Martin Fowler's latest fragments roundup flags a pattern worth taking seriously: Alphabet's rising capital investments, Oracle's 500% debt-to-equity ratio, and a wave of comparisons to the dot-com bubble. But the sharpest version of this argument isn't about the dot-com crash at all — it's in Groundbreaker's "The Second Derivative", which argues the AI infrastructure boom is structured more like the 2008 mortgage crisis than the 2000 tech bust, and that the trigger isn't falling demand — it's slowing growth.
This article covers what a "second derivative" means in plain terms, why the 2008 comparison is more precise than the dot-com one, what the actual capex and valuation numbers show, and what this means for anyone building on top of AI infrastructure that depends on this spending continuing.
Levels, velocity, and the thing nobody's watching
Markets are good at watching two things: levels (how big is a number right now) and velocity (how fast is it growing — the first derivative). What they're bad at watching is the second derivative — whether that growth rate itself is speeding up or slowing down. A number can still be rising in absolute terms while its rate of increase is quietly decelerating, and that deceleration is much easier to miss than an outright decline.
A borrower doesn't default when prices fall — they default when the pace of price appreciation slows, because the refinancing plan was built on the pace continuing, not just the direction.
That's the actual mechanism Groundbreaker points to in 2008: mortgage delinquencies started rising in 2006, while home prices were still going up. The popular narrative — prices fell, so people defaulted — has the causality backwards. Homeowners had refinanced based on an assumption of continued acceleration in home values. When the rate of appreciation merely slowed, the refinancing treadmill that depended on ever-larger equity gains broke, well before prices actually dropped.
Why AI infrastructure spending looks like a loan book, not a budget
The argument applied to AI is structural: frontier labs, especially OpenAI, are financing enormous compute commitments largely through continuous equity refinancing at escalating valuations — not through revenue covering the cost. Every funding round is, in effect, a refinancing event, and each one has needed a bigger "step-up" in valuation than the market can indefinitely keep supplying.
The step-up multiple — how much bigger each new valuation is than the last — has been trending down, and the projected IPO target represents the smallest step-up in the entire sequence. That's the second derivative made visible: the valuation is still rising every round, but the rate at which it's rising is decelerating, which is exactly the pattern that broke the mortgage refinancing treadmill in 2006.
The capex numbers tell the same story
Hyperscaler capital expenditure — the spending by Microsoft, Google, Amazon, and Oracle on data centers and chips to serve AI workloads — shows the identical deceleration pattern, one step further along.
Growth accelerated into 2025, then the acceleration itself turned negative — 2026 and 2027 both show slower growth than 2025, even though total spending keeps climbing every year. Layered on top of that: capex as a share of operating cash flow has climbed from roughly 30% in 2022 to a projected 100% in 2026 — the point at which hyperscalers stop being able to fund this spending from their own cash flow and cross fully into external financing dependence, which Groundbreaker calls the end of the "fortress balance sheet" defense that bulls have leaned on.
Why the backlog concentration makes this a credit story, not just a valuation one
The most concerning number isn't a growth rate — it's concentration. Total hyperscaler contracted backlog (revenue promised under existing contracts) sits at roughly $2.1 trillion, and OpenAI and Anthropic together account for about half of it. Oracle's backlog is 54% exposed to these two companies, with roughly $300 billion of that owed by OpenAI alone. That's not diversified counterparty risk — it's a small number of borrowers whose ability to pay depends on their own continued access to refinancing.
This is what makes the 2008 comparison more useful than the dot-com one: dot-com companies mostly burned through equity they'd already raised and then failed as businesses. This structure has a genuine credit component — contracted backlog functioning like a loan book, with the hyperscalers' compute buildout as the funded principal — which means a slowdown doesn't just compress valuations, it can freeze credit the way mortgage-backed securities did.
What this means for builders
Anyone building a product that depends on cheap or abundant frontier-model API access should treat compute pricing and availability as exposed to this cycle, not insulated from it. If the refinancing treadmill that funds frontier lab compute purchases slows — not stops, just slows — the practical effect for a downstream builder could be tighter API rate limits, price increases, or reduced free-tier generosity well before any lab actually fails.
It's also worth diversifying model and infrastructure dependencies now, while switching costs are low, rather than during a credit event when everyone is trying to do the same thing at once. The Groundbreaker piece is explicit that it doesn't know the timing — "I do not know when" — which is itself a useful signal: this isn't a call to panic on a deadline, it's a case for not building single points of failure into a stack that rests on financing dynamics you don't control.
Conclusion
The comparison to the dot-com bubble has always been slightly wrong, because dot-com companies mostly failed on revenue, not financing structure. The more precise — and more useful — comparison is 2008, where the trigger wasn't a market turning negative, it was a market that kept growing while its growth rate quietly decelerated underneath the surface. Anthropic's and OpenAI's compute commitments sit inside a financing structure with exactly that shape. Nobody knows if or when the acceleration keeps decelerating into an actual reversal — but the numbers say it's already happening, and that's worth watching regardless of how the story ends.