RISK ALERT — READ THIS BEFORE YOU ADD TO ANY COMPUTE POSITION: The Bank for International Settlements has put a number on the thing the sell-side will not say out loud. 55.2% of AI funding comes from other AI firms. Not pension capital. Not sovereign wealth. Not retail flows. Other AI firms. That is not a market pricing a future. That is a mirror financing its own reflection.
The chart lied. The story you have been sold says AI demand is broad, secular, and independently funded. The number says more than half of the money entering the sector originates inside the sector. I have spent twelve years watching capital markets dress circular flows in the costume of organic growth. In 2017 I audited whitepapers for fifty ICOs and found a re-entrancy bug in a contract hours before its mainnet launch — nine figures raised, one flaw, and almost nobody reading the code. In 2022 I traced FTX's eight billion across four chains in three connected threads. Every time, the pattern held. Follow the money until it either reaches a customer or reaches itself. This one reaches itself.
That is the hook. Now the forensics.
Context
The BIS is the central bank of central banks. It does not publish for clicks. When it names a specific financing structure — vendor financing, circular funding, intra-industry capital — it is not writing an academic footnote. It is drafting the regulatory agenda. Alpha moves before the charts confirm the truth, and a BIS naming is the kind of pre-chart signal that historically precedes disclosure requirements, leverage caps, and concentration limits. Read it as a precursor, not a postmortem. And note where this landed first: a crypto outlet. Crypto Briefing ran the story. That is not an accident. The crypto audience has spent a decade learning to smell a circular economy, because DeFi Summer was one, and the crypto crowd is the first to recognize the smell of a market paying itself. The AI trade has now acquired that smell, and the news reached the people most primed to detect it.
Now the mechanism, stripped of the marketing.
A chip designer invests in a model lab. The lab commits to buying that designer's silicon for three years. The designer books the commitment as revenue visibility. A hyperscaler invests in the same lab. The lab commits to renting that hyperscaler's compute. The hyperscaler books the commitment as backlog. The lab, now holding both fresh capital and booked demand, raises again — frequently from the same two counterparties — at a higher mark. The higher mark validates the previous round. The previous round funded the next procurement. Loop closed. Investment became procurement. Procurement became revenue. Revenue justified the round that funded the procurement.
That is vendor financing. It has a name because we have run the experiment before. Lucent and Nortel financed the competitive carriers that bought their switches. The equipment shipped, revenue was recognized, and when the carriers went bankrupt the vendors ate billions in write-downs on receivables that were never really external. The mechanism was identical. The only open question in 2000 — and the only open question now — is whether terminal demand can catch up to financed capacity before the financing matures. In 2000 it could not. The telecom buildout was financed against a demand curve that arrived a decade too late.
One more piece of context the source buried. The number arrived without a denominator. No report title, no sample definition, no time series, no baseline. In a normal market, that omission would itself be a red flag. In a bull market, it gets rounded to "AI is fine." I have watched this exact editorial behavior before — during DeFi Summer, when TVL was quoted without distinguishing recycled leverage from net deposits, and the number went up until it didn't. Data lies, but volume never cheats. A number without a denominator is a number without a denominator. Treat it as a structural signal, not a precise measurement. The structure is the story.
Core
Here is where the forensics matter. Strip the 55.2% down and you get a set of structural facts, each more uncomfortable than the last. I will walk them in order, because the order matters — each fact makes the next one worse.
Fact one: the number describes self-referential demand validation. When more than half of the capital entering an industry comes from incumbents inside that industry, the "demand" you are measuring is partly a function of the capital itself. Cloud credits sold to a lab you invested in. API resale between affiliated entities. Compute procurement booked as revenue by the vendor that funded the buyer. Model licensing between companies with overlapping cap tables. Every one of these transactions is real in the accounting sense and questionable in the economic sense. None of them requires fraud. That is what makes the pattern so durable — nobody has to break a rule for the loop to spin. The terminal consumer penetration rate — the share of revenue that ultimately traces to a non-AI customer paying cash — is the only metric that separates a business from a balance-sheet shuffle. Almost nobody in the sector reports it. That silence is the signal.
