GenAI hyperscalers’ Q2 reports prove one thing: spending must go up

Over the past week, Meta, Alphabet, Microsoft and Amazon reported results covering the second calendar quarter of 2026. There was plenty of revenue growth, cloud boasting, AI usage statistics and executive waffle about unprecedented demand, but one particularly funny detail kept appearing across every call.

These are trillion-dollar companies employing some of the best-paid financial, technical and infrastructure experts on the planet, yet none of them can establish a credible ceiling for what this shit is going to cost.

Every forecast is temporary. Every estimate is revised upwards. We are not talking about accidentally overlooking the office biscuit budget either. The lower end moves by billions of dollars at a time, sometimes tens of billions, and management still emerges three months later to announce that demand is even stronger, capacity is even tighter and spending must rise again.

Meta began 2026 expecting capital expenditure of $115 billion to $135 billion. It later increased that to $125 billion to $145 billion, and has now raised the minimum again to $130 billion while keeping the $145 billion maximum. Alphabet started at $175 billion to $185 billion, moved to $180 billion to $190 billion after one quarter, then jumped to $195 billion to $205 billion after the next. Amazon lasted six months with its approximately $200 billion estimate before casually adding another $20 billion.

Microsoft has been slightly cleverer with the presentation. Its underlying 2026 investment expectations remain unchanged, but it extended the estimated useful lives of data centres and office buildings from 15 to 25 years. This means more future data-centre leases will be classified as operating leases instead of finance leases, moving them outside the company’s reported capital expenditure figure. The physical investment has not shrunk, but the accounting column containing part of it has changed. Microsoft still expects capital expenditure to grow again during its next financial year.

Using the current figures, these four companies are preparing to spend roughly $730 billion during 2026. That number will probably be obsolete before some of the data centres it supposedly funds have planning permission.

Every spending forecast expires before the next earnings call:

There is something genuinely deranged about the consistency of these revisions. Management repeatedly claims to possess excellent visibility into customer demand, long-term contracts, capacity requirements, hardware orders and expected returns. Amazon says many AI contracts last at least five years. Microsoft has hundreds of billions of dollars in remaining performance obligations. Google says it takes a multiyear view of infrastructure requirements. All of them insist that demand exceeds supply.

Yet apparently none can predict the cost of satisfying that visible, contracted, multiyear demand for longer than a quarter or two.

The explanation is always another variation of the same story. Demand accelerated. Deliveries were brought forward. Memory prices increased. Customers consumed newly installed capacity faster than expected. More third-party compute was needed while internal facilities were constructed. The company became even more confident about the opportunity after spending another few dozen billion dollars on it.

Google’s latest increase is especially ridiculous. Its original 2026 range of $175 billion to $185 billion became $180 billion to $190 billion in April, then $195 billion to $205 billion in July. The company added between $15 billion and $25 billion to the forecast in three months, then warned that spending would increase “significantly” again during 2027. Alphabet is already using more expensive third-party capacity as a bridge because its own enormous construction programme remains insufficient.

Amazon increased its estimate from $200 billion to $220 billion, largely blaming higher memory costs, while saying it still expects insufficient capacity during 2026 and probably 2027. Amazon claims it is already seeing strong demand extending into 2028. Evidently it can see customer requirements several years into the future, but the price of the memory needed to serve those customers appeared from behind a bush carrying a $20 billion invoice.

This is less a collection of financial forecasts than a quarterly ritual. Announce an already obscene number, spend towards it, discover that every competitor is doing the same thing, then raise the number because slowing down first might imply that somebody else is winning the imaginary race to machine god.

No executive wants to be remembered as the sensible adult who preserved cash while a competitor supposedly built AGI. It is professionally safer to waste another $20 billion alongside everyone else than risk being the only company that refused.

Meta’s advertising machine is being fed into the furnace:

Meta provides the weakest direct case for the spending because its disclosed business remains almost entirely advertising.

The company generated $60.8 billion of revenue during the quarter, of which $59.4 billion came from advertising. Costs and expenses rose 55% to $42 billion, operating profit fell from $20.4 billion to $18.8 billion, and Meta spent $31.1 billion on capital expenditure and finance-lease principal payments.

