Blackstone 2026 AI Investment Panorama: The Real Returns of Trillion-Dollar Infrastructure
Jon Gray
Blackstone President & COO
Original Statement
1. Core Q&A: "Where's the Beef?"
• Core controversy and core metaphor: Borrowing from the classic 1984 Super Bowl ad phrase "Where's the Beef?", directly confronts market skepticism about whether the trillion-dollar AI CapEx constitutes false prosperity or circular financing.
• Blackstone's qualitative conclusion:
• "Bun": Includes chips, hyperscale data centers, power infrastructure, and the tens of trillions of dollars of global capital expenditure.
• "Beef": The real return on investment (ROI) that is currently exploding on the enterprise side. The cost reduction and efficiency improvements of Blackstone's portfolio companies are not mere conceptual hype, but are being transformed into real cash flow and high multiples of returns, further stimulating and supporting long-term demand for computing power.
2. Key Data: Token Surge, Model Revenue, and CapEx Leap
• Token usage leaps in magnitude: Taking Google Token data as an example, monthly token usage has surged from an extremely low base to 3.2 Quadrillion (3.2 petabytes, or 3.2 trillion), maintaining a steep growth curve.
• Revenue explosion for leading model vendors: OpenAI and Anthropic's annualized run-rate revenue has reached $105 billion, with the combined valuation of the two giants expected to reach $3 trillion.
• Measured spending surge in Blackstone's portfolio companies: Among the 1,400 portfolio companies tracked by Blackstone, model procurement spending's annualized run-rate has skyrocketed 21 times from $25 million a year ago to $525 million.
• Doubling of hyperscaler capital expenditures:
• The CapEx of the world's five largest hyperscalers has doubled from $41.5 billion last year to $82 billion, equivalent to 2.5% of the nominal GDP of the United States.
• The leasing scale of Blackstone's largest global data center platform (such as QTS) has jumped from 1 GW in 2024, 2 GW in 2025, to an expected 6 GW or more in 2026, corresponding to nearly $100 billion in infrastructure capital expenditure (tenant chip investments require several hundred billion dollars more).
• Valuation differentiation in hardware and market caution: Storage chip manufacturers (Micron, SK Hynix) have seen stock price increases of 500%-600%, but SK Hynix's dynamic P/E ratio remains around 4 times. This indicates that the capital market has not experienced the irrational exuberance of Cisco's 150 times P/E during the 2000 internet bubble, and still retains a cautious attitude towards overheating cycles.
3. Historical Reflections: The Mirror Insights of the 1870 Industrial Revolution and Railway Bubble
• The thirty-year transformation from 1870 to 1900:
• In 1870, the United States was in an agricultural and early industrial stage characterized by wooden structures, horse-drawn transportation, and candle lighting; by 1900, it had evolved into a society with steel structures, electric lighting, a railway network, and urbanization.
• Economic multiplier effect: During these 30 years, U.S. labor productivity doubled, GDP grew fourfold, real manufacturing output increased sixfold, and the stock market achieved a sevenfold increase. The current AI revolution is on the brink of a technological explosion similar to that of 1870.
• The core differences between two types of infrastructure cycles:
• Historical railway crisis: In the 1870s, 200 railway companies went bankrupt, essentially due to high leverage and excessive supply ahead of demand (similar to the telecom fiber surplus in 2000).
• Current AI infrastructure reality: Demand is currently severely outpacing supply, and the core performance parties (major hyperscalers) have extremely healthy balance sheets and low leverage, driven by certain demand for the construction of physical assets.
4. Real Business ROI: Comprehensive Productivity Release
• Process automation and extremely high capital returns:
• Phoenix Tower (mobile communication towers): Invested $4 million to develop AI leasing review processes, improving processing efficiency by 5 times, directly contributing $4.5 million in savings annually, achieving over 100% annualized investment return.
• Enverus (energy data analysis): Applied large models in code maintenance and repair, achieving 18 times efficiency/economic returns on model procurement costs.
• Trion (residential leasing): Reduced application processing time by 90%, significantly reshaping customer experience.
• Digital elevation of traditional analog businesses (Chamberlain):
• A traditional garage door opener manufacturer has entered AI visual perception and intelligent access control (Secure View 3-in-1), with related business run-rate reaching $40-50 million, and the company expects to sprint to a $500 million run-rate in the next five years.
• Macroeconomic productivity and financial indicators:
• The average annual labor productivity in the U.S. over the past decade was about 1.5%, while in the past two and a half years, it has jumped to 2.6%.
• Revenue per employee for leading hyperscalers has increased by 65% in 3.5 years.
• The EBITDA profit margins of the S&P 500 index and Blackstone's portfolio companies have expanded by 500 and 700 basis points, respectively, over the past four years.
5. Core Constraint Links: Four Major Obstacles from Digital Space to Physical World
• Community and approval barriers (Entitlements & Moratoriums): Local communities' concerns about water consumption and grid pressure from data centers are spreading, with some areas experiencing policy moratoriums that extend site selection and project implementation cycles.
• Severe scheduling for power and heavy industrial equipment: The expansion of the power grid and gas turbine equipment is in extreme short supply, with core suppliers like GE Vernova having heavy turbine orders scheduled as far out as 2030-2031.
• Imbalance in supply and demand for advanced process chips: Hyperscalers' CapEx has increased ninefold in five years, but upstream semiconductor manufacturers' capital expenditure expansion remains restrained, leading to a continued tight supply of high-performance computing clusters.
• Surge in funding scale for individual projects: Building a 1 GW scale AI factory (including power distribution networks, data center buildings, and high-performance chips) typically requires mobilizing up to $55 billion in capital, heavily relying on sovereign-level institutional investors with ultra-large-scale capital deployment capabilities.
