Lambda, Runpod, and Nebius have all increased their publicly listed prices. AWS hiked its AWS Capacity Blocks offering from $5.19 to $5.97 for each H100-hour from August to October. (The H100 is Nvidia’s most popular AI chip and started shipping in 2022.)
In most other markets, compute prices follow a predictable pattern: every quarter, prices for last-generation chips come down. With AI compute, prices have done the opposite: last-gen chips have gone up, while cutting-edge chips are in such short supply that buyers are placing change orders to get them.
Amazon, Alphabet, Microsoft and Meta plan to spend more than $700 billion this year on data centers, chips, power and cooling, up from about $410 billion in 2025.
Just this past third quarter, those companies spent a whopping $208.7 billion, an 84% increase over the same quarter last year.
Why am I blathering on about these numbers?
If you’ll indulge me with a drum roll…
Compute is the new oil!
In circles that orbit AI, you often hear compute compared to oil. This is apt, and yet not really apt.
Here’s what I mean.
“Compute” is a slippery term because it can mean many things. Among AI engineers, though, it usually refers to how much math a computer can do. Often, that math is measured in floating-point operations, aka FLOPs. When you prompt a model, for example, its operator runs lots of multiplications (among other operations) for every word it produces in response. That’s called inference. Training a model works similarly, except the AI lab’s engineers do many, many more multiplications (and other operations) on mountains of data, text, images, code, and so on.
Herein lies a difference between compute and oil: If compute means FLOPs, then on a given chip it’s a good approximation of energy usage. That’s because every multiplication and addition requires flipping bits, and flipping bits takes energy. So when an AI operator runs lots of FLOPs, they consume a lot of energy.
Oil can be burned to produce energy, sure. But that’s not the only thing oil is used for. It’s also a valuable chemical feedstock, among other things.
Compute, on the other hand, is essentially a proxy for energy use.
Which is somewhat important difference that has big implications for founders, investors, and even solo operators.
So why is everyone buying so much compute?
If you’re an engineer, you may have noticed a trend over the last several years: everyone seems to be spending more on compute. In particular, AI labs spend a lot on compute to train their models. And operators of those models spend a lot on inference.
The former is driven by scaling laws, which describe roughly how model quality increases with compute spent on training. The latter is driven by the fact that it turns out reasoning models, which take lots of steps before responding to a user’s prompt, are better at answering complicated questions than non-reasoning models.
Scaling laws and inference gains from reasoning are both important trends that have significantly increased the compute bill for AI labs and operators. But there’s a third trend that’s quietly gobbling up a mountain of compute: agentic AI.
If you’ve ever unleashed an agent on a task, you’ve likely noticed that, in seconds, it does a whole lot. It might read files, write code, check its own work, discover an error, and try again. In that time, it’s likely made hundreds of calls to a model.
That’s because agents are chock-full of models! And models gobble up compute.
It’s not uncommon for agents to use more compute for a single user prompt than hundreds of chatbot prompts did in 2023.
Compute is an input
Economists have a word for things consumed to produce something else: inputs. Compute is an input.
If you’ve deployed an agent, you know that every time you do, you consume compute and money. You’re buying compute so agents can do work for you. If you run an app with a model embedded in it, you’re buying compute to delight customers. If you’re a startup founder, you’re buying compute so you can ship features fast enough to keep up with your competition.
In fact, if you’re a startup founder, you should be delighted that compute is an input because that means your customers are buying inputs, too! Every startup selling AI-embedded products to businesses should ask, “How much is my customer spending on compute?” Because that’s your total addressable market.
In a way, that makes compute even more similar to oil. Businesses buy oil all the time, too.
Five similarities…
There are five ways compute is similar to oil, and I think all five are important.
Compute is used everywhere. Like oil, compute is used in nearly every industry. Lawyers use compute to review contracts. Marketers use compute to write ads. Pharma companies use compute to model proteins. Game developers use compute to create new graphics. Delivery companies use compute to optimize logistics. Banks use compute to catch fraud.
Pricing power. Whoever can set the price of compute effectively sets a cost floor for everyone else, just like whoever can set the price of oil.
