PodMagazine.
硅谷101

E215 | The Trillion-Dollar AI Infrastructure Play: Who’s Paying, and How Do We Power It?

11/21/2025
Current Segment

Bitcoin Miner Transition:A 15GW Power 'Great Migration'

When the 'Interruptible Power' of Crypto Mines Tries to Meet the 24/7 Rigid Demands of the AI Era

Continuing from our earlier discussion on the AI infrastructure bubble, the core issue has shifted from 'whether there's compute' to 'whether there's power.' Microsoft's Nadella put it very bluntly: AI doesn't lack GPUs, but rather the infrastructure supported by trillions of dollars in investment.

As an observer of frontier technology, I've noticed that capital markets, especially Morgan Stanley, have recently proposed a very interesting argument:Bitcoin mines are collectively pivoting to AIDCs (AI Data Centers). According to Morgan Stanley's estimates, this transition could release 15GW of power capacity in the next 18 to 24 months.

[ Context: What does 15GW mean? ] New York City's average annual power consumption is about 6GW. A 15GW release means that in the next two years, just through the transformation of mines, we can conjure up a power supply equivalent to '2.5 New York Cities.' Based on the current logic that every 1GW of AIDC requires a $50 billion investment, this is a market starting at the hundred-billion-dollar level.

The reason these mine transformations can happen so quickly is that the construction period usually only takes 9 to 12 months. Many miners started announcing their pivots last October and are already scrambling to order key components with extremely long lead times. If there are no issues with financing or the supply chain, we will see large-scale delivery of this capacity by 2026.

“Mining transformation is currently the most viable path because, aside from it, the three paths of nuclear, gas, and energy storage are almost impossible to rely on in the short term.”

To be honest, besides mine conversion, conventional paths are stuck on some very 'hard' physical constraints. Take gas turbines, for example: while the US has plenty of natural gas, manufacturers like GE and Siemens were burned badly by 'excess demand' in previous cycles. They are now extremely cautious and unwilling to expand production easily. How tight is it right now? To build power plants, some people are even scavenging used gas turbines from old airplanes to make do.

Mental Model

The 'Policy Trap' of Gas Turbines

Why aren't manufacturers expanding? Because the production cycle for gas turbines is extremely long. If you invest in a factory now and it’s finished in three or four years, and then the Democrats tighten environmental policies, those turbines might not sell at all. This 'cyclical mismatch' has led to the current severe shortage in equipment supply.

Two 'Non-Mainstream' Survival Paths

[ Path A: Offshoring Training ]

Since the US is short on power, we'll just move the training tasks elsewhere—to places like Singapore, Johor in Malaysia, or even Brazil, where power is abundant. This 'training diplomacy' could become a norm in the next two years.

[ Path B: Diesel Power Generation ]

This is a method that makes environmentalists 'want to scream.' If it weren't for emission constraints, as long as large-scale use of diesel generators as backup was allowed, the 80GW power gap in the US could be filled instantly. This is entirely a man-made, policy-driven bottleneck.

Question / Concerns

'Can power from mines really be used for AI? I've heard many industry analysts say the requirements for power in these two scenarios are completely different.'

The Reality Check / Reality

There is indeed a catch.Miners can use 'interruptible power' and can go offline at any time,which allows them to enjoy peak-shaving incentives from the grid. But AIDCs require extreme 24/7 stability. This means when you pivot, you need very high power redundancy. Even if Morgan Stanley says 15GW, the amount that can truly be converted into 'high-spec AI power' might only be half, or even less.

Next: A $50 Billion Cost Breakdown for a 1GW Data Center →
Phase II: The Monetization Crucible

Conquering the B-Side: AI Infrastructure's
The Life-and-Death Line

Transitioning from 'mining power' to 'AI power' is just a physical migration; the real challenge lies in whether B-side applications can support these multi-trillion-dollar valuations over the next two years.

