Summary

The U.S. AI data-center power shortfall over the next several years is estimated at 46—47GW, equivalent to the average load of roughly 8 New Yorks and $2.3T in capex. Nuclear, gas turbines, fuel cells and solar-plus-storage are all unlikely to scale within 2—3 years, while moving training workloads offshore and using diesel backup face diplomatic and environmental constraints; converting mining sites, despite contributing only a contentious 6—15GW, is still the fastest solution within 18—24 months. “Even at the low end, 6GW is still a lot of power.”
Zheng Di estimates that the all-in cost of a 1GW AI data center is $45B—$50B, with GPUs accounting for roughly 70%—80%; Hong Jun therefore sees Nvidia at the value center of the entire infrastructure cycle. A 1GW IT load requires 1.1—1.2GW of actual power supply, with liquid cooling, water plants and captive power generation built in parallel; scaled to 100GW, the total reaches $4T—$5T, making the real constraints not just chips, but whether power, sites and capital can arrive simultaneously.
Owning power does not mean a miner can convert it into AIDC cash flow, and the market is expressing that skepticism through valuations below the $11—$13/W construction cost. IREN would be valued at only about $6—$7/W if its full planned 2.9GW were included, while Bitdeer, Hut 8 and others are as low as $2—$3/W; the discount prices in the risks around 24×7 supply, construction delivery, customer contracting and financing dilution. “The market does not believe you can convert all your power into the kind of power an AIDC needs.”
IREN’s 5-year, $9.8B contract with Microsoft exposes the key dividing line in miner conversions: whoever buys the GPUs takes on greater financing and depreciation risk. About $5.8B goes toward GPUs, while Microsoft prepays only $1.95B; the remaining gap may need to be filled through bonds, preferred equity, project finance and ATM offerings. By contrast, miners that lease power and sites to CoreWeave do not need to buy GPUs and can match 12—15-year contracts with 20-year site depreciation.
AI infrastructure financing may ultimately migrate from corporate cash flow to investment-grade debt, high-yield debt, private credit, ABS, CDOs and REITs, with the business model increasingly resembling rolling real-estate development. JPMorgan estimates the investment-grade market can provide $300B next year and $1.5T over the next 5 years, but Zheng Di’s experience in 2008 carries a warning: securitization technology is neutral; the real danger lies in deteriorating underlying rents, ignored fat tails and correlations spiking in tandem with default rates.
“Underinvestment is more dangerous than overinvestment” is first a prisoner’s dilemma for CEOs, not proof that a bubble has already been disproven. Big-tech operating cash flow may already be consumed by capex, with the debt cycle only beginning; even if overinvestment is confirmed 3—5 years from now, the more likely initial outcome is years of debt repayment and depressed share prices. Whether this becomes a bubble and overcapacity still depends on whether Claude programming, OpenAI’s enterprise Token customers and industrial applications can generate real profits.
OpenAI is simultaneously Nvidia’s “catfish” and a “too-big-to-fail” node in U.S. AI, making the IPO window more important than the valuation debate. Zheng Di believes a $1T valuation supported by roughly $13B in annual revenue or $20B of ARR still looks stretched, but a 2027 listing could be too late to deliver on the $1.4T construction commitment; key opportunities among miners include Riot, CleanSpark and Hut 8, which have cheap power but no signed orders, as well as Bitdeer, whose prospects have a binary geopolitical dimension.

Deep dive

1. The 46—47GW Shortfall Turns AI Compute Buildout into a $2.3T Proposition

Zheng Di cited Morgan Stanley’s estimate that the U.S. could face a 46—47GW shortfall from AI data centers over the next several years. New York’s average load is roughly 6GW, putting the gap at about 8 New Yorks.
At $50B of investment per 1GW data center, the shortfall implies roughly $2.3T. Zheng Di also noted that Morgan Stanley is an adviser to Cipher Mining, so its 15GW mining-site conversion forecast “is likely subject to some bias.”

2. Mining-Site Conversions Are Fastest, but 15GW Is Only the Ideal Ceiling

Morgan Stanley’s optimistic case has mining sites releasing 15GW within 18—24 months. Conversions typically take 9—12 months; Core Scientific and Applied Digital are already ahead, while Riot and CleanSpark may not deliver until 2027.
Hong Jun used a lower range: different institutions arrive at 8.4GW or 6—10GW. Zheng Di acknowledged the dispute but stressed that “even at the low end, 6GW is still a lot of power,” with these projects potentially deliverable before 2028 if there are no delays and financing proceeds smoothly.
Mining sites can curtail at peak periods, soft-start and receive grid subsidies; AIDCs require 24×7 operation and substantial redundancy. Nominal miner power therefore cannot be converted one-for-one into compliant data-center power.

3. Nuclear, Gas and Solar-plus-Storage Cannot Close the Near-Term Gap

Zheng Di used “1GW is one nuclear plant” as a benchmark: conventional nuclear construction often takes more than 10 years, while even an optimistic SMR case points to commercial deployment in 2030—2035 and a pessimistic case to after 2035. Before 2030, it is essentially not a solution.
The U.S. does not lack natural gas; it lacks gas turbines. GE Vernova and Japanese manufacturers remain cautious after previous capacity expansions quickly created oversupply, and orders are already backlogged 2—4 years. Fermi even acquired an abandoned power plant for one second-hand turbine, while Musk’s data centers use 7 old turbines.
Hong Jun said the market broadly assumes Bloom Energy could contribute about 2GW. He also mentioned Sam Altman’s solar-plus-storage route but said it cannot solve the problem within 2—3 years. xAI’s 200,000-GPU Colossus 2 uses Megapack, but storage remains primarily backup capacity.

4. Moving Training Offshore and Diesel Backup Can Add Power, but Policy Comes First

The first unconventional route is to move training workloads to Singapore, Johor in Malaysia and South America—especially Brazil—but that requires “data-center diplomacy” and cannot be delivered immediately.
David Sacks cited energy experts who put actual U.S. grid utilization at roughly 50%, while the Abilene project and similar estimates are only in the low 30s. If environmental rules were relaxed and large-scale diesel backup allowed, the theoretical release could reach 80GW. Zheng Di judged the route politically difficult in an environment of closely divided parties and heightened sensitivity around emissions rules.

5. The Outcome in 2 Years Depends on Whether AI Applications Can Absorb Supply

Zheng Di sees the critical window for the power shortage in the next 2 years: if AI applications take off, new capacity can run continuously; if AI remains a small share of revenue, the industry will enter a phase of “overcapacity and survival of the fittest.”
Claude has opened a To B path through programming. Whether OpenAI generates roughly $13B in annual revenue or $20B of ARR, the key is not Tokens themselves but whether To B and then To C customers such as Duolingo can use them to make money, expand margins and continue increasing consumption.
Industrial applications remain constrained by hallucinations and may require specialist models and consulting firms. Zheng Di cited the GDPval test, which covers 9 industries and more than 40 sub-industries: Opus 4.1 could match or outperform human experts in roughly 47% of real-world cases, with GPT next but still within the realm of possibility. “If it can crack the B2B market, the chain works.”

6. GPUs Take the Lion’s Share of the $50B Cost of 1GW

Zheng Di estimates GPUs account for 70%—80% of the total cost of 1GW, or $35B—$40B. The GPU share in the Microsoft-IREN transaction is somewhat lower but still close to 60%. Hong Jun therefore sees Nvidia, AMD and inference-chip suppliers as the core beneficiaries.
He distinguished between pure site conversion and full AIDC construction. The Tier 3 site estimate is $1.1B—$1.3B, rising to $1.7B—$1.9B at the high end; the full construction cost used in miner valuations is roughly $11—$13/W.
Liquid cooling, power systems and water plants make up the remaining capital needs. Colossus 2 uses direct-to-chip liquid cooling and built its own water plant; PUE is typically 1.1—1.2, meaning a 1GW IT load requires 1.1—1.2GW of power supply. At 100GW, total cost can reach $4T—$5T.

7. EV/W Below Replacement Cost Prices in Delivery Risk

If all planned power is included, few miners trade above $12/W; Applied Digital is one of the exceptions. IREN is around $6—$7/W, versus only $1—$2/W several months ago, while many miners seeking conversion remain below the $11—$13/W construction cost.
The discount embeds 2 judgments: unbuilt power receives no value for now, and even power that is secured is not believed to be fully convertible into AIDC capacity and signed orders. The market may simply be saying, “build the power and we will value that much”; the rest may never become orders or data centers.
The discount also weakens miners’ financing capacity and willingness to convert. After CoreWeave cut its revenue guidance, related miners broadly fell 20%—30%, showing that funding conditions and market beta directly determine whether planned power can be delivered.

