Cummins AI Data Centres : 55 Gigawatts And A Prayer

Cummins does not make artificial intelligence. It does not train large language models, manufacture GPUs or have Sam Altman wandering around Columbus asking whether anybody has seen another trillion dollars down the back of the sofa. What Cummins does have is something the AI boom desperately needs: power. Lots of it. Specifically, enormous diesel and gas generator sets capable of keeping data centres alive when the grid coughs, falls over or simply cannot provide enough electricity in the first place.

And business is fucking magnificent. Cummins says its full-channel data-centre revenue grew to around $3.5 billion in 2025. For 2026, the midpoint of its guidance puts that at roughly $5 billion. By 2030, Cummins wants more than $9 billion. To service that demand, it has announced another $450 million of investment in high-horsepower engine and generator capacity, adding 20 gigawatts and taking planned capacity to 55 gigawatts by 2030.

Fifty-five gigawatts and a prayer, because underneath all those generators, server halls, substations and forecasts sits one fairly important question: what happens if the buildings don’t come?

The Power Behind The Machine

Forget the finance jargon for a minute. A modern AI data centre is a staggeringly expensive warehouse full of computers that eat electricity like a pissed-up stag party attacks a kebab shop. The chips cost fortunes, cooling them costs fortunes, the buildings cost fortunes, and transformers, switchgear, grid connections, land, cabling and backup systems pile billions more onto the invoice before the first clever answer appears on anybody’s screen.

Then comes another minor inconvenience: the power cannot go off. That is where Cummins walks through the door carrying the diesel. Standby generators are the industrial insurance policy bolted to the outside of the building. If grid power disappears, Cummins wants its engines waiting there to roar into life before several billion dollars of computing equipment starts having a very bad afternoon.

It is lucrative work. In the second quarter of 2026, Cummins’ Power Systems business recorded $2.3 billion of sales, up 19%. EBITDA hit $552 million, giving the segment a 24.5% margin. Cummins attributed the increase primarily to power-generation demand, particularly data centres, while its Distribution business is getting a piece of the feast too. This is not some experimental side project operating from a cupboard behind the engine factory. Data-centre power has become one of the hottest parts of Cummins.

Nine Billion Dollars Is Not A Side Hustle

Here is where the numbers start getting interesting. Cummins generated $33.7 billion of total revenue in 2025, with data centres representing roughly $3.5 billion of full-channel revenue. The company now wants overall revenue of around $45 billion to $50 billion by 2030, with more than $9 billion coming from data centres.

Strip away the investor-day perfume and that means something rather simple. If Cummins hits the lower end of its 2030 company target, roughly one dollar in every five of revenue could come from the data-centre market. More revealing still is the growth. Between $33.7 billion in 2025 and $45 billion to $50 billion in 2030, Cummins is asking investors to envisage roughly $11.3 billion to $16.3 billion of additional annual revenue. Data centres alone need to add about $5.5 billion over the same period for that part of the plan to land.

Do the crude arithmetic and data centres account for somewhere between roughly one-third and one-half of the additional revenue Cummins expects to add by 2030. That is not a footnote. That is a fucking load-bearing wall.

Enter Ed Zitron With A Crowbar

Ed Zitron has spent years arguing that the generative-AI economy contains a rather ugly contradiction. The technology industry is spending astonishing amounts of money building computing infrastructure, while the revenue required to justify all that spending remains a much harder bastard to find. His argument is not that artificial intelligence has no users or no value. It is that usefulness and economic sustainability are not the same thing.

A restaurant can be packed every night and still go bankrupt if every £20 meal costs £35 to serve. That is the basic smell coming from Zitron’s AI analysis. He has specifically asked what happens if organisations cannot keep spending, hyperscalers cannot build quickly enough, GPUs sit uninstalled, financing becomes strained or data-centre construction demand collapses.

That last phrase matters enormously to Cummins, because Cummins does not need AI to disappear. It needs infrastructure to keep multiplying. A planned server campus might represent thousands of processors, acres of concrete, cooling towers, transformers and banks of standby generators. Delete the campus from the spreadsheet and the Cummins generators disappear with it. No apocalypse required, just fewer purchase orders.

The Numbers Above Cummins Are Getting Fucking Weird

You do not have to subscribe to Zitron’s darkest interpretation to find the wider economics eyebrow-raising. The Financial Times recently examined estimates putting US hyperscaler capital expenditure at around $800 billion in 2026 and $1.1 trillion in 2027. The analysis suggested approximately $300 billion of annual AI revenue would cover only a basic economic break-even on that investment, with around $636 billion annually required to produce a return comparable with the hyperscalers’ historical economics.