Consider what this does to the growth narrative. If 55% of the funding is internal, then the "AI adoption curve" quoted in every pitch deck is partly a curve the industry drew on itself. That does not mean the technology is fake. The models work. It means the revenue validating the technology is, in significant part, the industry paying itself to keep the validation going. A real product can still be financed by a circular flow. The circularity does not disprove the product. It just means the product's market validation is weaker than the headline growth implies, and the gap between the two is exactly where the risk lives.
Fact two: the money lands almost entirely in the compute layer. Follow the flow and it terminates in GPUs, high-bandwidth memory, optical interconnect, data center shells, and power. That is not diversification. That is concentration in the single most capital-hungry, fastest-depreciating layer of the entire stack. The circular capital is fuel for the compute supply chain, which means the compute supply chain is the epicenter of the fragility, not a safe harbor from it. When analysts tell you the "picks and shovels" trade is the conservative way to play AI, they are describing the exact asset class that carries the circular risk. The shovel-sellers are the ones holding the IOUs. They are also the ones with the longest lead times, the heaviest capex, and the thinnest ability to stop — a fab or a data center shell does not pause on a whim. The supply chain's order visibility depends entirely on the loop continuing, and the loop depends on the supply chain booking the orders. Mutual hostage-taking, dressed as a supply chain.
Fact three: the collateral is depreciating while the debt is not. A growing share of this buildout is financed against GPU and data center assets. GPUs are close to the worst possible collateral for a multi-year loan. They depreciate in three to five years, they lose value the moment a faster generation ships, and their resale market is thin, correlated, and dominated by the same buyers who are already overextended. A data center shell has a longer physical life but a far narrower set of tenants — you build for a hyperscale anchor or you do not build. Pair a fast-depreciating asset with a slow-maturing liability and you have a textbook duration mismatch. In a rising-demand environment, nobody notices. In a flattening-demand environment, the asset is worth less than the loan before the loan is due. That is how "asset-light" becomes "asset-stranded." I have seen this exact mismatch before, in a smaller form, in 2020 — the DeFi Summer loans collateralized by tokens whose price assumed a liquidity depth that evaporated the moment two large holders moved. The collateral was real until it wasn't, and then it was a number nobody would honor.
Fact four: utilization is the lie detector. The real question is not how much compute was built. It is how much is actually rented, at what price, and for how long. I built a detection tool in 2025 to map AI-driven manipulation in decentralized exchange volumes, and I found a bot network controlling roughly fifteen percent of trading activity on a niche layer-2. The lesson transferred directly to physical compute: nominal capacity and real utilization diverge, and the gap is exactly where the manipulation hides. If data center occupancy and GPU rental prices are holding, the circular flow is supporting genuine demand. If occupancy is propped up by affiliated tenants signing below-market leases to keep the headline full, then the revenue is a costume. Watch the rental price trend, not the capacity announcement. Capacity is a press release. Utilization is a confession. And utilization is the one number in this entire structure a counterparty cannot fake for long, because it requires an actual workload to run on actual silicon.
Now layer on the opacity. This is the part the BIS actually flagged, and it is the part the market is ignoring.
Most of the largest AI players are private. Their valuations are set by funding rounds, not by market trades. That means the marks are stale by construction. A private lab's last round might have cleared six months ago at a valuation the next round will quietly reset — or, more commonly, that nobody resets at all because no secondary trade forces the question. This is the same stale-mark problem that hid losses inside FTX's balance sheet, except FTX was a single entity and this is an asset class. Combine stale private marks with GPU-backed debt, data center securitization, and special purpose vehicles, and you get an asset base with no observable price and no disclosure obligation. The circular flow runs through structures designed to be unobservable. Chaos is where the institutional money hides, and this is the most orderly-looking chaos the industry has ever built. It is not a coincidence that the BIS used the word "opacity." Regulators choose their words. Opacity is a technical term, and it points straight at the private-market and off-balance-sheet layer where the marks do not move until someone forces them to.
Then there is the multiplier. This is the most underrated point in the entire debate. A dollar of circular investment does not correspond to a dollar of revenue. It corresponds to several dollars of booked revenue, procurement commitments, and derived compute demand. The loop magnifies itself on the way up. Here is the part that should keep you up: the arithmetic does not reverse on the way down — it magnifies in the opposite direction too. If circular funding contracts by a dollar, the associated revenue and demand do not fall by a dollar. They fall by a multiple, because the commitments unwind faster than the cash did. A three-year procurement contract does not shrink smoothly; it cancels, and the cancellation removes the entire future revenue stream from the vendor's book in a single revision. The downside of circular financing is nonlinear, and the nonlinearity is invisible while the loop is expanding. That is not a reason to panic. It is a reason to know exactly what you are holding before the loop turns.
