Betting $400 Billion on the Weather
I try to keep this blog pointed at technology, and away from the things I am merely passionate about, on the grounds that nobody subscribed to hear my views on energy policy. Every so often a subject crosses the line far enough that following it becomes the technology story rather than a digression from it. This is one of those, so I am going to ask your indulgence for a post that starts with electricity and ends up, as these things tend to, back at what we are actually committing to when we sign a contract.
The question I would put to anyone building in this space is a simple one. How would you like to bet four hundred billion dollars on the weather?
Not on whether it rains tomorrow. On whether the grid, the transmission network, the substations, the transformers, the cooling and the rest of the physical apparatus required to turn that money into productive compute will all be working on the day you need them. Because that is increasingly what the AI infrastructure trade looks like, and the uncomfortable part is that you do not need an AI winter, a breakthrough out of China, a rogue superintelligence or a spectacular financial collapse to make the model creak. You just need the grid to behave the way old grids have always behaved.
The happy story, which is genuinely a good one
I want to be fair to the bull case, because it is not a fantasy and the numbers behind it are real.
Oracle reported $19.3 billion of revenue in Q1 FY2027, with cloud infrastructure up 121% year on year, remaining performance obligations of $664 billion after more than $30 billion of additional AI cloud contracts booked in the quarter, 850 MW of new data centre capacity delivered and more than 300,000 GPUs handed to AI cloud customers. That is not imaginary demand and it is not an accounting trick. It is an enormous commercial opportunity that a lot of very capable people have worked very hard to win.
The difficulty is not the demand. It is the part between the contract and the cash, because the customer does not buy a remaining performance obligation. The customer buys compute, and compute requires electricity in quantities that make the rest of the chain look trivial. Demand becomes a contract, the contract justifies capital expenditure, the expenditure becomes a data centre, the data centre needs a grid connection, the connection delivers electricity, the electricity feeds the GPUs, the GPUs produce compute, and only then does any of it become revenue and eventually cash. There is a rather important dependency sitting in the middle of that sequence, and it is not one that any of the companies involved control.
Now roll the dice
Picture a finished AI campus. The buildings are up, the GPUs are racked, the customer is ready, and the transmission upgrade is late. Nothing has gone catastrophically wrong and nobody has been negligent; the project simply took longer than the plan assumed, which is the single most ordinary thing that happens in infrastructure. The customer waits. The capital does not wait, the financing does not wait, the construction contracts do not wait, the land does not wait, and the interest bill certainly does not wait. The revenue just arrives later.
Roll again and the connection works, but an extreme heat event pushes regional demand towards its limit and the site is curtailed rather than destroyed, so the customer simply receives less compute than they contracted for. Roll again and a transformer fails. Again and a storm takes out transmission. Again and a wildfire closes a corridor. Again and a major grid upgrade runs over budget and behind schedule.
None of those is exotic, and that is precisely the point. The risk is not that America suffers a once-in-a-century civilisational event. The risk is that thoroughly ordinary physical events interact with an infrastructure system that is already being asked to do something extraordinary.
The US Department of Energy now describes the need for new transmission as pressing, citing hyperscale AI data centres, new large loads, generation interconnection and congestion by name. S&P Global reported that congestion is intensifying as data centres add to existing constraints, with renewable curtailment across the major US markets running 17.5% higher through July 2026 than the same period in 2025, and Texas alone recording 7.9 million MWh of curtailment and $2.4 billion of congestion cost in 2025. What those numbers describe is not a national shortage of electrons. It is a shortage of the infrastructure required to get the electrons to the right place at the right time, which is a different problem with a much longer lead time.
The grid is the dependency nobody puts on the diagram
This is the part of the AI story that gets remarkably little attention, and I think it is because it sits just outside the boundary of the system we are used to drawing.
A hyperscale data centre is an extraordinary piece of engineering. Redundant cooling, redundant networking, UPS, backup generation, multiple fibre paths, sophisticated monitoring, thousands of GPUs, all of it designed by people who think carefully about failure. And upstream of every bit of it there is still a transmission corridor, a substation and a transformer that simply have to work. You can make the data centre as redundant as you like. You cannot make the regional grid redundant by adding another UPS to the server room, and no amount of architectural diligence inside the fence changes what is happening outside it.
The scale of what is being asked is the part that stops you. Texas now has more than 470 GW of data centre and other large load projects seeking grid connection, which is over five times the state's peak demand, and the state has responded by halting new data centre permits pending a grid audit. That is not evidence that the AI industry is about to collapse. It is evidence that the physical system has become a constraint on the growth model, which is a far more interesting problem and a much harder one to engineer around.
And then there is the money
Oracle spent $28.5 billion of capex in a single quarter and reported roughly $5.4 billion of negative free cash flow in Q1 FY2027. Annualise the capex and you are north of a hundred billion dollars a year.
Run the happy story and it works beautifully. Projects land on time, connections are ready, electricity is available, customers consume what they contracted for, revenue compounds, and the capital turns into cash-generating infrastructure. Now change one assumption and let the projects take six months longer, so billions sit in infrastructure that is not yet producing the expected revenue. Change another and let financing costs rise. Another and let construction costs rise. Another and delay a transmission upgrade. Another and let an extreme weather event temporarily reduce available capacity.