After generating $31.9 billion of operating cash flow, it finished with $784 million of free cash flow, down from $8.5 billion a year earlier. Reality Labs supplied its usual contribution by losing another $4.6 billion on $431 million of revenue.

There were sizeable legal and severance charges during the quarter, so the deterioration was not entirely caused by AI. That does not fix the cash-flow problem. Meta has now raised the bottom of its capital expenditure forecast twice while continuing to fund Reality Labs, expensive AI hiring and the broader superintelligence fantasy from the same Facebook and Instagram advertising engine.

Meta can claim that AI improves recommendations, engagement and advertising performance, and there is obviously some truth to that. Better ad ranking has economic value. What it cannot show is a clean standalone AI business remotely proportionate to the infrastructure being built in its name.

The company is spending like a cloud provider while earning like an advertising platform. Its AI monetisation story consists largely of making the existing adverts work better today, followed by vague possibilities involving subscriptions, business agents, APIs and perhaps renting excess compute tomorrow.

This is the company asking investors to accept up to $145 billion of annual capital expenditure while almost every dollar of current revenue still arrives from adverts placed beside posts, videos and photos. Zuckerberg already has one industrial-scale money furnace called Reality Labs. He has now positioned another beside it and connected both to the same advertising pipe.

Alphabet had a great quarter and still spent more cash than it generated:

Alphabet is more complicated because its actual operating performance was excellent.

Revenue rose 24% to $119.8 billion, operating income increased 30% to $40.8 billion, and Google Cloud revenue rose 82% to $24.8 billion. Cloud operating income more than tripled to $8.8 billion, producing a 35.6% operating margin. Search revenue also continued growing strongly.

Google is not collapsing. The underlying businesses are working extremely well. The problem is that even an outstanding quarter from one of the greatest cash-generating companies ever built could not keep pace with its infrastructure bill.

Alphabet generated $39.1 billion of operating cash flow and spent $44.9 billion on capital expenditure, leaving negative free cash flow of $5.9 billion. It then raised the full-year capex forecast to as much as $205 billion and warned that higher depreciation, energy use and data-centre operating costs would continue pressuring the income statement. Spending will rise significantly again in 2027.

The reported profit figure also arrived wearing almost $100 billion of accounting makeup. Alphabet recorded $98 billion of other income, primarily from unrealised gains in its equity securities portfolio. Net income therefore exploded upwards because private and public investments were marked higher, not because Google suddenly discovered another $98 billion of cash beneath a sofa.

The actual operations were already impressive. There was no need to confuse them with imaginary investment wealth, especially while the real cash was disappearing into data centres faster than the company generated it.

Alphabet can at least demonstrate meaningful direct AI and cloud revenue. Cloud is growing rapidly, TPU systems are being sold to customers, subscriptions are benefiting from paid AI plans and Google owns the entire stack from custom chips through to Gemini, Search and enterprise distribution.

It still cannot show where the spending ceiling sits. A forecast raised twice in six months is not a forecast. It is a placeholder management updates whenever the latest bill arrives.

Amazon has the strongest demand story and still cannot produce free cash flow:

Amazon delivered the strongest evidence that customers genuinely want the infrastructure.

AWS revenue rose 37% to $42.2 billion, its fastest growth in 18 quarters, while AWS operating income reached $16.6 billion. Amazon’s total revenue rose 20% to $200.6 billion and operating income increased 43% to $27.5 billion. Management says both its AI business and custom chip business have exceeded annual revenue run rates of $25 billion.

Those are actual products, actual customers and actual cloud revenue. Amazon’s case is substantially stronger than Meta promising that somebody may eventually pay a subscription for superintelligence.

It also makes the cash-flow result harder to ignore.

Amazon spent $53.1 billion in cash capital expenditure during the quarter, primarily on AWS and generative AI. Trailing 12-month operating cash flow reached $161.4 billion, yet trailing free cash flow swung from a positive $18.2 billion to an outflow of $7.6 billion because property and equipment purchases increased by $66.1 billion.