6. Investment Risk Assessment and Scarcity Defensive Asset Allocation
• Valuation declines in disrupted industries: The software and professional information services sectors have seen significant valuation compression (P/E compression) in the face of expectations for agent replacement, with private equity software M&A transactions experiencing a cliff-like decline (large transactions down over 66%).
• Comparison of retail industry changes and winner-loser differentiation:
• In the face of e-commerce shocks, barrier-free retailers (Kmart, Sears, Toys "R" Us) have faced extinction;
• Players with deep moats and supply chain barriers (Costco, Walmart, TJ Maxx) have successfully broken through.
• In the face of AI, only enterprises with core business factual databases (System of Record) that shift their charging model from "per seat" to "based on outcomes" can build lasting barriers.
• "Irreplaceable physical and experiential assets" that are not afraid of AI disruption:
• While Blackstone bets on the most fundamental compute assets, it continues to allocate scarce counter-cyclical targets completely decoupled from AI: such as the Indian cricket team with 1.4 billion audiences but only 10 teams, the core hub airport in Rome, scarce first-line beach properties, and high-loyalty consumer chains (Seven Brew).
• Four major systemic macro tail risks:
• Regulatory forced brakes triggered by unknown cybersecurity incidents affecting critical financial infrastructure;
• Divergence in technological evolution paths (such as breakthroughs in space-based data centers, edge computing, but ground-based centralized data centers remain the mid-term mainstay);
• Geopolitical single-point dependencies (about 90% of advanced process chips are concentrated in Taiwan);
• Extreme irrational bubbles in the primary market without business support (such as some defense and AI concept companies that achieve valuations in the tens of billions despite having no revenue).
ABAB AI Insight
Blackstone's Jon Gray: The Real Returns Behind the Trillion-Dollar AI CapEx, Infrastructure Wars, and a New Capital Order
If you only look at the surface, today’s AI investment boom does indeed resemble a massive bubble.
Chip companies' market values are skyrocketing.
Data centers are expanding rapidly.
Tech giants are investing hundreds of billions of dollars each year.
Electricity, natural gas, optical communication, nuclear energy, and storage chips are all driven by AI.
The primary market is continuously seeing valuations in the tens of billions and hundreds of billions.
Thus, the market begins to ask an increasingly important question:
How will this money ultimately be earned back?
This is precisely the core question posed by Blackstone President and COO Jon Gray in his latest speech in September 2026:
Where’s the Beef?
This phrase comes from Wendy’s classic advertisement in 1984.
The hamburger bun looks very big.
But when you open it:
Where’s the beef?
Gray uses this metaphor to ask:
The world is currently investing trillions of dollars in AI.
Data centers are the bun.
GPUs are the bun.
Power systems are the bun.
Fiber optics and networks are also the bun.
The real "beef" must be the economic returns generated by this capital in the end.
And Blackstone's answer is very clear:
They believe that the "beef" of AI has already begun to appear.
Not at some point in the future.
But it is already reflected in corporate profit statements, labor productivity, software development efficiency, business processes, and real-world infrastructure demands. Blackstone officially summarizes the core of this speech as: the deployment of AI by companies has already brought quantifiable improvements in productivity, revenue, and operations, and these returns are in turn driving demand for computing power.
But what is truly worth our study is not that "Blackstone is bullish on AI."
Blackstone itself is one of the largest providers of AI infrastructure capital in the world.
It certainly has its own position.
The truly important question is:
What does it see correctly? What might it see incorrectly?
This is what investors should study.
────────────────
1. The real issue in the AI bubble debate has never been whether the investment is too large.
In the past two years, when the market discusses the AI bubble, it often makes a mistake.
Everyone focuses on one number:
CapEx.
This year spending $500 billion.
Next year spending $800 billion.
In the future possibly trillions of dollars.
So many people directly conclude:
With such large investments, it must be a bubble.
But the scale of capital expenditure itself cannot determine a bubble.
What should really be looked at are three variables:
Demand;
Cost of capital;
The cash flow generated by the assets in the end.
Railroads are expensive.
Power grids are expensive.
Communication networks are expensive.
Highways are expensive.
Data centers are also very expensive.
But "expensive" never equals "wrong."
The real danger is:
The speed of supply construction far exceeds real demand.
This is the root cause of the vast majority of infrastructure bubbles bursting.
────────────────
2. What investors should really ask is not "how much has been spent," but which side of Demand/Supply is leading.
This may be one of the most important frameworks for understanding today’s AI infrastructure cycle.
Infrastructure investments generally exhibit two states.
The first:
Demand > Supply.
Demand appears first.
Infrastructure cannot keep up.
At this time, prices rise;
Rents rise;
Equipment delivery cycles extend;
Companies scramble for resources;
Capital only then begins to enter.
This is a typical:
Shortage Cycle.
The second:
Supply > Demand.
Everyone sees profit opportunities and simultaneously expands production.
Then:
Capacity comes online in large quantities;
Demand is not as high as expected;
Prices fall;
Utilization rates decline;
Debt begins to have problems.
This is:
Overbuild Cycle.
All large infrastructure cycles ultimately oscillate around these two phases.
────────────────
3. Blackstone's biggest judgment today: AI is still in a Shortage Cycle.
Why does Gray dare to continue investing in data centers?
Because Blackstone sees not media reports.
But real leasing.
Blackstone officially disclosed that its global data center platform expects:
To sign about 1GW in 2024;
About 2GW in 2025;
And over 6GW in 2026.
Gray stated in May 2026 that just this approximately 6GW of data centers corresponds to about $100 billion in data center construction capital, and the chips that tenants put in may require an additional investment of about $200 billion.
This is very exaggerated.
What does 6GW mean?
It is not office electricity consumption.
But a continuous power load at the level of a medium-sized city or even a large industrial area.
So AI has officially transitioned from:
Software issues
To:
Industrial infrastructure issues.
────────────────
4. The real big change: AI has moved from Silicon Valley to the power grid.
In the past, the most important production material for internet startups was:
Servers.
Today, the most important production materials for AI have changed to:
GPUs;
HBM;
Optical communication;
Transformers;
Data centers;
Gas turbines;
Transmission lines;
Land;
Water;
Cooling equipment.
This indicates that AI is undergoing a very important phase transition:
First phase
Model Revolution.
Who can train the smartest model?
Second phase
Compute Revolution.
Who has enough GPUs?
The third phase
Is currently happening:
Infrastructure Revolution.
Who can truly provide these GPUs with:
Power;
Cooling;
Networking;
Building data centers;
Acquiring land;
Obtaining permits?
So the biggest problem AI ultimately encounters is not mathematics.
But:
The physical world.
────────────────
5. This is why Blackstone may be better positioned than many tech VCs to earn money in the second phase of AI.
Understanding Blackstone is very important.
Blackstone is not Andreessen Horowitz.
Nor is it Sequoia.
Its most powerful capability is not:
To bet on the next app.
What it truly excels at is:
Real estate;
Infrastructure;
Credit;
Energy;
Large mergers;
Massive capital organization.
And at this stage, AI just happens to need these capabilities.
Thus, AI is gradually transforming from a:
VC Opportunity
To:
Infrastructure Capital Opportunity.
This is actually very critical.
In the early stages of a technological revolution, money:
Is usually earned by tech entrepreneurs and VCs.
Once the technological revolution enters the infrastructure stage:
Money begins to flow to:
Real estate;
Private credit;
Infrastructure funds;
Energy companies;
Industrial equipment;
Large asset management companies.
So Blackstone's real advantage is not:
That it understands models better than OpenAI.
But rather:
When AI transitions from code to rebar, cement, electricity, and debt, that is precisely Blackstone's world.
────────────────
6. The real watershed in the AI capital cycle is whether "tokens have generated income."
Gray presented several very exaggerated data points in his speech.
Google's monthly token usage:
Starting close to zero in 2024;
Previously about 480 trillion;
By May 2026, it had reached approximately:
3.2 Quadrillion.
That is:
32 trillion tokens.
Gray also stated that as of July 2026, the combined annualized revenue run rate for OpenAI and Anthropic had reached about:
$105 billion.
Whether we fully accept these metrics or not,
The most important trend is already very clear:
AI usage is transitioning from experimentation to consumption.
This is a significant difference in terms of investment.
────────────────
7. The real issue during the internet bubble was not the lack of traffic, but that the traffic was not fully monetized.
During the 2000 internet bubble:
Internet users were real.
Web page visits were real.
Traffic growth was also real.
So was "the internet a scam?"
Of course not.
The internet later changed the entire world.
But the problem was:
Many companies' valuations at that time had overdrawn decades of commercialization.
This is the most difficult part of understanding bubbles:
A great technological revolution can completely coexist with a huge asset bubble.
These two things are not contradictory.
Railroads are real.
Railroad companies can also go bankrupt.
The internet is real.
Pets.com can also go to zero.
AI is the same.
AI changes the world, but that does not mean:
All AI companies are worth their current prices today.
This is one of the most important principles for understanding this round of capital markets.
────────────────
8. Therefore, to judge the AI bubble, it is essential to separate "technology correctness" from "price correctness."
An asset can have four possibilities.
The first:
Technology is wrong.
Price is also wrong.
The worst.
The second:
Technology is right.
Price is wrong.
This is the most common form of bubble.
The third:
Technology is right.
Price is also right.
The most ideal.
The fourth:
The market undervalues the technological value.
This is the biggest investment opportunity.
So professional investors should never just ask:
"Is AI the future?"
The answer is likely:
Yes.
But this is not very helpful for investment.
What should really be asked is:
How much am I paying for this future today?
This is asset pricing.
────────────────
9. The most important data for Blackstone is not OpenAI's revenue, but that its corporate clients are starting to spend more.
Blackstone has a very special observation window.
It spans:
Private equity;
Real estate;
Credit;
GP stakes;
Infrastructure.
Thus, it can observe a large number of real enterprises.
Gray disclosed that among about 1,400 related companies, GP portfolio companies, and borrowing companies observed by Blackstone, the annual spending related to Anthropic rose from about:
$25 million in September 2025
To:
$525 million.
An increase of:
21 times in about a year.
Here, a detail must be noted:
This does not mean that Blackstone's 1,400 companies have "total AI spending of only $525 million."
What Gray presented is a more specific perspective on the annual spending of relevant companies on Anthropic.
But it still illustrates one thing:
Companies have moved from trial to deployment.
────────────────
10. The most important threshold in the history of enterprise software: Pilot → Production
The tech circle loves to say:
Pilot.
Trial.
But before enterprise technology can really make money, it must cross a hurdle:
Production Deployment.
Because:
Pilot budgets are very small.
Once it enters the production environment:
Data will be integrated;
Workflows will change;
Employee time will be replaced;
Customers will be affected;
It will enter the annual budget.
Blackstone's own survey in 2026 shows:
Nearly 50% of surveyed Portfolio Company CEOs stated that AI projects have entered production or full deployment stages;
Less than 30% a year ago.
At the same time, about 80% of CEOs indicated plans to increase AI spending in the next 12 months.
This is more worthy of capital attention than ChatGPT download numbers.
Because:
Enterprise budgets are the long-term cash flow.
────────────────
11. The real "meat" of AI can ultimately only have two types
Any enterprise purchasing AI cannot avoid two ROIs.
The first type:
Cost Reduction
Lower costs.
Fewer hires;
Fewer working hours;
Faster process completion;
Fewer errors;
Lower customer service costs.
The second type:
Revenue Enhancement
Increase revenue.
Sell more;
Improve conversion;
Launch new products;
Enhance customer value;
Increase prices.
If neither exists:
AI is just a toy.
