The build-out of artificial intelligence infrastructure is the defining market story of 2025 and 2026: the largest, fastest and most capital-intensive corporate spending cycle in modern history. The numbers run into the hundreds of billions a year, the supply chain is straining at every link, and the share-price swings have moved whole indices. This page is Khan Capital’s running guide to the supercycle, layer by layer, from the companies writing the cheques down to the single scarcest component that gates the entire thing. Each section links to our detailed analysis, so you can follow the parts that matter to you.
Latest in this story
Broadcom AI Revenue Tripled to $16.7 Billion. The Stock Fell Anyway. · 3 Sep 2026
Start here: the scale of the spend
Begin with the number that frames everything else. When Microsoft, Alphabet, Meta and Amazon reported within a single 36-hour window, they laid out a combined capital-expenditure plan approaching three-quarters of a trillion dollars, a cycle without modern precedent. Our anchor piece, AI Capex Hits $725bn: Wall Street Splits on the Hyperscaler Trade, explains where the money is going and why the market is divided on whether it will pay off. The bill’s arrival on the income statement is covered in Alphabet’s $205 billion capex quarter, the first negative free cash flow quarter of the AI era at a hyperscaler. A week later the whole cohort faced the same audit: the earnings week that graded AI capex line by line, splitting Microsoft and Amazon from Meta and Apple by nineteen percentage points. The software counterpoint arrived days later in Palantir’s 93 per cent quarter, the print that split the AI trade into hardware sceptics and software believers. The bellwether itself reported at the end of August: Nvidia’s $96 billion quarter, the widest beat-and-raise of the cycle, ended a year-long run of sold prints. The index-level view of the same week, including the milestone that capped it, is in Amazon’s $3 trillion market cap. The build-out’s financing leg found its clearest listed expression in CoreWeave’s second quarter, where a $104 billion backlog sits against a $640 million quarterly interest bill. The custom-silicon leg of the same build-out was priced a week later, when Broadcom tripled AI revenue to $16.7 billion and the market sold it on a guidance figure seven tenths of one per cent light.
The chips at the centre
At the heart of the build-out sits the accelerator, and Nvidia remains the company the whole trade is read through. Our analysis of Nvidia’s $91 billion quarter shows how the bar has risen as fast as the results. The fragility of that leadership was exposed in the $1 trillion AI semiconductor selloff, when a record week broke on a single Friday and good news briefly became bad news. The July 2026 turn, when Meta moved to sell its excess capacity and the supply premise cracked, is covered in Meta’s AI Cloud Pivot. The foundry that actually builds them reported in July, and TSMC’s Q2 2026 earnings raised full-year growth guidance above 40 per cent while lifting 2026 capital spending to $60bn to $64bn: the build-out funded, not just described. And the challenger’s answer landed on 23 July: Intel’s Q2 2026 earnings, the fastest growth since 2011, with AI server CPUs sold out and the 18A foundry case gathering evidence.
The memory bottleneck
An accelerator is only as fast as the memory feeding it, and memory has become the supply chain’s binding constraint. The NVIDIA SK hynix memory deal examines how reserving a supplier’s roadmap years in advance has become the new operating model, and why a memory shortage is now, in effect, an AI compute shortage. Micron’s record quarter then confirmed that thesis in hard numbers, with high-bandwidth memory sold out, gross margins above 80 per cent, and a guide to fifty billion dollars in a single quarter. The squeeze has since reached consumers, as the AI memory price shock pushed Apple and Microsoft into their first memory-driven price rises. Samsung’s record ₩89 trillion quarter shows where that constraint has taken pricing, and why the market sold the news. The constraint has now reached the capital markets too: SK Hynix’s record $26.5bn Nasdaq listing saw the dominant HBM supplier raise the largest foreign US share sale on record to fund the next leg of capacity. Policy is now a live input too: the first licensed H200 shipments to China reopened a market the controls era had closed.
The Semiconductor Bear Market: Three Weeks That Unwound a 105% Rally
The Philadelphia Semiconductor Index fell 20 per cent from its June record in fifteen sessions, against the strongest results in the sector’s history. What the flows, the leverage unwind and the rotation into value say about the AI trade’s first bear market.
Down the stack: servers, networking and the cloud
The economics ripple outward from the chip. Dell’s $51.3 billion AI server backlog shows the build-out moving into enterprise hardware, while Cisco’s networking surge is the other, quieter AI trade. At the top of the stack, Oracle’s $638 billion cloud backlog reveals how the operators who house and rent the hardware are pre-selling capacity years before it is built.
The application layer: who profits, who pays
Further up, the question becomes who turns all this compute into products and profit. The Apple Siri Gemini deal shows a platform owner choosing to rent the model rather than build it, a theme we first traced in Apple after Cook. And Tesla’s capex surprise opened the Magnificent Seven earnings reckoning, where the cost of the AI bet began to show up on the cash-flow statement. And IBM’s worst day on record showed who pays for it: a 3.7 per cent revenue miss cost a quarter of the company’s value, after clients funded AI hardware by delaying the software they buy from everyone else.
The market’s verdict
For all the spending, the market has not rewarded the trade evenly. The great tech divergence sets out why software broke while silicon soared, and why capital intensity and supply security, rather than the simple demand story, have become the swing variables in how the whole complex is priced.
The through-line
Read together, these pieces tell one story: the AI build-out is a physical supply chain with hard constraints, and the advantage is shifting from whoever has the best model to whoever has secured the inputs to build it. We update this hub as the cycle develops; the latest analysis always appears first on our analysis page.