$1 trillion. That is the number Big Tech has committed to artificial intelligence infrastructure โ data centers, GPU clusters, power delivery, and the industrial machinery to support it โ spread over roughly two years. As a share of the U.S. economy, that is around 3.6% of annual GDP. That is not a corporate line item. That is a systemic demand shock wearing a quarterly earnings call.
The Federal Reserve is currently asking whether this spending is inflationary. That is the wrong question. The right question is whether the Fed's transmission mechanism can reach this investment cycle at all โ and the answer, based on decades of evidence about who actually borrows at the margin, is likely no.
I have spent 28 years auditing systems that claim to be stable. I traced transaction hashes across the Ethereum Classic blockchain after the 51% attack in 2017, when 'community governance' turned out to be a phrase used to paper over a $3.6 million structural failure. I reverse-engineered OlympusDAO's bonding contracts in 2021 and found the recursive yield loop that inevitably drained liquidity. I watched Terra's UST arbitrage engine in 2022 fail not because of bad actors but because its geometry was mathematically incapable of holding under stress. In every case, the narrative led. The code followed. The collapse came third. The macro economy is no different. The models lag. The incentives diverge. The data arrives late and is interpreted later.
The situation, in brief: Microsoft, Google, Meta, Amazon, and a handful of others are deploying a combined trillion dollars into compute infrastructure. OpenAI and Anthropic are effectively capped-out purchasers of that compute. The Fed, meanwhile, sits in a 'wait and see' posture, hoping the inflation wave from 2021-2023 is genuinely finished.
Enter the political layer. The administration wants rate cuts, tax cuts, and record equity prices โ simultaneously. The fiscal deficit sits at levels that would historically require higher long-end rates. And the private sector is stepping on the accelerator exactly when the public sector should be braking. This is the macro equivalent of a smart contract with an owner-only emergency pause: technically, the tools exist. Practically, nobody wants to call the function.
For crypto, the stakes are existential. I have said it before and I am saying it again: I measure risk in gas units, not in hope. The gas units of the U.S. macro economy are electricity prices, copper prices, and transformer lead times. When those move, every risk asset moves โ including assets whose entire promise is independence from the fiat system. Bitcoin trades as a duration asset whether its holders like it or not. Stablecoin treasuries are a bet on the Fed funds rate. DeFi leverage is volatility around real-rate expectations. The AI capex cycle is now the principal driver of those expectations, and the market has not priced the mechanics. That is what this analysis is for: a pre-mortem on an investment cycle that has already, structurally, begun to show its cracks.
Failure mode one: The patient cannot feel the medicine
The textbook rationale for higher interest rates is that borrowing costs rise, marginal investment projects go negative, and aggregate demand cools. That chain works for mortgage borrowers, small business owners, and leveraged industrial firms. It does not work on a set of companies that sit on tens of billions in net cash, generate quarterly free cash flow in the tens of billions, and can issue investment-grade debt at rates their own return on invested capital dwarfs by an order of magnitude.
The four largest spenders in this cycle โ Microsoft, Alphabet, Amazon, Meta โ hold far more cash and marketable securities than debt. When the Fed holds rates high, their internal cost of capital barely registers the change. Their investment decisions are driven by competitive positioning and fear of disintermediation, not by the discount rate. This is the rate-insensitivity problem.
The Fed is being asked to cool a boom it cannot reach. Its primary tool acts on credit demand at the margin, but the marginal borrower in this boom is the sovereign balance sheet, not the commercial lender. The result is that monetary policy transmits to the weakest parts of the economy โ housing, small business, consumer credit โ while the strongest part, Big Tech's capex machine, simply continues. That produces an asymmetric slowdown: rate-sensitive sectors contract while the AI sector keeps inflating the demand base. For a central bank, that is exactly the wrong mix.
Failure mode two: The inflation is physical
Let me break the transmission into channels.