Enterprises will eventually cut budgets.
────────────────
12. Why is the case of Phoenix Tower very important?
The case of Phoenix Tower provided by Blackstone is very simple.
It is a mobile communication tower business company.
The company invested about:
$4 million
to build an AI leasing review process.
The result:
The speed of processing leases increased by about 5 times;
Generating about:
$4.5 million in economic benefits in a year.
In other words:
It basically recouped the investment in about a year.
The annualized ROI exceeds 100%.
This is a very typical enterprise AI application.
Note:
It did not train AGI.
It did not create robots.
It did not change human civilization.
It simply:
Automated an expensive, repetitive, slow process.
But enterprises love such projects.
Why?
Because CFOs can calculate the numbers.
────────────────
13. The AI that enterprises are truly willing to spend on continuously is not necessarily the coolest AI
This is something entrepreneurs must understand.
Consumers like:
Wow Factor.
Enterprises like:
Payback Period.
If an AI product tells the CFO:
Our model's IQ has increased by 20%.
It means little.
But if it tells the CFO:
Originally 30 people did it;
Now 8 people do it.
Or:
Originally the process took 3 days;
Now it takes 3 hours.
Or:
Saving $5 million a year;
Software costs $1 million.
Then it is:
5× ROI.
The probability of closing the deal is completely different.
So the real golden formula for B2B AI is not:
AI Capability.
But rather:
Economic Value Created / Price Charged.
────────────────
14. This is also why Enverus's 18× is very worth studying
Blackstone also mentioned the energy data company Enverus.
After participating in code maintenance and software engineering processes through AI, its model costs relative to efficiency gains formed about:
18 times economic return.
Gray listed software engineering as one of the areas where its portfolio companies generally gain AI benefits.
Why is coding commercializing so quickly?
Because code naturally meets several conditions that AI loves:
Input is digitized;
Output is digitized;
Results are verifiable;
Can iterate extensively;
Errors can be redone;
Data is rich.
So Coding Agents are likely just the beginning.
In the future, AI will move from:
Code
to:
Accounting;
Legal;
Insurance;
Healthcare Admin;
Procurement;
Customer Support.
The commonality is:
A large number of rules + a large amount of text + verifiable results.
────────────────
15. The real big opportunity is to turn "labor costs" into "computing costs"
This is one of the most important statements in AI economics.
In the past, a workflow:
Costs mainly came from:
Labor.
In the future, part of the workflow:
Costs will begin to turn into:
Compute.
For example, previously:
100 people × $100,000 annual salary
= $10M.
In the future:
30 people + $2M AI costs
= $5M.
Thus, the company creates:
$5M cost advantage.
AI companies can extract:
A portion of the Value from it.
This explains why the AI market may be far larger than traditional SaaS.
────────────────
16. The market for SaaS is the IT Budget, while the ultimate market for AI may be Payroll
This is a very important magnitude change.
Traditional software companies sell:
Software.
So customers pay from:
IT Budget.
If AI Agents can truly complete work,
Customers are actually purchasing:
Digital Labor.
This means AI can theoretically enter:
Payroll Budget.
The global annual labor costs for enterprises
Far exceed software budgets.
Therefore, the largest TAM for AI may not be:
Software Market.
But rather:
Labor Market.
This is also why capital is willing to invest such terrifying amounts.
────────────────
17. After understanding this, one can understand why the demand for computing power may last a very long time
If AI is just:
Writing emails;
Summarizing articles;
Generating images.
Then the current construction of data centers may indeed be oversupplied.
But if AI begins to take on:
Customer service;
Programming;
Finance;
Legal analysis;
Scientific research;
Robotics;
Autonomous driving;
Manufacturing;
Healthcare administration;
Then:
Every time "AI work"
Will consume Tokens.
Tokens ultimately become:
GPU computations.
GPU computations become:
Electricity.
So the future chain may be:
More AI labor → More Tokens → More Compute → More electricity.
This is how the digital economy reconnects with the physical economy.
────────────────
18. AI may ultimately become a new "energy converter"
This is a very advanced but very important understanding.
What is the essence of the Industrial Revolution?
To convert:
Coal;
Oil;
Natural gas;
Electricity
into:
Mechanical labor.
The AI revolution may add a new link:
Electricity → Computing → Intelligent labor.
In the past:
1 MWh of electricity
drives machines.
In the future:
1 MWh of electricity can also generate:
Tokens;
Reasoning;
Code;
Design;
Automated decision-making.
In other words:
Energy is beginning to be converted into artificial intelligence.
This will fundamentally change the strategic value of energy assets.
────────────────
19. Therefore, future AI competition may ultimately evolve into energy competition
Many countries today talk about:
AI Sovereignty.
What everyone thinks of first is:
Models.
But true AI sovereignty includes:
Chips;
Wafer fabs;
HBM;
Data centers;
Electricity;
Power grids;
Cooling;
Talent;
Capital.
Especially:
Power.
If you don't have enough electricity,
Buying more GPUs is useless.
So whether the US can maintain its AI advantage in the next decade largely depends on:
Whether it can quickly increase:
Natural gas;
Nuclear energy;
Renewable energy;
Energy storage;
Transmission;
Transformers.
AI has upgraded from a software issue to:
A national industrial capability issue.
────────────────
20. Why have companies like GE Vernova suddenly become part of the AI chain?
This is something many ordinary investors could not have imagined a few years ago.
Everyone buys AI:
Buying NVIDIA.
But after the increase in data center construction:
Demand for gas turbines increases;
Demand for grid equipment increases;
Demand for transmission and transformation equipment increases.
Thus, second-order beneficiaries begin to emerge.
This is a very important concept in investing:
Second-Order Effects.
The first layer is visible to everyone:
GPUs.
Truly excellent capital begins to study:
After GPU growth, who will inevitably benefit?