Concentrated supply. Nvidia sells most of the GPUs that AI labs buy to train models. TSMC, mostly in Taiwan, makes most of the world’s cutting-edge chips. Three companies make most of the high-bandwidth memory that those chips pair with. A few hyperscalers own most data centers.
Expensive and slow to produce. Oil fields cost billions of dollars to find and years to start producing oil. Data centers cost billions of dollars to build and years to start producing compute. CreditSights recently estimated that Alphabet, Meta, and Microsoft will spend 46%, 54%, and 47% of their respective revenues on capex this year. Not exactly the asset-light businesses they once were!
Strategic resource. Oil has been strategically important enough that countries have gone to war over who controls it. Compute is now strategically important enough that countries are interfering with its supply, as the US has restricted Nvidia’s ability to sell certain GPUs to China, China has subsidized its domestic chip industry, and France, India, Japan, the UK, and the EU have launched compute initiatives (the EU’s are “AI Gigafactories”).
When countries start controlling the supply of a resource, it’s a pretty good sign that the resource matters a lot.
Oil countries are investing in becoming compute countries.
The UAE recently agreed to build a major OpenAI-related campus. Saudi Arabia set up Humain, a sovereign AI company, focused on building data centers. Countries that have used oil to amass wealth are now investing that wealth to control the next key resource.
If the compute-oil analogy is wrong, the rulers of Saudi Arabia and the UAE are terrible strategists.
Wait; maybe Compute is ~not the new Oil?
There’s a limit to how much you can compress the energy in a barrel of oil, and a hundred years from now, a barrel of oil will probably still contain roughly the same amount of energy it does today. It seems unlikely, for instance, that engineers will figure out how to extract double the energy from a barrel of oil every two years.
Compute is different. Every year or so, Nvidia releases new chips that offer significant improvements in computation per dollar and per watt. Models and model training have gotten more efficient too. As recently as late 2024, Andreessen Horowitz reported the cost of achieving a given level of model capability was falling by 10x per year. And because LLMs consume and produce “tokens” (roughly ¾ of a word) and are priced by the API provider in units of millions of tokens, that means over the last three years, the cost of state-of-the-art LLMs has fallen off a cliff.
So far, compute looks like the polar opposite of oil: rather than seeing rising costs for a given quantity of compute, we’ve seen plummeting costs for a given level of capability. So why would anyone expect to see anything resembling the situation we have with oil - why are people paying more to rent Nvidia’s 2022-vintage chips in 2026 than they did in 2025? Why are a handful of companies projecting CAPEX this year in excess of $700 billion?
There’s a clue to answering that question in an observation made in 1865 by the English economist William Stanley Jevons: as steam engines became more efficient at extracting work from burning coal, Britain used more coal rather than less. Jevons’ Paradox, as it’s since become known, can help explain what’s going on with compute. When coal became cheaper to burn, new uses for it emerged that previously weren’t economical. The same thing seems to be happening with compute.
Each time LLMs get cheaper, we see new applications for tokens - when inference got cheap enough, applications emerged that took advantage of longer context windows, deeper reasoning, agentic loops, LLMs checking the work of other LLMs, etc. I pay a lot less for tokens than I did two years ago, but I pay a lot more per month for having LLMs do things for me that I wouldn’t have hired a human to do.
Jevons’ Paradox highlights one way the oil analogy breaks down: aggregate spend on compute can keep rising even as the price of a given level of computation falls. But it also highlights another way in which the analogy breaks down. GPUs aren’t like oil reserves; if anything, they’re more like fishing boats: assets whose value depreciates rapidly as newer, better assets are introduced. Betting on sustained high GPU rental prices is like betting your fishing fleet will be tied up year after year filling other people’s orders at high prices. And borrowing against your fleet of H100s on that assumption isn’t so much betting on the future of AI as it is betting on shortages.
Betting against sustained growth in compute spend because the price of a given amount of computation keeps falling would be making the opposite mistake: demand for compute could keep growing faster than prices fall, in which case total spending would keep growing. Whether that happens indefinitely remains to be seen.
