Wednesday 18 Nov 2026
main news image

This article first appeared in The Edge Malaysia Weekly on November 9, 2026 - November 15, 2026

ON Nov 3, the Democratic Party swept the US midterm Congressional elections by taking control of both the House of Representatives as well as the Senate, ousting the Republicans who had won control of the White House and two houses of legislature in 2024 when President Donald Trump returned to power. The biggest drubbing was in the contests for governors of Texas, Georgia, Ohio, Kansas, Alaska and other states which flipped to Democratic candidates who ran against tech giants and billionaires.

A key target was trillionaire electric vehicle (EV) pioneer Elon Musk who reportedly spent US$500 million (RM2.03 billion) to prop up pro-data centres candidates. Little wonder, then, that tech stocks took a hammering as election results began streaming in. 

Among the issues surrounding artificial intelligence (AI), computing power and infrastructure capital expenditure (capex) is the widening “monetisation gap” between upfront spending and realised revenue. Investors are concerned whether the returns on investment will ever justify the sheer scale of the massive AI buildout. If you follow the markets closely, you may have noticed that Wall Street has stopped rewarding companies just for splurging a lot of money on AI and indeed now aggressively punishes those that are unable to show rapid monetisation.

That brings me to the most valuable currency in the market: compute or computational power. Think of compute as the raw processing capacity to run calculations. Or the total capacity to do the mathematical work that training and running AI models requires. It’s often used as a noun, like “water” or “fuel”. You can have more or less of it, you can buy it, or you might rent it and obviously because it is scarce commodity, at some point you might even run out of it.

Compute will soon become a tradable asset class just like stocks, bonds and commodities. AI is likely to generate US$1 trillion of chip revenue within the next five years and the top five hyperscalers — Google’s parent Alphabet, Microsoft, Amazon Web Service or AWS, Meta Platforms and Oracle — will spend over US$700 billion on AI capex this year. Total global graphics processing unit (GPU) rental market this year is likely to be around US$52 billion. Within four years it is forecast to exceed US$200 billion.

Compute basically comes from the GPU chips or xPUs like Google-designed or tensor processing units often called TPUs. A single Nvidia B300 is a unit of compute. A rack of 72 is a lot more. And a data centre full of those GPU chips is huge. Think of the whole chain like this: GPU chips and high bandwidth memory chips or HBMs are the raw ingredient of compute. Think of data centres as the buildings that house and power thousands of chips. You might think of a data centre as a factory where all the ingredients or raw materials or components are put together into a finished product called compute. Training a frontier model such as OpenAI’s GPT or Anthropic’s Claude or Moonshot AI’s Kimi K3 can consume an enormous amount of compute up front running trillions of calculations over months. There is also Inference or using the model to answer a query which also consumes compute, in smaller amounts but continuously, every time someone uses it.

Scarcity drives volatility. One-year contract and rental rates for Nvidia H100 data-centre GPUs have surged 60% over the past year from a low of about US$1.70 per hour up to over US$2.80 per hour while spot rental rates for Nvidia’s top-of-the-line Blackwell B200 and B300 GPUs were up nearly 70% over the past few months, leaving hyperscalers and neoclouds with problems like what airlines face when oil price suddenly spikes each time someone mentions turmoil in the Strait of Hormuz. In an AI-driven world, everyone wants to buy or rent compute. When your biggest input cost moves 70% over a few months, you need a way to lock in prices. That’s exactly what turned oil, electricity and wheat into traded commodities. Now it’s happening to compute.

But, let’s step back a bit to look at why compute is such a scarce resource. Mainly, because demand is exploding, while supply is throttled by not one but by an entire chain of bottlenecks. A shortage anywhere in the chain limits the whole thing. Here’s another way to look at it. AI models get better mainly by using more compute. Training a frontier model takes vastly more computation than the last generation. The amounts multiply with each cycle. So every frontier AI lab wants a lot more compute. Then there is inference demand. Once an AI model exists, every user query consumes more compute. As hundreds of millions of people use AI, that’s a massive, growing, always-on demand separate compute needed for training. When every hyperscaler, frontier lab and AI start-up is racing for the same finite supply, because being short of compute means falling behind, it turns strong demand into a stampede. Everyone wants more compute instantly.

On top of that, there are all these supply choke points that limit supply as demand keeps growing. Let’s look some of these choke points. Almost all of the AI chips are made by Taiwan Semiconductor Manufacturing, or TSMC, which until recently had a monopoly on leading-edge 2 nanometres nodes. It’s not easy to build a cutting-edge foundry. It can take up to US$50 billion and four to five years to build one of them. South Korea’s Samsung Electronics now has the capability to almost match TSMC but at leading-edge nodes, it has 5% of the capacity that TSMC has. Intel is trying to build a leading-edge foundry but could be a few years away. Elon Musk is building his own Terafab but that foundry could be at least five years away since its still in the design stages.

There are also bottlenecks in advanced packaging or Chip-on-Wafer-on-Substrate (CoWoS). AI chips need a special technique to bond the Nvidia GPUs to its memory. CoWoS capacity has been one of the hardest constraints, even when you can make the silicon, you might not be able to assemble it fast enough.

