The economic event people keep predicting is already scheduled. On September 16, 2026, the Federal Reserve announces a rate decision that futures markets price at a 92.5% chance of a hike. It lands on top of the most debt financed corporate capital expenditure cycle on record. AI spending stopped being funded out of cash flow this year. That is the actual structural change, and it is measurable.
Summary
- Five hyperscalers are on track for roughly $660 to $690 billion of capex in 2026, against about $380 billion in 2025.
- AI linked bond issuance hit $344 billion by early August 2026. All of 2025 was $182 billion.
- Oracle now sits at BBB minus, the lowest investment grade rung. Alphabet posted negative free cash flow for the first time since its IPO.
- Markets price a 92.5% probability the Fed raises rates on September 16, up from 69.4% the previous Friday.
- The depreciation bear case is not confirmed by the rental market. A100 rental prices are up about 17% year over year.
- Crypto is not moving with this. Stablecoins are flat at $302.9 billion, tokenized real world assets flat at $38.86 billion.
- AI and crypto are not one trade. One is consuming credit at record scale. The other is a settlement upgrade that is currently sideways.
Table Of Contents
- The Event Is On The Calendar
- What Actually Changed: Cash To Credit
- The Depreciation Argument And Why The Rental Market Disagrees
- Meanwhile, Crypto Is Flat
- What Is Actually Being Reshaped
- Measurable Versus Narrative
- My Last Call Is Being Tested Right Now
- What I Am Watching
- Frequently Asked Questions
- The Bottom Line
The Event Is On The Calendar
There is a genre of prediction that cannot be wrong, because it never says when or what. “A huge economic event is coming” belongs to it. So does most of what gets posted about AI and crypto reshaping everything.
Here is the version you can check.
The federal funds target has sat at 3.50% to 3.75% since December 2025. The Federal Open Market Committee announces its decision on Wednesday, September 16, 2026 at 2:00 PM Eastern. As of September 15, the CME FedWatch tool put the probability of a move to 3.75% to 4.00% at 92.5%. That number was 86.5% the day before and 69.4% the previous Friday. The market repriced hard, fast, and in one direction.
The reason is the chair. Kevin Warsh was sworn in on May 22, 2026. At Jackson Hole on August 28 he pointed at twelve month PCE inflation of 3.7% and called progress “modest,” saying the committee needs to see inflation moving to target “clearly and at sufficient speed.” July headline CPI was 3.4% year over year with core at 2.5%. August payrolls came in at 162,000 with unemployment at 4.1% and average hourly earnings up 0.3% on the month. That is not an economy asking for relief.
So the event is a hike, not a crash, and it is tomorrow. What makes it matter is what it is landing on.
What Actually Changed: Cash To Credit
For most of the last decade, big technology capex was a rounding error against operating cash flow. The companies building data centres were the most cash generative businesses in history and they paid for their buildings out of pocket. That is no longer true, and 2026 is the year it stopped being true.
| Company | 2026 Capex | 2025 Capex | Change |
|---|---|---|---|
| Amazon | $200B | ~$147B | +36% |
| Alphabet | $175B to $185B | $71B to $73B | +139% to +160% |
| Meta | $115B to $135B | Not disclosed | Sharp increase |
| Microsoft | $120B+ | Not disclosed | Accelerating |
| Oracle | $50B | ~$21B | +136% |
| Combined | $660B to $690B | ~$380B | +73% to +82% |
Now the financing side, which is the part that gets less attention and carries the risk.
- Data centre related debt issuance in all of 2025 was $182 billion, per S&P Global.
- AI infrastructure bond issuance reached $344 billion by early August 2026, with Bank of America Global Research projecting roughly another $100 billion before year end.
- In August 2026, Nvidia announced a $500 billion financing package with six major asset managers, structuring chip purchases through asset backed securities.
Read that last one twice. Graphics processors are being securitised. The collateral is a depreciating semiconductor with a contested useful life. That is not a moral failure, it is just a structure, and structures have failure modes worth naming.
The credit market has already started pricing it. Oracle was downgraded to BBB minus, one notch above junk, with five year credit default swaps around 215 basis points. Nvidia’s own five year CDS hit a record 82 basis points on July 27, 2026. Alphabet posted negative free cash flow for the first time since it went public. The IMF and the Bank for International Settlements have both flagged AI related circular financing as a systemic downside risk.
Circular financing is the part that should make an operator uncomfortable, because it is a pattern you can recognise from much smaller businesses. Nvidia has committed up to $100 billion to OpenAI, roughly $30 billion of it deployed, tied to OpenAI buying Nvidia systems. Microsoft has put over $13 billion into OpenAI, and OpenAI has committed to buying $250 billion of Microsoft cloud. Amazon has invested $15 billion with a further $35 billion pledged conditionally. AMD committed up to $5 billion to Anthropic in July 2026, contingent on Anthropic deploying two gigawatts of AMD Instinct hardware.
Every one of those may be a rational commercial deal. But vendor financed revenue is revenue that depends on the vendor continuing to finance it. Any founder who has ever extended terms to a customer so the customer could afford the invoice already knows the shape of this.
