Aluminum to Silicon
Part III: Duck Curves and Frequent Flier Miles Come for Compute
Alright so we all knew this was coming: time to bring in yet another securitization market to explain a big concept in GPU financing optimization, and a missing gap in the compute markets today.
So earlier in my career I traded a bunch of solar ABS, and one of the firms I worked for actually managed one of the largest utility scale solar PV portfolios globally at the time. The numbers were massive at the time, but now look quaint in the age of AI :)

Anyone who has spent time in the solar market has seen the "duck curve". It's a chart of net electricity demand over the grid over the course of a day (total demand minus what solar is producing), and its shape looks like a duck: a sagging belly at midday, when rooftop and utility-scale solar flood the grid, followed by a steep neck in the early evening, when the sun goes down just as everyone comes home and turns on the lights. The belly means midday power can be worth very little (and at times even negative!), while the neck forces the grid to ramp up enormous amounts of dispatchable capacity in a matter of hours.
The underlying asset didn't change, but the shape of when it produced economic value did, and therefore anyone underwriting solar PV loans had to start underwriting the shape of the curve as well as the volume under it. It ain't sunny everywhere all the time.
Compute has the same problem, except upside down. In a dedicated GPU deployment, supply is a flat line (the GPUs are there 24 hours a day whether you use them or not), while demand swings with human working hours. Consequently, an AI application serving real-time traffic has to provision for its daily peak, and because nobody can tolerate a queue during that peak, it can't even run the peak anywhere close to full saturation. Think about how usage spikes at OpenEvidence or Vera Health during clinic hours (or at Harvey and Legora during the legal workday).
In solar PV, this is a largely solved problem. In GPUs, we're not quite there yet and it's a massive opportunity in a couple different respects.
1. Why the spikes are so expensive
The reason this matters is that you can't simply run GPUs hotter to make up for it. Inference latency behaves like a queue, and Hebbia's engineering team showed in a post this spring that expected wait time scales with 1 / (1 - utilization). In practice, that means a fleet at 50% utilization sees roughly 2x the baseline wait, whereas the same fleet at 80% sees roughly 5x, and at 95% roughly 20x, so the last few points of utilization are extraordinarily expensive in latency terms. Furthermore, since roughly 85% of a GPU's cost is fixed (Part I), cost per token scales roughly inversely with utilization, and therefore a deployment averaging 20% utilization pays roughly 3.5x more per token than the same deployment averaging 70%.
2. The grid already solved this
The power industry didn't solve the duck curve by building more peaker plants for every evening ramp. Instead, it built a set of market mechanisms that make demand itself flexible. Demand response programs pay large users to shift or curtail load when the grid is tight. Time-of-use tariffs price power by the hour so that flexible load migrates to cheap periods, and regional imbalance markets let utilities across multiple states net their surpluses and shortfalls against each other in close to real time. Underneath all of it sits a tiered system of reserves, ranging from frequency regulation that responds in seconds, to spinning reserves that are online and ready, to non-spinning reserves that can come up within minutes.
Each of those mechanisms has a fairly direct analog in compute:
- Demand response: speed-insensitive work like nightly evals, synthetic data generation, re-indexing, and offline document processing gets scheduled into idle GPU troughs and preempted the moment real-time traffic picks back up.
- Cross-tenant pooling (the imbalance market): dedicated capacity gets shared across customers whose peaks don't line up, for example a US clinical app, an APAC enterprise tool, and an evening consumer app, so that the aggregate curve is much flatter than any one customer's.
- Tiered reserves: the fastest tier keeps a shared base model pinned in GPU memory and hot-swaps small fine-tuned adapters per request; the middle tier parks a customer's full model weights in host memory and pages them back onto the GPU in seconds as traffic ramps; and the last tier spills any sudden spike onto pay-per-token serverless capacity, which plays the same role as a peaker plant.
- Almost like Hot, Warm, and Cold Back-Up Servicing.
3. Why it isn't solved yet
The large multi-tenant inference platforms (Fireworks, Baseten, Together, Modal) already pool demand across their own customers, so there is already some demand netting. However, AI native companies will inevitably all graduate into this problem in a big way once they or their customers start demanding stricter latency targets, and they've scaled across more than a single inference provider.
The problem persists for a few reasons.
- First, no single provider can pool a customer's commitments across its competitors, so a company with dedicated capacity on two or three platforms has idle slack sitting in silos that can't see each other.
- Second, although several platforms now bill by the second and can scale to zero, latency-sensitive customers generally keep warm replicas running at all times, which means they pay for idle capacity anyway, and the provider has little reason to help them buy less of it.
- Third, many of these deployments sit inside isolated environments for compliance reasons (HIPAA in healthcare, for example), or just because they are sensitive to cyber, privacy, and other risks, which rules out quietly running someone else's traffic on them.
- Finally, there's real physics involved: You can't instantly yank a GPU out from underneath an active workload and put it onto something new instantly.
Even with perfect orchestration, nobody volunteers their dedicated capacity, or agrees to let a batch job wait a few hours, unless someone pays them to, and nobody signs up for a pooled capacity commitment unless someone funds that commitment up front. The grid figured this out a long time ago. Spike management in compute is therefore as much an economic and working capital problem as it is an engineering one.