[ Recap: We discussed unconventional paths like offshoring training and diesel power, as well as market doubts regarding the reliability of power from mining sites. ]

I believe the core pressure of the power shortage is concentrated in these two years. If, in two years, we see that AI applications still account for a negligible share of overall revenue, then the entire industry will inevitably fall into a cycle of overcapacity and survival of the fittest.

The key now is the sentiment on the B-side. Although Claude has already blazed a trail in the programming field and OpenAI is transitioning from 2C to 2B, the core question remains: **Who can consistently make money and improve profit margins?** Only if the 2B2C or pure 2B paths are proven viable will this massive infrastructure be sustainable.

[ Deep Dive: Hallucination and the Enterprise Threshold ] Current programming scenarios work because code is verifiable. But in broader business contexts, "model hallucination" remains a chronic issue hindering large-scale enterprise adoption. However, OpenAI is attempting to tackle this by leveraging expert-level processing capabilities. If they can match or even exceed the efficiency of human experts, the model will be fully proven.

Breaking Down the "Bill" for a 1GW Computing Center

People might not have a concrete financial concept of what a 1GW data center looks like. According to Nvidia and industry-accepted estimates, building a site of this scale requires a total investment of roughly $45 billion to $50 billion

Estimated Cost Breakdown (Per 1GW AI Data Center)

Note: Units in 100 million USD. Data based on market estimates for Tier 3 AI data centers.

You'll notice that the lion's share goes to Nvidia and AMD.

Out of the $50 billion total package, GPU chips alone account for $32 billion to $35 billion. The rest is for construction and systems. And here’s a detail: when we say 1GW, we mean IT Load, which is the actual power supplied for computing. In reality, you have to build 1.1 to 1.2GW of physical power capacity—that extra 10%-20% is required to maintain the data center's own cooling and power systems.

"If the market continues to give these pivoting miners such low valuations—even below replacement costs—then they may have no incentive to release that expected 15GW of power. In the end, the AI infrastructure boom might just be prosperity on paper."
Next, we will dive deep into how IREN struggles within the asset-heavy valuation trap and how Nvidia dominates the ecosystem through its "turnkey" model.
Next: IREN & Nvidia's Hegemony →

Bridge: From Miner Pivot Models to the Capex Dilemma

IREN: The Debt-Fueled GPU Adventurer

Trapped in the "Asset-Heavy" Valuation Trap and Nvidia's "Turnkey" Hegemony

We previously discussed how the transition from miners to AI IDCs (AI Data Centers) seems logical. However, the real barrier isn't how many watts of power you have, but whether you have enough money to fill that terrifying "funding gap."

Take IREN (Iris Energy) as an example. If you look purely at its EV/Watt (Enterprise Value per watt), it looks ridiculously cheap compared to its peers. But why won't the valuation budge? Because it faces massive financing pressure. Subtracting the $2 billion already raised, it still has a funding gap of nearly $3.8 billion.

[ First-Person Perspective ]

"I think the reason its valuation won't go up right now is fundamentally due to the sense of unsustainable financing. You don't just have to buy the chips (GPUs); you have to build the house (the IDC facility). The company's current strategy is to 'have it all': issuing preferred stock, doing ATMs (At-The-Market offerings). This constant dilution of existing shareholders and the pursuit of high-interest debt makes the market very uneasy."

Even more critical is the depreciation assumption. Currently, most people calculate based on 5 years for GPUs and 20 years for the facility. But if AI iterations happen faster than we expect, or if new players like IREN can't secure long-term renewal contracts, will they have to accelerate depreciation? Once depreciation accelerates, the originally calculated IRR (Internal Rate of Return) of 10% to 12% will collapse immediately.

Financial Model: The Root of Valuation Pressure

Core Pain Point

ATM (At-The-Market) Financing

Raising capital by selling shares directly on the secondary market. While flexible, it exerts long-term downward pressure on the stock price.

Depreciation Trap
  • GPUs: 5 years (rendered obsolete if computing power jumps too quickly)
  • Infrastructure: 20 years
  • Risk: If contracts aren't renewed, cash flow won't cover the high upfront costs
“The rules of the game have changed: people no longer see NVIDIA as just a graphics card vendor; they see it as an ecosystem builder. It’s more like the ‘general contractor’ model in the real estate industry.”