AWS is booming. Advertising is growing. Retail operations are improving. Amazon is generating more operating cash, and the infrastructure programme is consuming it even faster.

Then there is the quarterly net income of $62.6 billion, an apparently magnificent figure containing $53.4 billion of non-operating pre-tax income, primarily from Amazon’s investment in Anthropic. The real operating income was already a healthy $27.5 billion, but the headline profit was transformed into something completely absurd because a privately owned AI laboratory received a higher theoretical value.

Amazon therefore supplied both the strongest defence of the AI infrastructure boom and one of the best demonstrations of its financial insanity. AWS demand is real, but the cost of serving it has pushed free cash flow negative, the spending estimate has risen by $20 billion, and reported profit was flattered by a vast gain on a company that Amazon itself is financing.

Apparently this was all wonderful news. Amazon’s shares jumped because investors decided that AWS growth made this the good kind of cash incineration.

Microsoft is winning because its old monopolies can absorb more punishment:

Microsoft is the designated winner of the quarter. Azure revenue grew 43%, Microsoft Cloud revenue reached $59.3 billion, total quarterly revenue reached $90 billion, and the company remained comfortably free-cash-flow positive.

After spending $41 billion on capital expenditure, Microsoft still produced $19.6 billion of free cash flow. That is considerably healthier than Amazon, Alphabet or Meta. Office, Windows, enterprise software and Azure form an extraordinary financial fortress.

This does not mean Microsoft has proven that GenAI possesses magnificent economics. It means Microsoft owns the strongest collection of existing businesses available to subsidise it.

The old Microsoft business model was almost offensively good. Build software, duplicate it at negligible cost, lock companies into the ecosystem and collect recurring payments forever. GenAI partially reverses those economics. Every Copilot request consumes compute, memory, networking and electricity. Greater customer usage creates another real bill that apparently came as a surprise to MS.

Microsoft’s cloud gross margin fell to 65%, driven by the shift towards Azure, continuing AI infrastructure investment and increased product usage. Microsoft 365’s gross margin was also slightly reduced by growing Copilot usage. GitHub Copilot consumption became expensive enough that Microsoft changed the business model in June to align pricing with usage, after which margins improved during the quarter.

That little detail says more than another hour of Satya Nadella describing agents. Customers using the product created enough additional cost that Microsoft needed to change how it charged them.

The company will tell investors that Microsoft 365 Copilot now has more than 30 million paid seats. It does not disclose standalone Copilot revenue or profit. We get user counts, token volumes, adoption statistics and carefully selected examples of GPU efficiency, but no clean answer to the obvious question: after allocating the full infrastructure cost, how profitable are these products?

Azure itself is also more dependent on frontier AI laboratories than the reassuring presentation initially suggests. Microsoft said nearly 90% of its full-year $214 billion Microsoft Cloud revenue came from customers outside frontier-model companies.

That does not mean 90% of Azure profit comes from normal customers, as the figure has sometimes been interpreted. Microsoft does not disclose Azure profit, and Microsoft Cloud includes Microsoft 365 and other products. What it does imply is that roughly a tenth of Microsoft’s entire cloud operation, somewhere around $20 billion or more of annual revenue, may already come from a very small collection of frontier-model companies.

Because those laboratories overwhelmingly purchase Azure compute rather than Office subscriptions, their share of Azure revenue could be materially higher than 10%. Microsoft does not provide enough information to calculate it precisely, which is presumably why it chose the broader Microsoft Cloud denominator.

OpenAI has contracted to purchase an additional $250 billion of Azure services. Anthropic has committed to purchase $30 billion of Azure compute capacity, while Microsoft agreed to invest up to $5 billion in Anthropic. Microsoft’s commercial backlog reached $678 billion, but growth was only 25% when OpenAI was excluded, compared with 84% on the reported figure.

This is the strongest company in the group, and a meaningful portion of its great cloud success is already connected to laboratories that it finances, invests in or depends upon for products.