Oil and compute do cycle similarly, however. As the price of oil rises, new wells get drilled. As the price of compute goes up, new data centers get built, and new chips get developed. But while new oil supply is ultimately limited by geology (Saudi Arabia benefits from a geological fluke: a place where lots of plankton died hundreds of millions of years ago), new compute supply is limited only by economics. TSMC could open new fabs. Nvidia could design new chips. Amazon could pour more concrete. Give them enough time and tens of billions of dollars, and there’s no physical limit to the compute they could supply. (Limits due to permitting, electricity, etc. aside.) This makes cartels harder. When OPEC strangled supply in 1973, they pushed oil prices up fourfold because they knew it would take years for new supply to respond. If the major suppliers of compute tried something similar, they’d face an industry with lots of well-capitalized competitors (especially once Nvidia’s margins get any higher) and new suppliers like Amazon (whose custom silicon business - Graviton CPUs, Trainium accelerators and Nitro chips - passed a $20B annual revenue run rate this year). Expect cycles of overbuilding and busts, consolidation and innovation-led demand booms, not a cartel propping up prices.
Oil is fungible; compute is less so.
A barrel of Brent in Rotterdam is roughly fungible with any other barrel of Brent. You can ship it around the world on tankers, and refineries will pay a bit more or less depending on exact location and oil quality. Oil’s fungibility is what allowed traders to develop liquid global spot and futures markets.
Compute is less fungible. An hour on an A100 in Finland isn’t fungible with an hour on an H100 in Virginia, let alone Google’s TPUs. If you’re training a frontier model, you want tens of thousands of GPUs in the same data center, connected with fat pipes (the GPUs spend most of their time talking to each other). You can’t assemble that from parts purchased from multiple suppliers. If you’re running a chatbot, you care about inference latency and want to be as close to your users as possible. If you’re running batch jobs, you’ll do it at night on whatever GPUs are sitting idle. And if you’re an AI startup, most of your engineers will be using CUDA, the software ecosystem Nvidia has been growing since 2007. You could port to AMD or TPUs, but your engineers would have to spend time doing so.
All of this means that when people talk about “the price of compute,” they’re really only talking about one small corner of a highly segmented market (like the H100 rental indices put out by various research shops). It’ll get more fungible over time as software gets more portable and contracts more standardized, but that day is still a ways off.
Power is the bottleneck.
When people think about who has power in AI, they tend to think about GPUs. But the new power struggle is all about power. AI data centers can consume as much electricity as small cities, and GPUs are being secured faster than electricity can keep up.
At the start of 2026, over 2,000 gigawatts of generation and storage are waiting in interconnection queues in the US (for context, total US generating capacity is ~1,300 gigawatts). In the two biggest data center markets (ERCOT and PJM), it takes three to four years to get connected to the grid. As of early 2026, ERCOT had ~233 gigawatts of applications from “large loads” (mostly data centers), which grew to over 400 gigawatts by spring. Developers are getting tired of waiting. By 2026, they’d announced ~100 gigawatts of on-site generation, placed orders for over half of it (mostly gas turbines, but also solar and batteries), and gas turbine orders since 2024 are at a twenty-year high.
Microsoft signed a twenty-year deal to restart Three Mile Island. Meta has signed several gigawatts worth of nuclear deals and has said it wants 6.6 gigawatts of nuclear by 2035.
The impact of all this demand is already showing up in electricity prices. For example, PJM’s capacity price went from $28.92 per megawatt-day in 2024/25 to $329.17 in 2026/27 (capacity prices have hit the auction’s cap in the last three auctions and PJM’s market monitor thinks data centers drove almost 40 percent of the last auction). These prices will be passed on to all electricity customers in PJM’s service territory.
In an age of energy, compute is less like oil and more like a refinery, a way of transforming a basic input into something customers will happily pay for (a lot more than they’d pay for just energy). In that case, instead of asking who will control compute, maybe we should ask who controls (or will be allowed to build next to) reliable, inexpensive energy. Tech firms that have never thought about electricity are now hiring nuclear engineers. American labs are eyeing sites in the Gulf states because they can get electricity quickly and cheaply. Microsoft’s CFO recently said that some of its contracted-for data center capacity from third parties has been delayed, which is why its capex grew slower than planned.