This is the actual transmission mechanism. When capex was cash funded, rates were close to irrelevant to the build out. Now that a meaningful share is credit funded, the cost of that credit is an input. A hiking Fed does not stop a data centre being built. It changes what the last tranche costs, which changes which projects clear, which shows up in spreads long before it shows up in a share price.
The Depreciation Argument And Why The Rental Market Disagrees
The loudest bear case is an accounting one. Michael Burry has argued that Meta, Amazon, Microsoft, Google and Oracle are depreciating Nvidia hardware over five to six years when the real economic life is closer to two to three, and that this produces roughly $176 billion of understated depreciation and overstated profit between 2026 and 2028.
The mechanics are real. Amazon shortened server useful life from six years to five in early 2025, citing the pace of AI development. Meta moved to 5.5 years. Both Microsoft and Google had previously extended to six years around 2023 and 2024. If Burry is right about the true life, a large impairment is arithmetic, not opinion.
Here is the problem with the argument, and I think it is a serious one.
If a GPU were economically dead at three years, five year old hardware would be close to unrentable. The opposite is happening.
| GPU | Rental Price, September 2026 | Direction |
|---|---|---|
| A100 | $1.65 per hour | Up about 17% year over year |
| H100 | $3.38 per hour median | Down from $7+ in early 2024, but rose through 2026 |
| H200 | $3.59 to $4.59 per hour | Current generation pricing |
| B200 | $5.63 to $5.66 per hour | Newest, spot |
The A100 launched in 2020. It is renting for more today than it was a year ago. H100 pricing rose through most of 2026 even after H200 and Blackwell parts were available. Google says seven and eight year old TPUs run at full utilisation.
The explanation offered is a memory supply crunch making newer silicon relatively expensive, which props up older silicon. That is a real caveat, because it means the price strength is partly a supply artefact rather than proof of durable demand. But it cuts both ways: an artefact that has now persisted for over a year is not something you can wave away while claiming certainty about a two year life.
So the honest read is that the depreciation bear case is currently unproven and the burden is on it. Second hand asset prices are the cleanest available test of an asset’s economic life, and right now they are pointing the other way. That does not make it wrong. It makes it a thesis waiting on evidence, and you should treat anyone stating it as settled fact accordingly.
Meanwhile, Crypto Is Flat
The standard version of this argument bundles AI and crypto into a single civilisational shift. The data does not support the bundle.
As of September 15, 2026, the total stablecoin market capitalisation is $302.9 billion, up 0.09% over seven days. Tether is $183.3 billion of that at 60.52% share, USDC is $74.1 billion at 24.46%. Tokenized real world assets stand at $38.86 billion across roughly 4.24 million holders, up 1.00% over thirty days.
| Category | On Chain Value |
|---|---|
| US Treasuries | $15.9B |
| Commodities | $4.9B |
| Active strategies | $3.6B |
| Asset backed credit | $2.56B |
| Tokenized stocks | $2.52B |
Two observations that matter more than the totals.
First, growth has stalled. Stablecoins moved 0.09% in a week. Tokenized assets moved 1.00% in a month. Whatever is reshaping capital markets this quarter, it is not on chain volume, and I say that having written the bull case for both stablecoin adoption and on chain listings recently. The infrastructure argument still holds. The near term growth argument does not, right now.
Second, Pantera Capital’s analysis found 77.6% of tokenized assets sit at the “wrapper” level, meaning they are digital representations of off chain instruments rather than natively issued on chain. A wrapper is a distribution improvement. It is not a change in what the asset is or who is liable for it. Most of the $38.86 billion is a better delivery mechanism for something that already existed.
And prices are behaving like a risk asset ahead of a hike, not like a hedge against one. Bitcoin is near $76,000 as this publishes, down about 3.2% on the day and 2.6% on the week, with a 24 hour range of $75,595 to $79,530 and a market cap around $1.526 trillion. It remains roughly 40% below the October 6, 2025 high of $126,080. Ethereum is near $2,480. Both were rallying last week as AI warnings hit equities, then sold off into the Fed meeting alongside everything else.
That sequence is the tell. For a few days crypto traded like an AI hedge. The moment a rate decision came into view it traded like a long duration risk asset, which is what it has been for years.
What Is Actually Being Reshaped
Strip out the narrative and two real things are happening. They are unrelated, and conflating them is what makes most commentary on this useless.
One: AI Turned Into A Credit Event Waiting For A Catalyst
Not a bubble claim. A funding claim. The build out has migrated from equity and cash onto balance sheets and securitisation vehicles. That migration is documented in issuance volumes, credit ratings and CDS spreads. It means the AI trade now has an interest rate sensitivity it did not have eighteen months ago, and that the first damage, if damage comes, appears in credit rather than in headlines about chatbots.
Two: Settlement Is Being Rebuilt Quietly And Slowly
Three hundred billion dollars of stablecoins is a real payment rail regardless of this week’s growth rate. Tokenized treasuries at $15.9 billion are a real product. But this is plumbing, and plumbing compounds on a decade clock, not a quarter clock. It does not need a catastrophe to justify it and it does not produce one.
