Power Is the New GPUs
40% of announced datacenter projects are dying. Bloom Energy just printed its first billion-dollar quarter. TSMC raised capex because tool prices are inflating. And the semicap selloff is fighting its biggest customer.
The AI buildout story has a math problem. Announced gigawatts massively overstate what actually gets built, and the selloff in semiconductor equipment stocks this week is pricing in the opposite of what their largest customer just told the world.
The power model puts the US grid shortfall at 34 gigawatts by 2028 in the base case. That is before you account for the 40% mothball rate Chamath Palihapitiya flagged on the July 18 All-In Podcast: "About 40% of all these projects are getting mothballed and stopped. It's creating this massive deficit of available energy to actually drive the use of AI." The same episode noted a recent PJM auction drew only 156 megawatts against a 7-to-8-gigawatt need. The auction market is the price-discovery mechanism, and it is screaming scarcity.
Chamath has a book on this: he owns datacenter power assets. But the data is confirming the thesis. Bloom Energy reported Q2 revenue of $1.07 billion Tuesday, up 166% year over year, against a FactSet consensus of $826 million. Adjusted EPS of $0.78 beat the $0.41 estimate by 91%. The company raised full-year guidance to $3.9 billion to $4.2 billion in revenue and $2.55 to $2.85 in adjusted EPS. The stock closed at $188 and held its gain through a broad market selloff.
Bloom's fuel cells are behind-the-meter generation: on-site power that sidesteps the grid interconnection queue. That is the exact asset class Chamath described as commanding "extreme pricing at the front end of the curve." The Brookfield financing framework expanding from $5 billion to $25 billion says the capital is following the thesis.
But the permitting friction is real. SemiAnalysis reported on July 18 that Oracle's Project Jupiter, a behind-the-meter datacenter in New Mexico planned to use Bloom Energy fuel cells, faces a possible one-to-two-year delay on permitting and pipeline construction. The same dynamic that makes live megawatts scarce also makes new projects harder to deliver. That is the bull case in a nutshell: the scarcity is structural, not cyclical.
TSMC's C.C. Wei made the power-constraint story concrete on the supply side. On the July 16 Q2 earnings call, asked whether leading-edge demand exceeds supply by 30% to 50%, Wei answered: "The gap is very big." He then raised TSMC's 2026 capex to $60 billion to $64 billion, up from $52 billion to $56 billion, and cited tool inflation as one reason: "Now we buy the tools with inflation price." The shortage is feeding back into equipment pricing itself.
That is what makes the semicap selloff notable. Applied Materials closed down 7.8% Monday and is down 23% over the last 21 sessions. Lam Research is down 27% over the same window. KLA Corporation is down 21%. The PHLX Semiconductor Index entered correction territory.
KLA then reported results after the close Monday. Fiscal Q4 revenue of $3.66 billion came in above the midpoint of guidance. The company guided Q1 fiscal 2027 revenue to $4.0 billion, a nearly 10% sequential jump. CEO Rick Wallace said the trends driving growth "are strengthening" and "momentum across our business is accelerating in the second half of calendar 2026 and continuing through 2027." The print was strong, and the outlook was stronger.
The semicap tape is fighting the semicap fundamentals. TSMC, the industry's largest customer, is raising capex. KLA, the process-control leader, is guiding up. The selloff is a positioning unwind and a rotation out of AI, not a demand signal. As we showed last week, NVIDIA's datacenter revenue leads hyperscaler capex by about three quarters, with a correlation of 0.37 at that lag. The capex cycle is still accelerating.
The power names are telling the same story from a different angle. Vistra trades at 15.6 times forward earnings with 18 analysts rating it a strong buy and a mean price target of $223, implying 50% upside from the current $149. Constellation Energy trades at 20.3 times forward earnings. GE Vernova, with $10.3 billion in net cash and $5.1 billion in free cash flow last quarter, is the turbine supplier to the same buildout. These are not expensive stocks for companies sitting on the binding constraint of the entire AI economy.
The Bargo power model projects the US grid shortfall grows from 34 gigawatts in 2028 to 65 gigawatts by 2030. Every project that gets mothballed widens the gap. Every live megawatt becomes more valuable. The 40% failure rate Chamath cited means the effective delivered capacity is far below the announced pipeline, and the market is only beginning to price that.
The AI buildout is not a GPU story anymore. It is a power story, and the power story is a scarcity story.
Sources: All-In Podcast, July 18, 2026 (Chamath Palihapitiya); SemiAnalysis, July 18, 2026; TSMC Q2 2026 Earnings Call, July 16, 2026; Bloom Energy Q2 2026 Earnings Release, July 28, 2026; KLA Corporation FQ4 2026 Earnings Release, July 28, 2026; Bargo power model, base case, July 28, 2026.
More research at bargo.ai/research.