That is a lot of subscriptions to a chatbot that occasionally tells you there are three Rs in “strawberry”. Meanwhile, Moody’s says the five largest hyperscalers carry around $2.8 trillion of commitments that do not currently appear as liabilities on their balance sheets, while highlighting the complexity and limited transparency surrounding the arrangements.

Read that slowly. The companies driving this infrastructure boom are attaching trillions of dollars of future commitments to an industry still arguing over how much money the end product can actually make. Cummins sits several floors below that argument. It does not need to prove that generative AI deserves the spending. It needs the spending to continue.

What If The Concrete Never Gets Poured?

This is where the story gets beautifully grubby. Cummins is expanding physical manufacturing capacity because it sees physical demand for physical generators intended for physical buildings. But an announced data centre is not a completed data centre. A PowerPoint slide does not need backup power, a press release does not need a QSK95, and a proposed 500-megawatt campus sitting in a planning document will consume precisely fuck all diesel if nobody ever pours the concrete.

Goldman Sachs estimates that only around 50% to 60% of US data-centre capacity currently scheduled for the next one to two years will actually come online on time, with delays and cancellations among the reasons. Another recent model from Columbia Business School professor Stijn Van Nieuwerburgh goes considerably further. Using the enormous US project pipeline, it estimates that developers could ultimately abandon around 45% of currently planned capacity.

That latter figure is a model, not holy scripture carved into a transformer, but imagine being Cummins and looking down the production schedule. Your strategy says data-centre revenue marches from $3.5 billion to $5 billion to more than $9 billion. Factory expansion follows that expectation, while customers place orders for the largest machines years ahead. Then projects start slipping, substations take another year, financing packages stop working, hyperscalers trim capex and speculative campuses quietly disappear.

Nothing has to explode. Nobody needs to ring a bell and officially declare an AI bubble burst. Enough small deletions from enough construction schedules eventually arrive at the factory gate as missing demand.

“Orders Into 2028” Sounds Lovely

Cummins has a strong answer to the sceptics: demand is real. In September, Cummins Investor Relations executive Nicholas Arens said the company was taking orders for its 95-litre solution into the second half of 2028. Cummins remains confident in its $9 billion-plus exposure to the space by 2030, largely underpinned by diesel standby power.

That sounds reassuring, and it may well be reassuring. But TCAP enjoys reading the paperwork after the conference-room applause has died down, and something interesting happened between Cummins’ 2024 and 2025 annual reports. The 2024 10-K said Cummins had “firm orders from data center and electrolyzer customers”. The 2025 10-K instead said it had “strong demand for our data center products extending out six to eight quarters.”

“Firm orders” disappeared.

Perhaps there is a perfectly mundane explanation. Wonderful. What is it? The surrounding backlog disclosures matter because Cummins explains that a large portion of its wider business operates through open purchase orders. Historically, customers could cancel many of those orders on reasonable notice without paying cancellation charges, so Cummins does not treat them as firm.

None of that proves Cummins’ current data-centre backlog can simply evaporate without penalty. It does explain why the vocabulary matters. “Strong demand”, “order board”, “visibility”, “framework agreement” and “orders into 2028” all sound reassuring when projected six feet high behind an executive. They are not necessarily synonyms for money the customer is contractually trapped into paying.

Show Us The Bolts Holding It Together

Cummins does have substantial protected business. At 30 June 2026, it reported $7.6 billion of remaining performance obligations under longer-term arrangements, including contracts that customers cannot simply cancel without consequences or that impose cancellation penalties. Good. That is real contractual substance.

Unfortunately, Cummins does not tell investors how much of that $7.6 billion comes from data-centre orders. And therein lies a lovely little information gap. Management says data-centre demand stretches years into the future, describes multi-year customer arrangements and expects more than $9 billion of data-centre revenue by 2030. Executives also tell investors that customers continue placing orders for the company’s biggest machinery.

What the public disclosure does not let us do is draw a neat line from that giant future revenue ambition to the amount of data-centre business customers are actually financially locked into. So here is a perfectly reasonable question: of the data-centre order book underpinning Cummins’ $9 billion-plus 2030 target, what proportion is non-cancellable or carries economically meaningful cancellation penalties?