Electricity.
Then:
After electricity growth, who benefits?
Gas turbines;
Nuclear energy;
Transformers;
Transmission;
Copper;
Cooling.
Investment opportunities thus expand outward.
────────────────
21. One of the most profitable stages of AI investment may not be the models, but the Bottleneck
A very practical method in industrial investment is:
Do not chase the hottest things.
Look for:
Bottlenecks.
Because after demand explodes:
Pricing power usually flows to the most scarce links.
AI currently may have several huge bottlenecks:
Advanced GPUs;
HBM;
Advanced packaging;
High-speed optical communication;
Electricity;
Gas turbines;
Transformers;
Data center land;
Interconnection;
Approval.
Whoever controls the bottleneck,
Whoever has:
Pricing Power.
────────────────
22. This is why a 1GW AI Factory can be worth hundreds of billions of dollars
Gray discussed building a 1GW AI Factory in his speech.
If you include:
Data centers;
Electricity;
Chips;
Networks;
Supporting infrastructure
All counted in,
The capital demand can reach about:
$55 billion level.
This is no longer the ordinary CapEx of a tech company.
What does $55 billion mean?
It is close to:
A large energy project;
A large airport;
Even part of a national-level infrastructure.
This means AI has entered:
The Mega Project Economy.
────────────────
23. The Mega Project Economy will change the structure of capital markets
A $100M project:
VCs can handle it.
A $1B project:
Growth Equity + Debt can handle it.
A $50B project:
Already requires:
Hyperscalers;
Infrastructure Funds;
Private Credit;
Banks;
Bond markets;
Insurance funds;
Pensions;
Sovereign wealth funds.
So as AI develops further,
it is no longer just the solo act of startups.
It has become:
A global capital market engineering.
This is exactly why alternative asset giants like Blackstone, Apollo, Brookfield, and KKR are truly entering AI.
────────────────
24. This is also why Private Credit will become an important financial infrastructure for the AI revolution
Many people think:
AI = Stock Market.
In fact, what large data centers need most is:
Debt.
Because data centers typically:
Have extremely high CapEx;
Have long contract cycles;
Have relatively predictable cash flows.
This is very suitable for debt financing.
Especially after banks are constrained by capital regulations:
Private Credit can take on more and more large projects.
Thus, AI may indirectly promote:
Private Credit
Further expansion.
Behind technological revolutions,
financial innovations often occur simultaneously.
────────────────
25. But it must be noted here: this is also where risks begin to increase
In the early AI CapEx:
Meta;
Microsoft;
Alphabet;
Amazon
paid out of their own pockets.
These companies:
Have a lot of cash;
Have very little debt;
Have strong credit.
The risks are not high.
But as CapEx continues to rise from hundreds of billions:
More and more financing will enter:
SPVs;
Project financing;
Private Credit;
Asset securitization;
Equipment financing;
Guarantee structures.
By 2026, the market has already begun to discuss extensively the off-balance-sheet commitments, guarantees, and special financing structures in hyperscaler AI investments. FT reports that large tech companies are using structures like residual-value guarantees to support financing for data centers and chips; another market analysis estimates that the five major hyperscalers' CapEx in 2026 will be about $820 billion, already exceeding their operating cash flow of about $750 billion.
This means:
The AI CapEx cycle has begun to shift from:
Cash flow financing
to:
Credit expansion.
Historically, all large capital cycles reach this point,
and risks begin to rise.
────────────────
26. So saying "today is not like 2000" does not mean "it will never turn into 2000"
Gray's important argument is:
The capital market has not completely lost its mind.
For example:
Although Micron and SK Hynix's stock prices have skyrocketed,
SK Hynix's Forward P/E at that time was still below about 4 times.
While in 2000, Cisco's bubble peak:
The P/E ratio once reached extreme levels.
This indicates that at least some AI infrastructure stocks are not rising based on infinite valuation expansion.
This judgment makes sense.
But it needs to be added:
Bubbles do not necessarily first appear in stock P/E ratios.
They can also hide in:
Private Valuation;
Data center land prices;
Debt;
Long-term leases;
GPU Residual Value;
Power contracts;
Primary markets.
So judging bubbles cannot only look at:
Nasdaq P/E.
────────────────
27. The real indicators to monitor the AI bubble are six metrics
If I were a capital allocator, I would not ask every day:
"Did NVIDIA go up today?"
I would look at these six things.
First: Compute Utilization
Are GPUs really busy?
Or are they largely idle?
────────────────
Second: Inference Revenue / CapEx
For every additional dollar of infrastructure,
how much AI revenue can it ultimately generate?
────────────────
Third: AI Customer ROI
Why do customers continue to pay?
Is there real cost savings,
or is it just that the CEO demands "must AI"?
────────────────
Fourth: Debt / Cash Flow
Is CapEx increasingly reliant on debt,
or is it still supported by cash flow?
────────────────
Fifth: Supply Lead Time
Are GPUs, electricity, and data centers still in short supply,
or is inventory starting to appear?
────────────────
Sixth: Price per Unit of Compute
If the price of computing power plummets,
it means:
Technological efficiency has improved;
or:
Supply is starting to exceed demand.
It must be distinguished.
These six metrics are more important than any slogan of "Is AI a bubble?"
────────────────
28. The real lesson from the 1870 railroad story is not that "infrastructure won't bubble"
Gray compares today's AI to:
1870.
This analogy is very interesting.
The United States in 1870 and the United States in 1900 were almost two different worlds.
In 1870:
Wood structures;
Horse-drawn carriages;
Kerosene lamps;
A large agricultural population.
In 1900:
Steel;
Railroads;
Electrification;
Modern cities;
Mass industrialization.
In these 30 years, the productive capacity of the United States underwent a huge leap.
Gray uses this to illustrate:
AI may be similar to the starting point of the Second Industrial Revolution.