GPUs and memory are today’s bottleneck, but transformers, switchgear, turbines, transmission, and electricians will be tomorrow’s. Somewhere out there is a company that makes some incredibly boring electrical component that’s going to make its investors a lot of money. Probably more than one.
The bubble question
The bear’s argument
We’ve been here before.
In the 1840s and again at the turn of the last century, investors poured fortunes into what they thought would be transformative new technologies – railways in the 1840s and telecom fiber at the turn of the century. Investors were right that both technologies were transformative. Investors lost their shirts. In both cases, investors badly misjudged the pace of demand growth, resulting in gluts of underutilized capacity for years on end. Fiber optic networks were so overbuilt that, years after the bust, ‘dark fiber’ was still sitting idle.
Capex by the big techs is running at levels that would raise similar concerns today. Meta’s capex runs at more than half its revenue. Microsoft, Meta and Oracle all tapped the bond market in the first week of November to raise cash for capex. Torsten Slok, chief economist at Apollo, estimates capex by the big five (Amazon, Microsoft, Alphabet, Meta and Oracle) will run at around 2% of US GDP this year.
The bull’s rejoinder
This time is different.
The bear’s case rests on an analogy with past bubbles and busts. Bulls point out three key ways in which this time is different.
Unlike past busts, tech’s capex is being funded from cash generated within the group, not through excessive borrowing. (Although, as noted, borrowing has picked up this month, and capex now exceeds the group’s operating cash flow. Still, leverage at the big four is far below WorldCom levels.)
Unlike past busts, demand for compute is outstripping supply. When Alphabet increased its capex guidance, this was because it was bringing compute online faster than expected and customers were snapping it up. Backlog at Google Cloud topped $460bn earlier this year and has risen since. Telecoms built networks in advance of demand and found they badly overestimated the pace of adoption. Data centers are being built to fill an order book of customers clamoring for compute.
If supply really did outstrip demand, we’d expect compute prices to fall. Instead, four-year-old H100s are still commanding a price premium.
Of course, revenues to date for AI are small compared with the cost of building out compute. If AI revenues fail to live up to expectations, somebody’s going to write off a lot of silicon (and concrete). But at the moment demand appears to be outstripping supply.
My view
Demand for compute will grow for years to come as reasoning models and agents allow businesses to increasingly purchase ‘work’ from AI labs (albeit by the token, something businesses have always been willing to pay for). That said, not all compute spending will be equally successful. Some companies will build data centers in the wrong geographies, buy the wrong mix of chips, and finance capex in the wrong way.
They’ll be badly hurt when next-generation chips are released. Like the fiber busts of the last century, there will likely be painful write-downs, restructurings, and forced sales among AI infrastructure companies. But that won’t dent the longer-run growth in AI. Just as the fiber busts of the last century didn’t prevent the growth of the internet (Amazon, Netflix, and Google are beneficiaries of overbuilt, cheap fiber laid by others), so the AI infrastructure busts of this century won’t dent the longer-run growth in AI.
If pressed, I’d put the probability of annual growth in aggregate compute spending each year until 2030 at 70% and the probability of at least one significant AI infrastructure bust (write-down, restructuring, or forced sale) over the same period at around 50%. These aren’t contradictory, just as in oil there can be multiple boom-bust-bankrupt cycles within a decade-long upward trend.
Who made money from oil?
The oil industry can be usefully divided into three segments:
Upstream – producers of crude oil (akin to producers of semiconductors and memory, e.g., TSMC and Nvidia, and of power and energy used in AI)
Midstream – refiners of crude oil (akin to owners of data centers and producers of AI models, e.g., GPU clouds, AI labs, and the hyperscalers that both train and use AI models)
Downstream – producers of petroleum-based products and services (akin to everyone else using AI to produce products and services – yes, that includes most readers of this post!)
