chips. It is the number a grid planner, a utility or a regulator should have in mind when they think about this load.
Figure 2. Revenue per MWh consumed vs. a $150/MWh all-in generation, transmission and distribution cost.
|
If power costs… |
Share of H100 revenue |
Share of B300 revenue |
|
$150 / MWh (base case) |
10.4% |
5.3% |
|
$300 / MWh (2x) |
20.9% |
10.7% |
|
$450 / MWh (3x) |
31.3% |
16.0% |
Our earlier paper, Placing the Bust, laid out scenarios for how the chip market could unwind. Two matter for the October 20 evidence, and a third is the risk that links them.
|
Scenario |
What it means |
What October 20 says |
What to watch |
|
Universal scarcity |
Every chip vintage is short. Anything that can be powered earns. |
No GB300, GB200, TPU, Trainium or H100 capacity available to rent. Renewals at higher dollars per watt. |
H100 and B200 index rates; used-H100 prices; renewal terms. |
|
Forced hardware eviction |
Power or land is the binding limit, so the rational move is to replace older chips with newer ones at powered sites. Older chips move to weaker sites or sit idle. |
Power, cooling and turbine blades are the stated bottlenecks. A megawatt of B300 earns about twice a megawatt of H100. |
Eviction payback (about 16 months in the base case); B300 vs. H100 rate spread; secondary-market volume. |
|
Stranded silicon |
Chips arrive before the megawatts to run them, and Blackwells gather dust in warehouses. |
The CEO said in early 2026 that chip output would outrun the ability to turn chips on by year-end. Vulcan sees construction on 88% of units listed under construction, but projects 15.0 GW of the 36.8 GW listed for 2027 slipping to 2028 or later. |
Vulcan construction status and online dates against chip delivery schedules. |
|
Our read: both, in sequence and by site Universal scarcity sets the rental rate. Forced hardware eviction decides who gets the megawatt. Where power is abundant or self-supplied, H100s keep earning. Where power is the limit, they are evicted by the chip that earns twice as much from the same megawatt. The most dangerous place to be is holding silicon at a site that has not yet energized. |
Vulcan, SynMax’s construction-monitoring product, uses satellite imagery to detect whether announced projects are being built, records construction milestones, and projects online dates independently of official schedules. The figures below cover U.S. datacenter units with listed online dates from 2026 through 2028. Each unit is counted by its own status, not its campus’s status, and units that have left the IIR feed are excluded. Imagery runs through October 15, 2026; the unit list is current as of October 16.
Figure 3. U.S. datacenter capacity listed online in 2026–2028 by Vulcan status (left), and satellite detection on the 32.4 GW listed as under construction (right).
Planned and engineered capacity due in these years (98.3 GW) is 3.0 times the 32.4 GW construction pipeline, and satellites already see construction on one in four of those megawatts. Five findings stand out.
|
Finding |
Evidence from Vulcan |
Why it matters |
|
1. Listings vs. dirt |
Satellites detect construction on 88% (28.5 GW) of the 32.4 GW listed as under construction, but on only 39% (53.7 GW) of all 138 GW listed for 2026–2028. For 2028, 70% (58 of 84 GW) shows no activity. |
The 88% covers listed construction only. Most 2028 capacity has no ground activity yet. |
|
2. Where the construction is |
53.7 GW across 342 facilities has detected construction (28.5 GW at units listed under construction, 24.8 GW at units listed as planned or engineered, 63% of that land clearing only). By grid: PJM 15.4 GW, ERCOT 12.3, MISO 9.6, SPP 3.1. Top states: Texas 13.3 GW, Virginia 6.5, Ohio 3.4. |
Scarce chips will land first where the construction is. |
|
3. The 2027 forecast |
Vulcan expects 27.0 GW online in 2027 across 304 units: 20.0 GW that IIR also lists for 2027, plus 7.0 GW moving in from other years. Texas leads with 6.2 GW, then Virginia (3.1 GW) and Missouri (1.5 GW). |
Chip orders and lease starts for 2027 should be sequenced to these dates. |
|
4. Slower than the paper schedule |
IIR lists 36.8 GW for 2027. Vulcan sees 20.0 GW (54%) landing in 2027, 15.0 GW (41%) slipping to 2028 or later and 1.8 GW (5%) arriving in 2026. Net, the 2027 forecast is about a quarter below the IIR figure. |
Plan for fewer 2027 megawatts than the listings suggest. The slip is where silicon strands. |
|
5. Self-supply |
19 facilities (4.8 GW, 15% of capacity under construction) include behind-the-meter generation. |
The playbook management described is already visible. Self-supplied sites can keep older chips earning. |
Figure 4. 2027 only: from IIR’s 36.8 GW listing to Vulcan’s 27.0 GW forecast. Vulcan clients have access to the unit-level forecast for other years.