No single one of those kills the model, and I am not suggesting it does. Together they change the economics, and they do it quietly, over a period long enough that nobody has to make a decision about it. Which is why I do not think the interesting question is whether Oracle's $664 billion RPO is real. It plainly is. The question worth asking is how much capital has to be deployed, and for how long, before that number becomes cash.
Oracle is the clearest case, not the only one
I have used Oracle throughout because its reported numbers make the relationship between the physical and the financial unusually legible, not because it is uniquely exposed. The same structural issue runs through the entire stack.
Hyperscalers need power. Specialist GPU clouds need power. Data centre developers need power. Chip manufacturers need factories, and the factories need power. Cooling needs power and, depending on the design, water. Transmission needs investment, generation needs investment, transformers need manufacturing capacity, and construction needs steel, copper, aluminium, concrete, labour and diesel. Nearly all of it needs financing. The AI industry has quietly commissioned an enormous physical infrastructure programme underneath something that is still being discussed, and valued, primarily as a software story.
Where the weather gets a vote
We are used to treating weather as an operational nuisance. A hurricane, a heatwave, a wildfire, a flood, a winter storm; things you harden against and then recover from.
Infrastructure finance does not work that way, because finance works on a clock. Delay a project six months and the capital does not disappear, it just sits there. Delay it twelve and it sits there longer. If electricity availability is lower than assumed, utilisation falls, and if utilisation falls while the debt stays fixed, the economics deteriorate without anyone doing anything wrong. The building does not care that the weather was unusual, and neither does the bondholder, the interest bill or the customer contract. That asymmetry is the whole argument: physical infrastructure operates in the real world, and financial obligations operate on a schedule.
None of which means the utilities are asleep
I want to be careful here, because it would be easy to read all of this as a complaint that nobody has been investing, and that is not true.
EIA data shows real annual utility spending on electricity infrastructure rose from $287 billion in 2003 to $320 billion in 2023, with transmission spending close to tripling and distribution investment up substantially. Utilities have been replacing ageing equipment, hardening against storms and fires, adding generation and upgrading networks throughout.
The difficulty is that the same investment now has to serve several purposes at once: maintain the existing system, replace what is worn out, harden it against worsening weather, connect new generation, build new transmission, connect manufacturing, support electrification, and now absorb enormous AI loads as well. That is a materially different proposition from the claim that America needs more electricity. America needs more reliable electricity, in specific places, at specific times, delivered through infrastructure that takes years to permit and build. The distinction sounds pedantic right up until you are the one waiting on a connection.
The boring questions are the ones that matter
The industry has become very good at telling stories about risks several steps removed from the infrastructure it is actually building. What happens if a model escapes, what happens if China catches up, what happens if the next system is vastly more capable, what happens when an agent changes the world. Those are genuinely interesting and some of them may turn out to matter enormously. They also make spectacular headlines, which is not nothing.
The unglamorous questions get far less airtime. Who pays for the transmission line, the transformer, the late substation upgrade, the grid reinforcement. Who pays when demand arrives faster than the infrastructure, when a community's electricity costs rise, when a facility has to run on backup generation, when the project is simply delayed. And who absorbs the consequences when the physical assumptions underneath hundreds of billions of dollars of investment turn out to be wrong.
This is where the phrase about socialised costs and privatised profits starts to get uncomfortable, and I want to be accurate rather than rhetorical about it. It does not follow that every AI project is being subsidised by everyone else. Oracle says explicitly that it will fund the grid upgrades it needs or build its own generation for new AI data centres, and its Project Jupiter design includes behind-the-meter generation, which is precisely the responsible answer. But the broader question survives the specific case. When an industry suddenly requires infrastructure at a scale that moves electricity markets, transmission planning, water, roads, land, construction capacity and public policy, the boundary between private investment and public infrastructure stops being obvious, and somebody eventually has to decide where it sits.
Which makes this a rather beautiful experiment
Oracle has the demand, the customers, the contracts, the GPUs, the buildings and the capital programme. What it cannot manufacture on its own timetable is the physical environment all of that has to operate in, and the company is candid enough to list risks around its ability to anticipate, plan for, secure and manage data centre capacity.
So the useful question is not whether AI is real, because it obviously is, and it is not whether AI will beat us, because nobody knows. It is something considerably more mundane. How much are we willing to spend building an AI economy that depends on infrastructure being built at the same time, at extraordinary speed, by people who do not work for us.
At some point the most consequential model in the world may not be running on an NVIDIA GPU at all. It may be running in the control room of a regional grid operator, deciding which load gets the electricity tonight.
If you are putting hundreds of billions into AI infrastructure, you are not only betting on artificial intelligence. You are betting on copper, transformers, transmission, construction schedules and interest rates, and on hurricanes, heatwaves, wildfires and storms. Which is to say you are betting on the weather, and the weather has never once cared what was in the contract.