Oil production was risky; successful wells were the exception, not the rule. By contrast, refining and especially transportation were key bottlenecks on the path to profitability. This is why John D. Rockefeller focused on refining and transportation, not production. He let other people drill risky exploration wells, extracted secret rebates from railroads, and acquired pipelines, allowing him to control the price of kerosene across the US. During downturns, he’d acquire competitors that had over-invested and were highly leveraged. By 1880, Standard Oil controlled ~90% of oil refining in the US.
The lesson from Rockefeller’s Standard Oil for AI is that value accrues to those that control bottlenecks, not to those involved in the (risky!) production of AI’s analog of crude oil (chips and energy). The key bottlenecks have varied over time: first chips, then memory and chip packaging, and now electricity and planning permissions, and investors that pile into AI’s ‘it’s all about X this year’ bottleneck tend to buy just as it stops being the key constraining factor!
Are you Ford or an airline?
So if AI is akin to oil, are you more likely to be Henry Ford or an airline?
Ford’s cars encouraged people to buy petrol; Ford profited from rising demand for oil without being exposed to swings in oil prices. By contrast, airlines are heavily exposed to swings in oil prices; when oil prices rise, so do operating costs, squeezing already thin margins (many airlines have gone bust for this very reason!). Unlike Ford, airlines don’t pass on the full impact of fuel cost increases to customers since fares are set weeks in advance.
You’re probably an airline if…
Your product is a service
You pay for each unit of service provided (tokens!)
Your customers could do what you’re offering themselves by accessing AI models directly
If this describes your business, then when token prices rise, so will your costs, squeezing margins and potentially putting you out of business. You’re also exposed to customer churn if larger customers negotiate better pricing with AI providers, or if those providers simply add your offering as a feature. Sound familiar? Sadly, this describes many AI startups founded in 2023-2026!
You’re probably more akin to Ford if…
You sell an outcome
That outcome is significantly more valuable than the tokens required to deliver it
You have an asset AI providers don’t (customers, data, brand, reputation, distribution, workflow, taste, etc.)
If this describes you, then you’re likely to welcome falling token prices! Not only will it reduce your costs, but it may increase demand for your product. Falling token prices will also have little impact on your customers’ willingness to pay for your offering.
Most people moan when the price of something they buy increases (and some even base their startup around securing a discount!). Founders should think of tokens like restaurants think of electricity: monitor costs, but focus on the food!
So…so what?
When oil was cheap, owning a machine meant having a leg up over anyone else doing the same job with nothing but muscle. Cheap oil meant that a guy with a diesel-powered backhoe could do the work of dozens of guys with shovels. It meant that farmers could get more done with tractors than they ever could with horses. Muscle mattered less and less, and the ability to operate increasingly complex machinery mattered more and more.
Compute is now being applied to cognitive labor in much the same way that (at one time) cheap oil was applied to physical labor. Employers will expect any work you do that’s easily distilled down into instructions for a computer, and can be accomplished by one for a few cents, to be replaceable by a few cents’ worth of compute. I’m not saying this to be alarmist; I’m saying it because I wish someone had said it to me sooner rather than later.
That said, just as oil-powered machines created more jobs than they destroyed, I suspect compute will too. And as with oil-powered machines, those jobs will disproportionately accrue to (a) those that operate said machines and (b) those that perform tasks made possible by said machines. Truck drivers, mechanics, airplane pilots, petrochemical engineers, interstate highway builders - all their jobs are downstream effects of cheap oil. I expect similar downstream effects from cheap compute. People who learn how to effectively delegate work to AI agents, oversee their progress, and take accountability for their (the agents’) outputs will have an advantage over people who shun the technology altogether, or naively copy-paste outputs into deliverables without so much as skimming them.
This week, someone reposted Karpathy’s tweet, “The hottest new programming language is English,” onto my timeline.