With that being said, let's go back to the aviation again.
4. Finally! Back to the most profitable business in all of aviation: frequent flyer miles
I've teased this enough by this point and as you can already guess, the most lucrative business in flying planes is actually the points and miles. You've probably heard the pithy remark that "Delta/United/American isn't really an airline, it's a credit card company that just happens to own planes." In many ways airline loyalty programs are easier to securitize and lend against than the airframes themselves!
Flying people from point A to point B is a brutal, capital-intensive commodity business. We covered some of those reasons in Parts I and II. But suffice it to say that over a full business cycle, the metal-moving business barely earns its cost of capital.
A loyalty program like Delta SkyMiles, by contrast, is effectively a shadow bank. Delta mints a currency at almost no cost and sells it to American Express for hard dollars as cardholders earn miles on everyday spend, which means Delta collects cash up front and delivers the product months or years later (if ever). When those miles are finally redeemed, Delta largely hands over seats that would otherwise have flown empty, and because it controls the exchange rate between miles and seats, it rarely displaces a cash-paying passenger. This is why SkyMiles was valued at roughly $26 billion when Delta borrowed against it in 2020 (more than Delta's entire market cap at the time!).
The part of SkyMiles that matters most for this blog post, though, is how the redemption side actually works. Airlines don't just use miles to reward loyalty; they also use them to move demand around. Award seats get released on flights that aren't selling, off-peak dates price at fewer miles, and a particularly generous redemption on a half-empty Tuesday flight costs the airline almost nothing, since that seat would otherwise have flown empty, while it frees up the Friday seat for a cash-paying customer. In effect, the airline runs its own demand response program, and it pays for it in a currency it mints for free.
Compute credits can work exactly the same way. A customer who lets its evals run overnight, or who accepts a few extra seconds of latency on non-urgent traffic, earns credits that redeem for burst capacity, fine-tuning runs, or frontier model access, all of which can be fulfilled out of capacity that would otherwise sit idle. Furthermore, the economics can be engineered so that every party opts in willingly: neoclouds sell forward and off-peak capacity at a discount in exchange for cash today, the credit and payment layer uses that discount to fund the rewards, customers get cheaper compute in exchange for flexibility they barely notice, and operators end up with smoother, higher utilization that's much easier to finance.
5. The fix is technical, financial, and incentive-driven
I think the solution has three parts, and none of them works without the other two:
- An orchestration layer that does the technical work: profiling each customer's demand curve, filling the troughs with flexible workloads, pooling tenants whose peaks don't line up, and spilling sudden spikes onto serverless capacity.
- A capital layer that funds it: extending credit judiciously at various rates to customers whose usage it can see in close to real time. And those lending rates would not just differing by the obligor's credit profile, but also based on individual deployments, hardware ratings, and Return on Token (which we will DEFINITELY revisit in a later post).
- An incentive layer that gets everyone to opt in: credits that reward customers for shifting flexible work into the valleys, priced so that the discount on otherwise idle capacity pays for the reward.
The layer that hasn't been built yet: a compute clearinghouse and credit layer. Most companies today buy compute across several model providers, a few inference clouds, and a pile of AI software, and they pay for all of it out of a checking account or a standard corporate card. Meanwhile, neoclouds and inference providers need cash up front to fund capex and fill their off-peak hours, and many would likely sell forward or off-peak capacity at a meaningful discount to someone willing to pay today. Whoever sits between those two sides, aggregating demand, collecting cash up front, extending credit against usage, and routing workloads into perishable spare capacity, would look a lot like the Amex and Delta relationship rebuilt for compute.
If I were building a company in the compute / GPU space, this would be one of my top 3 or 4 ideas :)
6. Why lenders should care
In Part I, I argued that a debt-financed GPU cluster needs something like 65% to 75% utilization just to cover its debt service, and in Part II, I argued that lenders only care about the worst outcome when a contract rolls off. Demand response attacks both problems at once, since it raises utilization at the margin and, more importantly, smooths the cash flows that a lender is actually underwriting. Furthermore, a pooled, demand-responsive cluster starts to look a lot less like a single-asset, single-borrower (SASB) deal and a lot more like the diversified pools from Part II. That matters financially, since diversification compresses the credit spread lenders require, supports higher advance rates, and therefore lowers the weighted average cost of capital for the underlying hardware.
The first phase of the AI infrastructure buildout went to whoever could buy the most GPUs. The next phase, I'd argue, goes to whoever figures out how to financialize the idle troughs, turning wasted capacity into a currency, a credit product, and ultimately a smoother cash flow that lenders can actually underwrite. In the next couple installments I'll talk about three other fairly significant discussions in the compute markets.
Part IV is up: Know Your Index.
I appreciate all the feedback (both positive and negative) on my write-ups so far. In particular, I also want to express my gratitude to my friends Pooja Jotwani (ex Cursor, Millennium), and John Hwang (Enterprise AI Trends) for inspiring and clarifying this discussion!