8. IREN’s Large Microsoft Contract Leaves GPU Risk on Its Balance Sheet

IREN’s 2.9GW is not all operational: Canada accounts for about 100MW; Texas Childress is 750MW, of which only 200MW is contracted to Microsoft; Sweetwater totals 2GW, with 1.4GW guided to come online in April of the following year and the remaining 600MW targeted for 2027.
The Microsoft contract covers 200MW for 5 years and $9.8B. By comparison, Applied Digital has a fixed 250MW contract worth $7B over 15 years, plus a 150MW option; Core Scientific has a 12-year, $10B contract, and Galaxy a 15-year, $9B contract.
The difference is GPU ownership. CoreWeave buys its own GPUs, while miners simply lease sites and power; IREN signed directly with Microsoft but must fund roughly $5.8B of GPU purchases. Microsoft prepays $1.95B, and IREN has about $600M of cash and $1B of convertible debt, still leaving a potential gap of more than $2B.
Analysts are therefore focused on the mix of debt, preferred equity, project finance and ATM issuance. GPUs are depreciated over 5 years, versus typically 6 years at CoreWeave and Nebius and roughly 20 years for the site. If Microsoft does not renew after 5 years, a 10%—12% unlevered IRR could fall to the single digits because of accelerated depreciation.

9. CoreWeave Sells Financing Capacity; Nvidia Sells the Ecosystem

Based on the second-quarter data cited on the program, CoreWeave had more than $11.1B of debt, about $1.15B of cash and high interest costs. Zheng Di called it “a financing game”: the GPUs are procured, and even if Nvidia supports the company, it will not offer particularly steep discounts.
His analogy is that Nvidia has become “the general contractor of the GPU industry,” while AMD was previously closer to a pure card vendor. AMD entered the OpenAI ecosystem through a 6GW partnership; OpenAI owns up to 10% of AMD, subject to conditions including AMD reaching a $1T market capitalization for the final tranche.
Zheng Di viewed Core Scientific’s rejection of CoreWeave’s all-stock acquisition offer as sensible: CoreWeave traded around $160 when the offer was made, then fell to $90, and miner shareholders did not want to exchange “real money” for an overvalued stock. Core Scientific faces less financing pressure than IREN, but CoreWeave is less creditworthy than Microsoft—an uneven trade-off.

10. Crusoe Has Turned Data Centers into Rolling Real-Estate Development

Crusoe’s Series D the previous year was led by Founders Fund, raising $600M at a $2.7B valuation. It subsequently sought funding at a $10B valuation, though the transaction has not closed, and brought Blue Owl Capital into the first phase of Stargate Abilene.
The Abilene project is roughly 1.2—1.3GW with $15B of total investment, planned across 8 buildings and 400,000 GB200 GPUs, or about 50,000 per building. Blue Owl manages about $40B and specializes in REITs; Zheng Di speculated that Crusoe may be bringing it in to package, tranche and sell the project to secondary-market or senior investors.
Zheng Di’s envisioned path is “build one building, lease one building, turn one building into a REIT,” then roll development forward through milestone-based cash receipts. For CoreWeave, Crusoe and the miners, delivered megawatts unlock the next tranche of cash flow. “It is exactly the same as real estate.”

11. Securitization Is Not the Original Sin; Mispricing Tail Risk Is

Hong Jun asked what fundamentally distinguishes this from the subprime securitization of 2005—2008. Zheng Di, who worked on a CDO desk in 2006—2007, replied that ABS, CDOs, CMBS and RMBS are neutral technologies; the core question remains whether data centers can be leased and rents sustained.
The subprime problem began after the good assets had been securitized: bonuses and commissions drove originators to lower underwriting standards, eventually producing zero-down-payment loans with almost no interest for the first 2 years, followed by high interest and principal repayment from year 3. “Financial engineering had been taken to an extreme,” hollowing out the underlying assets.
The first modeling error was Black-Scholes’ assumption of a normal distribution, while reality contains skew and fat tails. At least at the time, many investment banks and brokers failed to properly account for extreme losses.
The second error was treating default rates and correlation as independent. Both sat at historical lows in 2007, and AAA tranches could earn several dozen basis points more, with models showing that roughly 7 defaults among 100 names were needed to erase the yield. In 2008, the market learned that “when default rates rise, correlation rises as well.”