To laymen, this probably reads as good news for non-programmers, since they too can now “talk” computers into doing things for them. But it’s also sort of ominous, because if anyone can instruct computers to do stuff by simply telling them what to do in English, that diminishes the value of being able to instruct computers to do stuff by simply telling them what to do in English. Instead, value will accrue to those who (a) know what stuff needs doing and (b) can tell good computer-performed stuff from bad computer-performed stuff. There’s no assurance you’ll keep getting paid for work computers can do better and cheaper than you, but the good news is you get to choose what “stuff” you’re irreplaceably good at.
Some thoughts on positioning yourself (or your business) well in this new economy:
Know your costs. How much do your offerings cost you, in tokens? Many founders and executives I speak with don’t have a firm grasp on this, and are in for rude awakenings when their EAs or engineers come to them with bills.
Model (pun intended) out scenarios in which the cost of tokens either (a) doubles in an environment of high competition and thin margins, or (b) decreases by 90% (a likely outcome for businesses whose moat was access to models). Would your business still make sense?
Have something a model provider couldn’t easily offer their customers, even if they wanted to. Model providers can offer features, but they can’t offer your brand, your insights, your audience, your data, your track record, your customers’ trust.
Develop an intuitive understanding of what’s going on under the hood. I’m not saying everyone should learn how to train models from scratch, but having a basic understanding of why LLMs hallucinate, what determines how much a token costs you, how to tell when delegating a task to an agent will save you hours vs. cost you your afternoon, etc., will set you apart from your peers. People with an intuitive understanding of internal combustion engines thrived in the oil economy; people who simply bought gas at the pump thrived less.
Predictions I’d prefer you not take to Kalshi
By the end of 2027, I give it an 80% chance that the effect data centers have on electricity rates will be salient to voters in multiple states, and will thus be a salient topic in elections there (this is already starting to happen). Politicians respond to what’s salient to voters, and electricity rates are salient to voters.
By the end of 2027, I give it an 85% chance that at least two oil- or gas-rich countries will have announced compute megaprojects subsidized by their governments. If you’re a leader in a country with a comparative advantage in energy production, it wouldn’t take a genius to figure out who your potential customers are.
Through 2028, I give it a 75% chance that the cost of a given unit of model capability will continue to decrease by at least 5x per year. It’s been this way for years, and model providers are incentivized to compete aggressively on price.
By 2030, I give it a 70% chance that aggregate spending on compute will continue to increase year over year. I give it a 50% chance that at least one well-known company in the AI infra space will have gone through either an embarrassing write-down or fire sale. Already covered both of these horse races above.
Less than three years from now:
60% probability that futures contracts on GPU hours will be traded on one or more exchanges. Analysts are already calculating indices for H100 rental prices. For context, futures contracts on crude oil weren’t traded on the New York Mercantile Exchange until 1983, by which point oil had long since become a critical input into economic activity. Futures contracts tend to emerge once enough market participants are interested in hedging the price of something they don’t control. We’re approaching that inflection point with compute.
65% probability that the binding constraint on building out compute infra in the US will be generation capacity and grid-related hardware, not chips. Some builders have already been vocal about this. Semiconductor fabs seem to scale up faster than electrical utilities.
Is compute the new oil?
Yes, if by “oil” you mean “a critical input into nearly every sector of the economy, produced by a handful of suppliers that take years and enormous capex to scale up, and thus likely to be the subject of geopolitical strife.”
No, if by “oil” you mean “something that appreciates in value while sitting in the ground, the suppliers of which are guaranteed to capture value as a result.” Compute will likely keep getting cheaper, thanks to talented engineers. Compute is produced with the express purpose of producing other things. Compute is more constrained by electrons than by any other element.
Those investors who can keep both halves of that answer in their heads at once will make better investments. Founders who can figure out whether they’re building Fords or airlines will build better businesses. Workers who figure out how to best complement AI will have better careers than workers who demonize or deify it.
I think about my monthly compute spend the same way my grandparents thought about their monthly fuel spend. I’m fine with the total dollar amount going up each month, as long as the amount of work I can get done on that compute goes up by more.
🍕 I’m JA Westenberg. I write essays about technology, philosophy, productivity, and what it means to be human. If you’re an optimist, you’re in the right place. You can read my AI disclosure here
Work with me:





