Ajay Banga wants you to sleep soundly. The World Bank president recently stepped in front of a microphone to deliver a comforting piece of conventional wisdom: capital availability is not really an issue, even as artificial intelligence infrastructure demands push borrowing costs upward.
It is a soothing narrative meant for central bankers and Davos panels. It is also completely detached from the physical reality of silicon, copper, and power generation.
I have spent the last two decades watching institutional capital misallocate billions into shiny narrative boxes while the actual unit economics of industrial scaling broke down. When leaders of global financial institutions claim that liquidity is infinite while interest expenses climb, they are confusing financial engineering with physical capacity.
Capital is not some abstract fluid that pools wherever a PowerPoint deck looks pretty. It has a cost, a gravity, and a threshold of pain. And right now, the infrastructure buildout required for modern machine learning is hitting a brick wall of debt service that no amount of optimism can wave away.
Let us dismantle the lazy consensus.
The Liquidity Delusion
The core argument from the multilateral banking establishment runs like this: global savings pools are deep, sovereign balance sheets are flexible, and private credit markets are overflowing with dry powder. Therefore, data center buildouts costing tens of billions of dollars each will find funding regardless of macroeconomic tightening.
This logic ignores how capital actually flows under high interest rates.
Imagine a scenario where a hyperscale cloud provider needs to raise fifteen billion dollars to finance a new cluster of training clusters. Six percent debt on fifteen billion dollars demands nine hundred million dollars in annual interest payments before a single token is generated or a single model is monetized.
When borrowing costs rise, the hurdle rate for projects changes. CFOs do not look at an expensive credit market and shrug because the World Bank says liquidity exists. They look at projected return on invested capital and they cut the fringe projects.
Banga is looking at aggregate global liquidity aggregates while ignoring credit transmission friction. There is plenty of money sitting in money market funds and sovereign wealth vaults. None of that capital is walking into high-risk compute clusters without demanding a risk premium that makes early-stage unit economics look mathematically impossible.
The Power Constraint Is the Real Balance Sheet
The mistake most analysts make is treating compute as a software problem wrapped in a financial package. It is not. Compute is a heavy industrial commodity.
You cannot run a cluster of advanced accelerators on good intentions and venture debt. You need dedicated substations, transformer units, cooling loops, and sustained gigawatts of base-load electricity.
When borrowing costs spike, the capital expenditure required to secure energy generation assets becomes the primary bottleneck. Utilities are not lending money to tech companies to build dedicated nuclear or gas-peaker plants out of charity. They demand long-term power purchase agreements backed by ironclad credit ratings.
If borrowing costs are elevated, the cost of capital for energy infrastructure scales non-linearly. We are not just facing expensive servers; we are facing expensive electrons. When the cost of funding the physical grid rises, the entire downstream intelligence ecosystem feels the squeeze.
The market has priced in exponential software growth while ignoring linear thermodynamic limits.
The Venture Capital Hangover
To understand why the current funding narrative is broken, look at how private markets treat enterprise software versus industrial AI infrastructure.
For a decade, tech companies operated in a zero-interest-rate environment where capital was essentially free. You could lose money on every transaction and make it up in volume because some venture fund would refill your coffers at a higher valuation the following year. That era is dead.
Today, every dollar of capital carries an explicit opportunity cost represented by risk-free yields hovering near five percent. When risk-free assets pay you to sit on your hands, venture capital and institutional debt demand actual, immediate cash flows.
Yet, the primary buyers of massive compute clusters are burning cash at unprecedented rates to train frontier models that face severe monetization friction. Enterprise adoption is real, but it is slow, methodical, and heavily integrated into existing legacy software stacks. It does not move at the speed of a venture capital burn rate.
When the financial ledger does not balance, the narrative pivots. Suddenly, we hear that capital is abundant, but the projects themselves are being throttled not by lack of money, but by lack of discipline. The truth is harsher: the money is expensive precisely because the risk of structural failure is too high for rational lenders to ignore.
What People Get Wrong About Compute Economics
People ask: "If demand for intelligence is infinite, why would capital ever dry up?"
This question contains a fundamental category error. Infinite demand does not imply infinite willingness to fund unprofitable delivery mechanisms.
If it costs three dollars in power, amortization, and debt service to deliver two dollars of enterprise value, scaling that operation globally does not create an empire; it creates a bankruptcy.
The financial intermediaries cheering on this infrastructure cycle are incentivized to keep the music playing because advisory fees, debt underwriting commissions, and cloud credits generate revenue regardless of whether the end-user achieves profitability.
We saw this movie play out during the fiber-optic boom of the late nineteen-nineties. Billions of dollars of capital were poured into laying dark fiber across oceans and continents because the demand for data was obviously going to explode. And the demand did explode—surpassing even the most aggressive projections.
Yet, most of the companies that laid the cable went bankrupt because the capital costs outpaced near-term pricing power. The infrastructure ended up being pennies on the dollar for the next wave of operators.
We are hurtling toward the exact same reckoning in machine learning hardware.
Unconventional Strategy for the Next Credit Cycle
If you are running a technology enterprise, a venture fund, or an infrastructure fund right now, you need to stop listening to global financial officials who view markets from thirty thousand feet.
First, purge long-duration debt assumptions from your financial models. Assume that capital will remain expensive, selective, and demanding of near-term cash generation.
Second, shift your focus from training frontier base models to vertical inference optimization. The real margin is not in building the biggest possible brain; it is in making specialized models run on cheap, existing hardware with minimal energy overhead.
Third, treat energy procurement as your primary financial strategy. If you do not lock in power purchase agreements with predictable, long-term pricing structures, interest rate fluctuations will destroy your operating margins before your competitors ever lay a glove on you.
The era of cheap, speculative capital funding infinite compute experiments is over. Pretending otherwise is a luxury reserved for bureaucrats who never have to make payroll when the debt comes due.
Stop waiting for the liquidity cavalry to save your unit economics. Build something that actually makes money today.