I spent a Tuesday afternoon in Cape Town reading convertible bond prospectuses, which is not how most people measure the temperature of a bull market. But that is where the truth lives. Between the optimism of a product launch and the arithmetic of a debt covenant, there is a paper trail. And in 2025, that paper trail points somewhere uncomfortable: the most technically ambitious industry on Earth is financing itself with the oldest instrument in the creditor's playbook.
Meta lifted its 2025 capital expenditure guidance to between $60 billion and $65 billion, almost all of it for AI infrastructure. Microsoft, Amazon, and Alphabet have issued investment-grade debt at scales that would have been unthinkable a decade ago. OpenAI and Anthropic, still private, are anchoring valuations that now function as credit signals. Convertible bonds — low coupon, equity upside, downside cushion — have become the instrument of choice. US convertible issuance has hit record highs. The financial press called it AI "devouring capital markets." That framing is lazy. AI is not devouring anything. It is borrowing against a promise, and the promise is compute. Investment banks underwrite the trend with enthusiasm because fees scale with issuance, and nobody on that desk gets paid for restraint.
Convertible bonds are not new. What is new is the scale, and what the scale reveals about where the AI industry believes value will live. A convertible bond carries a coupon that is often near zero and converts to equity when the share price crosses a threshold. For the issuer, it is delayed equity financing at debt prices. For the buyer, it is a call option with a bond floor — you participate in the upside, and if the story collapses, you collect principal plus a modest yield. Structurally, it is a bet that volatility will rise but not catastrophically. Right now, everyone is making that bet on AI.
The ledger tells us more than the earnings call. When I audited ERC-20 token standards in 2017, the lesson was that the structure of a financial instrument encodes who holds power when things go wrong. A token with a hidden mint function is not neutral; it is a promise with an escape hatch. A convertible bond is the same. It looks like cheap capital, but it transfers optionality to the lender and refinancing risk to the borrower. If AI commercialization arrives later than debt maturities, the escape hatch opens the other way.
Here is the arithmetic the cheerleaders skip. Compute is not a moat. It is a depreciating asset. A GPU cluster bought today is worth less in eighteen months, and the model trained on it may be obsolete in twelve. When you finance a depreciating asset with debt, you are betting that revenue outpaces both interest and obsolescence. That is a narrow window. Frontier training runs now cost well over $100 million each. You do not finance a $100 million run from operating cash flow when your product is still being priced by the market. You borrow. That is what the record issuance means.
There is a mechanical detail most coverage ignores. Convertible bonds are bought heavily by arbitrage funds that simultaneously short the underlying stock. This delta hedging means the bond does not simply reflect belief in AI; it reflects volatility demand. When the arbitrage community piles in, the bond price is driven by implied vol, not by fundamentals. That is why record issuance can coexist with flat share prices. The market is not voting on AI's future; it is pricing a hedging instrument. Traders are not believers. They are engineers of variance.
And where does the borrowed money actually go? Not to abstract "AI." It goes to Nvidia GPUs, HBM memory from SK Hynix, advanced packaging from TSMC, optical modules, and the power and cooling for data centers that consume electricity at the scale of small cities. The convertible bond market has become, quietly, the financing arm of the physical compute supply chain. When you buy that bond, you are not buying software. You are buying a claim on silicon, copper, and megawatts.
This is where my community should pay attention, and not for the reason the newsletter hacks claim. The bears say: traditional markets are cracking, so money will flee into Bitcoin. That is a comforting story for people who need a reason to hold. The real signal is deeper. The AI boom is recreating, in physical infrastructure, the exact centralization pattern that decentralized systems were built to resist. A handful of firms control the capital, the compute, the talent, and now the debt that funds all three. Open source models run on rented clusters. Independent researchers queue for API access. The compute layer is becoming as permissioned as the banking layer it claimed to disintermediate.
I have watched this movie before. In 2020, I ran DeFi workshops in Cape Town for people losing money to impermanent loss they did not understand. The problem was never liquidity. The problem was that a few products monopolized liquidity and branded the fragmentation elsewhere a crisis. The same trick is running now. "Liquidity fragmentation" was a manufactured narrative to sell the next product; the AI capital "shortage" is a manufactured urgency to justify leverage. When capital concentrates, the story about scarcity is always told by the people holding the supply.
From the creator's side, none of this is abstract. The platforms building on borrowed compute set the terms for everyone downstream — the royalty rates, the licensing terms, the API prices. In 2021, I worked with ten indigenous South African artists to build royalty-enforcing smart contracts because major platforms silently dropped automatic secondary-sale payments. Sixty percent of secondary sales lacked enforcement. The lesson was not that platforms are evil; it was that whoever controls the rails sets the toll. When the rails are financed by convertible debt, the toll goes up to service the interest. Artists own their pixels; we just hold the keys — and someone just mortgaged the keyring.
Everyone is worried about the wrong thing. The pundits warn of a debt crisis — AI firms unable to repay their convertible bonds, credit markets freezing, the bubble popping. Maybe. But bubbles never destroy value evenly, and the real cost of this cycle is not default risk. It is governance lock-in.
A convertible bond sold in 2024 matures in 2029. The company that survives will have spent five years building centralized infrastructure with no incentive to open it. The more compute they borrow against, the more they must protect it — which means closed weights, proprietary tooling, locked APIs, and the quiet conversion of "AI safety" into a licensing regime. Open source is not a license; it is a promise. And promises are the first thing to go when the interest payment comes due.
The maturity schedule is the number to tattoo on your arm. If a meaningful share of 2024's and 2025's convertible debt matures between 2027 and 2029, it collides with the period when the largest capex commitments convert from cost into expectation. Companies will face a choice: refinance at higher rates, dilute shareholders through forced conversion, or cut the research that made them interesting. Historically, they cut the research. The history of leveraged technology is a history of deferred safety.
The utilization question is the one nobody wants to answer out loud. Capital expenditure is a leading indicator; utilization is a lagging one, and it rarely appears in a prospectus. A data center financed at scale needs sustained demand to justify its depreciation schedule. If every major lab is training similar models on similar architectures toward similar benchmarks, the marginal cluster adds less value than the one before it. Redundant compute is not a moat; it is a cost center wearing a moat's uniform. The bondholders will discover this at maturity, which is precisely when the discovery is most expensive.
I have been told I am too philosophical about balance sheets. But every debt structure is a philosophy. A convertible bond says: we believe in the upside, but we refuse to pay for it upfront. That is not conviction. That is a hedge dressed in ambition's clothing. The AI industry is not devouring capital markets; it is hedging its own uncertainty and calling the hedge a moat.
So what should you actually track? Not the headline issuance figures. Watch three things: the maturity calendar, the free cash flow of the issuers, and the utilization rates of the data centers being financed. If the first runs ahead of the second while the third declines, the record convertible issuance was never a sign of strength. It was the last cheap round before the door closed.
Tracing the code back to the conscience behind it, I keep arriving at the same place. The defining question of 2026 is not whether AI will be powerful. It will be. The question is who owns the layers beneath it — the compute, the data, the identity — and whether an open path survives for the people who were never invited to the bond sale. Education is the only true decentralized currency. Spend it carefully, and teach someone else to read the ledger. The ledger does not care about your conviction. It only cares about the date.