|
# |
Facility |
State |
Units |
MW online in 2027 |
|
1 |
Santa Teresa (Project Jupiter) |
NM |
2 |
1,050 |
|
2 |
Saline (Project Mitten) |
MI |
2 |
922 |
|
3 |
Rayville (Hyperion) |
LA |
4 |
888 |
|
4 |
McCarran (Reno) |
NV |
1 |
810 |
|
5 |
Port Washington (Lighthouse) |
WI |
3 |
750 |
|
6 |
Abilene (Frontier) |
TX |
5 |
700 |
|
7 |
Pampa (Meitner) |
TX |
3 |
630 |
|
8 |
Cedar Rapids (Hawkeye) |
IA |
6 |
600 |
The ten largest facilities hold 28% of the 27.0 GW, and 12 facilities exceed 500 MW. This paper shows 2027 only; Vulcan clients can see the forecast for every other year, including 2026, 2028 and beyond.
|
Symbol: three clocks Three clocks decide when a megawatt starts earning. None of them is short. Completion is not energization, so the build clock is a floor.
|
Notes: capacity is listed unit capacity in MW, counted by each unit’s own status, for U.S. units on the current IIR list with online dates in 2026–2028. “Construction detected” means Vulcan status Confirmed or Identified; the two differ only in when construction was first spotted. Median imagery date on these sites is September 29, 2026. The 2027 forecast is Vulcan’s unit-level projection as of October 23, 2026.
The cost of getting this wrong is specific. Underwrite paper megawatts and you buy chips for sites that will not energize on time. Hold the wrong vintage in a power-capped site and you pay for a megawatt you cannot fill with the chip that earns the most. Both are avoidable with three steps.
|
1 |
Price your megawatt |
Replace this paper’s inputs (GPU, rental rate, PUE, utilization, power cost) with your own. Compute revenue per MWh and the eviction payback for the chips you actually hold or plan to buy. |
|
2 |
Verify the supply |
Check every site you finance, lease or supply against satellite evidence: construction status, milestones and the date of the latest image. |
|
3 |
Time the energization |
Map Vulcan’s projected online dates against chip deliveries and lease starts. Choose chips for the power you will actually have, and when. |
|
See which megawatts are real. Vulcan tracks U.S. datacenter construction by satellite: construction status, milestones and projected online dates, updated as new imagery arrives. In a briefing we will show you:
Request a Vulcan datacenter briefing → David Bellman dbellman@synmax.com |
You can turn any GPU rate into dollars per megawatt-hour. You can say which chip earns the most from a constrained megawatt, and what that does to older hardware. And you can separate the megawatts being built from the megawatts being announced.
The shortage is not in the shop. It is in the power. No power, no cheese.
IMPORTANT DISCLOSURES
Not investment advice. This material is provided by SynMax Inc. for general informational and educational purposes only. It is not investment, financial, legal, accounting, or tax advice, and it is not an offer, solicitation, or recommendation to buy, sell, or hold any security or financial instrument. SynMax is not a registered investment adviser, broker-dealer, or research analyst and has not considered the objectives or circumstances of any reader.
Data and sources. Market data and company information come from the public and third-party sources listed above. They are believed to be reliable, but SynMax does not guarantee their accuracy or completeness. Stock prices and index levels are as of the October 14, 2026 close and will not be updated after publication. Some chart values are approximations, and percentage declines marked as ours are SynMax calculations. SynMax Vulcan data are subject to revision, and land-clearing figures for recent months are preliminary and typically rise as dates are added to the database later.
Opinions and forward-looking statements. Statements about what may happen, including possible changes to moratoriums after the November elections and the future direction of land clearing, friction, or first structure, are SynMax's opinions and hypotheses. They are not predictions or guarantees, and actual results may differ materially. A relationship between Vulcan data and market prices does not establish causation, and past data relationships may not continue.
No affiliation. Company, product, and index names are the property of their respective owners and are used only to identify them. SynMax is not affiliated with, and is not endorsed by, any company named.
Investing in securities involves risk, including the possible loss of principal. Consult a qualified professional before making investment decisions.