But history has another half.
────────────────
29. Railroads changed America, but railroad investors could still go bankrupt
This must be understood.
Railroads indeed:
Changed America;
Created a national market;
Reduced transportation costs;
Promoted urbanization;
Boosted steel;
Expanded commerce.
But the 19th-century American railroad industry also experienced a lot of:
Bankruptcies;
Debt crises;
Overbuilding.
So history does not tell us:
"Big technological revolutions should be blindly bought."
But rather:
The infrastructure itself may be great, but the capital returns on infrastructure may still be poor.
This is the biggest difference between the history of technology and the history of investment.
────────────────
30. The most common mistake investors make is confusing social value with shareholder returns
Railroads generated:
Huge social value.
Aviation generated:
Huge social value.
Telecommunications generated:
Huge social value.
The internet generated:
Huge social value.
But many industries have long:
Poor capital returns.
Why?
Because competition will drive prices down.
This is called:
Consumer Surplus.
Consumers take away a lot of technological dividends.
Companies do not necessarily capture all the profits.
So to judge AI companies, one must ask:
How much value does AI create?
Then ask:
Who can capture this value?
These are two completely different questions.
────────────────
31. The biggest battle for wealth in the future of AI is actually Value Capture
After AI creates value,
who takes it away?
It could be:
Model companies;
Chip companies;
Cloud computing companies;
Energy companies;
Data centers;
Software applications;
Enterprise customers;
Consumers.
The final profit distribution depends on:
Who is the most scarce;
Who is the hardest to replace;
Who has pricing power.
Currently, the strongest Value Capture:
Is clearly concentrated in:
Advanced chips;
HBM;
Some cloud infrastructure;
Leading models.
But the future may not necessarily be permanent.
────────────────
32. A strong model does not mean the model layer will always make the most money
Historically:
Internet protocols were very important.
TCP/IP did not capture all the profits of the internet.
Smartphone OS was very important.
But real wealth was distributed among:
Apple;
Google;
Meta;
Uber;
Airbnb;
TikTok.
So the future AI industry may also see:
Basic models gradually commoditized.
Real profits shifting to:
Companies with customer relationships;
Proprietary data;
Workflows;
Brands;
Distribution.
This is also why a company's System of Record remains very important.
────────────────
33. The real danger for the software industry is not that AI will eliminate software, but that AI will eliminate "Seats"
The core economic model of past SaaS:
The more employees,
The more software Seats.
Thus, revenue increases.
But one core result of AI may be:
Fewer customer employees.
So originally:
1000 Seats
will later only have:
500 Seats.
The traditional SaaS model will be under pressure.
Thus, software must shift from:
Per Seat
to:
Per Outcome.
Customers ultimately are unwilling to pay more and more money for:
"Using software"
They are willing to pay for:
"Completing work"
────────────────
34. The truly strong software companies in the future will evolve from System of Record to System of Action
Traditional Salesforce:
Records sales.
In the future:
AI directly finds customers;
Sends emails;
Schedules meetings;
Updates CRM;
Tracks transactions.
Traditional ServiceNow:
Records work orders.
In the future:
AI directly resolves work orders.
Traditional ERP:
Records business.
In the future:
Agents directly execute procurement and financial processes.
So the biggest opportunity for software in the future is not:
Store Information.
But rather:
Act on Information.
────────────────
Thirty-five, why do many traditional software companies still have huge advantages?
Because AI needs the most:
Context.
Where is the enterprise Context hidden?
CRM;
ERP;
Financial systems;
Email;
Slack;
HR systems;
Customer databases.
In other words:
AI is the brain.
The System of Record is:
Enterprise memory.
So the real danger is not all traditional software.
The danger is:
No proprietary data;
No core workflows;
No high Switching Cost;
Just software that provides a UI.
Such products are most easily consumed by Agents.
────────────────
Thirty-six, why is Blackstone buying AI while also buying "things that AI can never fully replace"?
This is a very advanced capital allocation logic behind this speech.
If Blackstone truly believes AI is the future,
why not invest all the money in AI?
Because truly large asset management institutions never have only one script.
On one side, buying:
AI Infrastructure.
On the other side, buying:
Airports;
Sports assets;
Scarce real estate;
Consumer experiences.
Why?
Because the value of these assets comes from:
Scarcity.
────────────────
Thirty-seven, one of the biggest economic consequences of AI is making "replicable things" cheaper and cheaper.
AI excels at:
Replicating knowledge;
Replicating text;
Replicating images;
Replicating code;
Replicating digital products.
Thus:
The supply of digitizable things
explodes.
Economics is very simple:
Supply explosion → Price decline.
So where will the money go?
To:
Things that cannot be replicated.
────────────────
Thirty-eight, the most expensive things in the AI era may ironically be "the scarcity of the real world."
For example:
Land in Manhattan;
Beaches;
Airports;
Sports leagues;
Exclusive licenses;
Access to quality power grids;
Data center locations;
Housing in popular cities;
Global brands.
AI can generate:
1 billion beach images.
But it cannot generate:
A second Malibu coastline.
AI can generate:
10,000 cricket match simulations.
But cannot replicate:
The emotional and cultural networks of real sports leagues.
So as AI gets stronger,
some real-world scarce assets may actually become more expensive.
This is very important:
Digital Abundance → Physical Scarcity Premium.
Unlimited supply in the digital world.
Increased scarcity premium in the physical world.
────────────────
Thirty-nine, this explains Blackstone's "dual-barbell" strategy.
On one end:
Owning the data centers, energy,
computing power,
and infrastructure that AI must rely on.
On the other end:
Owning things that are hard to replicate with AI:
Airports;
Land;
Experiences;
Brands;
Sports.
The most dangerous in the middle is:
Ordinary businesses that are easily commoditized by AI
and have no physical scarcity.
This is a very mature asset allocation mindset.