The rest of Microsoft is being expected to maintain the appearance of corporate discipline while Azure swallows progressively more capital. Total headcount fell 2% over the year. Xbox revenue fell 10%, with content and services also down 10%, while Microsoft recorded Xbox impairment and severance charges. The company had announced another Xbox restructuring on 6 July, after the quarter closed but before reporting the results.

The quarter then received a $3.2 billion gain from Microsoft’s investment in Anthropic, partly offsetting those Xbox charges and other severance expenses. Microsoft invested real money into Anthropic, Anthropic promised to spend $30 billion on Azure, Microsoft recorded the resulting cloud commitment, Anthropic’s private valuation increased, and Microsoft booked a paper gain on its investment.

Workers and gaming assets produced real charges. Anthropic produced imaginary profit. The imaginary profit won.

Microsoft also extended the estimated useful lives of data centres and office buildings from 15 to 25 years. The company says this will have only a minimal effect on FY2027 operating income, but it will shift more future data-centre leases from finance leases into operating leases. Finance leases count towards reported capex; operating leases do not.

Microsoft’s underlying investment expectation remains the same. The revised presentation simply lowers the capex figure to approximately $175 billion by moving more of the spending outside that particular metric.

It is an excellent example of what the supposed winner now requires. Higher spending, lower cloud margins, usage-based pricing to recover inference costs, repeated restructuring elsewhere, fewer employees, weaker consumer businesses and an accounting change that makes part of the same infrastructure programme appear less capital-intensive.

Microsoft is not proving that GenAI has restored software’s old high-margin economics. It is showing how long the world’s greatest enterprise software monopoly can prevent those economics from collapsing visibly.

The OpenAI and Anthropic private valuation machine:

The most ridiculous part of the whole system is how private valuations, cloud commitments and capital expenditure continually validate one another.

Amazon has invested $13 billion in Anthropic and established a facility making up to another $20 billion available as Amazon reaches compute-capacity delivery milestones. Anthropic has committed more than $100 billion over ten years to AWS, securing up to five gigawatts of capacity using Amazon’s chips.

Amazon builds the capacity, reaching delivery milestones that make more Amazon money available to Anthropic. Anthropic can then use capital from Amazon and other investors to pay AWS. AWS records the revenue and backlog, which Amazon presents as evidence of overwhelming AI demand. That demand justifies another $20 billion increase in Amazon’s capital expenditure.

Anthropic’s next private funding round values it more highly. Amazon marks up its existing investment and records tens of billions of dollars of non-operating income. That enormous paper gain makes Amazon’s reported earnings look wonderful while its free cash flow is negative because it is constructing the infrastructure Anthropic agreed to consume.

There is your fucking perpetual motion machine.

Amazon has now constructed a similar arrangement with OpenAI. It invested an initial $15 billion and agreed to purchase another $35 billion of shares when certain conditions are met. At the same time, AWS and OpenAI expanded an existing $38 billion commercial arrangement by another $100 billion over eight years.

Microsoft’s version is equally elegant. Microsoft invests in OpenAI. OpenAI commits to purchasing another $250 billion of Azure. Azure reports stronger bookings and backlog. Microsoft spends more on data centres because customer demand exceeds capacity. The increased compute supply allows OpenAI to expand, supporting a higher private valuation. Microsoft’s own announcement valued its OpenAI holding at approximately $135 billion after the company’s recapitalisation.

Microsoft then invested in Anthropic, which committed $30 billion to Azure. Anthropic’s valuation increased again, allowing Microsoft to record a $3.2 billion quarterly gain.

The transactions are real in the narrow accounting sense. Real money changes hands. Real chips are purchased. Real data centres are built. Real compute is consumed.

The problem is that each stage is presented as independent evidence of success when it is frequently another part of the same loop.

The investment supposedly proves confidence in the laboratory. The laboratory’s cloud commitment supposedly proves organic customer demand. The cloud demand supposedly proves that the hyperscaler must spend more. The infrastructure supposedly proves that the laboratory can grow into its valuation. The higher valuation produces paper gains for the hyperscaler, which supposedly proves that the original investment was brilliant.

Then everyone repeats the process with larger numbers.