NVIDIA’s ‘General Contractor’ Hegemony and AMD’s ‘Ticket to the Game’

NVIDIA is extremely dominant in this layout. Even for a 'golden child' like CoreWeave, NVIDIA won't give you a discount. They nurture you and give you priority shipping rights just to cement their ecological hegemony.

What's interesting is the relationship between AMD and OpenAI. OpenAI isn't just buying chips; it holds a 10% stake in AMD. But there’s a catch: you have to ‘pay your dues’ first. OpenAI can only sell that final tranche of shares once AMD’s market cap hits the trillion-dollar mark. This isn’t just buying chips; it’s deep-rooted ecological symbiosis.

[ Deep Dive: The Real Estate-ification of Compute ]

If you look at the GPU industry as real estate, you'll find the logic is strikingly consistent:

  • • Land: Mines/sites with 1.3GW power permits.
  • • Buildings: NVIDIA’s H100/H200 compute clusters.
  • • Leases: 15-year long-term contracts signed by Microsoft and OpenAI.
  • • Rolling Development: Reinvesting returns immediately into the next data center.

GPU Securitization:
‘AI Gold Mine’ or ‘Tech Subprime’?

The wildest part right now is the financing. CoreWeave recently raised funds from Blue Owl Capital. This firm, which manages $40 billion, is essentially playing the asset securitization game.They bundle future lease revenues (backed by the credit of big players like Microsoft) into things like ABS (Asset-Backed Securities)

or even CDO (Collateralized Debt Obligations).“I was on a subprime CDO desk at an investment bank back in ’06 and ’07. I lived through the story before the 2008 crisis. The core issue isn't the technology (CMBS or RMBS); it’s the quality of the underlying assets.”Back then it was real estate; now it’s GPUs. If these IDC operators, driven by bonuses and scale, keep pushing down asset quality and renting compute to startups with poor credit ratings, and then package those returns into ‘senior tranches’ to sell to the world...

Isn’t this just the tech version of ‘subprime’?

At this stage, everyone still believes AI returns will cover it all, much like in 2005 when everyone believed home prices would always go up. The million-dollar question is: can GPU lease returns truly support this kind of complex financial leverage?

Comparison Table

Element

2008 Subprime

2024 AI Financing Underlying Assets Residential Mortgages
GPU Lease Revenue Financing Instruments CDO / RMBS
ABS / Private Debt Core Driver Expected House Price Growth
AI Scaling Laws Risk Factors Interest Rate Hikes / Default
Model Performance / Compute Oversupply Next: Even Nobel Economic Theories Get It Wrong? An Insider Recaps the Financial Theory Failures of Yesteryear To be continued...
The Ghost Reappears:
When Nobel-level Models Hit ‘Fat Tails’ and ‘Chain Reactions’

Vol. 04 / Financial Engineering & AI Risk

Bridge

After discussing how ‘GPU Securitization’ turns AI compute into an asset, we must face a ghost: what if the risk assessment of the underlying assets was wrong from the very beginning?

I want to share a lesson from my time working on the CDO (Collateralized Debt Obligation) desk between 2006 and 2007. From a pure financial engineering perspective, the biggest problem we faced was over-reliance on mathematical models.

At the time, investment banks and brokerages widely used the Black-Scholes (B-S) model for pricing. But this model assumes returns follow a normal distribution, while reality is skewed and has very obvious ‘Fat Tails.’ These fat tails mean that when losses occur, the scale is so massive it blows past everyone's warning systems.