Channel one: electricity. AI data centers consume somewhere between 2% and 3% of total U.S. electricity today and are projected to reach 8% to 10% by 2030. That is not a trend. That is a tripling of a major demand segment in under a decade, layered on a grid that has been under-invested since the 1970s. Utilities are already raising rates in data-center corridors. Northern Virginia, Texas, Ohio, and the Pacific Northwest are seeing regional energy cost shifts that feed directly into consumer prices. This is a documented line item in utility earnings reports, not a forecast.
Channel two: capital goods. The AI boom is not software consuming nothing. It is steel, copper, aluminum, transformers, substations, and construction labor. Chip fabs cost between ten and twenty billion dollars each. Data centers cost hundreds of millions to billions. The bidding pressure on these physical inputs goes straight into producer prices and spreads through industrial supply chains. If you want a leading indicator, look at transformer lead times: what used to be a standard procurement item now has a multi-year waitlist. That is physical inflation.
Channel three: labor. AI engineers are the most expensive labor category on the planet. The wage signals spill into adjacent technical disciplines, and from there into regional housing, services, and everything those workers consume. Standard wage-push inflation, amplified by winner-take-all labor markets.
Channel four, and the one most forecasts miss: the power grid itself. You cannot build thousands of megawatts of compute demand without building new transmission lines, substations, and generation. The U.S. needs trillions in grid investment just to support electrification and AI loads combined. That construction program is itself inflationary, and it is not in anyone's near-term CPI forecast because it is not in anyone's leading indicators. It is coming. It is structural. And it operates on a multi-year lag.
Failure mode three: The time-lag trap
This is where the Fed's dilemma becomes structurally dangerous rather than merely uncomfortable. AI investment creates inflation now and deflation later โ if it works. The inflation side is immediate: construction, energy, equipment, all of it hits this quarter and next. The deflation side is delayed: automation-driven productivity, cheaper marginal production, lower costs across service industries. That payoff arrives in five to ten years, not in the Fed's two-year projection window.
The dual mandate is set up to look 12 to 24 months ahead. It cannot price a five-year productivity breakthrough. It sees near-term inflation and tightens. But tightening now makes the interim more painful while the long-term gains arrive too late to affect today's decisions.
This is the geometry of Terra, scaled to the macro economy. The mechanism looked rational from one side of the trade and broke catastrophically when time horizons diverged. The arbitrage worked until it didn't, and the reserve assumption turned out to be the single point of failure. For the Fed, the single point of failure is the Phillips curve itself โ an empirical regularity built on decades of data that contains no epochal investment shifts. The model has not seen this input before. The output will be wrong.
The most dangerous scenario is the 1960s rerun: a Fed so focused on supporting growth that it tolerates rising inflation until the pattern is entrenched, then over-corrects hard enough to trigger a layoff cycle that kills both consumption and the productivity story. Chaos is just data waiting to be compiled โ but the compilation is happening in real time, without a debugger.
Failure mode four: The political overlay and the debt spiral
The policy mix in Washington is contradictory on its face. Fiscal policy is expansionary: the deficit remains above $1.5 trillion on a base of roughly $36 trillion in federal debt, with tax cuts still on the agenda. Private sector capex is a separate expansionary force, adding demand on top. If the labor market is near full employment, the combination is straightforwardly inflationary.
The Fed then has to fight the Treasury's fiscal stance with interest rates. Higher rates raise the cost of servicing existing debt. That raises the deficit. That requires more issuance. More issuance pushes long-end rates up. The fiscal-monetary feedback loop tightens even if the Fed never moves its own policy rate again.
Political pressure to cut rates makes it worse. A Fed that caves loses credibility, and a Fed that loses credibility sees inflation expectations de-anchor. The moment the market prices a politically captured Fed, the term premium rises, and tightening happens against the Fed's stated intentions. We are not there yet. But the path is visible, and the infrastructure along that path is being built now.