My tech columns over the past year have covered memory chip shortages in detail. There is an acute shortage of High Bandwidth Memory or HBM. There are only three global suppliers, South Korea’s SK Hynix and Samsung Electronics and US-based Micron, which has a growing Singapore footprint. Lead times for HBM run up to 26 weeks, and prices are still rising. SK Hynix recently warned the shortage may persist past 2030. Every AI chip needs HBM, so if HBM is in short supply, chip output has to be rationed, no matter how many new versions of GPUs Nvidia can design.

Don’t forget power and grid which are the backbone of compute. Without power there are no chips. But even after you have manufactured the AI chips, you still need enormous electricity to run them. Around the world more communities are saying they love innovation and new technology but don’t want data centres or power plants in their backyards. Grid connections take years and power has become the binding constraint on deploying compute in many places. Even in communities where data centres are welcomed, it can take unto a year or two just to build and fit out a data centre. According to one report nearly 40% of planned capacity eventually gets cancelled.

Where there is crisis and a crunch there is an opportunity. Neoclouds like CoreWeave have long rented GPUs by the hour. A spot market for GPUs already exists. Now a derivatives and futures market for GPUs is about to be launched. Silicon Data, a New York-based financial intelligence and data infrastructure platform, has been building a benchmark layer to hedge that risk, converting GPU data into indices and forward curves. It approached Chicago Mercantile Exchange, or CME Group, which now plans to launch the first listed GPU rental futures using Silicon Data indices later this month, pending regulatory approvals.

Intercontinental Exchange, the parent of New York Exchange, and a start-up called Ornn, are launching cash-settled GPU compute futures benchmarked to the Ornn Compute Price Index The index, which will be available on Bloomberg Terminal, tracks spot rental rates for Nvidia H100, H200, B200, and even RTX 5090 GPUs. So just as stocks are traded on a bunch of exchanges, you could trade compute futures on different exchanges.

The launch of CME’s AI compute futures, which track Nvidia GPU rental rates like H100 and B200 chips, was actually delayed because in late September the US Commodity Futures Trading Commission (CFTC) extended its review period by 45 days to Nov 9. The regulator had stepped in to examine several core concerns regarding compute futures trading just a few weeks before compute trading was to begin in early October. Clearly, computing power had never been traded as a standardised financial asset class before, prompting the CFTC to take a closer look at how the derivative contracts are designed. Regulators also want to assess potential vulnerabilities, fragmentation and transparency within the underlying cash market for GPU rentals.

How will compute trade? Like electricity rather than oil. Compute is in a way a weird sort of commodity. For one thing, it is non-storable. You can stockpile a barrel of oil but you cannot stockpile a “GPU-hour”. An idle GPU-hour is gone forever, like electricity or airline seats. That makes GPU a flow, rather than a stock listed on an equity exchange, so it trades more like power markets than oil.

For another, compute is a depreciating and rapidly obsolescing asset. An Nvidia H100 chip’s value falls as new B200/B300 arrive in the market. So unlike crude oil, each compute contract is tied to a specific chip generation that is decaying. I know not all oil is the same. There is light crude and heavy crude, there is sweet crude and sour crude. There are also grades like West Texas Intermediate, or WTI, and Brent but they are fairly close substitutes. With Nvidia’s H100 or B300 every new generation of chip is a huge technological leap forward. So you get separate contracts per chip, not one unified “compute” price. The compute will trade as cash-settled futures on rental-rate indices. You won’t have to take delivery of actual chips. You settle the difference against a published price index (Silicon Data’s or Ornn’s). That is the same model as electricity and other non-deliverable commodities.

So, who wins from compute trading? First, the exchanges, CME and Intercontinental Exchange with their trading fees will make a lot of money on volume regardless of whether compute prices go up or down. Secondly, the index and data providers like Silicon Data and Ornn will make some money as well as they get a cut of the overall revenues. There will also be a bunch of trading firms and market makers who will carve a niche for themselves with this new AI asset class. Finally, the AI companies themselves will be key beneficiaries. Compute trading will allow frontier AI labs, neoclouds and their enterprise customers to hedge their future compute costs, lock in prices, get visibility into future spending and turn what was once a wild variable cost into something more predictable. A neocloud like CoreWeave might hedge its rental income, a frontier lab like OpenAI might lock in its burgeoning training costs. Farmers, food companies, miners and steel makers dabble in derivative markets not because they love speculation but to better manage their risks.

Compute is on the verge of becoming commoditised, but whether it will evolve into a deep global market like oil anytime soon, or remain a fragmented collection of chip-specific hedging instruments, remains to be seen. What’s clear, however, is that at least initially, the market will be fairly thin, compute futures likely be illiquid and volatile and compute’s non-storable, depreciating, heterogeneous nature means it might never commoditise the way oil has. It might just remain like electricity, which trades but remains regional and fragmented.

Assif Shameen is a technology and business writer based in North America 

 

Save by subscribing to us for your print and/or digital copy.

P/S: The Edge is also available on Apple's App Store and Android's Google Play.

      Print
      Text Size
      Share