[ The Scenic Route: The Mathematical Illusion of 2007 ]

“A serious mistake everyone made back then was believing that 2007 default rates and asset correlations were at historic lows. We thought one company defaulting wouldn't affect another. According to the CDO senior tranche designs at the time, it would take seven companies going bust simultaneously to eat into the principal. Under the assumption of extremely low correlation, this was considered a ‘Black Swan’ event that was virtually impossible.”B-S 模型（Black-Scholes Model）来定价。但这个模型假设收益是正态分布的，现实却是偏态的，而且有非常明显的“肥尾（Fat Tail）”。这种肥尾意味着，在发生损失时，规模会大到超出所有人的预警。

[ The Scenic Route: 2007 年的数学幻觉 ]
“当时大家犯的一个严重错误是：认为 2007 年的违约率和资产相关性都是历史新低。我们觉得，一家企业违约并不影响另一家。按照当时的 CDO 优先级设计，需要同时有 7 家公司爆仓才会亏到本金。在相关性极低假设下，这被认为是几乎不可能发生的‘黑天鹅’。”

As soon as the Lehman crisis broke out, correlations spiked instantly. This wasn't just a butterfly effect; it was a collapse of the entire industry chain. If you are in the AI industry chain where upstream and downstream customers are the same group of people, this "coupling" leaves you with nowhere to hide when a crisis hits.

"The current AI industry chain has formed a sort of 'chained ships' structure. If any core company has a problem, the entire industry chain will have a problem. This is no longer just a tech competition, but a high-stakes financial gamble that cannot be lost."
Mental Model

Fat Tail Effect

In financial probability distributions, the probability of extreme events (such as financial crises) occurring is much higher than predicted by a normal distribution. When models ignore the "fat tail," they significantly underestimate systemic risk.

The Illusion of Asset Correlation

Normally, different assets move independently, but in times of crisis, correlations tend toward 1. In AI infrastructure, this means that if inference-side demand falls short of expectations, the financing chain from chips to the power grid will break collectively.

The Player Strategy

"Nvidia's strategy is to swap 'cards' for 'equity,' binding all the model companies to its own war chariot. This is an extremely clever defensive move."

Global Asset Allocation (Estimate)

260T
Total
100T
Bonds
20T
Corp

* Source: Goldman Sachs & Podcast Discussion. Global bonds total ~$100T, Corporate bonds ~$20T. AI infrastructure needs $1.5T (7.5% of total corporate bond market).

Who pays for this $1.5 trillion?

Banks are actually reluctant to provide long-term loans for this kind of frontier infrastructure. JPMorgan believes that over the next five years, the high-grade market needs to solve $1.5 trillion in financing needs for data centers.$1.5 trillionAt this scale, the burden will ultimately fall on the securitization market and global fixed-income investors. Why are rating agencies (Moody's, S&P, Fitch) so important? Because the world's largest institutional funds—insurance, pensions, banks—have it written in stone in their investment mandates: they can only invest in "Investment Grade" bonds.

Once the AI giants can structure their financing to achieve investment-grade ratings, they are tapping into the world's most stable and massive portion of capital—the "life savings" of the general public.

Final Strategy

Building "Too Big to Fail" Lego Bricks

Nvidia needs OpenAI. OpenAI is that "catfish" that keeps racing forward; as long as it keeps charging toward AGI, Meta, Google, and Amazon must follow suit. As long as this rhythm continues, everyone must buy chips and build power grids.

For the U.S. government, this is also a "baton." Through this deep binding, AI giants have already turned themselves into a part of U.S. national strategy.

"The more OpenAI cannot be allowed to fail, the hidden logic is: they want to turn themselves into that 'too big to fail' building block in the global AI order."

Previous: GPU Securitization: Are the underlying assets an "AI Gold Mine" or a "Tech Subprime"?

Next: Behavioral Finance’s Prisoner’s Dilemma: Betting everything for a ticket on the AI ship The AI Infrastructure Dossier / Part II

Behavioral Finance’s Prisoner’s Dilemma:

Betting everything for a ticket on the AI ship
When "Too Big to Fail" evolves into "Too Big to Not Gamble," Silicon Valley and Wall Street are engaged in a cross-century liquidity bet.