12. Banks Build the Bridge; the Global Fixed-Income Market Takes the Other Side

Large banks generally do not want to provide long-term loans, preferring short-term and bridge financing. Hyperscalers can issue investment-grade debt, while lower-rated companies rely on high-yield or private credit; roughly $26B of Meta’s $29B project came from private debt.
JPMorgan estimates the investment-grade market can provide $300B for data centers next year and $1.5T over the next 5 years. Big tech’s roughly $700B of annual operating cash flow can also be leveraged through low-cost debt.
BBB- and above is investment grade; below that is high yield. Banks, insurers and most fixed-income funds are primarily mandated to hold sovereign and investment-grade corporate debt, so ratings directly determine financing capacity and cost.
Zheng Di cited Goldman Sachs’ figures: global financial assets total roughly $260T, with bonds accounting for 37%, or about $100T. The U.S. represents more than $40T, while global corporate bonds total roughly $20T and are mostly investment grade. That pool is far larger than the high-yield market.

13. OpenAI’s “Catfish” Has Bound Nvidia and the U.S. Government Together

Asked about Nvidia’s precise future share, Zheng Di’s candid answer was that there is “not much research” to go on; he nevertheless believes its ecosystem advantage will be difficult to dislodge over the next 2—3 years. Nvidia binds model companies through equity-for-GPU deals, while Anthropic may gradually move toward the Google-Amazon camp.
OpenAI’s continued push toward AGI forces other major companies to bear the enormous tail risk of not participating, keeping both training and inference demand alive. Zheng Di calls it Nvidia’s “catfish,” while also crediting it with pushing the U.S. power grid, data centers and manufacturing chain—stagnant for years—into an investment cycle.
The result is an “iron-chain linkage” among Nvidia, OpenAI and the 7 giants. Zheng Di believes OpenAI CFO’s proposal for government guarantees is not logically wrong, only unsuitable for public discussion. The deeper the linkage, the more OpenAI resembles a “too-big-to-fail” institution, and the closer U.S.-China AI competition moves toward a government-involved “total war.”

14. Overinvestment Starts as a Prisoner’s Dilemma; the Debt Bubble Is Still Early

On the dispute over whether underinvestment is more dangerous than overinvestment, Zheng Di used behavioral finance: if a CEO alone refuses to participate and is wrong, they may lose their job; if everyone invests and is wrong together, individual accountability is limited. Zuckerberg therefore also refuses to miss the boat “at any cost.”
His stage assessment is that big-tech cash flow may only just have been fully absorbed by capex, with the debt cycle beginning now. It is therefore “too early” to declare a giant bubble. If overinvestment is proven 3—5 years from now, roughly $700B of annual operating cash flow makes years of debt repayment and depressed share prices more likely.
Hong Jun cited roughly $2B of debt at one of Musk’s power plants at a 12% interest rate to challenge the idea that debt remains low. Zheng Di distinguished project-company high-yield financing from the overall balance sheets of Hyperscalers, acknowledging that debt will rise but arguing that the aggregate is not yet extreme.
Meta promised Trump $600B of investment over the next several years, while OpenAI proposed $1.4T. Zheng Di applies a discount to numbers announced in political settings: the UAE said it would invest $1T, and Uzbekistan even claimed it would invest $30B over 3 years—roughly 10% of annual GDP. “How it is ultimately invested still depends on circumstances.”

15. OpenAI Must List Before the Liquidity and Political Windows Close

Zheng Di believes OpenAI “must IPO” to deliver on its $1.4T plan; even 2027 may be too late, with 2026 preferable. Roughly $13B of annual revenue is insufficient to support a $1T valuation, while Sam countered that ARR has reached $20B—though ARR is still not the same as actual revenue for the year.
The favorable window he envisages includes further rate cuts, the probability of a December cut briefly rising above 60% after the ADP data, Treasury spending being released as the TGA falls from roughly $1T to $950B, and the Fed expanding its balance sheet as early as the first quarter of the following year.
The biggest variable is the midterm election. Democrats are not opposed to AI or Web3, but could slow legislation while retaining emissions constraints. Markets “trade acceleration and the second derivative of acceleration”; if the expected policy slope declines, valuations can adjust first. Zheng Di is therefore bullish on the first half, viewing the previous shutdown-related correction as a “squat before the jump” ahead of a bubble phase.