────────────────
Forty, the truly dangerous assets in the future are those that are "neither scarce nor own data."
For example, certain:
Ordinary information services;
General consulting;
Low-value-added software;
Intermediaries;
Basic content production.
In the past, these industries relied on:
Information asymmetry
to make money.
AI is bringing information costs down to nearly zero.
So their economic rents will constantly be under pressure.
────────────────
Forty-one, the impact of AI on Private Equity may even be greater than on VC.
VC invests in AI startups.
This is explicit.
The changes in PE are more interesting.
PE owns mature companies.
AI can directly affect:
EBITDA.
For example:
A company:
Revenue = $1B.
EBITDA Margin = 20%.
EBITDA = $200M.
AI reduces costs by:
$30M.
Then EBITDA becomes:
$230M.
If the market values:
10× EBITDA.
The enterprise value directly increases by:
$300M.
This is why Blackstone is so actively promoting AI deployment in portfolio companies.
────────────────
Forty-two, PE will increasingly resemble an "industrialized AI transformation company."
The classic playbook of Private Equity in the past:
Cut costs;
Change management;
Mergers and acquisitions;
Optimize capital structure;
Improve sales.
In the future, it will also include:
AI Transformation.
If a PE platform has 100 companies,
and each company improves its margin by:
100–300bps through AI,
the combined value creation will be enormous.
So AI is not just a track for PE investment.
It will become:
PE Operating System.
────────────────
Forty-three, this may redefine what it means to be "the best private equity firm."
In the past, the advantages of excellent PE came from:
Deal Sourcing;
Financial Engineering;
Operational Expertise.
In the future, it may also include:
AI Deployment Capability.
Whoever can replicate the same set of:
Agents;
Data platforms;
Procurement capabilities;
Technical teams
across hundreds of companies,
will have a scale advantage.
Blackstone has already begun to build AI as a shared capability across its portfolio companies. Officially, its data science team has estimated that the cumulative bottom-line impact generated in collaboration with portfolio companies is about $200 million.
────────────────
Forty-four, but Blackstone's data must be viewed with an important bias.
This is what professional investors must do.
Blackstone tells you:
AI ROI is very high.
You cannot take that at face value.
Because Blackstone:
Owns AI infrastructure;
Invests in AI companies;
Finances AI projects;
Needs LPs to continue investing.
It naturally has:
Bullish Incentive.
This does not mean the data is false.
It means:
Investors must actively seek counter-evidence.
────────────────
Forty-five, truly mature investment research always asks, "Under what circumstances am I wrong?"
If Blackstone's AI Thesis goes wrong, there are probably five paths.
First: AI Revenue growth slows
Customers are no longer willing to continue increasing spending.
────────────────
Second: Model Efficiency grows too quickly
The compute needed for the same tasks drops significantly.
────────────────
Third: Supply suddenly releases in large quantities
GPUs;
HBM;
Data centers;
Electricity
all come online in large quantities.
Prices drop.
────────────────
Fourth: Application-side ROI is insufficient
Companies find:
AI looks good,
but cannot really reduce costs.
────────────────
Fifth: Financial leverage expands too quickly
The assets themselves are fine.
The financing structure has issues.
Historically, many crises have actually been of the fifth kind.
────────────────
Forty-six, why is "models becoming more efficient" both a benefit and a risk?
This is a very interesting paradox.
Model efficiency increases:
The cost per Token decreases.
On the surface:
Data center demand should decrease.
But there is a well-known phenomenon in economics:
Jevons Paradox.
When resource use efficiency improves,
because prices drop,
total demand may actually increase.
In the 19th century:
Steam engines became more coal-efficient.
As a result:
Total demand for coal skyrocketed.
AI may be similar.
If each Token becomes 10 times cheaper,
but humans use:
100 times more Tokens.
Ultimately, compute demand:
still grows 10 times.
So:
Model efficiency improvement
do not automatically mean a decrease in compute demand.
The key is:
Price Elasticity of Demand.
────────────────
Forty-seven, this may be the most core economic variable for future AI infrastructure.
Assuming:
Inference Cost ↓ 90%.
If AI usage only increases:
2×,
compute demand decreases.
But if AI usage increases:
100×,
total compute demand still skyrockets.
So all AI data center investments,
are essentially betting:
AI demand elasticity is high enough.
Blackstone's entire Thesis is largely built on this point.
────────────────
Forty-eight, robots and autonomous driving may become the next round of Token Explosion.
Currently, the vast majority of Tokens:
come from humans.
Humans type;
AI responds.
In the future:
Machines will call models themselves.
A robot:
may perform multiple perceptions and judgments per second.
An autonomous vehicle:
continuously reads cameras;
radar;
maps;
environment.
Thus, compute users are no longer just:
5 billion people.
But may include:
Cars;
Robots;
Drones;
Factory equipment;
Cameras;
Smart systems.
So once Physical AI erupts,
compute demand may again change by an order of magnitude.
────────────────
Forty-nine, once AI truly enters the physical world, the market size will be redefined again.
Currently:
AI mainly replaces part of knowledge work.
In the future:
AI + Robot
will begin to enter:
Physical labor.
The global Labor Market:
is enormous.
Manufacturing;
Warehousing;
Transportation;
Construction;
Caregiving;
Agriculture.
Any industry where 10% of labor is repriced by intelligent machines,
could generate a market of hundreds of billions behind it.
So the true endgame of AI is not just:
Chatbots.
But may be:
Machines gaining economic agency.
Fifty, what the capital market is really betting on today is a very long chain.
It’s not simply betting that:
OpenAI will succeed.
But rather betting that:
Stronger models
→ Improved corporate ROI
→ Increased AI usage
→ Token growth
→ Compute growth
→ Data center growth
→ Power growth
→ Infrastructure CapEx growth
→ Realization of investment returns.