Failure mode five: The balance-sheet bridge
The under-appreciated risk is the corporate credit accumulation underneath the boom. The same rate-insensitivity that makes Big Tech's capex unstoppable also lets it borrow more than it needs. Investment-grade technology issuance has marched upward for three consecutive years, financing buybacks, cloud expansions, and GPU purchases. Debt service is manageable now because cash flows are strong. But the entire proposition rests on a single assumption: that AI investment will generate returns above the cost of capital.
I have run this analysis at protocol level more times than I can count. The coupon is affordable until the day the revenue does not materialize. In crypto, the failure mode is called a death spiral: yield stays attractive, then the exit liquidity drains, then the price of the underlying collapses, then even solvent participants get caught. In the corporate credit market, the sequence is slower but no less mechanical. If a $10 billion data center generates $500 million of incremental annual revenue, the payback period is two decades. If utilization rates disappoint even modestly, those assets become stranded capital on leveraged balance sheets.
The rating agencies are not modeling this because their historical default data contains no AI-fab-capacity-overbuild scenario. The credit markets are not pricing it because bond demand is still chasing the same narrative as the equity market. That is precisely when the risk is cheapest to acquire and most expensive to carry.
What this means for crypto, specifically
Higher-for-longer is the base case. That is a hostile environment for duration assets โ which, in practice, includes nearly every speculative cryptocurrency. Assets with real cash flow compress less. Stablecoins, whose yield tracks short-term rates, benefit. The speculative middle of the market โ tokens with no earnings, no users, no structural demand โ gets repriced downward as the risk-free rate stays elevated. I have been through five cycles. This is structural, not editorial.
There is also a physical channel. AI data centers and Bitcoin miners compete for the same electricity. When AI demand outbids miners for power, the marginal mining operation goes negative. I have evaluated mining contracts where the power agreement was 80% of the risk profile. If AI keeps pushing energy prices up, that risk becomes structural. Miners get mined out of the grid by a bigger machine. The code doesn't lie. Neither do power bills.
Longer term, if AI delivers its productivity miracle, rates eventually fall, and that is a massive tailwind for every risk asset on the planet. If it does not, the over-investment collapses into the classic bust: capacity gluts, debt defaults, capital destruction. Both paths are volatile. The asymmetry is not obvious, and anyone who tells you they know which path the data will take is selling a narrative, not an analysis.
Contrarian: What the inflation hawks are missing
Now the part I stress-test myself on.
Every general-purpose technology in recorded history has been deflationary on net. Steam, electricity, the internal combustion engine, the internet โ each displaced labor and lowered the real cost of everything. AI is already doing this in code generation, customer service, logistics, and drug discovery. The inflation we are arguing about may be just the investment phase of a fundamentally deflationary technology.
Second, the capacity glut argument cuts both ways. The 2000 telecom bubble created a fiber network so oversupplied that bandwidth prices collapsed for two decades โ which enabled streaming, cloud, and every subsequent digital industry. If AI capacity overshoots by 2028, compute prices collapse, and intelligence-as-a-service becomes nearly free. That is profoundly deflationary.
Third, the productivity data may arrive faster than the models expect. There are early signals in U.S. labor productivity statistics that AI is moving operating costs, not just capex. If enough of those signals hit before the Fed over-tightens, the central bank gets cover to cut.
My own 2026 work on the AI-agent permit exploit cuts in a third direction. The exploit proved that autonomous agents lack contextual understanding, which means the productivity gains are slower and more constrained than the bulls claim. That supports the inflationary read โ but it also means the bubble bursts sooner. Both readings lead to volatility. That is the conclusion that matters.
Takeaway
Watch the data the models are not watching. Electricity prices per kilowatt-hour in data-center corridors. Transformer lead times. Tech capex guidance revisions, quarter over quarter. Five-year breakeven inflation. Those tell you which part of the geometry is loading stress before the headline CPI does.
The fork was inevitable; the error was optional. The Fed has modeled the past and priced the present. The next year will tell whether it understands the structural rewrite happening under its feet. Hold assets with real cash flow. Audit every yield claim. And remember: I measure risk in gas units, not in hope.
The code doesn't lie. It was never the code that got us here. It was the assumption that the model was right.