Context Bridge

Recap: We analyzed how tech giants use high-grade debt and national strategy alignment to build a "too big to fail" AI infrastructure landscape. But in this trillion-dollar race, the underlying logic driving CEOs to sign checks frantically isn't just a computing power race—it's a classic behavioral finance "Prisoner’s Dilemma."

When discussing why big tech is frantically building data centers, I've always believed there's a deep behavioral finance issue at play. As a CEO of a major firm, you are actually facing an asymmetric risk game.

If you decide to invest heavily in AI now and it eventually turns out to be over-investment, you might only lose some cash flow or incur depreciation costs—which isn't fatal considering these giants have tens of billions in annual operating cash flow. But if you don't follow suit and your competitors win the bet, you'll be nailed to the pillar of shame and lose your job entirely.

[ Deep Dive: 0 or 1 Strategy ]

[ Deep Dive: 0 or 1 Strategy ]

For Big Tech CEOs personally: if you bet right, you're a god; if you bet wrong, but everyone else is wrong too, the punishment is limited. But if you don't bet and miss out, that's an unforgivable strategic blunder. This 'must-get-on-board' mentality is driving the current reckless infrastructure frenzy.

Even Mark Zuckerberg (Zuck) admits that while it looks terrifying, you have to follow suit. The focus has shifted from 'over-investment' to 'can the cash flow last until dawn.' Right now, the debt ratios of Big Tech haven't reached exaggerated levels, but for a unicorn like OpenAI—which is calling for $1.4 trillion in funding—debt is practically the only way out.

“If you bet right, you're a god; if you bet wrong, but everyone else is wrong too, you won't be punished much.”

Speaking of OpenAI, Sam Altman is facing growing skepticism. A streamer even challenged him to his face: how can your projected $13 billion in revenue support a $1.4 trillion valuation? Sam looked very angry at the time. I believe OpenAI must go public by the first half of 2026, or even earlier. Because by 2027, with the Fed's potential balance sheet expansion and changes in the political cycle, the window for the funding game might narrow.

Financial Model

OpenAI’s Cash Flow Paradox

  • 2024 Projected Annual Revenue $13B
  • Long-term Infrastructure Needs $1.4T
  • Comparison (Uzbekistan's GDP) ~$100B
  • Implicit Conclusion IPO is the only lifeline

*Note: Investing capital equivalent to a mid-sized country's GDP annually is no longer just corporate behavior; it's a mobilization of resources at the national level.

Politics & Liquidity

2026: The Final Golden Window?

Trump might leave office after 2028, and Vance is currently leading. But the more crucial factor is the Fed's actions. If the Fed is forced to expand its balance sheet in Q1 next year to address low bank reserves, combined with a smooth political transition, 2026 could be the best—and last—opportunity for OpenAI to go public.

Risk: If the Democrats sweep both houses, the slope of AI/Web3 regulation will be adjusted.

Miners Pivoting to AI: The Real vs. Fake Power Play

Observing these companies pivoting from crypto mines to AI data centers (HPC/AI DC), we see the market giving a massive valuation premium to 'existing power capacity.'

Electricity Capacity Insight (2024-2027)

Data Reveals: Most companies are in a financing stress-test period between 'power on paper' and 'actual conversion.'

Risk Point A: Off-grid Power

“You can buy GPUs, but can you plug them in?” Even with on-site power plants, if they aren't grid-connected, the construction timeline could drag on until 2027.

Risk Point B: Conversion Financing

Converting from a mine to an AI data center requires massive CapEx. If the financing pace falls behind, power contracts will lapse.

Bitdeer: A Geopolitical Gray Balance

Bitdeer is a very unique entity. It's a Singaporean company, but its core team has a Chinese background. It’s securing power in Texas and is deeply tied to Tether through equity.

[ Tangent: The Tether Connection ]
Tether currently holds about 25.5% of Bitdeer. Even more subtle is Tether's relationship with Cantor Fitzgerald. Cantor is currently churning out reports supporting Bitdeer, which is actually a game of geopolitics and capital.

