Equity

Steve Hou on how GPU rentals become a Wall Street commodity

Steve Hou· Head of Research at Silicon Data at Silicon Data
·~31 min·English·TechCrunch
AI InfrastructureGPUBusiness Strategy
TL;DR

Silicon Data's head of research explains how GPU rentals are becoming a cash-settled Wall Street commodity, why that liquidity could unlock even more AI buildout, and how the price data cuts against the chip-depreciation doom narrative.

01The Big Idea

A reference price for a trillion-dollar buildout

<strong>Silicon Data wants to be the index a Wall Street futures contract settles against</strong> — turning GPU rentals, now the single biggest cost in AI, into a priced and tradeable commodity.

the Google's AWS and Amazons of the world are spending some $750 billion you know on compute right next year they're slated to spend something like a trillion

— Steve Hou, Equity
Key Insight
The prize is not the trade — it is owning the reference price everyone else settles against. That is the same position S&P holds for equities and the CME holds for corn: a benchmark is a quiet data monopoly that every downstream contract has to pay to touch.
02The Misconception

You do not need to store it to trade it

<strong>The objection that compute cannot be a commodity because chips go obsolete confuses the commodity with its settlement</strong> — like S&P 500 and most oil contracts, compute futures are cash-settled, so no GPU ever changes hands.

during COVID in March 2020 or whenever the oil futures became negative is it was exactly because people didn't want to take physical delivery so they were paying someone else to take the futures contracts away from them at a loss

— Steve Hou, Equity
Key Insight
Cash settlement means the contract never has to touch a physical GPU, so a chip aging out of the frontier is not a blocker — it is simply a number that feeds the index. Every storability objection is answered by the same three words: we settle P&L.
03Who Trades It

The three sides of the compute market

<strong>Data centers are natural sellers, AI labs are natural buyers, and market makers bridge the gap</strong> — enterprises may buy AI tokens rather than compute, leaving labs as the natural compute buyers.

it's actually not the enterprises that will necessarily buy compute. enterprises may actually just directly buy AI tokens and model tokens

— Steve Hou, Equity
Key Insight
The market only works because the two natural sides lean in opposite directions: data centers are long physical GPUs and want to sell them forward, while labs are short compute and want to buy it forward. That built-in tension is the liquidity — and it is why the labs, not the enterprises paying them, are the structural buyers.
04The Reframe

Hedging is not betting

<strong>A futures position is portfolio management, not a wager</strong> — you dial exposure up or down, and Hou admits people often hedge just to sleep better.

sometimes people do that just for psychological comfort. You know, they may not think that it's going to happen, but it helps them sleep better.

— Steve Hou, Equity
Key Insight
Reframing the trade from betting to exposure management is doing strategic work: it is how a compute-futures market gets sold to regulators and CFOs as risk-reduction infrastructure rather than a casino. And the sleep-better admission is a feature — comfort-driven hedging is steady volume that does not depend on anyone having a strong view.
05The Counterintuitive Part

Liquidity lets you build more, not less

<strong>The dilemma that caps the buildout is timing — build too much and you go bankrupt, too little and you miss the boom</strong> — and a liquid futures market helps operators hedge that risk and adjust exposure dynamically.

if you have liquid financial instruments you can actually build more and adjust you know that exposure dynamically subsequently with the help of financial derivatives

— Steve Hou, Equity
Key Insight
This is the same dilemma Dario Amodei described — get the timing wrong and you either make a fortune or go bankrupt — which is exactly the risk a liquid market removes. The implication runs against intuition: financial engineering should make the compute buildout larger and steadier, because hedging lifts the timing risk that was quietly capping how much anyone dared to commit.
06What The Data Says

The A100 is old — and still rented to the maximum

<strong>An ancient A100 still rented to the max is a demand signal, not a supply story</strong> — which cuts against the bear case that data centers are borrowing against fast-depreciating assets.

these things may not depreciate as quickly as people thought and may have actually a longer lifespan as people had previously budgeted for

— Steve Hou, Equity
Key Insight
If chips hold value longer than assumed, the whole data-centers-are-borrowing-against-depreciating-assets bear case weakens, because the collateral keeps its worth as long as demand keeps old silicon in service. The A100 staying fully rented is demand strong enough that buyers will run hardware that Hou flatly calls good at nothing.
07The Forward Curve

Long-term discounts — yet the entire curve moved upward

<strong>Long-term compute reservations carry discounts, yet the entire forward curve later moved upward</strong> — rental rates rose at every contract length as agentic-AI demand increased.

over the subsequent three to six months we see the entire curve essentially moved upward which is exactly consistent what we were saying before about rising agentic AI demand at every term length at every contract length the rental rate has gone up

— Steve Hou, Equity
Key Insight
The downward slope is not a forecast that compute gets cheaper — it is the discount you get for committing to a long reservation up front. The real signal is that the whole curve lifted over the following months, with rates rising at every tenor at once. That is Hou's second-derivative point: markets react to how fast demand is accelerating, not just its level, which is why even steady growth can swing equity sentiment.
08The Endgame

Toward cheapest-to-deliver compute

<strong>Silicon Data wants real physical delivery — grade individual GPUs, then settle against the cheapest available compute</strong> — a commodity that works only because inference can be split and routed anywhere.

you can basically say I can have cheapest to deliver a compute provided at point you know of expiration for futures contracts so that is actually what we believe to be the much much bigger you know plan for the future

— Steve Hou, Equity
Key Insight
Cheapest-to-deliver is the mechanism that turns thousands of different GPUs into one fungible commodity — the same trick that lets wheat from different farms settle a single wheat contract. It only works because inference, unlike training, can be chopped up and sent anywhere, which is why Silicon Data is betting inference becomes the bulk of demand.