And a note on geography the source completely skipped. Chinese AI funding looks structurally different — a large share runs through government guidance funds and state capital, which behaves nothing like American intra-industry circularity. State capital does not mark to market, does not demand a return on a quarterly clock, and does not unwind on a covenant breach. That does not make it healthier. It makes it differently fragile, with a different clock. The US model can unwind fast and loudly, in public, on a schedule the market can watch. The state-backed model can stay frozen for years and then adjust in a single administrative step, with no early warning and no observable price. Two systems, two failure modes, and a global supply chain — the same GPUs, the same HBM, the same optics — that depends on both. If the American loop contracts while the Chinese loop freezes, the compute supply chain absorbs both shocks at once, and the supply chain is the one part of this structure with no place to hide.
Let me also name what the number does not say, because the absence is load-bearing. It does not say what the denominator is — global AI venture funding, total sector capital expenditure, or a curated sample. It does not say whether hyperscaler cash-flow-funded investment is included, which changes the meaning entirely. It does not say how the ratio has moved over time, which is the only way to know whether 55.2% is a warning or a norm. And it does not separate equity from debt from multi-year procurement commitments, even though those three carry wildly different risk weights. Equity can absorb a loss. Debt and commitments are rigid. Fold a rigid commitment into a "funding" figure and you have quietly converted a liability into a headline strength. That is the single most important methodological gap in the whole release, and it is the one that decides whether 55.2% is alarming or routine.
Contrarian
Here is the angle almost nobody is trading.
The entire market is arguing about whether AI is a bubble. That is the wrong question, and it is the wrong question on purpose, because it keeps everyone focused on sentiment when the real risk is structural correlation. A bubble pops on emotion — a headline, a missed earnings call, a shift in narrative. A correlation unwind pops on a covenant. Those are not the same event, they do not share a timing, and they do not respond to the same hedges. If you are positioned for a sentiment-driven pop, you are positioned for the wrong failure.
The blind spot is exit. In a circular structure, the owners are the customers and the customers are the owners. The chip designer cannot cut the model lab without writing down its own investment and losing its own backlog. The hyperscaler cannot pull compute without destroying the revenue it just booked. The lab cannot stop procuring without losing the valuation that funds the procurement. Everyone is locked into everyone. This is not a market where smart money rotates out first. This is a market where the first mover to exit inflicts the most damage on itself. The exit is crowded by construction, and the crowd is made of the people who built the room.
That is the telecom lesson nobody internalized. In 2001, the vendors could not stop financing the carriers because stopping meant admitting the receivables were fake. They financed right up to the bankruptcy filings, and then they all wrote down together. The AI version has the same trap with a longer fuse. Multi-year commitments mean the losses will not surface as a single event. They will surface as a slow series of quarterly revisions, guidance cuts, and "strategic realignments" spread across two to three years. The unwind will look like a trend, not a crash. The trend is your friend until it ends abruptly — and this one ends in quarters, not days.
Which means the popular trade — short the bubble, wait for the pop — is mistimed by design. The dangerous position is not being long AI. The dangerous position is being long the circularity while believing you are long the demand. Those two look identical on a chart until the loop turns, and then they diverge by the exact width of the multiplier. Liquidity is the only religion in the DeFi temple, and in the AI temple the congregation is paying itself. When a congregation pays itself, the collection plate is the only real asset.
Takeaway
Stop watching the funding headline. Start watching three lines: the non-AI revenue share, the data center occupancy, and the GPU rental price. Those three will tell you whether the loop is still feeding terminal demand or just feeding itself. And keep the BIS file open — when the central bank of central banks names a structure, the disclosure requirements follow within quarters, not years. Patience is a luxury, but here it is a necessity: the unwind is a sequence, not an event, and the people who act on the right signal will be the ones who recognized it before it had a name. This is a yellow warning, not a red alarm — but yellow warnings are the ones you act on before the crowd sees red. The number is 55.2%. The question is who holds the other 44.8% — and whether they know they are holding a mirror.