Internal Logic

“Can a Singaporean company provide power for LLM training? Power might be negotiable, but data is highly sensitive. Advanced GPUs must stay in the hands of 'our own people.'”

— Regarding Bitdeer's Survival Boundaries

Geopolitical Risk List

  • ✕ Serving Chinese clients remotely (e.g., Tencent) is seen as a red line by short sellers.
  • ✕ Restrictions on H-series chip exports, even in overseas data centers.
  • ✓ Using offshore power for localized training might be the only exit in the gray zone.

"The first half of next year will be a great 'squat-to-jump' opportunity. With year-end adjustments and the Christmas holidays approaching, the financing game must continue because AI has become a battle for America's national destiny—too big to fail."

Previous

Big Tech Strategy: Linking Debt Issuance to National Strategy

Next Episode

Companies that successfully pivot from mining to AI are rarely the 'straight-A students' of the field.

Case Study: The Pivot Paradox

Why the pivot winners aren't the mining 'overachievers'?

Path Dependency and the Breakthrough of Power Hoarders

Context Bridge

Following the previous discussion, we covered the power supply gap in the crypto-to-AI pivot and Bitdeer's 'gray zone' in overseas geopolitics. However, what's even more dramatic in the capital markets is that companies once considered mediocre in the mining space have instead become the pioneers of this AI transformation.

I think the companies that successfully pivoted in the last wave weren't actually the standout leaders in the mining space. They were able to pivot because they entered a good track early on. It's a classic case of path dependency: the better a company performed in the mining world last time, the harder it is for them to pivot now.

Take Marathon (MARA) for example—people might think of it as a giant, but its management is actually pretty terrible, and its overhead has stayed sky-high. MARA's strategy has been all about 'chasing the hype.' Many people mistakenly think it has 1.2 GW of power, but that's an illusion. It doesn't even have enough of its own rigs, and a large chunk of its actual mining farms are leased. This 'heavy on hashrate, light on power' model has become a burden in the race to pivot to 'infrastructure-heavy' AI.

"For AIDC (AI Data Centers), the best model is 'rolling development.' It's like real estate: you keep mining to maintain cash flow while converting idle or inefficient mining farms into data centers."

Rolling Development: The Real Estate Logic of AIDC

The current situation is that Bitcoin's hashrate hasn't dropped due to the AI pivot because everyone is 'hedging their bets.' They want to capture the gross margins from Bitcoin mining while also betting on the future of AI.

[ Deep Dive: Financial Growing Pains ]

Take Core Scientific (CORZ) as an example: though they pivoted early, their short-term financial data isn't great. Last year's Q1 report showed that mining gross margins could reach 70-80%, while data center margins, after deducting depreciation, were actually lower than mining. This is because costs are upfront (buying GPUs, retrofitting power), while revenue lags behind. If you shut down mining to pivot to AI and cut off your cash flow, that's a very dangerous move.

So, the 'optimal solution' for miners pivoting right now is: use existing Bitcoin rigs as a cash flow backbone, then carve out a portion of the power capacity to build AI server rooms gradually through rolling development.

Disclosure & Disclaimer
[ Speaker 1 ]: I have personally bought into all the tickers mentioned above. [ Speaker 0 ]: I do not hold any of the companies mentioned.
Invest with caution. This content does not constitute investment advice and is simply a record of the podcast's content.

Related Episodes

Your Catchy English Title

Your Catchy English Title

Tesla vs. Waymo: The High-Stakes Battle for Self-Driving Supremacy

Tesla vs. Waymo: The High-Stakes Battle for Self-Driving Supremacy

AI 2025-2026: Consensus, Conflict, and the Next Frontier of Scaling Laws

AI 2025-2026: Consensus, Conflict, and the Next Frontier of Scaling Laws

e220-hyrox-4z8s5

The €14 Billion Hermès Mystery: Heirs, Gardeners, and the War for Luxury

The €14 Billion Hermès Mystery: Heirs, Gardeners, and the War for Luxury

e218-netflix-b97ly