No crystal ball required. Just tell shareholders how much of the scaffolding is actually bolted down.

AI Doesn’t Need To Die

This is perhaps the part most easily missed. Cummins does not require an AI apocalypse to get hurt. OpenAI does not need to collapse in a shower of unpaid invoices, NVIDIA does not need to discover its chips are useful only as elaborate drinks coasters, and ChatGPT does not need to vanish.

The growth rate merely has to disappoint.

Suppose Cummins reaches 2030 and data-centre revenue is still around today’s expected $5 billion rather than more than $9 billion. That leaves approximately $4 billion of anticipated annual revenue needing to come from somewhere else if the broader company target is to remain intact. A contraction in the data-centre market would widen the gap further.

Power Systems currently produces margins well above Cummins’ company average, so this is not necessarily just a question of losing sales. Take away enough of the good stuff and the profit mix gets uglier too. The future does not have to collapse. It merely has to turn up smaller than advertised.

Cummins Has Seen A Future Arrive Late Before

There is another reason Cummins should perhaps keep the bottle of champagne corked: hydrogen. Cummins spent years presenting hydrogen and electrolyzers as part of the glorious multi-solution future. Then adoption expectations changed, demand weakened and in 2025 Cummins recorded $458 million of charges associated with its electrolyzer business, including goodwill impairment, inventory write-downs, property and equipment impairments, contract termination costs and severance.

Cummins ultimately said it intended to stop new commercial activity in the electrolyzer space after the market deteriorated. Markets change, forecasts change and technologies arrive more slowly than PowerPoint decks suggested. Companies “pace” capital, accountants impair assets, management language migrates from revolution to discipline and suddenly everyone speaks as though some bloke who left three restructurings ago wrote the previous brochure.

AI data centres are obviously not hydrogen electrolyzers. The economics and customer base are different. The lesson is simpler: Cummins itself has recent experience of building strategy around a technology-adoption curve that subsequently bent in the wrong fucking direction.

This Isn’t A Balance-Sheet Suicide Pact

There is an important limit to the attack. Cummins has not shoved the company pension pot through a shredder. The extra manufacturing investment announced for this expansion is around $450 million, not $9 billion. Its high-horsepower factories can produce engines for mining, marine, oil and gas and other power-generation customers, and Cummins specifically emphasised this manufacturing flexibility when questioned at Analyst Day.

So if AI infrastructure cools, Cummins possesses something rather more useful than a warehouse full of obsolete GPUs. It owns factories that can make other enormous machines for other enormous dirty industries. That matters.

The wager is therefore not survival. The wager is growth. Cummins is telling investors that data-centre revenue can grow beyond $9 billion, helping drive a company targeting $45 billion to $50 billion of sales and EBITDA margins above 20%. That is materially different from saying Cummins goes bankrupt if Claude stops writing everyone’s emails.

But a growth story does not have to kill a company when it fails. It merely has to disappoint the market that paid for it.

The Generator At The End Of The Chain

Cummins occupies a fascinating position in the AI boom. At the glamorous end are models, GPUs, hyperscalers and breathless keynote speeches about intelligence changing civilisation. Travel far enough down the chain and eventually somebody has to build an actual shed. Another contractor must connect that shed to electricity, while somebody else has to make sure the electricity does not disappear.

That is where Cummins lives.

For now, Cummins can remain wonderfully removed from the argument about whether somebody’s AI assistant will ever generate enough revenue to justify the infrastructure sitting behind it. Economics eventually travel backwards, though. Weak AI returns can make capex committees nervous; nervous capex committees delay projects; prolonged delays can kill planned developments; and enough dead developments eventually leave fewer generator orders at the Cummins factory gate.

Suddenly the company promising more than $9 billion of annual data-centre revenue finds its growth machine depending upon a chain of assumptions stretching from a Cummins production line all the way to somebody convincing investors that the next hundred-billion-dollar AI buildout will eventually pay.

Cummins calls its expansion “aligned to visible customer demand.” Ed Zitron looks beyond that demand and asks what financial machinery is creating it. TCAP has a simpler question: if the AI building boom develops cracks, how much of Cummins’ $9 billion future is standing on the concrete above them?

The casino can stay in the title.

The real story is what happens when the builders stop turning up.

Lee Thompson – Founder, The Cummins Accountability Project

Cummins Confidential : 55 Gigawatts And A Prayer – Cummins Takes A Seat At The AI Casino