16. Miner Rankings Depend on Delivered Power, Not Gigawatts on a Pitch Deck

Excluding Galaxy and TeraWulf, Zheng Di places Applied Digital in the high-valuation tier: it does not need to buy GPUs, faces relatively low financing pressure and has its first building close to delivery. Core Scientific also benefits from having started construction earlier.
IREN and Cipher are in the second tier. Bitdeer, Riot, CleanSpark and Hut 8 typically trade at $2—$3/W based on contracted power, but still lack AIDC orders. In a strong market, “power is king”; in a weak market, investors recognize only delivery and financing.
“Secured power” also has levels. IREN paid tens of millions of dollars in deposits to ERCOT for Sweetwater, indicating that the feasibility study, construction materials and generation model are in late stages, but not that the site is energized. Being able to connect is different from actually receiving power.
Nadella said 2 data centers in Santa Clara may sit idle for several years, waiting for grid upgrades in 2028 before connecting. Zheng Di therefore suspects that a meaningful share of projects under construction nationwide have not truly secured power, leaving purchased GPUs idle first.

17. Bitdeer’s Discount Is a Binary Geopolitical Bet

Bitdeer was founded by Chinese nationals and incorporated in Singapore. Tether first privately invested $100M around $5, then continued buying at $7—$10; by around April 17, its 2 subsidiaries held a combined 25.5%. Tether has close ties to Cantor, and Zheng Di said Cantor had been consistently publishing reports and pushing up Bitdeer’s share price.
The company claims to control 1.6GW, but delivery expectations after earnings and ATM pressure drove the stock lower. The “million-dollar question” is whether it can build AIDCs in the U.S. and win orders from a Hyperscaler or CoreWeave. If it breaks through, Zheng Di thinks “doubling would be perfectly normal”; otherwise, it will develop along an ordinary path.
Geopolitical constraints also affect offshore training. The program mentioned a Japanese-listed company shorted after allegedly using advanced GPUs to provide remote services to Tencent; H20 may be permitted, while B-series chips are not. Even U.S. foundation-model synthetic-data suppliers may need to register in the U.S. Zheng Di could only say “maybe”: power may be less constrained than data and GPUs, but whether Singaporean or Johor power controlled by Chinese nationals can win orders remains untested.

18. Hoarding Power Makes Conversion Easier Than Hoarding Compute

CoreWeave and Nebius were not significant miners before their previous conversions and therefore had limited legacy baggage. Applied Digital mined only about 400MW but entered the market early, with its first building potentially delivering by the end of the year cited on the program.
MARA is the extreme case of hoarding compute: roughly 60EH/s, with an original year-end target of 75EH/s, but only 640MW of owned power—not the 1.2GW commonly cited by the market. Another 180MW of miners are hosted at Applied Digital’s Ellendale. Even after deducting roughly 50,000 BTC, its EV/W is still about $5, so it is not cheap.
IREN is the power hoarder: about 810MW is used for mining, while roughly 2.1GW of additional power has been accumulated. It can preserve mining cash flow while building Sweetwater. The optimal route is to “mine while converting into data centers,” which is why total industry hash rate has not yet declined.
Core Scientific’s mining gross margin was about 7% in the first quarter, versus only 5% for data centers. Its 86%—88% EBITDA margin excludes depreciation; Zheng Di called part of it an “illusion,” with true unlevered IRR still around 10%—12%. Asked how falling crypto prices affect collateralized financing, Zheng Di acknowledged the risk: with 50,000 BTC held at $100,000 each, 64% LTV could support roughly $3B of borrowing; CleanSpark and Riot could also borrow about $800M and $1.2B—$1.3B, respectively. Zheng Di had no current positions but could trade at any time; Hong Jun also held none of the companies mentioned.

Peter Thiel and Softbank Sell NVIDIA - Why? & Why VC Will Hit $1TRN and The Opening of Retail What Everyone Missed About Gemini 3 w/ Salim, Dave & Alexander Wissner-Gross | EP#209