Anthropic raised funding in May at a $965 billion post-money valuation. Amazon’s quarterly accounts subsequently contained $53.4 billion of non-operating income, primarily from Anthropic. Microsoft booked another $3.2 billion. None of that represents customers paying Amazon or Microsoft $56.6 billion in cash for useful AI products during the quarter. It represents investors agreeing upon a higher price for a privately owned company whose enormous cloud bills are partly supported by the same hyperscalers benefiting from that price.

This does not mean AWS or Azure demand is fake. Anthropic and OpenAI consume very real quantities of compute, and both have actual paying customers.

It means cloud demand is not necessarily independent, self-financing proof that the broader AI economy works. A supplier financing its customer, recognising revenue when that customer spends the money back with the supplier, then recording an investment gain when the customer’s private valuation increases is circular as fuck.

At some point, somebody should ask whether OpenAI and Anthropic can pay these gigantic cloud commitments from profitable operations rather than the next funding round supplied by cloud companies, chip companies, sovereign funds and investors chasing the last valuation.

Instead, every new funding round is treated as confirmation that the previous one was sensible.

How many times can they repeat this?:

The hyperscalers are not about to run out of cash next quarter. Their existing businesses remain too powerful. Microsoft has Office and enterprise software. Google has Search and YouTube. Meta has its advertising empire. Amazon has AWS, advertising and retail.

They can continue raising prices, issuing debt, reducing buybacks, cutting staff, cancelling projects, disposing of weaker businesses and shifting leases between accounting categories. They will destroy almost everything around the AI programme before admitting the programme itself has become too expensive.

Those levers are finite.

An employee can only be fired once. A studio can only be closed once. A product can only be cancelled once. Customers will only accept so many price increases before looking for alternatives or reducing usage. An asset’s useful life can be extended on paper, but the underlying server still becomes obsolete. A private valuation can rise repeatedly, but eventually somebody needs liquidity or an actual public market willing to buy the shares at that price.

The costs are also accumulating rather than replacing one another. These companies are paying for new infrastructure today while depreciation, electricity, leases, maintenance and financing costs from earlier construction increasingly reach the income statement. Much of Microsoft’s quarterly spending went towards short-lived CPUs and GPUs. That hardware may need replacing before the original investment has earned the returns used to justify it.

Investors currently tolerate this because Azure, AWS and Google Cloud are growing quickly. The market has decided that any spending level is acceptable provided the cloud growth percentage remains exciting enough.

The breaking point arrives when cloud growth slows while capital expenditure and depreciation remain elevated. It arrives when a major frontier laboratory requires another funding round mainly to service contracts already signed with its investors. It arrives when a private valuation stops rising, turning paper gains into write-downs. It arrives when data-centre capacity booked during the panic is renegotiated, delayed or left underused.

Nobody needs to announce that AGI has failed. AGI can remain five years away forever. The language will gradually change from imminent machine intelligence to a long-term platform transition, then to disciplined investment in durable cloud infrastructure.

The financial retreat will receive a similarly flattering name. Spending cuts will become capacity optimisation. Cancelled facilities will become portfolio prioritisation. Obsolete hardware will become a fleet-modernisation charge. Another few thousand workers will be removed in the name of efficiency while the executives responsible explain that the company remains more confident than ever.

For now, all four companies can claim some form of victory. Meta’s advertising improved. Google Cloud exploded. AWS had its best growth in years. Azure remains the strongest-looking business of the lot.

They are also consuming more capital than previously forecast, sacrificing free cash flow, accepting pressure on margins and using private AI valuations to manufacture accounting gains from companies that promise to spend much of their funding back with the investors.

These are not four independent demonstrations that the AI economy is thriving. They are increasingly interconnected companies investing in one another, renting one another compute, signing colossal long-term commitments, firing workers to preserve margins and marking private shares upwards while real cash disappears into chips and concrete.

Every quarter the cloud backlog grows, the private valuation rises and the capital expenditure forecast moves upwards again.

The only figure nobody seems eager to provide is how much genuine profit remains once the same money has finished travelling around the circle.