As long as one link breaks,
the entire valuation chain will be recalculated.
This is why AI is a huge opportunity,
but also a huge systemic risk.
────────────────
Fifty-one, the most important task for investors in the coming years is not to be "bullish or bearish on AI."
"AI Bull"
"AI Bear"
These labels are meaningless.
What should really be done is:
To find the best Risk/Reward positions across the entire value chain.
It may not be the hottest model.
It could be:
Power;
HBM;
Networks;
Private Credit;
Application Software;
Scarce assets.
The answers vary at different stages.
────────────────
Fifty-two, what Blackstone truly teaches us is how capital follows bottlenecks.
An excellent capital allocator never falls in love with a particular industry.
What they love is:
Return.
In 2023:
GPUs are the bottleneck.
In 2025:
Data centers become the bottleneck.
In 2026:
Power, equipment, and capital begin to become bottlenecks.
In the future:
It could be robotic data;
Licensing;
Physical labor deployment.
Money always moves towards:
Scarce links.
This is the real core of industrial investment.
────────────────
Fifty-three, understanding "Where’s the Beef?" ultimately returns to the most basic principles of capitalism.
All great stories must ultimately return to:
Cash Flow.
AI can change the world.
But:
If a company has no cash flow,
capital will eventually leave.
Data centers can be very sought after.
But if:
Rental returns are lower than financing costs,
it is still a bad investment.
OpenAI can have the best model.
But if:
The cost of computation is long-term higher than what customers are willing to pay,
the business model still has issues.
So truly professional capital always asks:
Where is the revenue?
Where is the profit?
Where is the cash?
Where is the return?
────────────────
Fifty-four, this is why "Where’s the Beef?" is actually one of the most important questions of the entire AI era.
It’s not about:
How smart the model is.
It’s not about:
How many parameters there are.
It’s not about:
Who releases the benchmark first.
But rather:
How much economic value AI actually creates, and how much of that can be converted into real cash flow?
If the answer continues to expand,
what seems like exaggerated CapEx today,
may appear very cheap in ten years.
If the answer does not expand,
today’s trillion-dollar investment
will ultimately become:
Another version of the:
Railroad bubble;
Telecom bubble;
Fiber optic bubble.
────────────────
Fifty-five, what should ordinary entrepreneurs really learn?
First:
Don’t sell AI, sell ROI.
Customers don’t care what model you used.
Customers care about:
How much money they save.
────────────────
Second:
Look for high-cost, repetitive, quantifiable workflows.
Such scenarios are most likely to generate real returns.
────────────────
Third:
Don’t just be an AI wrapper.
As models continue to upgrade,
pure functionality can easily disappear.
The real barriers come from:
Customers;
Workflows;
Data;
Brand;
Distribution.
────────────────
Fourth:
Products must enter production.
Demos have no value.
Only when integrated into the core systems of customers,
does it have value.
────────────────
Fifth:
Enable customers to calculate.
The best sales language is not:
"We used the latest model."
But rather:
Invest $1,
generate $5 in a year.
That’s what business is.
────────────────
Fifty-six, what should investors really learn?
Don’t just look at:
NVIDIA.
Look down the industry chain.
After AI usage increases:
What is most scarce?
Who has pricing power?
Who has long-term contracts?
Who can finance safely with debt?
Whose assets are irreplaceable?
Whose returns depend on a risky assumption?
This is true industrial investment.
────────────────
Fifty-seven, Blackstone’s greatest wisdom: simultaneously betting on "technological expansion" and "real scarcity."
If I were to compress Jon Gray’s capital allocation logic into one sentence, I would say:
AI will make the digital world cheaper, but will make the physical scarce resources that support the digital world increasingly important.
Thus, Blackstone buys on both sides.
On one side:
Compute.
On the other side:
Scarcity.
Data centers;
Energy;
Chips;
Networks.
And also:
Airports;
Sports;
Land;
Experiences;
Brands.
This is not a contradiction.
But rather two sides of the same AI era.
────────────────
Fifty-eight, the real big money is ultimately made in "places everyone has to go through."
This is one of the most important rules in the entire history of industrial capital.
In the railroad era:
Railroad nodes.
In the oil era:
Pipelines and refining.
In the internet era:
Cloud computing and platforms.
In the mobile internet:
App Stores and advertising platforms.
In the AI era, the most profitable assets
are likely to be:
Toll booths everyone has to pass through.
It could be:
GPUs;
HBM;
Cloud;
Power;
Data Centers;
Systems of Record;
or some future AI Agent platform.
What investors should really be looking for is not:
"What’s the coolest thing?"
But rather:
If the entire industry continues to expand, who owns the necessary pathways?
────────────────
The most memorable line
Jon Gray asks:
Where’s the Beef?
My answer can be pushed one step further:
The truly important question of the AI era is not how much the world is willing to invest in AI.
But rather:
For every dollar of capital consumed by AI, how many dollars of sustainable economic value can it actually create?
As long as this number continues to be greater than 1,
the capital cycle can continue:
Companies make money;
Increase AI spending;
Model companies increase revenue;
Continue to purchase computing power;
Data centers continue to be built;
Power continues to expand.
This forms:
The AI Economic Flywheel.
But one day:
If the growth rate of capital investment
long-term exceeds
the speed of economic value creation,
this flywheel will reverse.
At that time:
CapEx becomes inventory;
Inventory becomes discounted assets;
Debt becomes pressure;
High valuations become write-downs;
Bull markets turn into capital liquidation.
So a truly mature view on AI investment is neither:
"AI must be a bubble."
Nor is it:
"AI changes the world, so prices are all reasonable."
But rather always focuses on one of the oldest, simplest, and ultimately unavoidable questions of capitalism:
Did the money invested ultimately earn back?
This is what Jon Gray refers to:
Beef.
It is also the true North Star that all AI investors should continue to track in the coming years.
J