The mistake almost everyone makes is treating the second as evidence for the first, or the first as a reason to be positioned in the second. They run on different timescales and different mechanisms.
Measurable Versus Narrative
A claim you cannot check is entertainment. Here is the sorting.
| Claim | Status | How You Check It |
|---|---|---|
| AI capex is increasingly debt funded | Measurable, confirmed | Bond issuance volumes, CDS spreads, credit ratings |
| Rates are rising into that funding need | Measurable, pending | FOMC decision, September 16 |
| GPUs depreciate faster than booked | Measurable, not confirmed | Second hand and rental prices |
| Stablecoins are becoming core payment rails | Measurable, slow | Market cap and settlement volume |
| Tokenization is transforming markets | Partly narrative | Native issuance share versus wrappers |
| A huge economic event is coming | Not measurable | Cannot be checked, so cannot be useful |
My Last Call Is Being Tested Right Now
On September 7 I published a constructive case on Bitcoin and Solana. I refused to call the market primed, and I named the level that would break the thesis: a weekly close below $78,700.
Bitcoin is currently trading below that level. The weekly close has not happened yet, so the thesis is not dead, it is under test, and Sunday decides it. If it closes the week below $78,700 I was wrong on that call and I will say so in plain language rather than moving the level.
I am including this because it is the whole point of writing down an invalidation in the first place. A forecast with no exit condition is not analysis, it is identity. The people who told you an event was coming will tell you they were right regardless of what happens, because they never specified what would prove them wrong.
What I Am Watching
- The Fed decision, September 16, 2:00 PM Eastern. A hold against 92.5% odds would be a larger surprise than the hike.
- Oracle’s CDS spread. At 215 basis points and BBB minus, it is the most leveraged pure play on the thesis. Widening from here is the early signal.
- Hyperscaler free cash flow next quarter. Alphabet going negative once is a data point. Two in a row is a pattern, and a second name joining it is a trend.
- A100 and H100 rental prices. If these roll over meaningfully, the depreciation bear case stops being unproven.
- Native versus wrapped issuance on chain. That 77.6% wrapper share falling is the only honest measure of tokenization maturing.
- Bitcoin’s weekly close. $78,700, per my own published level.
What I am deliberately not doing is assigning a probability to a crash. I do not have one, nobody credible has one, and manufacturing a number to sound rigorous is the failure mode this entire post is arguing against.
Frequently Asked Questions
Is the AI boom a bubble?
That question is usually unanswerable because “bubble” is not defined. The checkable version is narrower: is the spending funded by cash flow or by credit, and is the collateral holding value? The answer in 2026 is that funding has shifted substantially to credit, with AI linked bond issuance at $344 billion by early August against $182 billion for all of 2025, while collateral values measured by GPU rental prices are holding or rising. That is a leverage story with an intact asset base, which is a different risk profile than a classic bubble.
Will a Fed rate hike crash AI stocks?
Not mechanically, and equity prices are the wrong place to look first. A higher policy rate raises the cost of the debt funding the data centre build out, which shows up in credit spreads and in which projects get approved. Credit markets typically reprice before equities do. Watching Oracle’s credit default swap spread will tell you more than watching Nvidia’s share price.
Are AI and crypto the same trade?
No, and the September 2026 data makes that clear. Hyperscaler capex is growing about 80% year over year while stablecoin supply grew 0.09% in a week and tokenized real world assets grew 1.00% in a month. Crypto did briefly trade as an AI hedge during the mid September equity selloff, then sold off into the Fed meeting like any other long duration risk asset. Different mechanisms, different timescales.
What would prove the depreciation bear case right?
A sustained decline in rental and resale prices for one and two generation old hardware, specifically A100 and H100. Right now A100 rentals are up roughly 17% year over year at about $1.65 an hour and H100 pricing rose through 2026, which is the opposite of what a two to three year economic life would produce. If those prices roll over while supply constraints ease, Michael Burry’s $176 billion figure becomes a live risk rather than a contested projection.
The Bottom Line
AI is reshaping the economy, but not in the way the phrase usually implies. It is not doing it through productivity statistics or job displacement headlines. It is doing it by becoming the largest new borrower in corporate credit markets, at the exact moment a new Fed chair appears willing to raise rates into an inflation print that will not fully cooperate.
Crypto is reshaping something too, and it is slower and less dramatic: settlement, custody and distribution. Three hundred billion in stablecoins is a real rail. It is also flat this quarter, and pretending otherwise to make the story bigger is how you end up with readers who stop trusting you.
The operator lesson generalises past markets. A prediction without an invalidation condition is not a prediction, it is a personality. If you are building a business on a thesis, write down now what would prove you wrong, and put a date on it. Then check. Most people never do, which is why most people cannot tell the difference between being right and being early and being lucky.
This is analysis, not financial advice. I am not a licensed financial advisor. Every figure here is sourced and dated, market data moves fast, and you should verify anything before acting on it.
If you build systems and want to think about them the way this post thinks about markets, with named assumptions and stated failure conditions, that is most of what I do. More on how I work.
